From the Contents: Basic concepts of structural analysis and graph theory -- Optimal force method of structural analysis

*308*
*18*
*43MB*

*English*
*Pages 445 p
[445]*
*Year 2014*

*Table of contents : From the Contents: Basic concepts of structural analysis and graph theory --Optimal force method of structural analysis --Optimal displacement method of structural analysis --Ordering for optimal patterns of structural matrices: graph theory methods.*

A. Kaveh

Computational Structural Analysis and Finite Element Methods

Computational Structural Analysis and Finite Element Methods

.

A. Kaveh

Computational Structural Analysis and Finite Element Methods

A. Kaveh Centre of Excellence for Fundamental Studies in Structural Engineering School of Civil Engineering Iran University of Science and Technology Tehran Iran

ISBN 978-3-319-02963-4 ISBN 978-3-319-02964-1 (eBook) DOI 10.1007/978-3-319-02964-1 Springer Cham Heidelberg New York Dordrecht London Library of Congress Control Number: 2013956541 © Springer International Publishing Switzerland 2014 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection with reviews or scholarly analysis or material supplied specifically for the purpose of being entered and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of this publication or parts thereof is permitted only under the provisions of the Copyright Law of the Publisher’s location, in its current version, and permission for use must always be obtained from Springer. Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations are liable to prosecution under the respective Copyright Law. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. While the advice and information in this book are believed to be true and accurate at the date of publication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect to the material contained herein. Printed on acid-free paper Springer is part of Springer Science+Business Media (www.springer.com)

Preface

Recent advances in structural technology require greater accuracy, efficiency and speed in the analysis of structural systems. It is therefore not surprising that new methods have been developed for the analysis of structures with complex configurations and large number of elements. The requirement of accuracy in analysis has been brought about by the need for demonstrating structural safety. Consequently, accurate methods of analysis had to be developed, since conventional methods, although perfectly satisfactory when used on simple structures, have been found inadequate when applied to complex and large-scale structures. Another reason why higher speed is required results from the need to have optimal design, where analysis is repeated hundred or even thousands of times. This book can be considered as an application of discrete mathematics rather than the more usual calculus-based methods of analysis of structures and finite element methods. The subject of graph theory has become important in science and engineering through its strong links with matrix algebra and computer science. At first glance, it seems extraordinary that such abstract material should have quite practical applications. However, as the author makes clear, the early relationship between graph theory and skeletal structures and finite element models is now obvious: the structure of the mathematics is well suited to the structure of the physical problem. In fact, could there be any other way of dealing with this structural problem? The engineer studying these applications of structural analysis has either to apply the computer programs as a black box, or to become involved in graph theory, matrix algebra and sparse matrix technology. This book is addressed to those scientists and engineers, and their students, who wish to understand the theory. The methods of analysis in this book employ matrix algebra and graph theory, which are ideally suited for modern computational mechanics. Although this text deals primarily with analysis of structural engineering systems, it should be recognised that these methods are also applicable to other types of systems such as hydraulic and electrical networks.

v

vi

Preface

The author has been involved in various developments and applications of graph theory in the last four decades. The present book contains part of this research suitable for various aspects of matrix structural analysis and finite element methods, with particular attention to the finite element force method. In Chap. 1, the most important concepts and theorems of structures and theory of graphs are briefly presented. Chapter 2 contains different efficient approaches for determining the degree of static indeterminacy of structures and provides systematic methods for studying the connectivity properties of structural models. In this chapter, force method of analysis for skeletal structures is described mostly based on the author’s algorithms. Chapter 3 provides simple and efficient methods for construction of stiffness matrices. These methods are especially suitable for the formation of wellconditioned stiffness matrices. In Chaps. 4 and 5, banded, variable banded and frontal methods are investigated. Efficient methods are presented for both node and element ordering. Many new graphs are introduced for transforming the connectivity properties of finite element models onto graph models. Chapters 6 and 7 include powerful graph theory and algebraic graph theory methods for the force method of finite element meshes of low order and high order, respectively. These new methods use different graphs of the models and algebraic approaches. In Chap. 8, several partitioning algorithms are developed for solution of multi-member systems, which can be categorized as graph theory methods and algebraic graph theory approaches. In Chap. 9, an efficient method is presented for the analysis of near-regular structures which are obtained by addition or removal of some members to regular structural models. In Chap. 10, energy formulation based on the force method is derived and a new optimization algorithm called SCSS is applied to the analysis procedure. Then, using the SCSS and prescribed stress ratios, structures are analyzed and designed. In all the chapters, many examples are included to make the text easier to be understood. I would like to take this opportunity to acknowledge a deep sense of gratitude to a number of colleagues and friends who in different ways have helped in the preparation of this book. Mr. J. C. de C. Henderson, formerly of Imperial College of Science and Technology, first introduced me to the subject with most stimulating discussions on various aspects of topology and combinatorial mathematics. Professor F. Ziegler and Prof. Ch. Bucher encouraged and supported me to write this book. My special thanks are due to Mrs. Silvia Schilgerius, the senior editor of the Applied Sciences of Springer, for her constructive comments, editing and unfailing kindness in the course of the preparation of this book. My sincere appreciation is extended to our Springer colleagues Ms. Beate Siek and Ms. G. Ramya Prakash. I would like to thank my former Ph.D. and M.Sc. students, Dr. H. Rahami, Dr. M. S. Massoudi, Dr. K. Koohestani, Dr. P. Sharafi, Mr. M. J. Tolou Kian, Dr. A. Mokhtar-zadeh, Mr. G. R. Roosta, Ms. E. Ebrahimi, Mr. M. Ardalan, and Mr. B. Ahmadi for using our joint papers and for their help in various stages of writing this book. I would like to thank the publishers who permitted some of our papers to be utilized in the preparation of this book, consisting of Springer-Verlag, John Wiley and Sons, and Elsevier. My warmest gratitude is due to my family and in particular my wife, Mrs. Leopoldine Kaveh, for her continued support in the course of preparing this book.

Preface

vii

Every effort has been made to render the book error free. However, the author would appreciate any remaining errors being brought to his attention through his email-address: [email protected] Tehran December 2013

A. Kaveh

.

Contents

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1.2 Structural Analysis and Design . . . . . . . . . . . . . . . . . . . . 1.2 General Concepts of Structural Analysis . . . . . . . . . . . . . . . . . . . 1.2.1 Main Steps of Structural Analysis . . . . . . . . . . . . . . . . . . 1.2.2 Member Forces and Displacements . . . . . . . . . . . . . . . . . 1.2.3 Member Flexibility and Stiffness Matrices . . . . . . . . . . . . 1.3 Important Structural Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.1 Work and Energy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.2 Castigliano’s Theorems . . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.3 Principle of Virtual Work . . . . . . . . . . . . . . . . . . . . . . . . 1.3.4 Contragradient Principle . . . . . . . . . . . . . . . . . . . . . . . . . 1.3.5 Reciprocal Work Theorem . . . . . . . . . . . . . . . . . . . . . . . . 1.4 Basic Concepts and Definitions of Graph Theory . . . . . . . . . . . . . 1.4.1 Basic Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.2 Definition of a Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.3 Adjacency and Incidence . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.4 Graph Operations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.5 Walks, Trails and Paths . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.6 Cycles and Cutsets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.4.7 Trees, Spanning Trees and Shortest Route Trees . . . . . . . . 1.4.8 Different Types of Graphs . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Vector Spaces Associated with a Graph . . . . . . . . . . . . . . . . . . . . 1.5.1 Cycle Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.2 Cutset Space . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.3 Orthogonality Property . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5.4 Fundamental Cycle Bases . . . . . . . . . . . . . . . . . . . . . . . . 1.5.5 Fundamental Cutset Bases . . . . . . . . . . . . . . . . . . . . . . . .

1 1 1 4 5 5 6 7 11 11 13 13 16 17 18 19 19 20 20 21 22 23 23 25 26 26 26 27 27

ix

x

Contents

1.6 Matrices Associated with a Graph . . . . . . . . . . . . . . . . . . . . . . . 1.6.1 Matrix Representation of a Graph . . . . . . . . . . . . . . . . . 1.6.2 Cycle Bases Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.3 Special Patterns for Fundamental Cycle Bases . . . . . . . . 1.6.4 Cutset Bases Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . 1.6.5 Special Patterns for Fundamental Cutset Bases . . . . . . . . 1.7 Directed Graphs and Their Matrices . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

. . . . . . . .

28 29 32 33 34 34 35 37

Optimal Force Method: Analysis of Skeletal Structures . . . . . . . . . 2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.2 Static Indeterminacy of Structures . . . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Mathematical Model of a Skeletal Structure . . . . . . . . . . . 2.2.2 Expansion Process for Determining the Degree of Static Indeterminacy . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3 Formulation of the Force Method . . . . . . . . . . . . . . . . . . . . . . . . 2.3.1 Equilibrium Equations . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.2 Member Flexibility Matrices . . . . . . . . . . . . . . . . . . . . . . 2.3.3 Explicit Method for Imposing Compatibility . . . . . . . . . . . 2.3.4 Implicit Approach for Imposing Compatibility . . . . . . . . . 2.3.5 Structural Flexibility Matrices . . . . . . . . . . . . . . . . . . . . . 2.3.6 Computational Procedure . . . . . . . . . . . . . . . . . . . . . . . . . 2.3.7 Optimal Force Method . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4 Force Method for the Analysis of Frame Structures . . . . . . . . . . . 2.4.1 Minimal and Optimal Cycle Bases . . . . . . . . . . . . . . . . . . 2.4.2 Selection of Minimal and Subminimal Cycle Bases . . . . . . 2.4.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.4 Optimal and Suboptimal Cycle Bases . . . . . . . . . . . . . . . . 2.4.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.6 An Improved Turn Back Method for the Formation of Cycle Bases . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.7 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.4.8 Formation of B0 and B1 Matrices . . . . . . . . . . . . . . . . . . . 2.5 Generalized Cycle Bases of a Graph . . . . . . . . . . . . . . . . . . . . . . 2.5.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.5.2 Minimal and Optimal Generalized Cycle Bases . . . . . . . . . 2.6 Force Method for the Analysis of Pin-Jointed Planar Trusses . . . . 2.6.1 Associate Graphs for Selection of a Suboptimal GCB . . . . 2.6.2 Minimal GCB of a Graph . . . . . . . . . . . . . . . . . . . . . . . . 2.6.3 Selection of a Subminimal GCB: Practical Methods . . . . . 2.7 Algebraic Force Methods of Analysis . . . . . . . . . . . . . . . . . . . . . 2.7.1 Algebraic Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

39 39 40 41 42 46 46 49 52 53 55 55 60 60 61 62 67 69 72 75 76 78 82 83 85 86 86 89 89 91 91 98

Contents

3

4

Optimal Displacement Method of Structural Analysis . . . . . . . . . 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.1 Coordinate Systems Transformation . . . . . . . . . . . . . . . . 3.2.2 Element Stiffness Matrix Using Unit Displacement Method . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.3 Element Stiffness Matrix Using Castigliano’s Theorem . . 3.2.4 The Stiffness Matrix of a Structure . . . . . . . . . . . . . . . . . 3.2.5 Stiffness Matrix of a Structure; an Algorithmic Approach . . . . . . . . . . . . . . . . . . . . . . . 3.3 Transformation of Stiffness Matrices . . . . . . . . . . . . . . . . . . . . . 3.3.1 Stiffness Matrix of a Bar Element . . . . . . . . . . . . . . . . . 3.3.2 Stiffness Matrix of a Beam Element . . . . . . . . . . . . . . . . 3.4 Displacement Method of Analysis . . . . . . . . . . . . . . . . . . . . . . . 3.4.1 Boundary Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.4.2 General Loading . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.5 Stiffness Matrix of a Finite Element . . . . . . . . . . . . . . . . . . . . . 3.5.1 Stiffness Matrix of a Triangular Element . . . . . . . . . . . . 3.6 Computational Aspects of the Matrix Displacement Method . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

xi

. . . .

101 101 101 102

. 105 . 109 . 111 . . . . . . . . . . .

116 118 118 120 122 124 125 128 129 132 135

Ordering for Optimal Patterns of Structural Matrices: Graph Theory Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2 Bandwidth Optimisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Preliminaries . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.4 A Shortest Route Tree and Its Properties . . . . . . . . . . . . . . . . . 4.5 Nodal Ordering for Bandwidth Reduction . . . . . . . . . . . . . . . . 4.5.1 A Good Starting Node . . . . . . . . . . . . . . . . . . . . . . . . 4.5.2 Primary Nodal Decomposition . . . . . . . . . . . . . . . . . . 4.5.3 Transversal P of an SRT . . . . . . . . . . . . . . . . . . . . . . 4.5.4 Nodal Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.5.5 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.6 Finite Element Nodal Ordering for Bandwidth Optimisation . . . 4.6.1 Element Clique Graph Method (ECGM) . . . . . . . . . . 4.6.2 Skeleton Graph Method (SkGM) . . . . . . . . . . . . . . . . 4.6.3 Element Star Graph Method (EStGM) . . . . . . . . . . . . 4.6.4 Element Wheel Graph Method (EWGM) . . . . . . . . . . 4.6.5 Partially Triangulated Graph Method (PTGM) . . . . . . 4.6.6 Triangulated Graph Method (TGM) . . . . . . . . . . . . . 4.6.7 Natural Associate Graph Method (NAGM) . . . . . . . . 4.6.8 Incidence Graph Method (IGM) . . . . . . . . . . . . . . . . 4.6.9 Representative Graph Method (RGM) . . . . . . . . . . . . 4.6.10 Computational Results . . . . . . . . . . . . . . . . . . . . . . . 4.6.11 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

137 137 138 140 142 142 143 145 146 146 147 147 149 149 150 151 152 153 153 155 156 157 158

xii

Contents

4.7

Finite Element Nodal Ordering for Profile Optimisation . . . . . . 160 4.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 160 4.7.2 Graph Nodal Numbering for Profile Reduction . . . . . 162 4.7.3 Nodal Ordering with Element Clique Graph (NOECG) . . . . . . . . . . . . . . . . . . . . . . 164 4.7.4 Nodal Ordering with Skeleton Graph (NOSG) . . . . . . 165 4.7.5 Nodal Ordering with Element Star Graph (NOESG) . . . 166 4.7.6 Nodal Ordering with Element Wheel Graph (NOEWG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 4.7.7 Nodal Ordering with Partially Triangulated Graph (NOPTG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 4.7.8 Nodal Ordering with Triangulated Graph (NOTG) . . . 167 4.7.9 Nodal Ordering with Natural Associate Graph (NONAG) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168 4.7.10 Nodal Ordering with Incidence Graph (NOIG) . . . . . 168 4.7.11 Nodal Ordering with Representative Graph (NORG) . . . 168 4.7.12 Nodal Ordering with Element Clique Representative Graph (NOECRG) . . . . . . . . . . . . . . . . . . . . . . . . . . 170 4.7.13 Computational Results . . . . . . . . . . . . . . . . . . . . . . . 170 4.7.14 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170 4.8 Element Ordering for Frontwidth Reduction . . . . . . . . . . . . . . . 171 4.9 Element Ordering for Bandwidth Optimisation of Flexibility Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174 4.9.1 An Associate Graph . . . . . . . . . . . . . . . . . . . . . . . . . . 174 4.9.2 Distance Number of an Element . . . . . . . . . . . . . . . . . 175 4.9.3 Element Ordering Algorithms . . . . . . . . . . . . . . . . . . . 175 4.10 Bandwidth Reduction for Rectangular Matrices . . . . . . . . . . . . 177 4.10.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .177 .. 4.10.2 Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .178 .. 4.10.3 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .179 .. 4.10.4 Bandwidth Reduction of Finite Element Models . . . . . . .181 .. 4.11 Graph-Theoretical Interpretation of Gaussian Elimination . . . . . 182 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 5

Ordering for Optimal Patterns of Structural Matrices: Algebraic Graph Theory and Meta-heuristic Based Methods . . . . 5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Adjacency Matrix of a Graph for Nodal Ordering . . . . . . . . . . . 5.2.1 Basic Concepts and Definitions . . . . . . . . . . . . . . . . . . . 5.2.2 A Good Starting Node . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.3 Primary Nodal Decomposition . . . . . . . . . . . . . . . . . . . . 5.2.4 Transversal P of an SRT . . . . . . . . . . . . . . . . . . . . . . . . 5.2.5 Nodal Ordering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2.6 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . .

187 187 187 187 190 190 191 191 192

Contents

5.3 Laplacian Matrix of a Graph for Nodal Ordering . . . . . . . . . . . . 5.3.1 Basic Concepts and Definitions . . . . . . . . . . . . . . . . . . . 5.3.2 Nodal Numbering Algorithm . . . . . . . . . . . . . . . . . . . . . 5.3.3 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4 A Hybrid Method for Ordering . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.1 Development of the Method . . . . . . . . . . . . . . . . . . . . . . 5.4.2 Numerical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.4.3 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.5 Ordering via Charged System Search Algorithm . . . . . . . . . . . . 5.5.1 Charged System Search . . . . . . . . . . . . . . . . . . . . . . . . . 5.5.2 The CSS Algorithm for Nodal Ordering . . . . . . . . . . . . . 5.5.3 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

xiii

. . . . . . . . . . . . .

192 192 196 196 196 197 198 199 203 203 208 211 213

Optimal Force Method for FEMs: Low Order Elements . . . . . . . . 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Force Method for Finite Element Models: Rectangular and Triangular Plane Stress and Plane Strain Elements . . . . . . . . . . . . 6.2.1 Member Flexibility Matrices . . . . . . . . . . . . . . . . . . . . . . 6.2.2 Graphs Associated with FEMs . . . . . . . . . . . . . . . . . . . . . 6.2.3 Pattern Corresponding to the Self Stress Systems . . . . . . . 6.2.4 Selection of Optimal γ-Cycles Corresponding to Type II Self Stress Systems . . . . . . . . . . . . . . . . . . . . . 6.2.5 Selection of Optimal Lists . . . . . . . . . . . . . . . . . . . . . . . . 6.2.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate Bending Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 Graphs Associated with Finite Element Models . . . . . . . . 6.3.2 Subgraphs Corresponding to Self-Equilibrating Systems . . . 6.3.3 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.4 Force Method for Three Dimensional Finite Element Analysis . . . 6.4.1 Graphs Associated with Finite Element Model . . . . . . . . . 6.4.2 The Pattern Corresponding to the Self Stress Systems . . . . 6.4.3 Relationship Between γ(S) and b1(A(S)) . . . . . . . . . . . . . 6.4.4 Selection of Optimal γ-Cycles Corresponding to Type II Self Stress Systems . . . . . . . . . . . . . . . . . . . . . 6.4.5 Selection of Optimal Lists . . . . . . . . . . . . . . . . . . . . . . . . 6.4.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method: Brick Element . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.5.1 Definition of the Independent Element Forces . . . . . . . . . . 6.5.2 Flexibility Matrix of an Element . . . . . . . . . . . . . . . . . . . 6.5.3 Graphs Associated with Finite Element Model . . . . . . . . . 6.5.4 Topological Interpretation of Static Indeterminacy . . . . . .

215 215 215 216 220 221 224 225 227 230 233 233 240 244 244 245 248 251 252 254 257 258 259 261 263

xiv

Contents

6.5.5 6.5.6

Models Including Internal Node . . . . . . . . . . . . . . . . . . Selection of an Optimal List Corresponding to Minimal Self-Equilibrating Stress Systems . . . . . . . . . . . . . . . . . 6.5.7 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

8

. . 270 . . 271 . . 272 . . 279

Optimal Force Method for FEMS: Higher Order Elements . . . . . . 7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.2.1 Definition of the Element Force System . . . . . . . . . . . . . . 7.2.2 Flexibility Matrix of the Element . . . . . . . . . . . . . . . . . . . 7.2.3 Graphs Associated with Finite Element Model . . . . . . . . . 7.2.4 Topological Interpretation of Static Indeterminacies . . . . . 7.2.5 Models Including Opening . . . . . . . . . . . . . . . . . . . . . . . . 7.2.6 Selection of an Optimal List Corresponding to Minimal Self-Equilibrating Stress Systems . . . . . . . . . . . . . . . . . . . 7.2.7 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3 Finite Element Analysis of Models Comprised of Higher Order Rectangular Elements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.1 Definition of Element Force System . . . . . . . . . . . . . . . . . 7.3.2 Flexibility Matrix of the Element . . . . . . . . . . . . . . . . . . . 7.3.3 Graphs Associated with Finite Element Model . . . . . . . . . 7.3.4 Topological Interpretation of Static Indeterminacies . . . . . 7.3.5 Selection of Generators for SESs of Type II and Type III . . . 7.3.6 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.3.7 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method: Hexa-Hedron Elements . . . . . . . . . . . . . . . . . . . . 7.4.1 Independent Element Forces and Flexibility Matrix of Hexahedron Elements . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.2 Graphs Associated with Finite Element Models . . . . . . . . 7.4.3 Negative Incidence Number . . . . . . . . . . . . . . . . . . . . . . . 7.4.4 Pattern Corresponding to Self-Equilibrating Systems . . . . . 7.4.5 Selection of Generators for SESs of Type II and Type III . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7.4.6 Numerical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Decomposition for Parallel Computing: Graph Theory Methods . . . . 8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2 Earlier Works on Partitioning . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 Nested Dissection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.2 A Modified Level-Tree Separator Algorithm . . . . . . . . . . . 8.3 Substructuring for Parallel Analysis of Skeletal Structures . . . . . . 8.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.2 Substructuring Displacement Method . . . . . . . . . . . . . . . .

281 281 281 282 282 282 284 287 290 291 297 298 300 301 303 307 308 309 316 317 321 325 325 331 334 338 341 341 342 342 342 343 343 344

Contents

8.3.3 Methods of Substructuring . . . . . . . . . . . . . . . . . . . . . . . 8.3.4 Main Algorithm for Substructuring . . . . . . . . . . . . . . . . 8.3.5 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.3.6 Simplified Algorithm for Substructuring . . . . . . . . . . . . . 8.3.7 Greedy Type Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 8.4 Domain Decomposition for Finite Element Analysis . . . . . . . . . 8.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.2 A Graph Based Method for Subdomaining . . . . . . . . . . 8.4.3 Renumbering of Decomposed Finite Element Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.4 Computational Results of the Graph Based Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.5 Discussions on the Graph Based Method . . . . . . . . . . . 8.4.6 Engineering Based Method for Subdomaining . . . . . . . . 8.4.7 Genre Structure Algorithm . . . . . . . . . . . . . . . . . . . . . . 8.4.8 Example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.9 Computational Results of the Engineering Based Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4.10 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5 Substructuring: Force Method . . . . . . . . . . . . . . . . . . . . . . . . . . 8.5.1 Algorithm for the Force Method Substructuring . . . . . . . 8.5.2 Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

xv

. . . . . . . .

346 348 348 350 352 352 353 354

. 356 . . . . .

356 359 360 361 364

. . . . . .

367 367 370 370 373 376

Analysis of Regular Structures Using Graph Products . . . . . . . . . . 9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.2 Definitions of Different Graph Products . . . . . . . . . . . . . . . . . . . . 9.2.1 Boolean Operation on Graphs . . . . . . . . . . . . . . . . . . . . . 9.2.2 Cartesian Product of Two Graphs . . . . . . . . . . . . . . . . . . . 9.2.3 Strong Cartesian Product of Two Graphs . . . . . . . . . . . . . 9.2.4 Direct Product of Two Graphs . . . . . . . . . . . . . . . . . . . . . 9.3 Analysis of Near-Regular Structures Using Force Method . . . . . . 9.3.1 Formulation of the Flexibility Matrix . . . . . . . . . . . . . . . . 9.3.2 A Simple Method for the Formation of the Matrix AT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9.4 Analysis of Regular Structures with Excessive Members . . . . . . . 9.4.1 Summary of the Algorithm . . . . . . . . . . . . . . . . . . . . . . . 9.4.2 Investigation of a Simple Example . . . . . . . . . . . . . . . . . . 9.5 Analysis of Regular Structures with Some Missing Members . . . . 9.5.1 Investigation of an Illustrative Simple Example . . . . . . . . 9.6 Practical Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

377 377 377 377 378 380 381 383 385 388 389 390 390 393 393 396 406

xvi

10

Contents

Simultaneous Analysis, Design and Optimization of Structures Using Force Method and Supervised Charged System Search . . . . 10.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.2 Supervised Charged System Search Algorithm . . . . . . . . . . . . . 10.3 Analysis by Force Method and Charged System Search . . . . . . 10.4 Procedure of Structural Design Using Force Method and the CSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10.4.1 Pre-selected Stress Ratio . . . . . . . . . . . . . . . . . . . . . . 10.5 Minimum Weight . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

407 407 408 409 414 415 420 432

Chapter 1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

1.1

Introduction

This chapter consists of two parts. In the first part, basic definitions, concepts and theorems of structural mechanics are presented. These theorems are employed in the following chapters and are very important for their understanding. For determination of the distribution of internal forces and displacements, under prescribed external loading, a solution to the basic equations of the theory of structures should be obtained, satisfying the boundary conditions. In the matrix methods of structural analysis, one must also use these basic equations. In order to provide a ready reference for the development of the general theory of matrix structural analysis, the most important basic theorems are introduced in this chapter, and illustrated through simple examples. In the second part, basic concepts and definitions of graph theory are presented. Since some of the readers may be unfamiliar with the theory of graphs, simple examples are included to make it easier to understand the presented concepts.

1.1.1

Definitions

A structure can be defined as a body that resists external effects such as loads, temperature changes, and support settlements, without undue deformation. Building frames, industrial building, bridges, halls, towers, dams, reservoirs, tanks, retaining walls, channels, pavements are typical structures of interest to civil engineers. A structure can be considered as an assemblage of members and nodes. Structures with clearly defined members are known as skeletal structures. Planar and space trusses, planar and space frames, single and double-layer grids are examples of skeletal structures, Fig. 1.1.

A. Kaveh, Computational Structural Analysis and Finite Element Methods, DOI 10.1007/978-3-319-02964-1_1, © Springer International Publishing Switzerland 2014

1

2

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

3D view of a footbridge

Side view

Top and bottom view

b

Fig. 1.1 (continued)

c

1.1 Introduction

3

d

e

g

f

h

Fig. 1.1 Examples of skeletal structures. (a) A foot bridge truss (b) A planar frame. (c) A space frame. (d) A space truss. (e) A single-layer grid. (f) A double-layer grid. (g) A single-layer dome. (h) A double-layer barrel vault

4

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

b

Fig. 1.2 Examples of continua. (a) A plate. (b) A dam Structure

Loading

Structural Analysis

Redesign

Stress Analysis

Structural Design

Fig. 1.3 The cycle of analysis and design of a structure

Structures which must artificially be divided into members (elements) are called continua. Concrete dams, plates, and pavements are examples of continua, Fig. 1.2. The underlying principles for the analysis of other structures are more or less the same. Airplane, missile and satellite structures are of interest to the aviation engineer. The analysis and design of a ship is interesting for a naval architect. A machine engineer should be able to design machine parts. However, in this book only structures of interest to structural engineers are studied.

1.1.2

Structural Analysis and Design

Structural analysis is the determination of the response of a structure to external effects such as loading, temperature changes and support settlements. Structural design is the selection of a suitable arrangement of members, and a selection of materials and member sections, to withstand the stress resultants (internal forces) by a specified set of loads, and satisfy the stress and displacement constraints, and other requirements specified by the utilized code of practice. The diagram shown in Fig. 1.3 is a simple illustration for the cycle of structural analysis and design. In optimal design of structures this cycle should be repeated hundred and sometime thousands of times to reduce the weight or cost of the structure. Structural theories may be classified from different points of view as follows: Static versus dynamic; Planar versus space;

1.2 General Concepts of Structural Analysis

5

Linear versus non-linear; Skeletal versus continua; Statically determinate versus statically indeterminate. In this book, static analyses of linear structures are mainly discussed for the statically determinate and indeterminate cases. Here, both planar and space skeletal structures and continua models are of interest.

1.2 1.2.1

General Concepts of Structural Analysis Main Steps of Structural Analysis

A correct solution of a structure should satisfy the following requirements: 1. Equilibrium: The external forces applied to a structure and the internal forces induced in its members should be in equilibrium at each node. 2. Compatibility: The members should deform so that they all fit together. 3. Force-displacement relationship: The internal forces and deformations satisfy the stress–strain relationships of the members. For structural analysis two basic methods are in use: Force method: In this method, some of the internal forces and/or reactions are taken as primary unknowns, called redundants. Then the stress–strain relationship is used to express the deformations of the members in terms of external and redundant forces. Finally, by applying the compatibility conditions that the deformed members must fit together, a set of linear equations yield the values of the redundant forces. The stress resultants in the members are then calculated and the displacements at the nodes in the direction of external forces are found. This method is also known as the flexibility method and compatibility approach. Displacement method: In this method, the displacements of the nodes necessary to describe the deformed state of the structure are taken as unknowns. The deformations of the members are then calculated in terms of these displacements, and by use of the stress–strain relationship, the internal forces are related to them. Finally, by applying the equilibrium equations at each node, a set of linear equations is obtained, the solution of which results in the unknown nodal displacements. This method is also known as the stiffness method and equilibrium approach. For choosing the most suitable method for a particular structure, the number of unknowns is one of the main criteria. A comparison for the force and displacement methods can be made, by calculating the degree of static indeterminacy and kinematic indeterminacy. As an example, for the truss structure shown in Fig. 1.4a, the number of redundants is 2 in the force method, while the number of unknown displacements is 13 for the displacement approach. For the 3 ! 3 planar frame shown in Fig. 1.4b, the static indeterminacy and the kinematic indeterminacy are 27 and 36, respectively. For the simple six-bar planar truss of Fig. 1.4c, the

6

1

a

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

b

c

Fig. 1.4 Some simple structures. (a) A planar truss. (b) A planar frame. (c) A simple planar truss

number of unknowns for the force and displacement methods is 4 and 2, respectively. Efficient methods for calculating the indeterminacies are discussed in Chap. 2. The number of unknowns is not the only consideration: another criterion for choosing the most suitable method is the conditioning of the flexibility and stiffness matrices, which are discussed in Kaveh [1, 2].

1.2.2

Member Forces and Displacements

A structure can be considered as an assembly of its members, subjected to external effects. These effects will be considered as external loads applied at nodes, since any other effect can be reduced to such equivalent nodal loads. The state of stress in a member (internal forces) is defined by a vector, ! "t rm ¼ r1k r2k r3k . . . rnk ,

ð1:1Þ

! "t um ¼ u1k u2k u3k . . . unk ,

ð1:2Þ

rm ¼ km um or um ¼ f m rm ,

ð1:3Þ

and the associated member deformation (distortion) is designated by a vector,

where n is the number of force or displacement components of the kth member (element), and t shows the transposition of the vector. Some simple examples of typical elements, common in structural mechanics, are shown in Fig. 1.5. The relationship between member forces and displacements can be written as:

where km and fm are called member stiffness and member flexibility matrices, respectively. Obviously, km and fm are related as: km f m ¼ I:

ð1:4Þ

Flexibility matrices can be written only for members supported in a stable manner, because rigid body motion of the undefined amplitude would otherwise result from application of applied loads. These matrices can be written in as many ways as there are stable and statically determinate support conditions.

1.2 General Concepts of Structural Analysis

7

a

b

c

d

e

f

Fig. 1.5 Some simple elements. (a) Bar element. (b) Beam element. (c) Triangular plane stress element. (d) Rectangular plane stress element. (e) Triangular plate bending element. (f) Rectangular plate bending element

The stiffness and flexibility matrices can be derived using different approaches. For simple members like bar elements and beam elements, methods based on the principles of strength of materials or classical theory of structures will be sufficient. However, for more complicated elements the principle of virtual work or alternatively variational methods can be employed. In this section, only simple members are studied, and further considerations will be presented in Chaps. 2, 6, and 7.

1.2.3

Member Flexibility and Stiffness Matrices

Consider a bar element as shown in Fig. 1.6 which carries only axial forces, and has two components of member forces. From the equilibrium, NmL þ NmR ¼ 0,

ð1:5Þ

then only one end force need be specified in order to determine the state of stress throughout the member. The corresponding deformation of the member is simply the elongation, and hence: r1m ¼ NmR , and u1m ¼ δmR :

ð1:6Þ

8

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs L

R

L

Nm

Nm L + dR

m

Fig. 1.6 Internal forces and deformation of a bar element

a

b

MA

MB dA

L, EI z VA

VB

qA

dB qB

Fig. 1.7 End forces and deflected shape of a beam element

R From Hooke’s law NmR ¼ EA L δm , and therefore:

fm ¼

L EA and km ¼ : EA L

ð1:7Þ

Now consider a prismatic beam of a planar frame with length L and bending stiffness EI. The internal forces are shown in Fig. 1.7. This element is assumed to be subjected to four end forces, as shown in Fig. 1.7a, and the deflected shape and position is illustrated in Fig. 1.7b. Four end forces are related by the following two equilibrium equations: VA þ VB ¼ 0, MA þ MB þ VB L ¼ 0:

ð1:8Þ

Therefore, only two end-force components should be specified as internal forces. Some possible choices for rm are {MA,MB}, {VB,MB} and {VA,MA}. Using classical formulae, such as those of the strength of materials or slopedeflection equations of the theory of structures, the force-displacement relationships can be established. As an example, the flexibility matrix for a prismatic beam supported as a cantilever is obtained using the differential equation of the elastic deformation curve as follows:

1.2 General Concepts of Structural Analysis

9

d2 v M z 1 ¼ ¼ ½VB ðL & xÞ þ MB (: 2 EIz EIz dx Integrating the above equation leads to, i $ dv 1 h # 1 ¼ VB Lx & x2 þ MB x þ C1 , 2 dx EIz and integrating again results in: v¼

i 1 h #1 2 1 3 $ 1 VB Lx & x þ MB x2 þ C1 x þ C2 : 2 6 2 EIz

Using the boundary conditions at A as, %

dv dx

&

x¼0

¼ 0 and ½v(x¼0 ¼ 0,

results in: C1 ¼ 0 and C2 ¼ 0: Substituting these constants leads to: i 1 h #1 2 1 3 $ 1 VB Lx & x þ MB x2 , 2 6 2 EIz h i # $ dv 1 1 ¼ VB Lx & x2 þ MB x : 2 dx EIz

v¼

For x ¼ L, the displacement and rotation of end B are obtained as, δB ¼

V B L3 M B L2 V B L2 M B L þ and θB ¼ þ , EIz 3EIz 2EIz 2EIz

using Iz ¼ I, the above relationships in matrix form become, 2

L3 % & % 1 & 6 3EI 6 δB u ¼ m ¼6 2 6 L2 θB um 4 2EI

3 L2 % & 2EI 7 7 VB 7 , L 7 5 MB EI

10

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

Fig. 1.8 Two sets of end forces and displacements for a beam element

b

3 rm

3 rm

2 rm

1 rm

1 rm

2 rm

or % L2 2L fm ¼ 6EI 3

& 3 : 6=L

ð1:9Þ

Using a similar method, for a simply supported beam with two moments acting at the two ends, we have: 2

L 6 3EI 6 fm ¼ 6 4& L 6EI

3 L % & 6EI 7 L 2 &1 7 : ¼ L 7 5 6EI &1 2 3EI

&

ð1:10Þ

If the axial forces are also included as member forces, then rtm ¼ [NB VB MB] and rtm ¼ [NB MA MB], as shown in Fig. 1.8. The above matrices become: 2

L 6 EA 6 6 6 0 fm ¼ 6 6 6 6 4 0

0 3

L 3EI L2 2EI

3

2

L 0 7 6 EA 7 6 6 L2 7 7 6 7 and f m ¼ 6 0 6 2EI 7 7 6 7 6 L 5 4 0 EI

0

0

3

7 7 L L 7 7 & 7 3EI 6EI 7: 7 L L 7 5 & 6EI 3EI

ð1:11Þ

The corresponding stiffness matrices are: 2

EA 6 L 6 6 6 0 km ¼ 6 6 6 6 4 0

3

2

EA 7 6 L 7 6 6 12EI 6EI 7 7 6 & 2 7 and km ¼ 6 0 3 6 L L 7 7 6 7 6 6EI 4EI 5 4 0 & 2 L L 0

0

0 4EI L 2EI L

3 0 7 7 2EI 7 7 7 L 7: 7 4EI 7 5 L

ð1:12Þ

1.3 Important Structural Theorems Fig. 1.9 Forcedisplacement relationships. (a) A non-linear relationship. (b) A linear relationship

a

11

b

r

r

d W* dr ri

ri

W*

W*

dW

W

W O

u

u i du

O

ui

u

It should be mentioned that both flexibility and stiffness matrices are symmetric, on account of the Maxwell-Betti reciprocal work theorem proven in the next section. More general methods for the derivation of member flexibility and stiffness matrices will be studied in Chaps. 2, 3, 6, and 7.

1.3 1.3.1

Important Structural Theorems Work and Energy

The work, δW, of a force r acting through a change in displacement du in the direction of that force is the product rdu. Consider a general load–displacement relationship as shown in Fig. 1.9a. The area under this curve represents the work done, denoted by W. The area above this curve is the complementary work designated by W*. For a total displacement of u1, the total work is given by, W ¼

ð u1

rdu,

0

ð1:13Þ

and the complementary work is: W

)

¼

ð

r1

udr: 0

ð1:14Þ

For a linear case, as shown in Fig. 1.9b, we have: W ¼ W) :

ð1:15Þ

In this book, it is assumed that the loads are applied to a structure in a gradual manner, and attention is limited to linear behaviour. Thus the load–displacement relationship is as shown in Fig. 1.9b, and the relation can be expressed as,

12

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

Fig. 1.10 Stress-strain relationships. (a) A general stress-strain relationship. (b) Linear stress–strain relationship

a

b

s

s U* U*

U

U e

O

O

r ¼ ku,

e

ð1:16Þ

where k is a constant. The work in Fig. 1.9b can be written as: W ¼

1 ri ui : 2

ð1:17Þ

Forces and displacements at a point are both represented by vectors, and their work is represented as a dot product. In matrix notations, however, the work can be written as: W ¼

1 t r u: 2

ð1:18Þ

Using Eq. 1.3, 1 1 W ¼ ut kt u ¼ ut ku: 2 2

ð1:19Þ

Similarly, W* can be calculated as: W) ¼

1 t r fr: 2

ð1:20Þ

Consider the stress–strain relationship as illustrated in Fig. 1.10a. The area under this curve represents the density of the strain energy, and when integrated over the volume of the member (or structure) results in the strain energy U. The area to the left of the stress–strain curve is the density of the complementary strain energy, and by integration over the member (or structure) the complementary energy U* is obtained. For the linear stress–strain relationship as shown in Fig. 1.10b, U ¼ U*. Since the work done by external actions on an elastic system is equal to the strain energy stored internally in the system (work-energy law), therefore: W ¼ U and W) ¼ U) :

ð1:21Þ

1.3 Important Structural Theorems

1.3.2

13

Castigliano’s Theorems

Consider the force-displacement curve in Fig. 1.9a, and suppose an imaginary displacement δui is imposed on the system. The work done, δW, under the action of ri in moving through δui is equal to: δW ¼ ri δui :

ð1:22Þ

Using Eq. 1.21, and taking limit, leads to the first theorem of Castigliano as, ∂U ¼ ri , ∂qi

ð1:23Þ

which can be stated as follows [3]: The partial derivative of the strain energy with respect to a displacement, is equal to the force applied at the point and along the considered displacement.

Similarly, if the system is subjected to an imaginary force δri along the displacement ui, then the complementary work done δW) is equal to, δW) ¼ ui δri ¼ δU) ,

ð1:24Þ

and in the limit, the second theorem of Castigliano is obtained as: ∂U) ¼ ui : ∂ri

ð1:25Þ

The partial derivative of the complementary strain energy with respect to a force is equal to the displacement at the point where the force is applied and directed along the action of the force.

For the linear case, U) ¼ U and therefore Eq. 1.25 becomes as: ∂U ¼ ui : ∂ri

1.3.3

ð1:26Þ

Principle of Virtual Work

The principle of virtual work is a very powerful means for deducing the conditions of compatibility and equilibrium [4], and it can be stated as follows: The work done by a set of external forces P acting on a structure, in moving through the associated displacements v, is equal to the work done by some other set of forces R, which is statically equivalent to P, moving through associated displacements u, which is compatible with v. Associated forces and displacements have the same lines of actions.

14

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs P

P

P a

C [email protected]

Fig. 1.11 A planar truss

a

b

1

C

1

C

c

1

Fig. 1.12 Three different systems capable of supporting the dummy load

Using a statically admissible set of forces and the work equation, the compatibility relations between the deformations and displacements can be derived. Alternatively, employing a compatible set of displacements and the work equation, one obtains the equations of equilibrium between the forces. These approaches are elegant and practical. Dummy Load Theorem. This theorem can be used to determine the conditions of compatibility. Suppose that the deformed shape of each member of a structure is known, then it is possible to find the deflection of the structure at any point by using the principle of virtual work. For this purpose a dummy load (usually unit load) is applied at the point and in the direction of required displacement, which is why it is also known as the unit load method. The dummy load theorem can be stated as: 8 9 8 9 < applied = < actual displacement = dummy ! of structure where external : ; : ; load dummy load is applied 8 9 8 9 < internal forces = < actual = ¼ statically equivalent to ! deformation : ; : ; the applied dummy load of elements

1.3 Important Structural Theorems Fig. 1.13 Internal forces equivalent to unit dummy load

15 11 1

2

3

12

13

4

5

14 6

7

8

10

9 1

It should be noted that the dummy load theorem is a condition on the geometry of the structure. In fact, once the deformations of elements are known, one can draw the deflected shape of the structure, and the results obtained for the deflections will agree with those of the dummy load theorem. Example 1. Consider a truss as shown in Fig. 1.11. It is desired to measure the vertical deflection at node C, when the structure is subjected to a certain loading. A unit load is applied at C, and a set of internal forces statically equivalent to the unit load is chosen. However, for such equivalent internal forces, there exists a wide choice of systems, since there are several numbers of structural possibilities which can sustain the load at C. Three examples of such systems are shown in Fig. 1.12a–c. Obviously, system (a) will need more calculation because of being statically indeterminate. System (c) is used here, since it has a smaller number of members than (b), and symmetry is also preserved. Internal forces of the members in this system shown in Fig. 1.13 are: n ot pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃ r¼ &1=2, 2=2,&1=2, 2=2, 2=2,&1=2, 2=2,&1=2,1=2,1=2,&1=2,&1,&1,&1=2 : Measuring the elongation in members of this system containing 14 bars, and using the dummy-load theorem, results in:

0 1 pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃ ) * @1Að0Þ þ ð1Þðvc Þ þ 1 ð0Þ ¼ vc ¼ & 1 e1 þ 2 e2 þ & 1 e3 þ 2 e4 þ 2 e5 & 1 e6 2 2 2 2 2 2 2 2 pﬃﬃﬃ 2 1 1 1 1 1 e7 & e8 þ e9 þ e10 & e11 & e12 & e13 & e14 : þ 2 2 2 2 2 2

Dummy Displacement Theorem. This method is usually used to find the applied external forces when the internal forces are known. In order to obtain the external force at a particular point, one subjects the structure to a unit displacement at that point in the direction of the force and chooses any set of deformations compatible with the unit displacement. Then from the principle of work, the dummy displacement theorem can be stated as:

16

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

b r1

r1 r2 r2

C

d

d C

P

Fig. 1.14 Element deformations equivalent to unit dummy displacement

8 9 8 9 < dummy displacement applied = < actual = in the direction of unknowns ! external : ; : ; actual external forces forces 8 9 8 9 < deformation of elements = < actual = ¼ compatible with ! internal : ; : ; dummy displacement forces

This method is also known as the unit displacement method.

Example 2. For the truss studied in Example 1, it is required to find the magnitude of P by measuring the internal forces in the members of the truss. Again, many systems can be chosen; two of which are illustrated in Fig. 1.14a, b. In these systems, the internal forces to be measured are shown in bold lines. Due to the symmetry, in both cases only two measurements are needed. Applying the dummy-displacement theorem to system (a) yields: Pd ¼ r1 d

1.3.4

pﬃﬃﬃ 2 2

þ r2 d þ r1 d

pﬃﬃﬃ 2 2

¼d

#pﬃﬃﬃ $ 2r 1 þ r 2 :

Contragradient Principle

Consider two statically equivalent force systems R and P, related by a linear transformation as: R ¼ BP,

ð1:27Þ

R is considered to have more entries than P, i.e. there are solutions to R for which P is zero. Associated with R and P let there be two sets of displacements v and u, respectively. These are compatible displacements and therefore the work done in each system is the same, i.e. Pt u ¼ Rt v: From Eq. 1.27,

ð1:28Þ

1.3 Important Structural Theorems

17

R t ¼ Pt B t :

ð1:29Þ

Pt u ¼ Pt Bt v:

ð1:30Þ

u ¼ Bt v:

ð1:31Þ

Therefore:

Since P is arbitrary, hence:

Equations 1.27 and 1.31 will be used in the formulation of the force method. In a general structure, if member forces R are related to external nodal loads P, similar to Eq. 1.27, then according to the contragradient principle [4], the member distortions v and nodal displacement u will be related by an equation similar to Eq. 1.31. If two displacement systems u and v are related by a linear transformation as, v ¼ Cu,

ð1:32Þ

and R and P are statically equivalent forces, then equating the work done for compatible displacements results in: Pt u ¼ Rt v ¼ Rt Cu:

ð1:33Þ

P ¼ Ct R:

ð1:34Þ

Again u is arbitrary and:

Equations 1.32 and 1.34 are employed in the formulation of the displacement method. For a statically determinate structure, P ¼ B&1 R,

ð1:35Þ

Ct ¼ B&1 :

ð1:36Þ

and therefore:

1.3.5

Reciprocal Work Theorem

Consider a structure as shown in Fig. 1.15a subjected to a set of loads, {P1, P2, . . ., Pm}. The same structure is considered under the action of a second set of loads {Q1, Q2, . . ., Qn}, Fig. 1.15b. The reciprocal work theorem can be stated as:

18

1

a P1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

b

D1

P2

D Dn

Pm

Q1

d1

D2

Q2 Q3

d2

Qn

dm

Fig. 1.15 A structure subjected to two sets of loads

The work done by {P1, P2, . . ., Pm} through displacements {δ1, δ2, . . ., δm} produced by {Q1, Q2, . . ., Qn}, is the same as the work done by {Q1, Q2, . . ., Qn} through displacements {Δ1, Δ2, . . ., Δn} produced by {P1, P2, . . ., Pm}; i.e.

m X i¼1

Pi δi ¼

n X

Qj Δj :

j¼1

ð1:37Þ

When single loads P and Q are considered, Eq. 1.37 reduces to, Pδi ¼ QΔj ,

ð1:38Þ

and for the case where P ¼ Q, one obtains: δi ¼ Δj :

ð1:39Þ

Equation 1.39 is known as Betti’s law, and can be stated as follows: The deflection at point i due to a load at point j is the same as deflection at j when the same load is applied at i.

The proof of the reciprocal work theorem is constructed by equating the strain energy of the structure in two different loading sequences [5]. In the first sequence, both sets of loads are applied simultaneously, while in the second sequence, loads {P1, P2, . . ., Pm} are applied first, followed by the application of the second set of loads {Q1, Q2, . . ., Qn}.

1.4

Basic Concepts and Definitions of Graph Theory

Some of the uses of the theory of graphs in the context of civil engineering are as follows: A graph can be a model of a structure, a hydraulic network, a traffic network, a transportation system, a construction system, or a resource allocation

1.4 Basic Concepts and Definitions of Graph Theory

19

system, for example. In this book, the theory of graphs is used as the model of a skeletal structure, and it is employed also as a way of transforming the connectivity properties of finite element meshes to those of graphs. Many such graphs are previously defined in [6], and employed throughout the combinatorial optimisations performed for optimal analysis of skeletal structures and finite element models. This part of the chapter will also enable the readers to develop their own ideas and methods in the light of the principles of graph theory. For further definitions and proofs, the reader may refer to Harary [7], Berge [8], and West [9].

1.4.1

Basic Definitions

The performance of a structure depends not only on the characteristics of its components, but also on their relative location. On the other hand, in a structure, if the properties of one member are altered, the overall behaviour may be changed. This indicates that the performance of a structure depends on the detailed characteristics of its members. On the other hand, if the location of a member is altered, the properties of the structure may again be different. Therefore, the connectivity (topology) of the structure influences the performance of the whole structure and is as important as the mechanical properties of its members. Hence, it is important to represent a structure so that its topology can be understood clearly. The graph model of a structure provides a powerful means for this purpose.

1.4.2

Definition of a Graph

A graph S consists of a non-empty set N(S) of elements called nodes (vertices or points) and a set M(S) of elements called members (edges or arcs), together with a relation of incidence which associates each member with a pair of nodes, called its ends. Two or more members joining the same pair of nodes are collectively known as a multiple member, and a member joining a node to itself is called a loop. A graph with no loops and multiple members is called a simple graph. If N(S) and M(S) are countable sets, then the corresponding graph S is finite. In this book, only finite graphs are needed, which are referred to as graphs. The above definitions correspond to abstract graphs; however, a graph may be visualised as a set of points connected by line segments in Euclidean space; the nodes of a graph are identified with points, and its members are identified as line segments without their end points. Such a configuration is known as a topological graph. These definitions are illustrated in Fig. 1.16.

20

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

Fig. 1.16 Simple and non-simple graphs. (a) A simple graph. (b) A graph with loop and multiple members

a

Fig. 1.17 A graph, two of its subgraphs, their union, intersection and ring sum. (a) S. (b) Si. (c) Sj. (d) L Si [ Sj Sj. (e) Si \ Sj. (f) Si

a

b

c

d

e

f

1.4.3

b

Adjacency and Incidence

Two nodes of a graph are called adjacent if these nodes are the end nodes of a member. A member is called incident with a node if this node is an end node of the member. Two members are called incident if they have a common end node. The degree (valency) of a node ni of a graph, denoted by deg(ni), is the number of members incident with that node. Since each member has two end nodes, the sum of node-degrees of a graph is twice the number of its members.

1.4.4

Graph Operations

A subgraph Si of S is a graph for which N(Si) * N(S) and M(Si) * M(S), and each member of Si has the same ends as in S. k

The union of subgraphs S1, S2, . . ., Sk of S, denoted by Sk ¼ [ Si ¼ S1 [ S2 [ i¼1

k

k

i¼1

i¼1

. . . [ Sk, is a subgraph of S with N(Sk) ¼ [ N(Si) and M(Sk) ¼ [ M(Si). The

intersection of two subgraphs Si and Sj is similarly defined using intersections of node-sets and member-sets of the two subgraphs. The intersection of two subgraphs does not need to consist only of nodes, but it is usually considered to do so in the L substructuring technique of structural analysis. The ring sum of two subgraphs Si Sj is a subgraph that contains the nodes and members of Si and Sj except those elements common to Si and Sj. These definitions are illustrated in Fig. 1.17. Two graphs S and K are called homeomorphic if one can obtain K from S, by suppressing or inserting nodes of degree 2 in the members.

1.4 Basic Concepts and Definitions of Graph Theory

21

a

Fig. 1.18 A walk, a trail and a path in S. (a) A walk w in S. (b) A trail t in S. (c) A path P in S

n4

3 n

1

b

n4

n1

1.4.5

6

9

8

1

n2

c

n5

n2

n5

4

n3

7 2

5 n

3

n4

n1

n5

n2

n3

Walks, Trails and Paths

A walk w of S is a finite sequence w ¼ {n0, m1, n1,. . ., mp, np} whose terms are alternately nodes ni and members mi of S for 1 + i + p, and ni&1 and ni are the two ends of mi. A trail t in S is a walk in which no member of S appears more than once. A path P is a trail in which no node appears more than once. The length of a path Pi, denoted by L(Pi), is taken as the number of its members. Pi is called the shortest path between the two nodes n0 and np, if for any other path Pj between these nodes, L(Pi) + L(Pj). The distance between two nodes of a graph is defined as the number of the members of a shortest path between these nodes. As an example, in Fig. 1.18, w ¼ ðn1 , m3 , n4 , m4 , n5 , m9 , n2 , m2 , n3 , m7 , n4 , m4 , n5 Þ is a walk between n1 and n5 in which member m4 and nodes n4 and n5 are repeated twice. t ¼ ðn1 , m3 , n4 , m4 , n5 , m9 , n2 , m2 , n3 , m7 , n4 Þ is a trail between n1 and n4 in which node n5 is repeated twice. P ¼ ðn1 ; m3 ; n4 ; m4 ; n5 ; m5 ; n3 Þ is a path of length 3 in which no node and no member is repeated. The path (n1, m6, n5, m5, n3) is a shortest path of length 2 between the two nodes n1 and n3, where the length of each member is taken as unity. Two nodes ni and nj are said to be connected in S if there exists a path between these nodes. A graph S is called connected if all pairs of its nodes are connected. A component of a graph S is a maximal connected subgraph, i.e. it is not a subgraph of any other connected subgraph of S.

22

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

Fig. 1.19 Cycles of S. (a) A cycle of S. (b) A hinged cycle of S

a

5

4 3 1

b

6 6

c

7 5 4

2

1

b

2

3

Fig. 1.20 Cutsets of S. (a) A cutset of S. (b) A prime cutest. (c) A cocycle of S

a

b

c

d

e

f

n0 Fig. 1.21 Different trees, a cotree and a forest of S. (a) A graph S. (b) A tree of S. (c) A spanning tree of S. (d) An SRT rooted from n0. (e) The cotree of (c). (f) A forest with two trees

1.4.6

Cycles and Cutsets

A cycle is a path (n0, m1, n1, . . ., mp, np) for which n0 ¼ np and p , 1; i.e. a cycle is a closed path. Similarly, a closed trail (hinged cycle) and a closed walk can be defined, Fig. 1.19. A cutset is a collection of members whose removal from the graph increases the number of its components. If a cutset results in two disjoint subgraphs S1 and S2, then it is called a prime cutset. Notice that no proper subsets of a cutset have this property. A link is a member which has its ends in S1 and S2. Each S1 and S2 may or may not be connected. If both are connected, the cutset is called prime. If one of S1 or S2 consists of a single node, the cutset is called a cocycle. These definitions are illustrated in Fig. 1.20.

1.4 Basic Concepts and Definitions of Graph Theory

1.4.7

23

Trees, Spanning Trees and Shortest Route Trees

A tree T of S is a connected subgraph of S which contains no cycle. A set of trees of S forms a forest. Obviously a forest with k trees contains N(S) & k members. If a tree contains all the nodes of S, it is called a spanning tree of S. Henceforth, for simplicity it will be referred to as a tree. A shortest route tree (SRT) rooted at a specified node n0 of S, is a tree for which the distance between every node nj of T and n0 is a minimum. An SRT of a graph can be generated by the following simple algorithm: Algorithm. Label the selected root n0 as “0” and the adjacent nodes as “1”. Record the members incident to “0” as tree members. Repeat the process of labelling with “2” the unnumbered ends of all the members incident with nodes labelled as “1”, again recording the tree members. This process terminates when each node of S is labelled and all the tree members are recorded. This algorithm has many applications in engineering, and it is called a breadth-first-search algorithm. A graph is called acyclic if it has no cycle. A tree is a connected acyclic graph. Any graph without cycles is a forest, thus the components of a forest are trees. The above definitions are illustrated in Fig. 1.21. It is easy to prove that, for a tree T, MðTÞ ¼ NðTÞ & 1,

ð1:40Þ

where M(T) and N(T) are the numbers of members and nodes of T, respectively. The complement of T in S is called a cotree, denoted by T*. The members of T are known as branches and those of T* are called chords. For a connected graph S, the number of chords is given by: MðT) Þ ¼ MðSÞ & MðTÞ:

ð1:41Þ

Since N(T) ¼ N(S), hence, MðT) Þ ¼ MðSÞ & NðSÞ þ 1,

ð1:42Þ

where M(S) and N(S) are the numbers of members and nodes of S, respectively. Notice that for a set and its cardinality the same notation is used and the difference should be obvious from the context.

1.4.8

Different Types of Graphs

In order to simplify the study of the properties of graphs, different types of graphs have been defined. Some important ones are as follows:

24

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

b

c

Fig. 1.22 Wheel graph W6. (a) Star graph S6. (b) Cycle graph C5. (c) Wheel graph W6

K1

K2

K3

K4

K5

Fig. 1.23 Five complete graphs

a

b

N1

N2

N1

N2

Fig. 1.24 Two bipartite graphs. (a) A bipartite graph. (b)A complete bipartite graph K3, 4

a

b

Fig. 1.25 A simple graph and its line graph. (a) A graph S. (b) The line graph L(S) of S

A null graph is a graph that contains no members. Thus, Nk is a graph containing k isolated nodes. A cycle graph is a graph consisting of a single cycle. Therefore, Ck is a polygon with k members. A path graph is a graph consisting of a single path. Hence, Pk is a path with k nodes and (k&1) members.

1.5 Vector Spaces Associated with a Graph

25

A wheel graph Wk is defined as the union of a star graph with k&1 members and a cycle graph Ck&1, connected as shown in Fig. 1.22, for k ¼ 6. Alternatively, a wheel graph Wk can be obtained from the cycle graph Ck&1 by adding a node O and members (spokes) joining O to each node of Ck&1. A complete graph is a graph in which every two distinct nodes are connected by exactly one member, Fig. 1.23. A complete graph with N nodes is denoted by KN. It is easy to prove that a complete graph with N nodes has N(N&1)/2 members. A graph is called bipartite if the corresponding node set can be split into two sets N1 and N2 in such a way that each member of S joins a node of N1 to a node of N2. This graph is denoted by B(S) ¼ (N1, M, N2). A complete bipartite graph is a bipartite graph in which each node N1 is joined to each node of N2 by exactly one member. If the numbers of nodes in N1 and N2 are denoted by r and s, respectively, then a complete bipartite graph is denoted by Kr,s. Examples of bipartite and complete bipartite graphs are shown in Fig. 1.24. A graph S is called regular if all of its nodes have the same degree. If this degree is k, then S is k-regular graph. As an example, a triangle graph is 2-regular and a cubic graph is 3-regular. Consider the set M of members of a graph S as a family of 2-node subsets of N (S). The line graph L(S) of S has its vertices in a one-to-one correspondence with members of S, and two vertices are connected by an edge if the corresponding members in S are incident. Thus the vertices of L(S) are the members of S, with two vertices of L(S) being adjacent when the corresponding members of S are incident. As an example, the line graph of Fig. 1.25a is illustrated in Fig. 1.25b. For the original graph S, the terms nodes and members are used, and for the line graph L(S), the terms vertices and edges are employed. In this book, many new graphs are defined and employed for transforming the connectivity properties of the original models to those of the induced new graphs.

1.5

Vector Spaces Associated with a Graph

A vector space can be associated with a graph by defining a vector, the field and the binary operations as follows: Any subset of the M(S) members of a graph S can be represented by a vector x whose M(S) components are elements of the field of integer modulo 2, where component xi ¼ 1 when the ith member is an element of the subset, and xi ¼ 0 otherwise. The sum of two subset vectors x and y is a vector z with entries defined by zi ¼ xi + yi, representing the symmetric difference of the original subsets. The scalar product of x and y defined by Σxiyi is 0 or 1 according as the original subsets have an even or an odd number of members in common. Although this vector space can be constructed over an arbitrary field, for simplicity the field of integer modulo 2 is considered, in which 1 + 1 ¼ 0.

26

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

As an example, consider x ¼ {0, 0, 0, 1, 1, 1, 0}t and y ¼ {0, 0, 1, 1, 1, 0, 0}t representing two subgraphs of S. Then, their symmetric difference is obtained as z ¼ {0, 0, 1, 0, 0, 1, 0}t, and the scalar product Σxiyi ¼ 0 (mod 2), since these subgraphs have two members in common. Two important subspaces of the above vector space of a graph S are the cycle subspace and cutset subspace, known as the cycle space and the cutset space of S.

1.5.1

Cycle Space

Let a cycle set of members of a graph be defined as a set of members which form a cycle or form several cycles having no common member, but perhaps common nodes. The null set is also defined as a cycle set. A vector representing a cycle set is called a cycle set vector. It can be shown that the sum of two cycle set vectors of a graph is also a cycle set vector. Thus, the cycle set vectors of a graph form a vector space over the field of integer modulo 2. The dimension of a cycle space is given by: nullity ðSÞ ¼ νðSÞ ¼ b1 ðSÞ ¼ MðSÞ & NðSÞ þ b0 ðSÞ,

ð1:43Þ

where b1(S) and b0(S) are the first and zero Betti numbers of S, respectively. As an example, the nullity of the graph S in Fig. 1.16a is ν(S) ¼ 9 & 6 + 1 ¼ 4.

1.5.2

Cutset Space

Consider a cutset vector similar to that of a cycle vector. Let the null set be also defined as a cutset. It can be shown that the sum of two cutset vectors of a graph is also a cutset vector. Therefore the cutset vectors of a graph form a vector space, the dimension of which is given by: rank ðSÞ ¼ ρðSÞ ¼ NðSÞ & b0 ðSÞ:

ð1:44Þ

As an example, the rank of S in Fig. 1.16a is ρ(S) ¼ 6 & 1 ¼ 5.

1.5.3

Orthogonality Property

Two vectors are called orthogonal if their scalar product is zero. It can be shown that a vector is a cycle set (cutset) vector, if and only if it is orthogonal to every vector of a cutset (cycle set) basis. Since the cycle set and cutset spaces of a graph S containing M(S) members are both subspaces of the M(S)-dimensional space of all

1.5 Vector Spaces Associated with a Graph

27

vectors which represent subsets of the members, therefore the cycle set and cutset spaces are orthogonal components of each other.

1.5.4

Fundamental Cycle Bases

A maximal set of independent cycles of a graph is known as its cycle basis. The cardinality of a cycle basis is the same as the first Betti number b1(S). A special basis known as a fundamental cycle basis can easily be constructed corresponding to a tree T of S. In a connected S, a chord of T together with T contains a cycle known as a fundamental cycle of S. Moreover, the fundamental cycles obtained by adding the chords to T, one at a time, are independent, because each cycle has a member which is not in the others. Also, every cycle Ci depends on the set of fundamental cycles obtained by the above process, for Ci is the symmetric difference of the cycles determined by the chords of T which lie in Ci. Thus the cycle rank (cyclomatic number, first Betti number, nullity) of graph S, which is the number of cycles in a basis of the cycle space of S, is given by, b1 ðSÞ ¼ MðSÞ & NðSÞ þ 1,

ð1:45Þ

and if S contains b0(S) components, then: b1 ðSÞ ¼ MðSÞ & NðSÞ þ b0 ðSÞ:

ð1:46Þ

As an example, the selected tree and three fundamental cycles of S are illustrated in Fig. 1.26.

1.5.5

Fundamental Cutset Bases

A basis can be constructed for the cutset space of a graph S. Consider the tree T and its cotree T*. The subgraph of S consisting of T* and any member of T (branch) contains exactly one cutset known as a fundamental cutset. The set of cutsets obtained by adding branches of T to T*, one at a time, forms a basis for the cutset space of S, known as a fundamental cutset basis of S. The cutset rank (rank of S) is the number of cutsets in a basis for the cutset space of S, and it can be obtained by a similar reasoning to that of the cycle basis as, ρðSÞ ¼ NðSÞ & 1, and for a graph with b0(S) components:

ð1:47Þ

28

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

Fig. 1.26 A graph S and a fundamental cycle basis of S

a

b

c

Fig. 1.27 A graph S and a fundamental cutset basis of S. (a) A graph S. (b) A tree T of S. (c) Cotree T* of T

ρðSÞ ¼ NðSÞ & b0 ðSÞ:

ð1:48Þ

A graph S and a fundamental cutset basis of S are shown in Fig. 1.27.

1.6

Matrices Associated with a Graph

Matrices play a dominant role in the theory of graphs and especially in applications to structural analysis. Some of these matrices conveniently describe the connectivity properties of a graph and others provide useful information about the patterns of the structural matrices, and some reveal additional information about transformations such as those of equilibrium and compatibility equations. In this section various matrices are studied which reflect the properties of the corresponding graphs. For simplicity, all graphs are assumed to be connected, since

1.6 Matrices Associated with a Graph

29 5

⎡0 ⎢1 ⎢ A = ⎢1 ⎢ ⎢1 ⎢⎣0

1 1 1 0⎤ 0 1 1 0⎥⎥ 1 0 0 1⎥ . ⎥ 1 0 0 1⎥ 0 1 1 0⎥⎦

6

(A-10)

3

7 4

4

3

5

2 1

1

2

Fig. 1.28 A graph S

the generalisation to non-connected graphs is trivial and consists of considering the direct sum of the matrices for their components.

1.6.1

Matrix Representation of a Graph

A graph can be represented in various forms. Some of these representations are of theoretical importance, others are useful from the programming point of view when applied to realistic problems. In this section six different representations of a graph are described. Node Adjacency Matrix. Let S be a graph with N nodes. The adjacency matrix A is an N ! N matrix in which the entry in row i and column j is 1 if node ni is adjacent to nj, and is 0 otherwise. This matrix is symmetric, and the row sums of A are the degrees of the nodes of S. The adjacency matrix of the graph S, shown in Fig. 1.28, is a 5 ! 5 matrix as: It can be noted that A is a symmetric matrix of trace zero. The (i, j)th entry of A2 shows the number of walks of length 2 with ni and nj as end nodes. Similarly, the entry in the (i, j) position of Ak is equal to the number of walks of length k with ni and nj as end nodes. The polynomial, f ðλÞ ¼ det ðIλ & AÞ,

ð1:49Þ

is called the characteristic polynomial of S. The collection of N(S) eigenvalues of A is known as the spectrum of S. Since A is symmetric, the spectrum of S consists of N(S) real numbers. The sum of eigenvalues of A is equal to zero. Node-Member Incidence Matrix. Let S be a graph with M members and N nodes. The node-member incidence matrix B is an N ! M matrix in which the entry in row i and column j is 1 if node ni is incident with member mj, and is 0 otherwise. As an example, the node-member incidence matrix of the graph in Fig. 1.28 is a 5 ! 7 matrix of the form:

30

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

2

1 61 6 B¼6 60 40 0

1 0 1 0 0

0 1 1 0 0

1 0 0 1 0

0 1 0 1 0

3 0 07 7 07 7: 15 1

0 0 1 0 1

ð1:50Þ

Obviously, the pattern of an incidence matrix depends on the particular way in which its nodes and members are labelled. One incidence matrix can be obtained from another by simply interchanging rows (corresponding to re-labelling the nodes) and columns (corresponding to re-labelling the members). The incidence matrix B and the adjacency matrix A of a graph S are related by, t

B B ¼ A þ V,

ð1:51Þ

where V is a diagonal matrix of order N(S) whose typical entry vi is the valency of the node ni of S for i ¼ 1, . . ., N(S). For the example of Fig. 1.28, Eq. 1.51 becomes: 2

0 61 6 t BB ¼ 6 61 41 0

1 0 1 1 0

1 1 0 0 1

1 1 0 0 1

3 2 3 0 6 07 7 6 6 17 7þ6 15 4 0

3

3 3 3 2

7 7 7: 7 5

ð1:52Þ

The rows of B are dependent, and one row can arbitrarily be deleted to ensure the independence of the rest of the rows. The node corresponding to the deleted row is called a datum (reference) node. The matrix obtained after deleting a dependent row is called an incidence matrix of S, and is denoted by B. Although A and B are of great theoretical value, the storage requirements for these matrices are high and proportional to N ! N and M ! N, respectively. In fact, a large number of unnecessary zeros is stored in these matrices. In practice, one can use different approaches to reduce the storage required, some of which are described in the following. Member List: This type of representation is a common approach in structural mechanics. A member list consists of two rows (or columns) and M columns (or rows). Each column (or row) contains the labels of the two end nodes of each member, in which members are arranged sequentially. For example, the member list of S in Fig. 1.28 is:

1.6 Matrices Associated with a Graph

% m1 m2 m3 m4 m5 m6 m&7 1 1 2 1 2 3 4 ML ¼ : 2 3 3 4 4 5 5

31

ð1:53Þ

It should be noted that a member list can also represent orientations on members. The storage required for this representation is 2 ! M. Some engineers prefer to add a third row containing the member’s labels, for easy addressing. In this case, the storage is increased to 3 ! M. A different way of preparing a member list is to use a vector containing the end nodes of members sequentially; e.g. for the previous example this vector becomes: ð1; 2; 1; 3; 2; 3; 1; 4; 2; 4; 3; 5; 4; 5Þ:

ð1:54Þ

This is a compact description of a graph; however, it is impractical because of the extra search required for its use in various algorithms. Adjacency List. This list consists of N rows and D columns, where D is the maximum degree of the nodes of S. The ith row contains the labels of the nodes adjacent to node i of S. For the graph S shown in Fig. 1.28, the adjacency list is: 3 2 n1 2 3 4 7 n2 6 61 3 47 6 AL ¼ n3 6 1 2 5 7 ð1:55Þ 7 n4 4 1 2 5 5 n5 3 4 N!D The storage needed for an adjacency list is N ! D. Compact Adjacency List. In this list, the rows of AL are continually arranged in a row vector R, and an additional vector of pointers P is considered. For example, the compact adjacency list of Fig. 1.28 can be written as: R ¼ ð2; 3; 4; 1; 3; 4; 1; 2; 5; 1; 2; 5; 3; 4Þ, P ¼ ð1; 4; 7; 10; 13; 15Þ:

ð1:56Þ

P is a vector (p1, p2, p3, . . .) which helps to list the nodes adjacent to each node. For node ni, one should start reading R at entry pi and finish at pi+1&1. An additional restriction can be put on R, by ordering the nodes adjacent to each node ni in ascending order of their degrees. This ordering can be of some advantage, an example of which is nodal ordering for bandwidth optimisation. The storage required for this list is 2M + N + 1.

32

1.6.2

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

Cycle Bases Matrices

The cycle-member incidence matrix C of a graph S, has a row for each cycle or hinged cycle and a column for each member. An entry cij of C is 1 if cycle Ci contains member mj, and it is 0 otherwise. In contrast to the node adjacency and node-member incidence matrices, the cycle-member incidence matrix does not determine a graph up to isomorphism; i.e. two totally different graphs may have the same cycle-member incidence matrix. For a graph S, there exist 2b1 ðSÞ & 1 cycles or hinged cycles. Thus C is a (2b1 ðSÞ & 1) ! M matrix. However, one does not need all the cycles of S, and the elements of a cycle basis are sufficient. For a cycle basis, a cycle-member incidence matrix becomes a b1(S) ! M matrix, denoted by C, known as the cycle basis incidence matrix of S. As an example, matrix C for the graph shown in Fig. 1.28, for the following cycle basis, C1 ¼ ðm1 , m2 , m3 Þ C2 ¼ ðm1 , m4 , m5 Þ

C3 ¼ ðm2 , m4 , m6 , m7 Þ is given by: 2 C1 1 C ¼ C2 4 1 C3 0

1 1 0 0 1 0

0 1 1

0 1 0

0 0 1

3 0 0 5: 1

ð1:57Þ

The cycle adjacency matrix D is a b1(S) ! b1(S) matrix, each entry dij of which is 1 if Ci and Cj have at least one member in common and it is 0 otherwise. This matrix is related to the cycle-member incidence matrix by the following relationship, CCt ¼ D þ W,

ð1:58Þ

where W is a diagonal matrix with wii being the length of the ith cycle, and its trace being equal to the total length of the cycles of the basis. For the above example: 2

0 CCt ¼ 4 1 1

3 2 1 1 3 0 15 þ 40 1 0 0

3 0 0 3 0 5: 0 3

ð1:59Þ

An important theorem can now be stated which is based on the orthogonality property studied in Sect. 1.5.3.

1.6 Matrices Associated with a Graph

33 10

Fig. 1.29 A graph with oriented members and cycles

3

2 1

C3 9 C2 8 C1

7

6 5

4

Theorem. Let S have an incidence matrix B and a cycle basis incidence matrix C. Then: CBt ¼ 0ðmod 2Þ:

ð1:60Þ

A simple proof of this theorem can be found in Kaveh [10]. Notice that Eq. 1.60 holds due to the orthogonality property discussed in Sect. 1.5.3. In fact, the above relation holds even if the cutsets or cycles do not form bases, or the matrices contain additional cutsets and/or cycle vectors.

1.6.3

Special Patterns for Fundamental Cycle Bases

Matrix C for a fundamental cycle basis, with special labels for its tree members and chords, finds a particular pattern. Let S have a tree T whose members are M (T) ¼ (m1, m2, . . ., mp) and a cotree for which M(T*) ¼ (mp+1, mp+2, . . ., mM(S)). Then there is a unique fundamental cycle Ci in S & M(T*) + mi, p + 1 + i + M(S), and this set of cycles forms a basis for the cycle space of S. As an example, for the graph S of Fig. 1.27a whose members are labelled as shown in Fig. 1.29, the fundamental cycle basis consists of, C1 ¼ ðm1 , m4 , m5 , m8 Þ,

C2 ¼ ðm2 , m1 , m4 , m5 , m6 , m9 Þ,

C3 ¼ ðm3 , m2 , m1 , m4 , m5 , m6 , m7 , m10 Þ, given by: 2 C1 1 C ¼ C2 4 1 C3 1

0 0 1 1 0 1 1 1 1 M ð TÞ

1 1 1

+ 3 0 0 ++ 1 0 0 1 0 ++ 0 1 0 5 ¼ ½CT jI(: 1 1+0 0 1 M ð T) Þ

ð1:61Þ

34

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

Fig. 1.30 A graph with oriented members and a cutset basis

b 10

3

9

2

8

7

3

7

6

2

6

1

5

5

1 4

1.6.4

4

Cutset Bases Matrices

) The cutset-member incidence matrix C for a graph S, has a row for each cutset of S ) and a column for each member. An entry c)ij of C is 1 if cutset C)i contains member mj, and it is 0 otherwise. This matrix, like C, does not determine a graph completely. ∗ Independent rows of C for a cutset basis, denoted by C*, form a matrix known as a cutset basis incidence matrix, which is a η(S) ! M matrix, η(S) being the rank of graph S. As an example, C* for the cutsets of Fig. 1.27 with members labelled as in Fig. 1.30a, is given below: 2

0 60 6 60 6 ) C ¼6 60 60 6 41 0

0 0 1 0 0 0 0

1 0 0 0 0 0 0

0 0 0 0 1 0 0

0 0 0 0 0 0 1

0 0 0 1 0 0 0

0 1 0 0 0 0 0

0 0 0 0 1 1 1

0 0 1 1 1 1 1

3 1 17 7 17 7 17 7: 17 7 15 1

ð1:62Þ

The cutset adjacency matrix D* is a η(S) ! η(S) matrix defined analogously to cycle adjacency matrix D.

1.6.5

Special Patterns for Fundamental Cutset Bases

For a fundamental cutset basis with appropriate labelling of the members in T and T*, as illustrated in Fig. 1.30b, if the cutsets are taken in the order of their generators (tree members), the matrix C* will have a particular pattern as:

1.7 Directed Graphs and Their Matrices

2

1 60 6 60 6 C)0 ¼ 6 60 60 6 40 0

0 1 0 0 0 0 0

0 0 1 0 0 0 0

0 0 0 1 0 0 0

35

0 0 0 0 1 0 0

0 0 0 0 0 1 0

+ 0 ++ 1 0 ++ 0 0 ++ 0 0 ++ 1 0 ++ 1 0 ++ 0 1+0

1 1 0 1 1 1 0

3 1 17 7 17 7 , + )+ 17 7 ¼ I Cc : 7 17 15 1

ð1:63Þ

From the orthogonality condition, C0C∗t 0 ¼ 0; i.e. %

I I ( ∗t Cc

½ CT

&

¼ 0:

ð1:64Þ

Hence CT þ C∗t c ¼ 0ðmod 2Þ, and : CT ¼ C)t c :

ð1:65Þ

Therefore, for a graph having C0, one can construct C∗ 0 and vice versa. There exists a very simple basis for the cutset space of a graph which consists of N&1 cocycles of S. As an example, for the graph of Fig. 1.28, considering n5 as a datum node, we have, 2

1 61 ∗ C ¼6 40 0

1 0 1 0

0 1 1 0

1 0 0 1

0 1 0 1

0 0 1 0

3 0 07 7, 05 1

ð1:66Þ

which is the same as the incidence matrix B of S. The simplicity of the displacement method of structural analysis is due to the existence of such a simple basis.

1.7

Directed Graphs and Their Matrices

An oriented or directed graph is a graph in which each member is assigned an orientation. A member is oriented from its initial node (tail) to its final node (head). The initial node is said to be positively incident on the member, and the final node negatively incident, as shown in Fig. 1.31a. The choice of orientation of members of a graph is arbitrary; however, once it is chosen, it must be retained. Cycles and cutsets can also be oriented as shown in Fig. 1.31b. As an example, m4 is positively oriented in cycle Ci, and m7 is negatively oriented in cutset C∗ i . All the matrices B, B, C and C* can be defined as before, with the difference of having +1, &1 and 0 as entries, according to whether the member is positively, negatively or zero incident with a cutset or a cycle.

36

1

Basic Definitions and Concepts of Structural Mechanics and Theory of Graphs

a

b

Ci

nj

c 5

4

3

5

Ci

2

ni

7

7

4

6

5

6

3 1

2

Fig. 1.31 An oriented member, a directed graph, and a directed tree (with chords shown in dashed lines)

As an example, for graph S in Fig. 1.31b, the matrix B with n1 as the datum node is formed: 2 n1 &1 n 6 0 B¼ 26 n3 4 0 n4 0

0 &1 0 0

1 0 &1 0

0 0 0 &1

1 &1 0 0

0 0 &1 0

3 0 0 7 7: 1 5 &1

ð1:67Þ

Consider a tree as shown in continuous lines, Fig. 1.31c. When the directions of the cycles are taken as those of their corresponding chords (dashed lines), the fundamental cycle basis incidence matrix can be written as: + 2 3 C1 1 &1 0 0 ++ 1 0 0 ð1:68Þ C ¼ C2 4 1 0 1 0 ++ 0 1 0 5: C3 1 &1 1 &1 + 0 0 1

It should be noted that the tree members are numbered first, followed by the chords of the cycles in the same sequence as their generation. Obviously, BCt ¼ CBt ¼ 0ðmod 2Þ,

with a proof similar to that of the non-oriented case. A cuset basis incidence matrix is similarly obtained as: + 2 3 1 0 0 0 ++ &1 1 &1 60 1 0 0+ 1 0 1 7 + 7 C) ¼ 6 4 0 0 1 0 + 0 1 &1 5, + 0 0 0 1+ 0 0 1 C)T C)c

ð1:69Þ

ð1:70Þ

where the direction of a cutset is taken as the orientation of its generator (the corresponding tree member).

References

37

It can easily be proven that: CT ¼ &C)t c :

ð1:71Þ

For a directed graph, Eq. 1.51 becomes: BBt ¼ A & V,

ð1:72Þ

Similarly, Eq. 1.59 for the directed case becomes: CCt ¼ D & W:

ð1:73Þ

References 1. Kaveh A (2004) Structural mechanics: graph and matrix methods, 3rd edn. Research Studies Press, Baldock, Herdfordshire, UK 2. Kaveh A (1979) Optimal structural analysis, 1st edn. Research Studies Press, Chicherster, Herdfordshire, UK 3. Castigliano A (1879) The´orie de l’e´quilibre des cutset e´lastiques et ses applications. AF Negro, Turin 4. Argyris JH, Kelsey S (1960) Energy theorems and structural analysis. Butterworth, London 5. Timoshenko S, Young DH (1945) Theory of structures. McGraw-Hill, New York 6. Kaveh A, Roosta GR (1998) Comparative study of finite element nodal ordering methods. Eng Struct 20(1–2):86–96 7. Harary F (1969) Graph theory. Addison-Wesley, Reading 8. Berge C (1973) Graphs and hypergraphs. North-Holland Publishing, Amsterdam 9. West DB (1996) Introduction to graph theory. Prentice-Hall, Upper Saddle River, NJ, USA 10. Kaveh A (1974) Application of topology and matroid theory to the analysis of structures. Ph.D. thesis, Imperial College, London University

Chapter 2

Optimal Force Method: Analysis of Skeletal Structures

2.1

Introduction

This chapter starts with presenting simple and general methods for calculating the degree of static indeterminacy of different types of skeletal structures, such as rigidjointed planar and space frames, pin-jointed planar trusses and ball-jointed space trusses. Then the progress made in the force method of structural analysis in recent years is presented, and the state of art is summarized. Efficient methods are developed leading to highly sparse flexibility matrices. The methods are mainly developed for frame structures, however, extensions are made to general skeletal structures. The force method of structural analysis, in which the member forces are used as unknowns, is appealing to engineers, since the properties of members of a structure most often depend on the member forces rather than joint displacements. This method was used extensively until 1960. After this, the advent of the digital computer and the amenability of the displacement method for computation attracted most researchers. As a result, the force method and some of the advantages it offers in optimisation and non-linear analysis, have been neglected. Six different approaches are adopted for the force method of structural analysis, which will be classified as follows: 1. 2. 3. 4. 5. 6.

Topological force methods, Combinatorial force methods, Algebraic force methods, Mixed algebraic-combinatorial force methods, Integrated force method, and Metaheuristic based methods.

Topological methods have been developed by Henderson [1], Maunder [2] and Kaveh [3]. Combinatorial force method is mainly developed by Kaveh [3] using different graph theoretical algorithms. Algebraic topology is employed extensively in the work of Langefors [4]. Algebraic methods have been developed by Denke [5], A. Kaveh, Computational Structural Analysis and Finite Element Methods, DOI 10.1007/978-3-319-02964-1_2, © Springer International Publishing Switzerland 2014

39

40

2 Optimal Force Method: Analysis of Skeletal Structures

Robinson [6], Topc¸u [7], and Kaneko et al. [8], and mixed algebraic-topological methods have been used by Gilbert et al. [9], Coleman and Pothen [10]. The integrated force method has been developed by Patnaik [11]. Meta-heuristic based methods are also developed for the formation of null basis in the work of Kaveh and Jahamshahi [12] and Kaveh and Daei [13].

2.2

Static Indeterminacy of Structures

Skeletal structures are the most common type of structures encountered in civil engineering practice. These structures sustain the applied loads mainly by virtue of their topology, i.e. the way members are connected to each other (connectivity). Therefore, topology plays a vital role in their design. The first step in design of such structures is to provide sufficient rigidity and make it reliable, but this depends in part on the degrees of static indeterminacy of the structures. One way to calculate the degree of static indeterminacy is to use classical formulae such as those given in Timoshenko and Young [14]; however, the application of these usually provides only a small part of the necessary topological properties. The methods presented in this chapter provide powerful means for understanding the distribution of the indeterminacy within a structure. The concepts presented are efficient in both the optimal force method of structural analysis, as will be discussed in the second part of this chapter. In the analysis of skeletal structures, three different properties are encountered, which are classified as topological, geometrical and material. Separate study of these properties results in a considerable simplification in understanding the structural behaviour leading to methods for efficient analysis. This chapter is confined to a study of those topological properties of skeletal structures needed to study force and displacement methods. The number of equations to be solved in the two methods may differ widely for the same structure. This number depends on the size of the flexibility and the stiffness matrices. The orders of this matrix are the same as the degree of static indeterminacy and the degree of kinematic indeterminacy of a structure, respectively. Obviously, the method that leads to the required results with the least amount of computational time and storage should be used for the analysis of a given structure. Thus, the comparison of the degree of static indeterminacy and the degree of kinematic indeterminacy may be the main criterion for selecting the method of analysis. The degree of kinematic indeterminacy of a structure, also known as its total number of degrees of freedom, can easily be obtained by summing up the degrees of freedom of its nodes. A node of planar and space trusses has two and three degrees of freedom, respectively. For planar and space frames, these numbers are 3 and 6, respectively. Single-layer grids have also three degrees of freedom for each node. For determining the degree of static indeterminacy of structures, numerous formulae depending on the kinds of members or types of joints have been given, e.g. Ref. [14]. The use of these classical formulae, in general, requires counting the

2.2 Static Indeterminacy of Structures

41

number of members and joints, which becomes a tedious process for multi-member and/or complex pattern structures; moreover, the count provides no additional information about connectivity. Henderson and Bickley [1] related the degree of static indeterminacy of a rigidjointed frame to the first Betti number of its graph model S. Generalising the Betti’s number to a linear function and using an expansion process, Kaveh [15] developed a general method for determining the degree of static indeterminacy and degree of kinematic indeterminacy of different types of skeletal structures. Special methods have also been developed to transform the topological properties of space structures to those of their planar drawings, in order to simplify the calculation of their degrees of static indeterminacy, Ref. [16]. It should be noted that various methods for determining the degree of static indeterminacy of structures are a by-product of the general methods developed by Kaveh [15]. The method of expansion and its control at each step, using the intersection theorem presented in this chapter, provides a powerful tool for further studies in the field of structural analysis.

2.2.1

Mathematical Model of a Skeletal Structure

The mathematical model of a structure is considered to be a finite, connected graph S. There is a one-to-one correspondence between the elements of the structure and the members (edges) of S. There is also a one-to-one correspondence between the joints of the structure and the nodes of S, except for the support joints of some models. For frame structures, shown in Fig. 2.1(a1) and (a2), two graph models can be considered. For the first model, all the support joints are identified as a single node called a ground node, as shown in Fig. 2.1(b1) and (b2). For the second model, all the joints are connected by an artificial arbitrary spanning tree, termed ground tree, Fig. 2.1(c1) and (c2). Truss structures shown in Fig. 2.2(a1) and (a2) are assumed to be supported in a statically determinate fashion (Fig. 2.2(b1) and (b2)), and the effect of additional supports can easily be included in calculating the degree of static indeterminacy (DSI) of the corresponding structures. Alternatively artificial members can be added as shown in Fig. 2.2(c1) and (c2) to model the components of the corresponding supports. For a fixed support, two members and three members are considered for planar and space trusses, respectively, and one member is used for representing a roller. The skeletal structures are considered to be in perfect condition; i.e. planar and space trusses have pin and ball joints only. Obviously, the effect of extra constraints or releases can be taken into account in determining their degrees of static indeterminacy and also in their analysis, Mauch and Fenves [17].

42

2 Optimal Force Method: Analysis of Skeletal Structures

a1

b1

c1

a2

b2

c2

Fig. 2.1 Frame structures and their mathematical models. (a1) A plane frame. (b1) First model with a ground node. (c1) Second model with a ground tree. (a2) A space frame. (b2) First model with a ground node. (c2) Second model with a ground tree

a1

b1

c1

a2

b2

c2

Fig. 2.2 Trusses and their graph models. (a1) A plane truss. (b1) First model without added members. (c1) Second model with replaced members. (a2) A space truss. (b2) First model without added members. (c2) Second model with replaced members

2.2.2

Expansion Process for Determining the Degree of Static Indeterminacy

The degree of kinematic indeterminacy of a structure is the number of independent displacement components (translations and rotations) required for describing a general state of deformation of the structure. The degree of kinematic indeterminacy is also referred to as the total degrees of freedom of the structure. On the other hand, the degree of static indeterminacy (redundancy) of a structure is the number of independent force components (forces and moments) required for describing a general equilibrium state of the structure. The DSI of a structure can be obtained by subtracting the number of independent equilibrium equations from the number of its unknown forces.

2.2 Static Indeterminacy of Structures

2.2.2.1

43

Classical Formulae

Formulae for calculating the DSI of various skeletal structures can be found in textbooks on structural mechanics, e.g. the DSI of a planar truss, denoted by γ(S), can be calculated from, γðSÞ ¼ MðSÞ $ 2NðSÞ þ 3,

ð2:1Þ

where S is supported in a statically determinate fashion (internal indeterminacy). For extra supports (external indeterminacy), γ(S) should be further increased by the number of additional unknown reactions. A similar formula holds for space trusses: γðSÞ ¼ MðSÞ $ 3NðSÞ þ 6:

ð2:2Þ

For planar and space frames, the classical formulae is given as, γðSÞ ¼ α½MðSÞ $ NðSÞ þ 1',

ð2:3Þ

where all supports are modelled as a datum (ground) node, and α ¼ 3 or 6 for planar and space frames, respectively. All these formulae require counting a great number of members and nodes, which makes their application impractical for multi-member and complex pattern structures. These numbers provide only a limited amount of information about the connectivity properties of structures. In order to obtain additional information, the methods developed in the following sections will be utilised:

2.2.2.2

A Unifying Function

All the existing formulae for determining DSI have a common property, namely their linearity with respect to M(S) and N(S). Therefore, a general unifying function can be defined as, γðSÞ ¼ aMðSÞ þ bNðSÞ þ cγ0 ðSÞ,

ð2:4Þ

where M(S), N(S) and γ0(S) are the numbers of members, nodes and components of S, respectively. The coefficients a, b and c are integer numbers depending on both the type of the corresponding structure and the property which the function is expected to represent. For example, γ(S) with appropriate values for a, b and c may describe the DSI of certain types of skeletal structures, Table 2.1. For a ¼ 1, b ¼ $1 and c ¼ 1, γ(S) becomes the first Betti number b1(S) of S, as described in Sect. 1.5.1.

44

2 Optimal Force Method: Analysis of Skeletal Structures

Table 2.1 Coefficients of γ(S) for different types of structures

2.2.2.3

Type of structure Plane truss Space truss Plane frame Space frame

a +1 +1 +3 +6

b $2 $3 $3 $6

c +3 +6 +3 +6

An Expansion Process

An expansion process, in its simplest form, has been used by Mu¨ller-Breslau [18] for re-forming structural models, such as simple planar and space trusses. In his expansion process, the properties of typical subgraphs, selected in each step to be joined to the previously expanded subgraph, guarantee the determinacy of the simple truss. These subgraphs consist of two and three concurrent bars for planar and space trusses, respectively. The idea can be extended to other types of structure, and more general subgraphs can be considered for addition at each step of the expansion process. A cycle, a planar subgraph, and a subgraph with prescribed connectivity properties are examples of these, which will be employed in this book. For example, the planar truss of Fig. 2.3a can be formed in four steps, joining a substructure Si with γ(Si) ¼ 1 as shown in Fig. 2.3b, sequentially, as illustrated in Fig. 2.3c.

2.2.2.4

An Intersection Theorem

In a general expansion process, a subgraph Si may be joined to another subgraph Sj in an arbitrary manner. For example, γ(Si) or γ(Sj) may have any arbitrary value and the union Si [ Sj may be a connected or a disjoint subgraph. The intersection Si [ Sj may also be connected or disjoint. It is important to find the properties of S1 [ S2 having the properties of S1, S2 and S1 \ S2. The following theorem provides a correct calculation of the properties of Si [ Sj. In order to have the formula in its general form, q subgraphs are considered in place of two subgraphs. Theorem (Kaveh [15]). Let S be the union of q subgraphs S1, S2, S3, . . ., Sq with the following functions being defined: γðSÞ γðSi Þ γðAi Þ

¼ aMðSÞ þ bNðSÞ þ cγ0 ðSÞ, ¼ aMðSi Þ þ bNðSi Þ þ cγ0 ðSi Þ i ¼ 1, 2, . . . , q, ¼ aMðAi Þ þ bNðAi Þ þ cγ0 ðAi Þ i ¼ 2, 3, . . . , q,

where Ai ¼ Si $ 1 \ Si and Si ¼ S1 [ S2 [ . . . [ Si. Then:

2.2 Static Indeterminacy of Structures

45

a

b

c S 1 = S1 S1 ÈS2 = S2 S2 ÈS3 = S3 S3 ÈS4 = S4 = S

Fig. 2.3 Process for the formation of a planar truss. (a) A planar truss. (b) Selected unit. (c) The process of expansion as S1 ¼ S1 ! S2 ! S3 ! S4 ¼ S

½γðSÞ $ cγ0 ðSÞ' ¼

q X i¼1

½γðSi Þ $ cγ0 ðSi Þ' $

q X i¼2

½γðAi Þ $ cγ0 ðAi Þ'

ð2:5Þ

For proof, the interested reader may refer to Kaveh [19]. Special Case. If S and each of its subgraphs considered for expansion (Si for i ¼ 1, . . ., q) are non-disjoint (connected), then Eq. 2.5 can be simplified as: γðSÞ ¼

q X i¼1

γðSi Þ $

q X i¼2

γðAi Þ,

ð2:6Þ

where γðAi Þ ¼ aMðAi Þ þ bNðAi Þ þ c: For calculating the DSI of a multi-member structure, one normally selects a repeated unit of the structure and joins these units sequentially in a connected form. Therefore, Eq. 2.6 can be applied in place of Eq. 2.5 to obtain the overall property of the structure.

2.2.2.5

A Method for Determining the DSI of Structures

Let S be the union of its repeated and/or simple pattern subgraphs Si (i ¼ 1, . . ., q). Calculate the DSI of each subgraph, using the appropriate coefficients from Table 2.1. Now perform the union–intersection method with the following steps:

46

2 Optimal Force Method: Analysis of Skeletal Structures

Step 1: Join S1 to S2 to form S2 ¼ S1 [ S2, and calculate the DSI of their intersection A2 ¼ S1 \ S2. The value of γ(S2) can be found using Eq. 2.5 or Eq. 2.6, as appropriate. Step 2: Join S3 to S2 to obtain S3 ¼ S2 [ S3, and determine the DSI of A3 ¼ S2 \ S3. Similarly to Step 1, calculate γ(S3). Step k: Subsequently join Sk+1 to Sk, calculating the DSI of Ak + 1 ¼ Sk \ Sk + 1 and evaluating the magnitude of γ(Sk+1). q

Repeat Step k until the entire structural model S ¼ [ Si has been reformed and i¼1

its DSI determined. In the above expansion process, the value of q depends on the properties of the substructures (subgraphs) which are considered for reforming S. These subgraphs have either simple patterns for which γ(Si) can easily be calculated, or the DSIs of which are already known. In the process of expansion, if an intersection Ai itself has a complex pattern, further refinement is also possible; i.e. the intersection can be considered as the union of simpler subgraphs.

2.3

Formulation of the Force Method

In this section, a matrix formulation using the basic tools of structural analysis— equilibrium, compatibility and load–displacement relationships—is described. The notations are chosen from those most commonly utilized in structural mechanics.

2.3.1

Equilibrium Equations

Consider a structure S with M members and N nodes, which is γ(S) times statically indeterminate. Select γ(S) independent unknown forces as redundants. These unknown forces can be selected from external reactions and/or internal forces of the structure. Denote these redundants by: n ot q ¼ q 1 ; q 2 ; . . . ; q γ ð SÞ :

ð2:7Þ

Remove the constraints corresponding to redundants, in order to obtain the corresponding statically determinate structure, known as the basic (released or primary) structure of S. Obviously, a basic structure should be rigid. Consider the joint loads as,

2.3 Formulation of the Force Method

47

p ¼ fp1 ; p2 ; . . . ; pn g t ,

ð2:8Þ

where n is the number of components for applied nodal loads. Now the stress resultant distribution r, due to the given load p, for a linear analysis by the force method can be written as, r ¼ B0 p þ B1 q,

ð2:9Þ

where B0 and B1 are rectangular matrices each having m rows, and n and γ(S) columns, respectively, m being the number of independent components for member forces. B0p is known as a particular solution, which satisfies equilibrium with the imposed load, and B1q is a complementary solution, formed from a maximal set of independent self-equilibrating stress systems (S.E.Ss), known as a statical basis. Example 1. Consider a planar truss, as shown in Fig. 2.4a, which is two times statically indeterminate. EA is taken to be the same for all the members. One member force and one component of a reaction may be taken as redundants. Alternatively, two member forces can also be selected as unknowns, as shown in Fig. 2.4b. Selecting the latter choice, the corresponding B0 and B1 matrices can now be obtained by applying unit values of pi (i ¼ 1, 2) and qj (j ¼ 1, 2), respectively: B0t ¼

"

$1 $2

0 $1

0 þ1

pﬃﬃﬃ pﬃﬃ2ﬃ 2

0 0

and B1t ¼

"

0 $1 0 $1

p0ﬃﬃﬃ 2

# 0 0 , 0 $1

# pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃ $1= 2 0pﬃﬃﬃ $1= 2 0pﬃﬃﬃ þ1 þ1 $1=p2ﬃﬃﬃ 0 0 0pﬃﬃﬃ : 0 $1= 2 0 $1= 2 0 0 $1= 2 þ1 þ1 $1= 2

The columns of B1 (rows of Bt1 ) form a statical basis of S. The underlying subgraph of a typical self-equilibrating stress system (for q2 ¼ 1) is shown in bold lines, Fig. 2.4b.

Example 2. Consider a portal frame shown in Fig. 2.5a, which is three times statically indeterminate. This structure is made statically determinate by an imaginary cut at the middle of its beam. The unit value of external load p1 and each of the bi-actions qi (i ¼ 1, 2, 3) lead to the formation of B0 and B1 matrices, in which the two end bending moments (Mi, Mj) of a member are taken as its member forces. Using the sign convention introduced in Chap. 1, B0 and B1 matrices are formed as: B0t ¼ ½ þ4

0

0

0 0

0 ',

48

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.4 A statically indeterminate planar truss. (a) A planar truss. (b) The selected unknown forces

p

a 6 5

1

Fig. 2.5 A statically indeterminate frame. (a) A portal frame S. (b) The basic structure of S

p

1

5

3 6

1 L

10 2 L

2

q1

9

8

7

b

2

4

4

a

L

3

b

2

p=10kN

q2

3

2 1

q3

q3

q

q1

1

3

1

q2

q2

4m

4 4m

and

2

þ4 B1t ¼ 4 $2 $1

0 þ2 þ1

0 $2 $1

0 $2 þ1

0 þ2 $1

3 $4 $2 5: þ1

The columns of B1 form a statical basis of S, and the underlying subgraph of each self-equilibrating stress system is a cycle, as illustrated in bold lines, Fig. 2.5b. Notice that three self-equilibrating stress systems can be formed on each cycle of a planar frame. In both of the above examples, particular and complementary solutions are obtained from the same basic structure. However, this is not a necessary requirement, as imagined by some authors. In fact a particular solution is any solution satisfying equilibrium with the applied loads, and a complementary solution is any maximal set of independent self-equilibrating systems. The latter is a basis of a vector space over the field of real numbers, known as a complementary solution space, Henderson and Maunder [20]. Using the same basic structure is equivalent to searching for a cycle basis of a graph, but restricting the search to fundamental cycles only, which is convenient but not efficient when the structure is complex or cycle bases with specific properties are needed. As an example, consider a three-storey frame as shown in Fig. 2.6a. A cut system as shown in Fig. 2.6b corresponds to a statical basis, containing three selfequilibrating stress systems formed on each element of the cycle basis shown in Fig. 2.6b. However, the same particular solution can be employed with a statical basis formed on the cycles of the basis shown in Fig. 2.6c.

2.3 Formulation of the Force Method Fig. 2.6 A three-storey frame with different cut systems

49

a

b

c

A basic structure need not be selected as a determinate one. For a redundant basic structure, one may obtain the necessary data either by analysing it first for the loads p and bi-actions qi ¼ 1(i ¼ 1, 2, . . ., γ(S)), or by using existing information.

2.3.2

Member Flexibility Matrices

In the force method of analysis, the determination of the member flexibility matrix is an important step. A number of alternative methods are available for the formation of displacement-force relationships describing the flexibility properties of the members. Four such approaches are: 1. 2. 3. 4.

Inversion of the force-displacement relationship; Unit load method; Castigliano’s theorem; Solution of differential equations for member displacements.

In the following, the unit load method is briefly described for the formation of the flexibility matrices: Consider a general element with n member forces, rmt ¼ fr1 ; r2 ; . . . ; rn g,

ð2:10Þ

umt ¼ fu1 ; u2 ; . . . ; un g:

ð2:11Þ

and member displacements:

A typical component of the displacement ui can be found using the unit load method as: ððð ui ¼ σit εdV, ð2:12Þ V

where σi represents the matrix of statically equivalent stresses due to a unit load in the direction of ri, and ε is the exact strain matrix due to all applied forces rm.

50

2 Optimal Force Method: Analysis of Skeletal Structures

The unit loads can be used in turn for all the points where member force are applied, and therefore, ððð um ¼ σt εdV, ð2:13Þ V

where, σ ¼ fσ 1 σ 2 . . . σ n gt :

ð2:14Þ

σ ¼ crm ,

ð2:15Þ

For a linear system,

where c is the stress distribution due to unit forces rm. The stress-strain relationship can be written as: ε ¼ ϕσ ¼ ϕcrm :

ð2:16Þ

Substituting in Eq. 2.13 leads to, um ¼

ððð

V

σt ϕcdV rm

ð2:17Þ

or, u m ¼ f m rm ,

ð2:18Þ

where, fm ¼

ððð

V

σt ϕcdV,

ð2:19Þ

represents the element flexibility matrix. The evaluation of σ representing the exact stress distribution due to the forces rm, may not be possible, and hence an approximate relationship should be used. Usually the matrix c is selected such that it will satisfy at least the equations of equilibrium. Denoting this approximate matrix by c, and using σ ¼ c: ððð fm ¼ ct ϕcdV: ð2:20Þ V

This equation will be used for the derivation of the flexibility matrices of some finite elements in the proceeding sections.

2.3 Formulation of the Force Method

51 y

Fig. 2.7 A beam element and selected independent member forces

r ,u 5 5 r2 ,u2 i

r ,u r4 ,u4 j 1 1

x

r3,u3 r6 ,u6

z

For a bar element of a space truss, however, the flexibility matrix can easily be obtained using Hooke’s law as already discussed in Chap. 1. For a beam element ij of a space frame, y and z axes are taken as the principal axes of the beams cross sections, Fig. 2.7. The forces of end j are selected as a set of independent member forces, and the element is considered to be supported at point i. The axial, torsional, and flexural behaviour in respective planes are uncoupled, and therefore, one needs only to consider the flexibility relationships for four separate members: 1. 2. 3. 4.

An axial force member (along x axis); A pure torsional member (about x axis); A beam bent about y axis; A beam bent about z axis.

Direct adaptation of the flexibility relationships derived in Chap. 1, gives the following 6 ( 6 flexibility matrix, 2

L 6 EA 6 6 6 6 0 6 6 6 6 6 0 6 6 fm ¼ 6 6 6 0 6 6 6 6 6 0 6 6 6 6 4 0

3

L3 3EIz

sym:

0

L3 3EIy

0

0

0 L2 2EIz

$

L GJ

L2 2EIy

0

L EIy

0

0

0

7 7 7 7 7 7 7 7 7 7 7 7 7, 7 7 7 7 7 7 7 7 7 7 L 7 5 EIz

ð2:21Þ

where G is the shear modulus, Iy and Iz are the moments of inertia with respect to y and z axes, respectively. J is the Saint-Venant torsion constant of the cross section.

52

2 Optimal Force Method: Analysis of Skeletal Structures

2.3.3

Explicit Method for Imposing Compatibility

The compatibility equations in the actual structure will now be derived. Using the displacement-load relationship for each member, and collecting them in the diagonal of the unassembled flexibility matrix Fm, one can write member distortions as: u ¼ Fm r ¼ Fm B0 p þ Fm B1 q:

ð2:22Þ

" # p ½u' ¼ ½Fm ' ½B0 B1 ' : q

ð2:23Þ

In matrix form:

From the contragradient principle of Chap. 1, "

# B0t ½u': ½v' ¼ B1t

ð2:24Þ

Combining Eqs. 2.23 and 2.24 results in, "

v0 vc

#

" # # p B0t ½Fm ' ½B0 B1 ' , ¼ q B1t "

ð2:25Þ

in which v0 contains the displacements corresponding to the force components of p, and vc denotes the relative displacements of the cuts for the basic structure. Performing the multiplication, "

v0 vc

#

"

B0t Fm B0 ¼ B1t Fm B0

B0t Fm B1 B1t Fm B1

#" # p : q

ð2:26Þ

Defining: D00 ¼ B0t Fm B0 , D01 ¼ B1t Fm B0 ,

D10 ¼ B0t Fm B1 , D11 ¼ B1t Fm B1 ,

ð2:27Þ

the expansion of Eq. 2.14 leads to: v0 ¼ D00 p þ D01 q,

ð2:28Þ

vc ¼ D10 p þ D11 q:

ð2:29Þ

and Consider now the compatibility conditions as: vc ¼ 0:

ð2:30Þ

2.3 Formulation of the Force Method

53

Equation 2.30 together with Eq. 2.29 leads to: q ¼ $D$1 11 D10 p ¼ Fp:

ð2:31Þ

Substituting in Eq. 2.22 yields, % & v0 ¼ D00 $ D01 D$1 11 D10 p,

ð2:32Þ

% & r ¼ B0 $ B1 D$1 11 D10 p:

ð2:33Þ

and the stress resultant in a structure can be obtained as:

2.3.4

Implicit Approach for Imposing Compatibility

A direct application of the work principle of Chap. 1, can also be used to impose the compatibility conditions in an implicit form as follows: Since the structure is considered to be linearly elastic, a linear relation exists between the unknown forces q and the applied forces p; i.e. q ¼ Qp,

ð2:34Þ

where Q is a transformation matrix which is still unknown. Equation 2.9 can now be written as: r ¼ B0 p þ B1 Qp ¼ ðB0 þ B1 QÞp ¼ Bp:

ð2:35Þ

Using the work theorem: Pt v ¼ rt u ¼ pt Bt u:

ð2:36Þ

Now a set of suitable internal forces, r*, is considered which is statically equivalent to the external loads. From work principle: t

pt v ¼ r) u,

ð2:37Þ

pt v ¼ pt B0t u:

ð2:38Þ

or

Comparison of the above two equations leads to:

54

2 Optimal Force Method: Analysis of Skeletal Structures

pt Bt u ¼ pt B0t u:

ð2:39Þ

Substituting u ¼ FmBP in the above equation: pt Bt Fm Bp ¼ pt B0t Fm Bp:

ð2:40Þ

This holds for any p, and therefore: Bt Fm B ¼ B0t Fm B:

ð2:41Þ

From Eq. 2.35 by transposition, Bt ¼ B0t þ Qt B1t ,

ð2:42Þ

( B0t þ Qt B0t Fm B ¼ B0t Fm B,

ð2:43Þ

Qt B0t Fm ðB0 þ B1 QÞ ¼ 0,

ð2:44Þ

' ( Qt B1t Fm B0 þ B1t Fm B1 Q ¼ 0:

ð2:45Þ

Qt ðD10 þ D11 QÞ ¼ 0,

ð2:46Þ

D10 þ D11 Q ¼ 0:

ð2:47Þ

Q ¼ $D$1 11 D10 ,

ð2:48Þ

q ¼ $D$1 11 D10 p,

ð2:49Þ

therefore, '

or

or

Using the notation introduced in Eq. 2.15 leads to,

or

Therefore,

and

and Eq. 2.20 is obtained as in the previous approach.

2.3 Formulation of the Force Method

2.3.5

55

Structural Flexibility Matrices

The overall flexibility matrix of a structure can be expressed as: v ¼ Fp:

ð2:50Þ

Pre-multiplying the above equation by pt, we have: pt Fp ¼ pt B0t Fm Bp:

ð2:51Þ

F ¼ B0t Fm B,

ð2:52Þ

F ¼ B0t Fm ðB0 þ B1 QÞ,

ð2:53Þ

F ¼ B0t Fm B0 $ B0t Fm B1 D$1 11 D10 :

ð2:54Þ

Since p is arbitrary,

or

or

Since Fm is symmetric, it follows that: t D10 ¼ B0t Fm B1 ¼ B0t Fmt B1 :

ð2:55Þ

Therefore, the overall flexibility matrix (known also as influence matrix) of the structure is obtained as, t D$1 F ¼ D00 $ D10 11 D10 ,

ð2:56Þ

and D11 ¼ Bt1 FmB1 ¼ G is also referred to as the flexibility matrix of the structure. In this book, properties of G will be studied, since its pattern is the most important factor in optimal analysis of the structure by the force method. Equation 2.34 can now be used to calculate the nodal displacements.

2.3.6

Computational Procedure

The sequence of computational steps for the force method can be summarized as: 1. Construct B0 and obtain Bt0 . 2. Construct B1 and obtain Bt1 . 3. Form unassembled flexibility matrix Fm.

56

2 Optimal Force Method: Analysis of Skeletal Structures

4. 5. 6. 7. 8. 9. 10. 11.

Form FmB0 followed by FmB1. Calculate D00, Dt10 , D10 and D11, sequentially. Compute $ D$1 11 . Calculate Q ¼ $ D$1 11 D10. Form B1Q and find B ¼ B0 + B1Q. Form Dt10 Q and find D00 + Dt10 Q. Compute the internal forces as r ¼ Bp. Compute nodal displacements as v0 ¼ Fp.

Example 3. In this example, the complete analysis of the truss of Example 1 will be given. B0 and B1 matrices are already formed in Example 1 of Sect. 2.2.1. The unassembled flexibility matrix can be constructed as: 2

3

1

6 6 6 6 6 6 L 6 6 Fm ¼ EA 6 6 6 6 6 6 4

1 0

1 1

0

pﬃﬃﬃ 2 pﬃﬃﬃ 2

1

pﬃﬃﬃ 2

pﬃﬃﬃ 2

1

7 7 7 7 7 7 7 7: 7 7 7 7 7 7 5

Using the above matrix and the matrices from Example 1, leads to: D11

" pﬃﬃﬃ L 2 2 þ 3=2 ¼ EA 1=2

and D10

# p1=2 ﬃﬃﬃ , 2 2þ2

" pﬃﬃﬃ pﬃﬃﬃ # L 2 þ 2= 2 2 þ 2= 2 p ﬃﬃ ﬃ pﬃﬃﬃ : ¼ EA 1= 2 2 þ 3= 2

Substituting in Eq. 2.25, results in: "

q1 q2

#

¼$

" pﬃﬃﬃ 2 2 þ 3=2 1=2

p1=2 ﬃﬃﬃ 2 2þ2

#$1 "

pﬃﬃﬃ pﬃﬃﬃ #" # 2 þ 2= pﬃﬃﬃ 2 2 þ 2=p2ﬃﬃﬃ p1 : 1= 2 2 þ 3= 2 p2

Taking p1 ¼ p2 ¼ P for simplicity, and solving the above equations gives: q1 ¼ $1:43P and q2 ¼ $1:17P: Equation 2.3 is then used to calculate the member forces as:

2.3 Formulation of the Force Method

57

r¼ f r1 r2 r3 r4 r5 r6 r7 r8 r9 r10 gt ¼ f $1:95P $0:17P 2:05P 0:83P 1:36P $1:44P $0:12P 0:24P $1:17P $0:17P gt : Nodal displacements can be found using Eq. 2.25. Example 4. In this example, the complete analysis of the frame in Example 2 is given. B0 and B1 matrices are already formed in Example 2 of Sect. 2.2.1. The unassembled flexibility matrix of the structure, using the sign convention introduced in Chap. 1, is formed as: 2

2 $1 6 $1 2 6 L 6 6 Fm ¼ 6EI 6 6 4

2 $1 $1 2

3

7 7 7 7: 7 7 2 $1 5 $1 2

Substituting in Eq. 2.21 leads to:

D11

2 64 0 L 4 0 56 ¼ 6EI $24 0

3 $24 0 5, 18

and

D10

2 3 32 L 4 $24 5: ¼ 6EI $12

The inverse of D11 is computed as, D$1 11

2 18=576 6EI 4 0 ¼ L 3=72

0 576 0

3 3=72 0 5, 1=9

and Q can be obtained as: Q ¼ $D$1 11 D10 Matrix B is now computed as,

2

3 $1=2 ¼ 4 þ3=7 5: 0

58

2 Optimal Force Method: Analysis of Skeletal Structures

2 3 2 þ4 4 607 6 0 6 7 6 607 6 0 7 6 B¼6 607 þ 6 0 6 7 6 405 4 0 $4 0

$2 þ2 $2 $2 þ2 $2

3 $1 2 3 þ1 7 7 $1=2 $1 7 74 þ3=7 5, þ1 7 7 0 $1 5 þ1

and finally by using Eq. 2.23 the member forces are obtained as: r ¼ f þ11:43

þ8:57

$8:57 $8:57

þ8:57

þ11:43 gt :

General Loading. When members are loaded in a general form, then it must be replaced by an equivalent loading. Such a loading can be found as the superposition of two cases; case 1 consists of the given loading but the ends of the member are fixed. The fixed end forces (actions), denoted by FEA, can be found using tables from books on strength of materials. Case 2 is the given structure subjected to the reverse of the fixed end actions only. Obviously, the sum of the loads and reactions of case 1 and case 2 will be the same effect as that of the given loading. This superposition process is illustrated in the following example: Example 5. A two-span beam is considered as shown in Fig. 2.8a. The fixed end actions are provided in b, and the equivalent forces are illustrated in Fig. 2.8c. The structure is twice indeterminate, and the primary structure is obtained by introducing two hinges as shown in d. The applied nodal forces and redundants are depicted in Fig. 2.8e, f, respectively. B0 and B1 matrices are formed as, 2

$1 6 0 B0 ¼ 6 4 0 0

3 0 0 þ1 0 7 7 0 0 5 0 þ1

2

$1 6 0 and B1 ¼ 6 4 0 0

3 0 þ1 7 7, $1 5 0

and the unassembled flexibility matrix of the structure is constructed as: 2

2 L 6 $1 6 Fm ¼ 6EI 4

$1 2 2 $1

Substituting in Eq. 2.27 leads: D11 ¼

" # L 2 1 , 6EI 1 4

3

7 7: $1 5 2

2.3 Formulation of the Force Method

a

59

b

20kN

12kN/m

4m

2m

c

6

e

6

10

6

16

2m 10

10

d

f

q2

q1

Fig. 2.8 A two-span beam with general loading. (a) A two-span beam. (b) Fixed end actions. (c) The equivalent loading. (d) The selected primary structure. (e) Applied force on primary structure. (f) Redundants on primary structure

and D10

" L 2 ¼ 6EI 1

1 2

# 0 : 1

The inverse of D11 is computed as, D$1 11

" # 1 6EI 4 $1 , ¼$ ( 7 L $1 2

and Q can be obtained as: Q¼

$D$1 11 D10

" 1 4 ¼$ 7 $1

$1 2

#"

2 1

# " 1 7 1 0 ¼$ 2 1 7 0

2 3

# $1 : 2

Now r is computed as, 02

$1 B6 0 0 6 r ¼B @4 0 0

0 1 0 0

3 2 0 1 2=7 6 0 $3=7 07 7þ6 0 5 4 0 3=7 1 0 0

31 2 3 2 3 $1=7 0:285 0 C 6 7 $2=7 7 7C4 6 5 ¼ 6 0:572 7, 5 A 4 2=7 5:428 5 10 0 10:00

a`dding the fixed end reaction, the final member forces are obtained as: r ¼ f 16:285

$15:428 15:428

0:000 gt :

60

2.3.7

2 Optimal Force Method: Analysis of Skeletal Structures

Optimal Force Method

For an efficient force method, the matrix G should be: (a) Sparse; (b) Well conditioned; (c) Properly structured, i.e. narrowly banded. In order to provide the properties (a) and (b) for G, the structure of B1 should be carefully designed, since the pattern of Fm for a given discretization is unchanged; i.e. a suitable statical basis should be selected. This problem is treated in different forms by various methods. In the following, graph theoretical methods are described for the formation of appropriate statical bases of different types of skeletal structures. The property (c) above has a totally combinatorial nature and is studied in Chaps. 5 and 6. Pattern Equivalence. Matrix B1 containing a statical basis, in partitioned form, is pattern equivalent to Ct, where C is the cycle-member incidence matrix. Similarly, Bt1 FmB1 is pattern equivalent to CICt or CCt. This correspondence transforms some structural problems associated with the characterization of G ¼ Bt1 FmB1 into combinatorial problems of dealing with CCt. As an example, if a sparse matrix G is required, this can be achieved by increasing the sparsity of CCt. Similarly for a banded G, instead of ordering the elements of a statical basis (self-equilibrating stress systems), one can order the corresponding cycles. This transformation has many advantages, such as: 1. The dimension of CCt is often smaller than that of G. For example, for a space frame the dimension of CCt is six-fold and for a planar frame three-fold smaller than that of G. Therefore, the optimisation process becomes much simpler when combinatorial properties are used. 2. The entries of C and CCt are elements of Z2 and therefore easier to operate on, compared to B1 and G which have real numbers as their entries. 3. The advances made in combinatorial mathematics and graph theory become directly applicable to structural problems. 4. A correspondence between algebraic and graph theoretical methods becomes established.

2.4

Force Method for the Analysis of Frame Structures

In this section, frame structures are considered in their perfect conditions; i.e. the joints of a frame are assumed to be rigid, and connected to each other by elastic members and supported by a rigid foundation. For this type of skeletal structure, a statical basis can be generated on a cycle basis of its graph model. The function representing the degree of static

2.4 Force Method for the Analysis of Frame Structures

61

indeterminacy, γ(S), of a rigid-jointed structure is directly related to the first Betti number b1(S) of its graph model, γðSÞ ¼ αb1 ðSÞ ¼ α½MðSÞ $ NðSÞ þ b0 ðSÞ',

ð2:57Þ

where α ¼ 3 or 6 depending on whether the structure is either a planar or a space frame. For a frame structure, matrix B0 can easily be generated using a shortest route tree of its model, and B1 can be formed by constructing 3 or 6 self-equilibrating stress systems on each element of a cycle basis of S. In order to obtain a flexibility matrix of maximal sparsity, special cycle bases should be selected as defined in the next section. Methods for the formation of a cycle basis can be divided into two groups, namely (a) Topological methods, (b) graph theoretical approaches. Topological methods useful for the formation of cycle bases by hand, were developed by Henderson and Maunder [20] and a complete description of these methods is presented in Kaveh [3]. Graph-theoretical methods suitable for computer applications were developed by Kaveh [21].

2.4.1

Minimal and Optimal Cycle Bases

A matrix is called sparse if many of its entries are zero. The interest in sparsity arises because its exploitation can lead to enormous computational saving, and because many large matrices that occur in the analysis of practical structures, can be made sparse if they are not already so. A matrix can therefore be considered sparse, if there is an advantage in exploiting its zero entries. The sparsity coefficient χ) of a matrix is defined * to be its number of non-zero entries. A cycle basis C ¼ C1 ; C2 ; C3 ; . . . ; Cb1 ðSÞ is called minimal, if it corresponds to a minimum value of: Lð C Þ ¼

bX 1 ð SÞ i¼1

LðCi Þ:

ð2:58Þ

Obviously, χ(C) ¼ L(C) and a minimal cycle basis can be defined as a basis which corresponds to minimum χ(C). A cycle basis for which L(C) is near minimum is called a subminimal cycle basis of S. A cycle basis corresponding to maximal sparsity of the CCt is called an optimal cycle basis of S. If χ(CCt) does not differ considerably from its minimum value, then the corresponding basis is termed suboptimal. The matrix intersection coefficient σi(C) of row i of cycle member incidence matrix C is the number of row j such that:

62

2 Optimal Force Method: Analysis of Skeletal Structures

(a) j ∈ {i + 1, i + 2, . . ., b1(S)}, (b) Ci \ Cj 6¼ ø, i.e. there is at least one k such that the column k of both cycles Ci and Cj (rows i and j) contain non-zero entries. Now it can be shown that: χðDÞ ¼ b1 ðSÞ þ 2

b1X ðSÞ$1 i¼1

σi ðCÞ:

ð2:59Þ

This relationship shows the correspondence of a cycle member incidence matrix C and that of its cycle basis adjacency matrix. In order to minimize χ(CCt), the b1X ðSÞ$1 σi ðCÞ should be minimized, since b1(S) is a constant for a given value of i¼1

structure S, i.e. γ-cycles with a minimum number of overlaps should be selected. In the force method, an optimal cycle basis is needed corresponding to the maximum sparsity of CCt matrix. However, because of the complexity of this problem, most of the research has been concentrated on minimal cycle basis selection, except those of Ref. [22], which minimize the overlaps of the cycles rather than only their length.

2.4.2

Selection of Minimal and Subminimal Cycle Bases

Cycle bases of graphs have many applications in various fields of engineering. The amount of work in these applications depends on the cycle basis chosen. A basis with shorter cycles reduces the time and storage required for some applications; i.e. it is ideal to select a minimal cycle basis, and for some other applications minimal overlaps of cycles are needed; i.e. optimal cycle bases are preferred. In this section, the formation of minimal and subminimal cycle bases is first discussed. Then the possibility of selecting optimal and suboptimal cycle bases is investigated. Minimal cycle bases were considered first by Stepanec [23] and improved by Zykov [24]. Many practical algorithms for selecting subminimal cycle bases have been developed by Kaveh [15]. In this section, the merits of the algorithms developed by different authors are discussed; a method is given for selection of minimal cycle bases, and efficient approaches are presented for the generation of subminimal cycle bases. Formation of a Minimal Cycle on a Member. A minimal length cycle Ci on a member mj, called its generator, can be formed by using the shortest route tree algorithm as follows: Start the formation of two SRTs rooted at the two end nodes ns and nt of mj, and terminate the process as soon as the SRTs intersect each other (not through mj itself) at say nc. The shortest paths between ns and nc, and nt and nc, together with mj, form

2.4 Force Method for the Analysis of Frame Structures Fig. 2.9 A minimal cycle on a member

63

nk

nc

Ci ns

mj

nt

a minimal cycle Ci on mj. Using this algorithm, cycles of prescribed lengths can also be generated. As an example, Ci is a minimal cycle on mj in Fig. 2.9. The SRTs are shown in bold lines. The generation of SRTs is terminated as soon as nc has been found. A minimal cycle on a member mj passing through a specified node nk can similarly be generated. An SRT rooted at nk is formed and as soon as it hits the end nodes of mj, the shortest paths are found by backtracking between nk and ns, and nk and nt. These paths together with mj form the required cycle. As an example, a minimal cycle on mj containing nk, is illustrated by dashed lines in Fig. 2.9. Different Cycle Sets for Selecting a Cycle Basis. It is obvious that a general cycle can be decomposed into its simple cycles. Therefore, it is natural to confine the considered set to only simple cycles of S. Even such a cycle set, which forms a subspace of the cycle space of the graph, has many elements and is therefore uneconomical for practical purposes. In order to overcome the above difficulty, Kaveh [15] used an expansion process, selecting the smallest admissible (independent with additional restriction) cycles, one at a time, until b1(S) cycles forming a basis had been obtained. In this approach, a very limited number of cycles were checked for being an element of a basis. As an example, the expansion process for selecting a cycle basis of S is illustrated in Fig. 2.10. Hubicka and Syslø [25] employed a similar approach, without the restriction of selecting one cycle at each step of expansion. In their method, when a cycle has been added to the previously selected cycles, increasing the first Betti number of the expanded part by “p”, then p created cycles have been formed. As an example, in this method, Steps 4 and 5 will be combined into a single step, and addition of cycle 5 will require immediate formation of the cycle 4. The above method is modified, and an efficient algorithm is developed for the formation of cycle bases by Kaveh and Roosta [26], Finally, Horton [27] proved that the elements of a minimal cycle basis lie in between a cycle set consisting of the minimal cycles on each member of S which passes through each node of S, i.e. each member is taken in turn and all cycles of minimal length on such a member passing through all the nodes of S are generated. Obviously, M(S) ( M(S) such cycles will be generated. Independence Control. Each cycle of a graph can be considered as a column vector of its cycle-member incidence matrix. An algebraic method such as the

64

2 Optimal Force Method: Analysis of Skeletal Structures

2 1

1

2

2 1

1 3

4

2 1

3

4

5

3

Fig. 2.10 A graph S and selected cycles

Gaussian elimination may then be used for checking the independence of a cycle with respect to the previously selected sub-basis. However, although this method is general and reduces the order dependency of the cycle selection algorithms, like many other algebraic approaches its application requires a considerable amount of storage space. The most natural graph theoretical approach is to employ a spanning tree of S, and form its fundamental cycles. This method is very simple; however, in general its use leads to long cycles. The method can be improved by allowing the inclusion of each used chord in the branch set of the selected tree. Further reduction in length may be achieved by generating an SRT from a centre node of a graph, and the use of its chords in ascending order of distance from the centre node, Kaveh [21]. A third method, which is also graph-theoretical, consists of using admissible cycles. Consider the following expansion process, with S being a 2-connected graph, C1 ¼ C1 ! C2 ! C3 ! . . . ! Cb1 ðSÞ ¼ S, where

k

Ck ¼ [ Ci . A cycle Ck+1 is called an admissible cycle, if for i¼1

Ck + 1 ¼ Ck [ Ck + 1:

' ( ' ( ' ( b1 Ckþ1 ¼ b1 Ck [ Ckþ1 ¼ b1 Ck þ 1:

ð2:60Þ

It can easily be proved that, the above admissibility condition is satisfied if any of the following conditions hold: 1. Ak + 1 ¼ Ck \ Ck + 1 ¼ ∅, where ∅ is an empty intersection; 2. b1 ðAkþ1 Þ ¼ r $ s, where r and s are the numbers of components of Ck+1 and Ck, respectively; 3. b1 ðAkþ1 Þ ¼ 0 when Ck and Ck+1 are connected (r ¼ s). In the above relations, b1 ðAi Þ ¼ Mi $ Ni þ 1, where Mi and Ni are the numbers of members and nodes of Ai, respectively. As an example, the sequence of cycle selection in Fig. 2.11 will be as specified by their numbers.

2.4 Force Method for the Analysis of Frame Structures

65

Fig. 2.11 A cycle and its bounded cycles

C k+3

C k+1 Ck

C k+2

A different approach suggested by Hubicka and Syslø, in which, ' ( ' ( b1 Ckþ1 ¼ b1 Ck þ p,

ð2:61Þ

is considered to be permissible. However, a completion is performed for p > 1. As an example, when C3 is added to Ck, its first Betti number is increased by 3 and therefore, cycles C1 and C2 must also be selected at that stage, before further expansion. Having discussed the mathematical concepts involved in a cycle basis selection, three different algorithms are now described. Algorithm 1 (Kaveh [15]) Step 1: Select a pseudo-centre node of maximal degree O. Such a node can be selected manually or automatically using the graph or algebraic graph theoretical methods discussed in Chap. 5. Step 2: Generate an SRT rooted at O, form the set of its chords and order them according to their distance from O. Step 3: Form one minimal cycle on each chord in turn, starting with the chord nearest to the root node. A corresponding simple path is chosen which contains members of the tree and the previously used chords, hence providing the admissibility of the selected cycle. This method selects subminimal cycle bases, using the chords of an SRT. The nodes and members of the tree and consequently the cycles are partially ordered according to their distance from O. This is the combinatorial version of the Turn Back method to be discussed in the section on algebraic force methods. Algorithm 2 (Kaveh [15]) Step 1: Select a centre or seudo-centre node of maximal degree O. Step 2: Use any member incident with O as the generator of the first minimal cycle. Take any member not used in C1 and incident with O, and generate on it the second minimal cycle. Continue this process until all the members incident with O are used as the members of the selected cycles. The cycles selected so far are admissible, since the intersection of each cycle with the previously selected cycles is a simple path (or a single node) resulting in an increase of the first Betti number by unity for each cycle. Step 3: Choose a starting node O0 , adjacent to O, which has the highest degree. Repeat a step similar to Step 2, testing each selected cycle for admissibility.

66

2 Optimal Force Method: Analysis of Skeletal Structures

If the cycle formed on a generator mk fails the test, then examine the other minimal cycles on mk if any such cycle exists. If no admissible minimal cycle can be found on mk, then, Form admissible minimal cycles on the other members incident with O0 . If mk does not belong to one of these subsequent cycles, then: Search for an admissible minimal cycle on mk, since the formation of cycles on other previous members may now have altered the admissibility of this cycle. If no such cycle can be found, leave mk unused. In this step more than one member may be left unused. Step 4: Repeat Step 3 using as starting nodes a node adjacent to O and/or O0 , having the highest degree. Continue the formation of cycles until all the nodes of S have been tested for cycle selection. If all the members have not been used, select the shortest admissible cycle available for an unused member as generator. Then test the minimal cycles on the other unused members, in case the formation of the longer cycle has altered the admissibility. Each time a minimal cycle is found to be admissible, add to Ci and test all the minimal cycles on the other unused members again. Repeat this process, forming other shortest admissible cycles on unused members as generators, until S is re-formed and a subminimal cycle basis has been obtained. Both of the above two algorithms are order-dependent, and various starting nodes may alter the result. The following algorithm is more flexible and less order-dependent, and in general leads to the formation of shorter cycle bases. Algorithm 3 (Kaveh [21]) Step 1: Generate as many admissible cycles of length 3 as possible. Denote the union of the selected cycles by Cn. Step 2: Select an admissible cycle of length 4 on an unused member. Once such a cycle Cn+1 is found, check the other unused members for possible admissible cycles of length 3. Again select an admissible cycle of length 4 followed by the formation of possible 3-sided cycles. This process is repeated until no admissible cycles of length 3 and 4 can be formed. Denote the generated cycles by Cm. Step 3: Select an admissible cycle of length 5 on an unused member. Then check the unused members for the formation of 3-sided admissible cycles. Repeat Step 2 until no cycle of length 3 or 4 can be generated. Repeat Step 3 until no cycle of length 3, 4 or 5 can be found. Step 4: Repeat similar steps to Step 3, considering higher-length cycles, until b1(S) admissible cycles forming a subminimal cycle basis are generated. Remark. The cycle basis C formed by Algorithms 1–3 can further be improved by exchanging the elements of the selected basis. In each step of this process, a shortest 0 0 cycle Ci independent of the cycles of C \Ci is replaced by Ci if L(Ci ) < L(Ci). This process is repeated for i ¼ 1, 2, . . ., b1(S). This additional operation increases the computational time and storage, and its use is recommended only when the formation of minimal cycle basis is required.

2.4 Force Method for the Analysis of Frame Structures

67

Algorithm 4 (Horton [27]) Step 1: Find a minimum path P(ni, nj) between each pair of nodes ni and nj. Step 2: For each node nk and member ml ¼ (ni, nj), generate the cycle having ml and nk as P(nk,ni) + P(nk,nj) + (ni,nj) and calculate its length. Degenerate cases in which P(nk, ni) and P(nk, nj) have nodes other than nk in common, can be omitted. Step 3: Order the cycles by their weight (or length). Step 4: Use the Greedy Algorithm, to find a minimal cycle basis from this set of cycles. This algorithm is given in in Kaveh [15, 20]. A simplified version of the above Algorithm can be designed as follows: Step 1: Form a spanning tree of S rooted from an arbitrary node, and select its chords. Step 2: Take the first chord and form N(S) $ 2 minimal cycles, each being formed on the specified chord containing a node of S (except the two end nodes of this chord). Step 3: Repeat Step 2 for the other chords, in turn, until [M(S) $ N(S) + 1] ( [N (S) $ 2] cycles are generated. Repeated and degenerate cycles should be discarded. Step 4: Order the cycles in ascending magnitude of their lengths. Step 5: Using the above set of cycles, employ the Greedy Algorithm to form a minimal cycle basis of S. The main contribution of Horton’s Algorithm is the limit imposed on the elements of the cycle-set used in the Greedy Algorithm. The use of matroids and the Greedy Algorithm, has been suggested by Kaveh [15], and they have been employed by Lawler [28] and Kolasinska [29].

2.4.3

Examples

Example 1. Consider a planar graph S, as shown in Fig. 2.12, for which b1(S) ¼ 18$11 + 1 ¼ 8. Using Algorithm 3, the selected basis consists of four cycles of length 3, three cycles of length 4 and one cycle of length 5, as follows: C1 ¼ ð1; 2; 3Þ, C2 ¼ ð1; 8; 9Þ, C3 ¼ ð2; 6; 3Þ, C4 ¼ ð2; 5; 6Þ, C5 ¼ ð1; 4; 5; 2Þ, C6 ¼ ð1; 7; 5; 2Þ, C7 ¼ ð8; 6; 2; 1Þ, C8 ¼ ð10; 8; 6; 3; 11Þ

The total length of the selected basis is L(C) ¼ 29, which is a counter example for minimality of a mesh basis, since, for any such basis of S, L(C) > 29. Example 2. In this example, S is the model of a space frame, considered as 27

S ¼ [ Si , where a typical Si is depicted in Fig. 2.13a. For Si there are 12 members i¼1

joining eight corner nodes, and a central node joined to these corner nodes.

68

2 Optimal Force Method: Analysis of Skeletal Structures 2

Fig. 2.12 A planar graph S

5

4

6

7

1

8

10

9

11 3

b

Fig. 2.13 A space frame S. (a) A typical Si (i ¼ 1, . . ., 27). (b) S with some omitted members

a

The model S is shown in Fig. 2.13b, in which some of the members are omitted for clarity in the diagram. For this graph, b1(S) ¼ 270. The selected cycle basis using any of the algorithms consists of 270 cycles of length 3, forming a minimal cycle basis of S. For Algorithm 3, the use of different starting nodes leads to a minimal cycle basis, showing the capability of this method. Example 3. S is a planar graph with b1(S) ¼ 9, as shown in Fig. 2.14. The application of Algorithm 3 results in the formation of a cycle of length 3 followed by the selection of five cycles of length 4. Then member {1, 6} is used as the generator of a six-sided cycle C7 ¼ (1,2,3,4,5,6,1). Member {2, 10} is then employed to form a seven-sided cycle C8 ¼ (2,11,12,13,14,15,10,2), followed by the selection of a five-sided cycle C9 ¼ (10,5,4,3,2,10). The selected cycle basis has a total length of L(C) ¼ 41, and is not a minimal cycle basis. A shorter cycle basis can be found by Algorithm 4 consisting of one three-sided and five four-sided cycles, together with the following cycles, C7 ¼ ð1; 2; 10; 5; 6; 1Þ, C8 ¼ ð2; 3; 4; 5; 10; 2Þ and C9 ¼ ð2; 11; 12; 13; 14; 15; 10; 2Þ, forming a basis with the total length of 40. However, the computation time and storage for Algorithm 3 is far less than that of Algorithm 4, as compared in Ref. [30].

2.4 Force Method for the Analysis of Frame Structures

69 8

Fig. 2.14 A planar graph S 3

4 13

12

11

14

7

2

18

17

5

15

16

9

10

1

2.4.4

6

Optimal and Suboptimal Cycle Bases

In what follows, a direct method and an indirect approach, which often lead to the formation of optimal cycle bases, are presented. Much work is needed before the selection of an optimal cycle basis of a graph becomes feasible.

2.4.4.1

Suboptimal Cycle Bases; A Direct Approach

Definition 1. An elementary contraction of a graph S is obtained by replacing a path containing all nodes of degree 2 with a new member. A graph S contracted to a graph S0 is obtained by a sequence of elementary contractions. Since in each elementary contraction k nodes and k members are reduced, the first Betti number does not change in a contraction, i.e. b1(S) ¼ b1(S0 ). The graph S is called homeomorphic to S0 , Fig. 2.15. This operation is performed in order to reduce the size of the graph and also because the number of members in an intersection of two cycles is unimportant; a single member is enough to render Ci \ Cj nonempty, and hence to produce a non-zero entry in CCt. Definition 2. Consider a member mi of a graph S. On this member, p minimal cycles of length q can be generated. P is called the incidence number and q is defined as the cycle length number of mi. In fact, p and q are measures assigned to a member to indicate its potential as a member in the elements of a cycle basis. In the process of expansion for cycle selection, an artificial increase in p results in the exclusion of this element from a minimal cycle, keeping the number of overlaps as small as possible. Space graphs need special treatment. For these graphs, when a member has p ¼ 1, then the next shortest length cycles with q0 ¼ q + l (l being the next smallest possible integer) are also considered. Denoting the number of such cycles by p0 , the incidence number and cycle length number for this type of member are taken as,

70

2 Optimal Force Method: Analysis of Skeletal Structures

a

Fig. 2.15 S and its contracted graph S0 . (a) S. (b) S0

b

0

Ijk ¼ p þ 1 and

+ , + , 0 0 0 Ijkc ¼ q þ p q = 1 þ p ,

ð2:62Þ

respectively. The end nodes of the considered member are j and k. Definition 3. The weight of a cycle is defined as the sum of the incidence numbers of its members. Algorithm A Step 1: Contract S into S0 , and calculate the incidence number (IN) and cycle length number (CLN) of all its members. Step 2: Start with a member of the least CLN and generate a minimal weight cycle on this member. For members with equal CLNs, the one with the smallest IN should be selected. A member with these two properties will be referred to as “a member of the least CLN with the smallest IN”. Step 3: On the next unused member of the least CLN with the smallest IN, generate an admissible minimal weight cycle. In the case when a cycle of minimal weight is rejected due to inadmissibility, the next unused member should be considered. This process is continued as far as the generation of admissible minimal weight cycles is possible. After a member has been used as many times as its IN, before each extra usage, increase the IN of such a member by unity. Step 4: On an unused member of the least CLN, generate one admissible cycle of the smallest weight. This cycle is not a minimal weight cycle, otherwise it would have been selected at Step 3. Such a cycle is called a subminimal weight cycle. Again, update the incidence numbers for each extra usage. Now repeat Step 3, since the formation of the new subminimal weight cycle may have altered the admissibility condition of the other cycles, and selection of further minimal weight cycles may now have become possible. Step 5: Repeat Step 4, selecting admissible minimal and subminimal weight cycles, until b1(S0 ) of these cycles are generated. Step 6: A reverse process to that of the contraction of Step 1, transforms the selected cycles of S0 into those of S. This algorithm leads to the formation of a suboptimal cycle basis, and for many models encountered in practice, the selected bases have been optimal.

2.4 Force Method for the Analysis of Frame Structures

2.4.4.2

71

Suboptimal Cycle Bases; an Indirect Approach

Definition 1. The weight of a member in the following algorithm is taken as the sum of the degrees of its end nodes. Algorithm B Step 1: Order the members of S in ascending order of weight. In all the subsequent steps use this ordered member set. Step 2: Generate as many admissible cycles of length α as possible, where α is the length of the shortest cycle of S. Denote the union of the selected cycles by Cm. When α is not specified, use the value α ¼ 3. Step 3: Select an admissible cycle of length α+1 on an unused member (use the ordered member set). Once such a cycle Cm+1 is found, control the other unused members for possible admissible cycles of length α. Again select an admissible cycle of length α+1 followed by the formation of possible α-sided cycles. This process is repeated until no admissible cycles of length α and α+1 can be found. Denote the generated cycles by Cn. Step 4: Select an admissible cycle Cn+1 of length α+2 on an unused member. Then check the unused members for the formation of α-sided cycles. Repeat Step 2 until no cycle of length α or α+1 can be generated. Repeat Step 3 until no cycles of length α, α+1 or α+2 can be found. Step 5: Take an unused member and generate an admissible cycle of minimal length on this member. Repeat Steps 1, 2 and 3. Step 6: Repeat steps similar to that of Step 4 until b1(S) admissible cycles, forming a suboptimal cycle basis, are generated. Using the ordered member set affects the selection process in two ways: 1. Generators are selected in ascending weight order, hence increasing the possibility of forming cycles from the dense part of the graph. This increases the chance of cycles with smaller overlaps being selected. 2. From cycles of equal length formed on a generator, the one with smallest total weight (sum of the weights of the members of a cycle) is selected. The cycle bases generated by this algorithm are suboptimal; however, the results are inferior to those of the direct method A. Remark. Once a cycle basis C is formed by Algorithm A or Algorithm B, it can further be improved by exchanging the elements of C. In each step of this process, a cycle Ck is controlled for the possibility of being exchanged by ring sum of Ck and a combination of the cycles of C \Ck, in order to reduce the overlap of the cycles. The process is repeated until no improvement can be achieved. This additional operation increases the computational time and storage, and should only be used when the corresponding effort is justifiable, e.g. this may be the case when a non-linear analysis or a design optimisation is performed using a fixed cycle basis.

72

2 Optimal Force Method: Analysis of Skeletal Structures

2.4.5

Examples

In this section, examples of planar and space frames are studied. The cycle bases selected by Algorithms A and B are compared with those developed for generating minimal cycle bases (Algorithms 1–4). Simple examples are chosen, in order to illustrate clearly the process of the methods presented. The models, however, can be extended to those containing a greater number of members and nodes of high degree, to show the considerable improvements to the sparsity of matrix CCt. Example 1. Consider a space frame as shown in Fig. 2.16a with the corresponding graph model S as illustrated in Fig. 2.16b. For this graph b1(S) ¼12, and therefore 12 independent cycles should be selected as a basis. Algorithm B selects a minimal cycle basis containing the following cycles, C1 ¼ ð1; 2; 3Þ, C2 ¼ ð1; 2; 5Þ, C3 ¼ ð1; 3; 4Þ, C4 ¼ ð1; 5; 4Þ, C5 ¼ ð2; 3; 6; 7Þ, C6 ¼ ð3; 4; 7; 8Þ, C7 ¼ ð4; 5; 8; 9Þ, C8 ¼ ð6; 7; 8; 9Þ, C9 ¼ ð7; 8; 11; 12Þ, C10 ¼ ð6; 7; 10; 11Þ, C12 ¼ ð9; 8; 12; 13Þ, C12 ¼ ð10; 11; 12; 13Þ

which corresponds to: χðCÞ ¼ 4 ( 3 þ 8 ( 4 ¼ 44, and χðCCt Þ ¼ 12 þ 2 ( 23 ¼ 58: Using Algorithm A leads to the formation of a similar basis, with the difference 0 that C8 ¼ (6,9,10,13) is generated in place of C8 ¼ (6, 7, 8, 9), corresponding to: + 0, χ C ¼ 4 ( 3 þ 8 ( 4 ¼ 44, + 0 0 t, χ C C ¼ 12 þ 2 ( 20 ¼ 52:

The CLNs and Ins of the members used in this algorithm are illustrated in Fig. 2.16b. Example 2. In this example, S is a space structure with b1(S) ¼ 33, as shown in Fig. 2.17a. Both Algorithms 3 and A select 33 cycles of length 4, i.e. a minimal cycle basis with χ(C) ¼ 4 ( 33 ¼ 132 is obtained. The basis selected by Algorithm 3 contains (in the worst case) all four-sided cycles of S except those which are shaded in Fig. 2.17a, with χ(CCt) ¼ 233.

2.4 Force Method for the Analysis of Frame Structures

a

73 13

b

3,4

10

3,4

11

9

8

3,4

6

3,4

12

7

5

4

3 , 3.67

2

3

2,4 1

Fig. 2.16 A space frame, and CLNs and Ins of its members. (a) A space structure. (b) The graph model S of the structure

a

b

Fig. 2.17 Minimal and suboptimal cycle bases of S. (a) A minimal cycle basis. (b) A suboptimal cycle basis

Algorithm A selects all three-sided cycles of S except those shaded in Fig. 2.17b, with χ(CCt) ¼ 190. It will be noticed that, for structures containing nodes of higher degrees, considerable improvement is obtained by the use of Algorithm A. Example 3. Consider a space frame as shown in Fig. 2.18, for which b1(S) ¼ 10. The minimal cycle basis selected by Algorithm 3 consists of the following cycles, C1 ¼ ð1; 2; 3Þ, C2 ¼ ð4; 5; 6Þ, C3 ¼ ð7; 8; 9Þ, C4 ¼ ð10; 11; 12Þ, C5 ¼ ð1; 2; 5; 4Þ, C6 ¼ ð2; 3; 6; 5Þ, C7 ¼ ð4; 5; 8; 7Þ, C8 ¼ ð5; 6; 9; 8Þ, C9 ¼ ð7; 8; 11; 10Þ, C10 ¼ ð8; 9; 12; 11Þ, corresponding to χ(C) ¼ 4 ( 3 + 6 ( 4 ¼ 36 and χ(CCt) ¼ 10 + 2 [0 + 0 + 0 + 2 + 3 + 3 + 4 + 3 + 4] ¼ 10 + 2 ( 19 ¼ 48. However, the following non-minimal cycle basis has a higher χ(C), and leads to a more sparse CCt matrix. The selected cycles are as follows,

74

2 Optimal Force Method: Analysis of Skeletal Structures 11

Fig. 2.18 A space frame S 8 5 10

2 7 1

12

4 9 6 3

C1 ¼ ð1; 2; 3Þ, C2 ¼ ð1; 2; 5; 4Þ, C3 ¼ ð2; 3; 6; 5Þ, C4 ¼ ð1; 3; 6; 4Þ, C5 ¼ ð4; 5; 8; 7Þ, C6 ¼ ð5; 6; 9; 8Þ, C7 ¼ ð4; 6; 9; 7Þ, C8 ¼ ð7; 8; 11; 10Þ, C9 ¼ ð8; 9; 12; 11Þ, C10 ¼ ð10; 11; 12Þ, for which χ(C0 ) ¼ 2 ( 3 + 8 ( 4 ¼ 38 corresponding to χ(C0 C0 t) ¼ 10 + 2 [1 + 2 + 3 + 1 + 2 + 3 + 1 + 2 + 2] ¼ 10 + 2 ( 16 ¼ 42. Therefore, the idea of having an optimal cycle basis in between minimal cycle bases is incorrect. Example 4. Consider the skeleton of a structure S, comprising of six flipped flags, as shown in Fig. 2.19a, for which b1(S) ¼ 6. After contraction, S0 is obtained as illustrated in Fig. 2.19b. Obviously, this is a planar graph. The CLNs for the members are 3 and IN for member (1, 2) is 6 and for the remaining members it is equal to 1., Algorithm 3 selects a minimal cycle basis for S0 , consists of six 3-sided cycles, corresponding to: χðCÞ ¼ 6 ( 3 ¼ 18 and χðCCt Þ ¼ 6 þ 2½0 þ 1 þ 2 þ 3 þ 4 þ 5' ¼ 6 þ 2 ( 15 ¼ 36 However, the following non-minimal cycle basis has a higher χ(S0 ), and leads to a lower sparsity, χ(C0 C0 t): C1 ¼ ð1; 3; 2; 4Þ, C2 ¼ ð1; 4; 2; 5Þ, C3 ¼ ð1; 2; 3Þ, C4 ¼ ð1; 2; 6Þ, C5 ¼ ð1; 6; 2; 7Þ, C6 ¼ ð1; 7; 2; 8Þ: For this basis, χ(C0 ) ¼ 4 ( 4 + 2 ( 3 ¼ 22, corresponding to χ(C0 C0 t) ¼ 6 + 2 [0 + 1 + 1 + 1 + 1 + 1] ¼ 6 + 2 ( 5 ¼ 16. After the back transformation from S0 to S, we have χ(C) ¼ 4 ( 6 + 2 ( 4 ¼ 32, corresponding to χ(CCt) ¼ 6 + 2 [0 + 1 + 1 + 1 + 1 + 1] ¼ 16.

2.4 Force Method for the Analysis of Frame Structures Fig. 2.19 A flipped flag before and after contraction. (a) S. (b) S0

a

75

b 1 5

4

3

6

7

8

2

2.4.6

An Improved Turn Back Method for the Formation of Cycle Bases

In this section, the combinatorial Turn-back method of Kaveh [15] is improved to obtain shorter cycle bases. This method covers all the counter examples, known for the minimality of the selected cycle bases. Step 1: Generate an SRT rooted from an arbitrary node O. Identify its chords, and order them according to their distance numbers from O. Step 2: Select the shortest length cycle of the graph on a chord and add this chord (generator) to the tree members. Repeat this process to all the chords, forming cycles of the least length containing the tree members and the previously used chords only. The selected cycles are all admissible, i.e. the addition of each cycle increases the first Betti number of the expanded part of the graph by unity. Store these cycles in C. Step 3: Form all the new cycles of the same length on the remaining chords, allowing the use of more than one unused chords in their formation. Step 4: Control the cycles formed in Step 3 to find only one cycle having a generator, which is in none of the other connected cycles formed in Step 3. When such a chord is found, add the corresponding cycle to C and include its generator in the tree members. Repeat this control until no such a cycle can be found. Step 5: Select a cycle of the next higher length in the graph containing only one chord. Add the selected cycle to C and its generator to the tree members. Step 6: Control the cycles formed in Step 3 to find a cycle containing only one unused chord. Add such a cycle to C and add its chord to the tree members. Repeat this control until no cycle of this property can be found. Step 7: Repeat Step 4. Step 8: Repeat Steps 5 and 6 and continue this repetition with the same length until no cycle in Step 5 can be found. Step 9: Repeat Steps 3 to 8, until b1(S) cycles forming a cycle basis is included in C.

76

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.20 Graph S and the selected SRT

4 3 2

1

2.4.7

5

1

9

13

8

12

16

7

11

15

6

10

14

5

9

13

1

4 3 2

1

Examples

Example 1. A graph is considered in the form of the 1-skeleton of a torus-type structure, Fig. 2.20. An SRT is selected, as shown in bold lines. The cycles selected in Step 2 are given in the following: C ¼ fð1; 2; 6Þ; ð1; 4; 5Þ; ð1; 5; 6Þ; ð1; 2; 13Þ; ð1; 4; 16Þ; ð1; 13; 16Þ; ð2; 3; 7Þ; ð2; 6; 7Þ; ð2; 3; 14Þ; ð2; 13; 14Þ; ð4; 5; 8Þ; ð4; 15; 16Þ; ð5; 6; 10Þ;

ð5; 9; 10Þ; ð5; 8; 9Þ; ð12; 13; 16Þ; ð11; 12; 16Þ; ð11; 15; 16Þg: The execution of Step 3 results in the following cycles:

ð3; 7; 8Þ, ð3; 4; 8Þ, ð7; 11; 12Þ, ð7; 8; 12Þ, ð8; 9; 12Þ, ð9; 13; 14Þ, ð9; 10; 14Þ, ð10; 14; 15Þ, ð10; 11; 15Þ, ð9; 12; 13Þ, ð3; 14; 15Þ, ð3; 4; 15Þ: Twelve cycles are generated, increasing the first Betti number by twelve. The control of Step 4, leads to generators {10, 11} and {7, 11} corresponding to the cycles (10, 11, 15) and (7, 11, 12), respectively. Thus no cycle is selected. In Step 5, a cycle of length 4 containing an unused chord is formed. On {3, 4}, cycle (1, 2, 3, 4) is generated and added to C. Then in Step 6, the following cycles are added to C: ð3; 4; 8Þ for f3; 8g, ð3; 7; 8Þ for f7; 8g, ð3; 4; 15Þ for f3; 15g, ð3; 14; 15Þ for f14; 15g: In Step 7 no cycle is found, but in Step 8, the execution of Step 5 leads to cycle (1, 5, 9, 13) on {9, 13}, and Step 6 leads to the following cycles completing C, and forming a minimal cycle basis of S:

2.4 Force Method for the Analysis of Frame Structures Fig. 2.21 A space graph and the selected SRT

77

1 5 4 3 2 1

11

6

16

21

10

15

20

25

9

14

19

24

8

13

18

23

7

12

17

22

6

11

16

21

26 30 29 28 27 26

ð9; 12; 13Þ for f9; 12g, ð9; 13; 14Þ for f9; 14g, ð8; 9; 12Þ for f8; 12g, ð7; 8; 12Þ for f7; 12g, ð7; 11; 12Þfor f7; 11g, ð9; 10; 14Þ for f10; 14g, ð10; 14; 15Þ for f10; 15g, and ð10; 11; 15Þ for f10; 11g:

Example 2. A space graph is considered as illustrated in Fig. 2.21. An SRT is selected as shown in bold lines. The application of Step 2, leads to the following cycle set: C ¼ fð1;2;6;7Þ; ð1;5;6;10Þ; ð2;3;7;8Þ; ð4;5;9;10Þ; ð6;7;11;12Þ; ð6;10;11;15Þ;

ð7;8;12;13Þ; ð9;10;14;15Þ; ð11;12;16;17Þ; ð11;15;16;20Þ; ð12;13;17;18Þ; ð14;15;19;20Þ; ð21;22;26;27Þ; ð21;25;26;30Þ; ð22;23;27;28Þ; ð24;25;29;30Þg:

In Step 3, the following cycles are generated: ð3; 4; 8; 9Þ, ð8; 9; 13; 14Þ, ð13; 14; 18; 19Þ, ð16; 17; 21; 22Þ, ð17; 18; 22; 23Þ, ð18; 19; 22; 23Þ, ð18; 19; 23; 24Þ, ð19; 20; 24; 25Þ, ð16; 20; 21; 25Þ, ð23; 24; 28; 29Þ: These cycles contain 11 unused chords. The control of Step 4 shows that {3, 4} and {28, 29} are included in one cycle, and therefore all the chords remain unused. In the next step, a cycle of length 5 including an unused chord is generated and added to C. Only with chord {3, 4}, the 5-sided cycle (1, 2, 3, 4, 5) is generated, and in Step 6 the following three-sided cycles are selected:

78

2 Optimal Force Method: Analysis of Skeletal Structures

a

5

6

b

11

10

9

8

7

i=5

4

3

1

j=11

2

Fig. 2.22 S and two of its SR subtrees

ð3; 4; 8; 9Þ, ð8; 9; 13; 14Þ, and ð13; 14; 18; 19Þ: Step 7 is carried out and cycle (23, 24, 28, 29) on {28, 29} is found, repetition of this control leads to cycle (18, 19, 23, 24) on {23, 24}. In the next step, no cycle is selected. The execution of Steps 3 and 4 in Step 9 results in no cycle. The execution of Step 5 in Step 9, forms cycle (1, 6, 11, 16, 21, 26) on chord {16, 21}, and the execution of Step 6 leads to the following cycles, ð16; 20; 21; 25Þ for f20; 25g, ð19; 20; 24; 25Þ for f19; 24g, ð16; 17; 21; 22Þ for f17; 22g, and ð17; 18; 22; 23Þ for f18; 23g: The selected cycles form a minimal cycle basis.

2.4.8

Formation of B0 and B1 Matrices

In order to generate the elements of a B0 matrix, a basic structure of S should be selected. For this purpose a spanning forest consisting of NG(S) SRTs is used, where NG(S) is the number of ground (support) nodes of S. As an example, for S shown in Fig. 2.22a, two SR subtrees are generated, Fig. 2.22b. The orientation assigned to each member of S is from the lower numbered node to its higher numbered end. For each SR subtree, the orientation is given in the direction of its growth from its support node. MATRIX B0: This is a 6M(S) ( 6NL(S) matrix, where M(S) and NL(S) are the numbers of members and loaded nodes of S, respectively. If all the free nodes are loaded, then NLðSÞ ¼ NðSÞ $ NGðSÞ, where NG(S) is the number of support nodes.

2.4 Force Method for the Analysis of Frame Structures

79

For a member, the internal forces are represented by the components at the lower numbered end. Obviously the components at the other end can be obtained by considering the equilibrium of the member. The coefficients of B0 can be obtained by considering the transformation of each joint load to the ground node of the corresponding subtree. [B0]ij for member i and node j is given by a 6 ( 6 submatrix as, 2

1

6 60 6 6 60 ½B0 'ij ¼ αij 6 60 6 6 6 Δz 4

0

0

0

0

1

0

0

0

0

1

0

0

1

0

$ Δx 0

1

0

0

$ Δz Δy 0

$Δy Δx

0

0

3

7 07 7 7 07 7, 07 7 7 07 5

ð2:63Þ

1

in which Δx, Δy and Δz are the differences of the coordinates of node j with respect to the lower numbered end of member i, in the selected global coordinate system, and αij is the orientation coefficient defined as: 8 < þ1 if member is positively oriented in the tree containing j, αij ¼ $1 if member is negatively oriented in the tree containing j, : 0 if member is not in the tree containing node j:

The B0 matrix can be obtained by assembling the [B0]ij submatrices as shown schematically in the following:

ð2:64Þ

MATRIX B1: This is a 6M(S) ( 6b1(S) matrix, which can be formed using the elements of a selected cycle basis. For a space structure, six self-equilibrating stress systems can be formed on each cycle. Consider Cj and take a member of this cycle as its generator. Cut the generator in the neighbourhood of its beginning node and apply six bi-actions as illustrated in Fig. 2.23. The internal forces under the application of each bi-action are a self-equilibrating stress system As for the matrix B0, a submatrix [B1]ij of B1 is a 6 ( 6 submatrix, the columns of which show the internal forces at the lower numbered end of member i under the application of six bi-actions at the cut of the generator j,

80

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.23 A cycle and the considered bi-action at a cut

j

y Cj

i

O

x

z

2

1

6 60 6 6 60 ½B1 'ij ¼ βij 6 60 6 6 6 Δz 4

0

0

0

0

1

0

0

0

0

1

0

0

1

0

$ Δx 0

1

0

0

$ Δz Δy 0

$Δy Δx

0

0

3

7 07 7 7 07 7, 07 7 7 07 5

ð2:65Þ

1

in which Δx, Δy and Δz are the differences of the coordinates x, y and z of the beginning node of the generator j and the beginning node of the member i. The orientation coefficient βij is defined as: 8 < þ1 if member i has the same orientation of the cycle generated on j, βij ¼ $1 if member i has the reverse orientation of the cycle generated on j, : 0 if member is not in the cycle whose generator is j: The pattern of B1 containing [B1]ij submatrices is shown in the following:

ð2:66Þ

Subroutines for the formation of B0 and B1 matrices are included in the program presented in Ref. [19].

2.4 Force Method for the Analysis of Frame Structures

81

Fig. 2.24 A four by four planar frame S

Fig. 2.25 Patterns of B1 and Bt1 B1 matrices for S. (a) Pattern of B1. (b) Pattern of Bt1 B1

Fig. 2.26 A simple space frame S

a

b

82

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.27 Patterns of B1 and Bt1 B1 matrices for S. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

Example 1. A four by four planar frame is considered as shown in Fig. 2.24. The patterns of B1 and Bt1 B1 formed on the elements of the cycle basis selected by any of the methods of the previous section are depicted in Fig. 2.25, corresponding to χ(B1) ¼ 241 and χ(Bt1 B1) ¼ 388. Example 2. A one-bay three-storey frame is considered as shown in Fig. 2.26. The patterns of B1 and Bt1 B1 matrices formed on the elements of the cycle basis selected by any of the graph theoretical algorithms of the previous Section are shown in Fig. 2.27, corresponding to χ(B1) ¼ 310 and χ(Bt1 B1) ¼ 562. Once B0 and B1 are computed, the remaining steps of the analysis are the same as those presented in Sect. 2.3.6. The interested reader may also refer to standard textbooks such as those of McGuire and Gallagher [31], Przemieniecki [32], or Pestel and Leckie [33] for further information.

2.5

Generalized Cycle Bases of a Graph

In this section, S is considered to be a connected graph. For γ(S) ¼ aM(S) + bN(S) + cγ0(S), the coefficients b and c are assumed to be integer multiples of the coefficient a > 0. Only those coefficients given in Table 2.1 are of interest.

2.5 Generalized Cycle Bases of a Graph

a

83

b

Fig. 2.28 Examples of (c) γ(S) ¼ M $ 3N + 6

a

γ-trees

(a)

c

γ(S) ¼ 3M $ 3N + 3.

(b)

γ(S) ¼ M $ 2N + 3.

b

Fig. 2.29 Structures satisfying γ(T) ¼ 0 which are not rigid. (a) γ(S) ¼ M $ 2N + 3. (b) γ(S) ¼ M $ 3N + 6

2.5.1

Definitions

Definition 1. A subgraph Si is called an elementary subgraph if it does not contain 0 0 a subgraph Si * Si with γ(Si ) > 0. A connected rigid subgraph T of S containing all the nodes of S is called a γ-tree if γ(T) ¼ 0. For γ(Si) ¼ b1(Si), a γ-tree becomes a tree in graph theory. Obviously a structure whose model is a γ-tree is statically determinate when γ(S) describes the degree of static indeterminacy of the structure. The ensuing stress resultants can uniquely be determined everywhere in the structure by equilibrium only. Examples of γ-trees are shown in Fig. 2.28. Notice that γ(T) ¼ 0 does not guarantee the rigidity of a γ-tree. For example, the graphs models depicted in Fig. 2.29 both satisfy γ(T) ¼ 0; however, neither is rigid. Definition 2. A member of S $ T is called a γ-chord of T. The collection of all γ-chords of a γ-tree is called the γ-cotree of S.

Definition 3. A removable subgraph Sj of a graph Si, is the elementary subgraph for which γ(Si $ Sj) ¼ γ(Si), i.e. the removal of Sj from Si does not alter its DSI. A γ-tree of S containing two chosen nodes, which has no removable subgraph is called a γ-path between these two nodes. As an example, the graphs shown in Fig. 2.30 are γ-paths between the specified nodes ns and nt. Definition 4. A connected rigid subgraph of S with γ(Ck) ¼ a, which has no removable subgraph is termed a γ-cycle of S. The total number of members of

84

2 Optimal Force Method: Analysis of Skeletal Structures

a

b

c

nt

nt

nt

ns

ns

ns

Fig. 2.30 Examples of γ-paths. (a) γ(S) ¼ α(M $ N + 1). (b) γ(S) ¼ M $ 2N + 3. (c) γ(S) ¼ M $ 3N + 6

a

b

c

Fig. 2.31 Examples of γ-cycles. (a) γ(S) ¼ α(M $ N + 1). (b) γ(S) ¼ M $ 2N + 3. (c) γ(S) ¼ M $ 3N + 6

a

b

Fig. 2.32 A planar truss S, and the elements of a GCB of S. (a) A planar truss S. (b) A generalized cycle basis of S

Ck, denoted by L(Ck), is called the length of Ck. Examples of γ-cycles are shown in Fig. 2.31.

2.5 Generalized Cycle Bases of a Graph

85

Definition 5. Let mi be a γ-chord of T. Then T [ mi contains a γ-cycle Ci which is defined as a fundamental γ-cycle of S with respect to T. Using the Intersection Theorem of Sect. 2.2.2, it can easily be shown that, γðT [ mi Þ ¼ 0 þ ða þ 2b þ cÞ $ ð2b þ cÞ ¼ a, indicating the existence of a γ-cycle in T [ mi. For a rigid T, the corresponding fundamental γ-cycle is also rigid, since the addition of an extra member between the existing nodes of a graph cannot destroy the rigidity. A fundamental γ-cycle can be obtained by omitting all the removable subgraphs of T [ mi.

Definition 6. A maximal set of independent γ-cycles of S is defined as a generalized cycle basis (GCB) of S. A maximal set of independent fundamental γ-cycles is termed a fundamental generalized cycle basis of S. The dimension of such a basis is given be η(S) ¼ γ(S)/a. As an example, a generalized cycle basis of a planar truss is illustrated in Fig. 2.32.

Definition 7. A generalized cycle basis-member incidence matrix C is an η(S) ( M matrix with entries $ 1, 0 and +1, where cij ¼ 1 (or $ 1) if γ-cycle Ci contains positively (or negatively) oriented member mj, and cij ¼ 0 otherwise. The generalized cycle adjacency matrix is defined as D which is an η(S) ( η(S) matrix when undirected γ-cycles are considered; then the negative entries of C become positive.

2.5.2

Minimal and Optimal Generalized Cycle Bases

A generalized cycle basis C ¼ {C1,C2, . . .,Cη(S)} is called minimal if it corresponds to a minimum value of: LðCÞ ¼

η ð SÞ X i¼1

LðCi Þ:

ð2:67Þ

Obviously, χ(C) ¼ L(C) and a minimal GCB can be defined as a basis which corresponds to minimum χ(C). A GCB for which L(C) is near minimum is called a subminimal GCB of S. A GCB corresponding to maximal sparsity of the GCB adjacency matrix is called an optimal generalized cycle basis of S. If χ(CCt) does not differ considerably from its minimum value, then the corresponding basis is termed suboptimal. The matrix intersection coefficient σi(C) of row i of GCB incidence matrix C is the number of row j such that:

86

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.33 A planar truss S and its associate graph A(S)

(a) j ∈ {i + 1, i + 2, . . ., η(S)}, (b) Ci \ Cj 6¼ ∅, i.e. there is at least one k such that the column k of both γ-cycles Ci and Cj (rows i and j) contain non-zero entries. Now it can be shown that: χðCCt Þ ¼ ηðSÞ þ 2

ηðX SÞ$1 i¼1

σj ðCÞ:

ð2:68Þ

This relationship shows the correspondence of a GCB incidence matrix C and that of its GCB adjacency matrix. In order to minimize χ(CCt), the value of ηðX SÞ$1 σj ðCÞ should be minimized, since η(S) is a constant for a given structure S, i¼1

i.e. γ-cycles with a minimum number of overlaps should be selected.

2.6

Force Method for the Analysis of Pin-Jointed Planar Trusses

The methods described in Sect. 2.5 are applicable to the selection of generalized cycle bases for different types of skeletal structures. However, the use of these algorithms for trusses engenders some problems, which are discussed in Ref. [34]. In this section, two methods are developed for selecting suitable GCBs for planar trusses. In both methods, special graphs are constructed for the original graph model S of a truss, containing all the connectivity properties required for selecting a suboptimal GCB of S.

2.6.1

Associate Graphs for Selection of a Suboptimal GCB

Let S be the model of a planar truss with triangulated panels, as shown in Fig. 2.33. The associate graph of S, denoted by A(S), is a graph whose nodes are in a one-toone correspondence with triangular panels of S, and two nodes of A(S) are connected by a member if the corresponding panels have a common member in S.

2.6 Force Method for the Analysis of Pin-Jointed Planar Trusses

87

Fig. 2.34 S with two cut-outs and its A(S)

a

Fig. 2.35 Two different types of cycles. (a) A type CI cycle. (b) A type CIII cycle

b

If S has some cut-outs, as shown in Fig. 2.34, then its associate graph can still be formed, provided that each cut-out is surrounded by triangulated panels. For trusses containing adjacent cut-outs, a cut-out with cut-nodes in its boundary, or any other form violating the above-mentioned condition, extra members can be added to S. The effect of such members should then be included in the process of generating its self-equilibrating stress systems. Theorem A. For a fully triangulated truss (except for the exterior boundary), as in Fig. 2.33, the dimension of a statical basis γ(S) is equal to the number of its internal nodes, which is the same as the first Betti number of its associate graph, i.e. γðSÞ ¼ Ni ðSÞ ¼ b1 ½AðSÞ':

ð2:69Þ

Proof. Let M0 and N0 be the numbers of members and nodes of A(S), respectively. By definition, 0

N ¼ RðSÞ $ 1, and M0 ¼ Mi(S) ¼ M(S) $ Me(S) ¼ M(S) $ Ne(S) ¼ M(S) $ [N(S) $ Ni(S)]. Thus: b1[A(S)] ¼ M0 $ N0 + 1 ¼ M(S) $ [N(S) $ Ni(S)] $ R(S) + 1 + 1 ¼ 2 $ R(S) + M(S) $ N(S) + Ni(S). By Euler’s polyhedron formula, we have: 2 $ RðSÞ þ MðSÞ $ NðSÞ ¼ 0: Therefore: For trusses which are not fully triangulated, we have:

88

2 Optimal Force Method: Analysis of Skeletal Structures

γðSÞ ¼ Ni ðSÞ $ Mc ðSÞ: A Cycle of A(S) and the Corresponding γ-Cycle of S. In Fig. 2.35a, a triangulated truss and its associate graph, which is a cycle, are shown for which γðSi Þ ¼ Ni ¼ 1 ¼ b1 ½AðSÞ': Since C1 of A(S) corresponds to one γ-cycle of S, it is called a type I cycle, denoted by CI. A γ-cycle of S is shown by continuous lines, and its γ-chords are depicted in dashed lines. Figure 2.35b shows a truss unit with one cut-out. In general, if a cut-out is an m-gon, then the completion of the triangulation requires m$3 members. Instead, m internal nodes will be created, increasing the DSI by m. Hence Eq. 2.68 yields, γðSÞ ¼ m $ ðm $ 3Þ ¼ 3, while: b1[A(S)] ¼ 1. However, in this case S contains three γ-cycles. A γ-path P and three γ-chords (dashed lines) are depicted in Fig. 2.35b. Obviously P[mi (i ¼ 1, 2, 3) form three γ-cycles which correspond to a cycle of type CIII of A(S). Thus two types of cycles CI and CIII should be recognized in A(S) and an appropriate number of γ-cycles will then be generated. Algorithm AA Step 1: Construct the associate graph A(S) of S. Step 2: Select a mesh basis of A(S), using an appropriate cycle selection algorithm. For fully triangulated S, Algorithms 1–3 generate cycle bases with three-sided elements. Step 3: Select the γ-cycles of S corresponding to the cycles of A(S). One γ- cycle for each cycle of type CI, and three γ-cycles for each cycle of type CIII should be chosen. Once a GCB is selected, on each γ-cycle one self-equilibrating stress system can easily be formed. Therefore, a statical basis with localized self-equilibrating stress systems will be obtained. Example. Let S be the graph model of a planar truss, as shown in Fig. 2.34, for which γ(S) ¼ 12. For A(S), six cycles of length 6 of type CI and two cycles of lengths 18 and 26 of type CIII are selected. Therefore, the total of 6 + 3 ( 2 ¼ 12 γ-cycles of S are obtained. On each γ-cycle one self-equilibrating stress system is constructed, and a statical basis consisting of localized self-equilibrating stress systems is thus obtained.

2.6 Force Method for the Analysis of Pin-Jointed Planar Trusses

2.6.2

89

Minimal GCB of a Graph

Theoretically a minimal GCB of a graph can be found using the Greedy Algorithm developed for matroids. This will be discussed in Kaveh [15, 20] after matroids have been introduced, and here only the algorithm is briefly outlined. Consider the graph model of a structure, and select all of its γ-cycles. Order the selected γ-cycles in ascending order of length. Denote these cycles by a set C. Then perform the following steps: Step 1: Choose a γ-cycle C1 of the smallest length, i.e. L(C1) < L(Ci) for all Ci ∈ C Step 2: Select the second γ-cycle C2 from C $ {C1} which is independent of C1 and L(C2) + L(Ci) for all γ-cycles of C $ {C1}. Step k: Subsequently choose a γ-cycle Ck from C $ {C1, C2, . . ., Ck$1} which is independent of C1, C2, . . ., Ck$1 and L(Ck) + L(Ci) for all Ci ∈ C $ {C1, C2, . . .,Ck $ 1}. After η(S) steps, a minimal GCB will be selected by this process, a proof of which can be found in Kaveh [19].

2.6.3

Selection of a Subminimal GCB: Practical Methods

In practice, three main difficulties are encountered in an efficient implementation of the Greedy Algorithm. These difficulties are briefly mentioned in the following: 1. Selection of some of the γ-cycles for some γ(S) functions. 2. Formation of all of the γ-cycles of S. 3. Checking the independence of γ-cycles. In order to overcome the above difficulties, various methods are developed. The bases selected by these approaches correspond to very sparse GCB adjacency matrices, although these bases are not always minimal. Method 1. This is a natural generalization of the method for finding a fundamental cycle basis of a graph, and consists of the following steps: Step 1: Select an arbitrary γ-tree of S, and find its γ-chords. Step 2: Add one γ-chord at a time to the selected γ-tree to form fundamental γ-cycles of S with respect to the selected γ-tree. The main advantage of this method is the fact that the independence of γ-cycles is guaranteed by using a γ-tree. However, the selected γ-cycles are often quite long, corresponding to highly populated CCB adjacency matrices. Method 2. This is an improved version of Method 1, in which a special γ-tree has been employed and each γ-chord is added to γ-tree members after being used for formation of a fundamental γ-cycle.

90

2 Optimal Force Method: Analysis of Skeletal Structures

Step 1. Select the centre “O” of the given graph. Methods for selecting such a node will be discussed in Chap. 5. Step 2: Generate a shortest route γ-tree rooted at the selected node O, and order its γ-chords according to their distance from O. The distance of a member is taken as the sum of the shortest paths between its end nodes and O. Step 3: Form a γ-cycle on the γ-chord of the smallest distance number, and add the used γ-chord to the tree members, i.e. form T [ m1. Step 4: Form the second γ-cycle on the next nearest γ-chord to O, by finding a γ-path in T [ m1 (not through m2). Then add the second used γ-chord m2 to T [ m1 obtaining T [ m1 [ m2. Step 5: Subsequently form the kth γ-cycle on the next unused γ-chord nearest to O, by finding a γ-path in the T [ m1 [ m2 [ . . . [ mk $ 1 (not through mk). Such a γ-path together with mk forms a γ-cycle. Step 6: Repeat Step 5 until η(S) of γ-cycles are selected. Addition of the used γ-chords to the γ-tree members leads to a considerable reduction in the length of the selected γ-cycles, while maintaining the simplicity of the independence check. In this algorithm, the use of an SRT, orders the nodes and members of the graph. Such an ordering leads to fairly banded member-node incidence matrices. Considering the columns corresponding to tree members as independent columns, a base is effectively selected for the cycle matroid of the graph, Kaveh [34]. Method 3. This method uses an expansion process, at each step of which one independent γ-cycle is selected and added to the previously selected ones. The independence is secured using an admissibility condition defined as follows. A γ-cycle Ck+1 added to the previous selected γ-cycles Ck ¼ C1 [ C2 [ . . . [ Ck is called admissible if, ( ' ( ' γ Ck [ Ckþ1 ¼ γ Ck þ a,

ð2:70Þ

where “a” is the coefficient defined in Table 2.1. The algorithm can now be described as follows. Step 1: Select the first γ-cycle of minimal length C1. Step 2: Select the second γ-cycle of minimal length C2 which is independent of C1, i.e. select the second admissible γ-cycle of minimal length. Step k: Subsequently, find the kth admissible γ-cycle of minimal length. Continue this process until η(S) independent γ-cycles forming a subminimal GCB are obtained. A γ-cycle of minimal length can be generated on an arbitrary member by adding a γ-path of minimal length between the two end nodes of the member (not through the member itself). The main advantage of this algorithm is avoiding the formation of all γ-cycles of S and also the independence control, which becomes feasible by graph theoretical methods.

2.7 Algebraic Force Methods of Analysis

91

The above methods are elaborated for specific γ(S) functions in subsequent sections, and examples are included to illustrate their simplicity and efficiency.

2.7

Algebraic Force Methods of Analysis

Combinatorial methods for the force method of structural analysis have been presented in previous sections. These methods are very efficient for skeletal structures and in particular for rigid-jointed frames. For general structures, the underlying graph of self-equilibrating stress systems will be discussed in Chaps. 6 and 7. Algebraic methods can be formulated in a more general form to cover different types of structures such as skeletal structures and finite element models. The main drawbacks of pure algebraic methods are the larger storage requirements, and the higher number of operations.

2.7.1

Algebraic Methods

Consider a discrete or discretized structure S, which is assumed to be statically indeterminate. Let r denote the m-dimensional vector of generalized independent element (member) forces, and p the n-vector of nodal loads. The equilibrium conditions of the structure can then be expressed as, Ar ¼ p,

ð2:71Þ

where A is an n ( m equilibrium matrix. The structure is assumed to be rigid, and therefore, A has a full rank, i.e. t ¼ m$n > 0, and rank A ¼ n. The member forces can be written as, r ¼ B0 p þ B1 q,

ð2:72Þ

where B0 is an m ( n matrix such that AB0 is an n ( n identity matrix, and B1 is an m ( t matrix such that AB1 is an n ( t zero matrix. B0 and B1 always exist for a structure and in fact many of them can be found for a structure. B1 is called a selfstress matrix as well as null basis matrix. Each column of B1 is known as a null vector. Notice that the null space, null basis and null vectors correspond to complementary solution space, statical basis and self-equilibrating stress systems, respectively, when S is taken as a general structure. Minimizing the potential energy requires that r minimize the quadratic form,

92

2 Optimal Force Method: Analysis of Skeletal Structures 1 t r Fm r, 2

ð2:73Þ

subject to the constraint as in Eq. 2.71. Fm is an m ( m block diagonal element flexibility matrix. Using Eq. 2.72, it can be seen that q must satisfy the following equation. '

( B1t Fm B1 q ¼ $B1t Fm B0 p,

ð2:74Þ

AP ¼ ½A1 ; A2 ',

ð2:75Þ

where Bt1 FmB1 ¼ G is the overall flexibility matrix of the structure. Computing the redundant forces q from Eq. 2.49, r can be found using Eq. 2.9. The structure of G is again important and its sparsity, bandwidth and conditioning govern the efficiency of the force method. For the sparsity of G one can search for a sparse B1 matrix, which is often referred to as the sparse null basis problem. Many algorithms exist for computing a null basis B1 of a matrix A. For the moment let A be partitioned so that,

where A1 is n ( n and non-singular, and P is a permutation matrix that may be required in order to ensure that A1 is non-singular. One can write: B1 ¼ P

"

# $A$1 1 A2 : I

ð2:76Þ

By simple multiplication it becomes obvious that: AB1 ¼ ½ A1

A2 '

"

$A$1 1 A2 I

#

¼ 0:

A permutation P that yields a non-singular A1 matrix can be chosen purely symbolically, but this says nothing about the possible numerical conditioning of A1 and the resulting B1. In order to control the numerical conditioning, pivoting must be employed. There are many such methods based on various matrix factorizations, including the Gauss-Jordan elimination, QR, LU, LQ and Turn-back method. Some of these methods are briefly studied in the following: Gauss-Jordan Elimination Method. In this approach one creates an n ( n identity matrix I in the first columns of A by column changes and a sequence of n pivots. This procedure can be expressed as, Gn Gn$1 . . . G2 G1 AP ¼ ½I; M',

ð2:77Þ

2.7 Algebraic Force Methods of Analysis

93

Fig. 2.36 Patterns of B1 and Bt1 B1 matrices for S using Gauss-Jordan elimination method, Example 1. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

Fig. 2.37 Patterns of B1 and Bt1 B1 matrices for S using Gauss-Jordan elimination method, Example 2. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

where Gi is the ith pivot matrix and P is an m ( m column permutation matrix (so Pt ¼ P) and I is an n(n identity matrix, and M is an n ( t matrix. Denoting GnGn$1 . . . G2G1 by G we have,

94

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.38 Patterns of B1 and Bt1 B1 matrices for S using LU decomposition method, Example 1. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

Fig. 2.39 Patterns of B1 and Bt1 B1 matrices for S using LU decomposition method, Example 2. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

or

GAP ¼ ½I; M', % & AP ¼ G$1 ½I; M' ¼ G$1 , G$1 M ,

which can be regarded as Gauss-Jordan factorization of A, and:

ð2:78Þ ð2:79Þ

2.7 Algebraic Force Methods of Analysis

"

G B0 ¼ P 0

#

95

and

"

$M B1 ¼ P I

#

ð2:80Þ

Example 1. The four by four planar frame of Fig. 2.24 is reconsidered. The patterns of B1 and Bt1 B1 formed by the Gauss-Jordan elimination method are depicted in Fig. 2.36, corresponding to χ(B1) ¼ 491 and χ(Bt1 B1) ¼ 1342. Example 2. The three-story frame of Fig. 2.24 is re-considered, and the GaussJordan elimination method is used. The patterns of B1 and Bt1 B1 matrices formed are shown in Fig. 2.37, corresponding to χ(B1) ¼ 483 and χ(Bt1 B1) ¼ 1592. LU Decomposition Method. Using the LU decomposition method, one obtains the LU factorization of A as, PA ¼ LU

and

UP ¼ ½U1 ; U2 ',

ð2:81Þ

P and P are again permutation matrices of order n ( n and m ( m, respectively. Now B0 and B1 can be written as: B0 ¼ P

"

$1 U$1 1 L P 0

#

and

B1 ¼ P

"

# $U$1 1 U2 : I

ð2:82Þ

Example 1. The four by four planar frame of Fig. 2.24 is re-considered. The patterns of B1 and Bt1 B1 formed by the LU factorization method are depicted in Fig. 2.38. The sparsity for the corresponding matrices are χ(B1) ¼ 408 and χ(Bt1 B1) ¼ 1248. Example 2. The three-storey frame of Fig. 2.24 is re-considered, and the LU factorization method is used. The patterns of B1 and Bt1 B1 matrices formed are shown in Fig. 2.39, corresponding to χ(B1) ¼ 504 and χ(Bt1 B1) ¼ 1530. QR Decomposition Method. Using a QR factorization algorithm with column pivoting yields, AP ¼ Q½R1 ; R2 ',

ð2:83Þ

where P is again a permutation matrix, and R1 is an upper triangular matrix of order n. B1 can be obtained as: "

# R2 $R$1 1 : B1 ¼ P I

ð2:84Þ

Turn-back LU Decomposition Method. Topc¸u developed a method, the so-called Turn-back LU procedure, which is based on LU factorization and often results in highly sparse and banded B1 matrices. Heath et al. [35] adopted this method for use with QR factorization. Due to the efficiency of this method, a brief description of their approach will be presented in the following.

96

2 Optimal Force Method: Analysis of Skeletal Structures

Fig. 2.40 Patterns of B1 and Bt1 B1 matrices for S using Turn-back LU decomposition method, Example 1. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

Fig. 2.41 Patterns of B1 and Bt1 B1 matrices for S using Turn-back LU decomposition method, Example 2. (a) Pattern of B1. (b) Pattern of Bt1 B1

a

b

Write the matrix A ¼ (a1,a2, . . .,an) by columns. A start column is a column such that the ranks of (a1, a2, . . ., as$1) and (a1, a2, . . ., as) are equal. Equivalently, as is a start column if it is linearly dependent on lower-numbered columns. The coefficients of this linear dependency give a null vector whose highest numbered non-zero is in position s. It is easy to see that, the number of start columns is m$n ¼ t, the dimension of the null space of A.

2.7 Algebraic Force Methods of Analysis

97

The start column can be found by performing a QR factorization of A, using orthogonal transformations to annihilate the subdiagonal non-zeros. Suppose that in carrying out the QR factorization we do not perform column interchanges but simply skip over any columns that are already zero on and below the diagonal. The result will then be a factorization of the form:

ð2:85Þ

The start columns are those columns where the upper triangular structure jogs to the right; that is, as is a start column if the highest non-zero position in column s of R, is no larger than the highest non-zero position in earlier columns of R. The Turn-back method finds one null vector for each start column as, by “turning back” from column s to find the smallest k for which columns as, as$1, . . ., as$k are linearly dependent. The null vector has a non-zero only in position s-k throughs. Thus if k is small for most of the start columns, then the null basis will have a small profile. Notice that the turn-back operates on A, and not on R. The initial QR factorization of A is used only to determine the start columns, and then discarded. The null vector that Turn-back finds from start column as may not be non-zero in position s. Therefore, Turn-back needs to have some way to guarantee that its null vectors are linearly independent. This can be accomplished by forbidding the leftmost column of the dependency for each null vector from participating in any later dependencies. Thus, if the null vector for start column as has its first non-zero in position s-k, every null vector for a start column to the right of as will be zero in position s-k. Although the term “Turn-back” is introduced in Ref. [7], the basic idea had also been used in Refs. [36]. Since this correspondence simplifies the understanding of the Turn-back method, it is briefly described in the following. For the Algorithm 1 of Sect. 2.3, the use of an SRT orders the nodes and members of the graph simultaneously, resulting in a fairly banded member-node incidence matrix B. Considering the columns of B corresponding to tree members as independent columns, effectively a cycle is formed on each ordered chord (start column) by turning back in B and establishing a minimal dependency, using the tree members and previously used chords. The cycle basis selected by this process forms a base for the cycle matroid of the graph, as it is described in Kaveh [37]. Therefore, the idea used in Algorithm 1 and its generalization for the formation of a generalized cycle bases in Ref. [38] seems to constitute a similar idea to that of the algebraic Turn-back method. Example 1. The four by four planar frame of Fig. 2.24 is re-considered. The patterns of B1 and Bt1 B1 formed by the Turn-back LU factorization method are depicted in Fig. 2.40, corresponding to χ(B1) ¼ 240 and χ(Bt1 B1) ¼ 408.

98

2 Optimal Force Method: Analysis of Skeletal Structures

Example 2. The four by four planar frame of Fig. 2.24 is re-considered, and the Turn-back LU factorization method is used. The patterns of B1 and Bt1 B1 matrices formed are shown in Fig. 2.41, corresponding to χ(B1) ¼ 476 and χ(Bt1 B1) ¼ 984. A comparative study of various force methods has been made in Ref. [30]. Many algorithms have been developed for selection of null bases, and the interested reader may refer to Refs. [38, 39].

References 1. Henderson JC de C, Bickley WG (1955) Statical indeterminacy of a structure. Aircr Eng 27:400–402 2. Maunder EWA (1971) Topological and linear analysis of skeletal structures. Ph.D. thesis, London University, IC 3. Kaveh A (1992) Recent developments in the force method of structural analysis. Appl Mech Rev 45:401–418 4. Langefors B (1961) Algebraic topology and elastic networks, SAAB TN49. Linko¨ping 5. Denke PH (1962) A general digital computer analysis of statically indeterminate structures, NASA-TD-D-1666 6. Robinson J (1973) Integrated theory of finite element methods. Wiley, New York 7. Topc¸u A (1979) A contribution to the systematic analysis of finite element structures using the force method (in German). Doctoral dissertation, Essen University 8. Kaneko I, Lawo M, Thierauf G (1982) On computational procedures for the force methods. Int J Numer Method Eng 18:1469–1495 9. Gilbert JR, Heath MT (1987) Computing a sparse basis for the null space. SIAM J Algebra Discr Method 8:446–459 10. Coleman TF, Pothen A (1987) The null space problem II; algorithms. SIAM J Algebra Discr Method 8:544–561 11. Patnaik SN (1986) Integrated force method versus the standard force method. Comput Struct 22:151–164 12. Kaveh A, Jahanshahi M (2006) An efficient program for cycle basis selection and bandwidth optimization. Asian J Civil Eng 7(1):95–109 13. Kaveh A, Daei M (2010) Suboptimal cycle bases of graphs using an ant colony system algorithm. Eng Comput 27(4):485–494 14. Timoshenko S, Young DH (1945) Theory of structures. McGraw-Hill, New York 15. Kaveh A (1974) Application of topology and matroid theory to the analysis of structures. Ph.D. thesis, London University, IC 16. Kaveh A (1988) Topological properties of skeletal structures. Comput Struct 29:403–411 17. Mauch SP, Fenves SJ (1967) Release and constraints in structural networks. J Struct Div ASCE 93:401–417 18. Mu¨ller-Breslau H (1912) Die graphische Statik der Baukonstruktionen. Alfred Kro¨ner Verlag, 1907, und Leipzig 19. Kaveh A (2004) Structural mechanics: graph and matrix methods, 3rd edn. Research Studies Press, Baldock 20. Henderson JC de C, Maunder EWA (1969) A problem in applied topology. J Inst Math Appl 5:254–269 21. Kaveh A (1976) Improved cycle bases for the flexibility analysis of structures. Comput Method Appl Mech Eng 9:267–272 22. Kaveh A (1988) Suboptimal cycle bases of graphs for the flexibility analysis of skeletal structures. Comput Method Appl Mech Eng 71:259–271

References

99

23. Stepanec GF (1964) Basis systems of vector cycles with extremal properties in graphs. Uspekhi Mat Nauk 19:171–175 (in Russian) 24. Zykov AA (1969) Theory of finite graphs. Nuaka, Novosibirsk (in Russian) 25. Hubicka E, Syslø MM (1975) Minimal bases of cycles of a graph. In: Fiedler M (ed) Recent advances in graph theory. Academia Praha, Prague, pp 283–293 26. Kaveh A, Roosta GR (1994) Revised Greedy algorithm for the formation of minimal cycle basis of a graph. Commun Numer Method Eng 10:523–530 27. Horton JD (1987) A polynomial time algorithm to find the shortest cycle basis of a graph. SIAM J Comput 16:358–366 28. Lawler EL (1976) Combinatorial optimization; networks and matroids. Holt, Rinehart and Winston, New York 29. Kolasinska E (1980) On a minimum cycle basis of a graph. Zastos Math 16:631–639 30. Kaveh A, Mokhtar-zadeh A (1993) A comparative study of the combinatorial and algebraic force methods. In: Proceedings of the Civil-Comp93, Edinburgh, pp 21–30 31. Brusa L, Riccio F (1989) A frontal technique for vector computers. Int J Numer Method Eng 28:1635–1644 32. Prezemieniecki JS (1968) Theory of matrix structural analysis. McGraw-Hill, New York 33. Pestel EC, Leckie FA (1963) Matrix methods in elastomechanics. McGraw-Hill, New York 34. Kaveh A (1993) Matroids applied to the force method of structural analysis. Z Angew Math Mech 73:T355–T357 35. Heath MT, Plemmons RJ, Ward RC (1984) Sparse orthogonal schemes for structural optimization using the force method. SIAM J Sci Stat Comput 5:514–532 36. Cassell AC (1976) An alternative method for finite element analysis; a combinatorial approach to the flexibility method. Proc R Soc Lond A352:73–89 37. Kaveh A (1979) A combinatorial optimization problem; optimal generalized cycle bases. Comput Method Appl Mech Eng 20:39–52 38. Coleman TF, Pothen A (1986) The null space problem I; complexity. SIAM J Algebra Disc Method 7:527–537 39. Plemmons RJ, White RE (1990) Substructuring methods for computing the null space of equilibrium matrices. SIAM J Matrix Anal Appl 11:1–22

Chapter 3

Optimal Displacement Method of Structural Analysis

3.1

Introduction

In this chapter, the principles introduced in Chap. 1 are used for the formulation of the general displacement method of structural analysis. Computational aspects are discussed and many worked examples are included to illustrate the concepts and principles being used. In order to show the generality of the methods introduced for the formation of the element stiffness matrices, the stiffness matrix of a simple finite element is also derived. Special attention is paid to the graph theory aspects of the displacement method for rigid jointed structures, where the pattern equivalence of structural and graph theory matrices is used. The standard displacement method employs cocycle bases of structural graph models; however, for general solutions a cutset basis of the model should be employed. This becomes vital, when solutions leading to well conditioned stiffness matrices are required. Methods for the selection of such cutset bases are described in this chapter. In the last half-century, considerable progress has been made in the matrix analysis of structures; see for example, Argyris and Kelsey [1], Livesley [2], McGuire and Gallagher [3], Przemieniecki [4], Zienkiewicz [5], and Kaveh [6, 7]. The topic has been generalized to finite elements, and extended to the stability, non-linear and dynamic analysis of structures. This progress is due to the simplicity, modularity and flexibility of matrix methods.

3.2

Formulation

In this section, a matrix formulation using the basic tools of structural analysis— equilibrium of forces, compatibility of displacements, and force-displacement relationships—is provided. The notations are chosen from those most often encountered versions in structural mechanics. A. Kaveh, Computational Structural Analysis and Finite Element Methods, 101 DOI 10.1007/978-3-319-02964-1_3, © Springer International Publishing Switzerland 2014

102

3 Optimal Displacement Method of Structural Analysis

a

b 2

2

2

3 1

0 2

c

0

3 0

3 0

3 0

2

6

6

6 0 0

6

6

6 0

0

0 0

Fig. 3.1 The degrees of freedom of the joints for three structures. (a) A planar truss. (b) A planar frame. (c) A space frame

3.2.1

Coordinate Systems Transformation

Consider a structure S with M members and N nodes; each node having α degrees of freedom (DOF). The degree of kinematic indeterminacy (DKI) of S may then be determined as, ηðSÞ ¼ αN $ β,

ð3:1Þ

where β is the number of constraints due to the support conditions. As an example, η(S) for the planar truss S depicted in Fig. 3.1a is given by η(S) ¼ 7 % 2 $ 3 ¼ 11, for the plane frame illustrated in Fig. 3.1b, it is calculated as η(S) ¼ 8 % 3 $ 4 % 3 ¼ 12, and for the space frame shown in Fig. 3.1c, it is calculated as η(S) ¼ 12 % 6 $ 6 % 6 ¼ 36. One can also calculate η(S) by simple addition of the degrees of freedom of the joints of the structure, i.e. for the truss S, η(S) ¼ 2 + 2 + 2 + 2 + 2 + 1 ¼ 11, and for the planar frame η(S) ¼ 4 % 3 ¼ 12, and for the space frame η(S) ¼ 6 % 6 ¼ 36. For a structure, the stiffness matrices of the elements should be prepared in a single coordinate system known as the global coordinate system, in order to be able to perform the assembling process. However, the stiffness matrices of individual members are usually written first in coordinate systems attached to the members, known as local coordinate systems. Therefore a transformation is needed, before the assembling process. Typical local and global coordinate systems are illustrated in Fig. 3.2. A global coordinate system can be selected arbitrarily, however, it may be advantageous to select this system such that the structure falls in the first quadrant of the plane, in order to have positive coordinates for the nodes of the structure. On the other hand, a local coordinate system of a member is so chosen that it has one of its axes along the member, the second axis lies in its plane of symmetry (if it has one) and the third axis is chosen such that it results in a right handed coordinate system. The transformation from a local coordinate to a global coordinate system can be performed as illustrated in Fig. 3.3, in which x, y, z is the global system and x2, y2, z2, often denoted by xyz, is the local system.

3.2 Formulation

103 y

y

x

x

y

x

O

Fig. 3.2 Local x, y and global coordinate x, y systems

a

b

y

y y1

y1

y2

x2 b

a

z1

b

x

a

z2 z1

x1

x1 z

z

c y2

x

d

y

y

y1

y3

x3 x2

yji

g x

g

z3 z1 z 2

zji

i

j

L

x ji L*

x

x1

z

z

Fig. 3.3 Transformation from local coordinate system to global coordinate system

For rotation about the y axis the relation between x1, y1, z1 and x, y, z can be expressed as: 2

3 2 x1 cosα 4 y1 5 ¼ 4 0 $sinα z1

0 1 0

32 3 sinα x 0 54 y 5: cosα z

ð3:2Þ

Similarly, for rotation about the z1 axis x2, y2, z2 and x1, y1, z1 are related by,

104

3 Optimal Displacement Method of Structural Analysis

2

3 2 x2 cosβ 4 y2 5 ¼ 4 $sinβ 0 z2

sinβ cosβ 0

32 3 0 x1 0 54 y1 5 1 z1

ð3:3Þ

and for rotation about the x2 axis x3, y3, z3 and x2, y2, z2 are related as: 2

3 2 x3 1 4 y3 5 ¼ 4 0 0 z3

32 3 0 x2 sinγ 54 y2 5 cosγ z2

0 cosγ $sinγ

ð3:4Þ

Combining the above transformations, results in: 2

3 ðcosαcosβÞ ðsinβÞ ðcosβsinαÞ T ¼ 4 $ðsinαsin γ þ cosαsinβcosγÞ ðcosβcosγÞ ðsinγcosα $ sinαsinβcosγÞ 5: $ðsinαcos γ $ cosαsinβsinγÞ ð$cosβsinγÞ ðcosαcosγ þ sinαsinβsinγÞ 2

2 3 x3 x 4 y3 5 ¼ ½T( 4 y 5: z z3

where :

3

ð3:5Þ

ð3:6Þ

The representation of a vector in the local coordinate system Γ and the global coordinate system Γ are related by: Γ ¼ T Γ:

ð3:7Þ

It can easily be proved that T is an orthogonal matrix, i.e. ½T($1 ¼ ½T(t :

ð3:8Þ

In the above transformation, γ represents the tilt of the member, which is quite often zero. Thus, T can be simplified as, 2

cosαcosβ T ¼ 4 $cosαsinβ $sinα

sinβ cosβ 0

3 sinαcosβ $sinαsinβ 5: cos α

ð3:9Þ

and for the two dimensional case and “α equal to zero”, T reduces to: T¼

!

cos β $sinβ

sinβ cosβ

"

ð3:10Þ

Equation 3.9 can easily be written in terms of the coordinates of the two ends of a vector. Considering Fig. 3.3b and using simple trigonometry, Eq. 3.9 becomes,

3.2 Formulation

105

2

xji =L T ¼ 4 $xji yji =L ) L $zji =L)

yji =L L ) =L 0

3 zji =L yji zji =L ) L 5, xji =L)

ð3:11Þ

where: xji ¼ xj $ xi yji ¼ yj $ yi zji ¼ zj $ zi # $12 # $12 L) ¼ z2ji þ x2ji and L ¼ z2ji þ y2ji þ x2ji :

ð3:12Þ

Notice that T transforms a 3-dimensional vector from a global to a local coordinate system and Tt performs the reverse transformation. However, if the element forces or element displacements (distortions) consist of p vectors, the block diagonal matrix with p submatrices should be used. As an example, for a beam element of a space frame, with each node having six degrees of freedom, the transformation matrix is a 12 % 12 matrix of the form: 2

6 T¼6 4

3.2.2

3

T T T T

7 7: 5

ð3:13Þ

Element Stiffness Matrix Using Unit Displacement Method

Consider a general element, as shown in Fig. 3.4, with n member forces, r m ¼ fr 1 r 2 . . . r n gt ,

ð3:14Þ

um ¼ fu1 u2 . . . un gt :

ð3:15Þ

and n member displacements:

A typical force component ri can be found by using the unit displacement method to be, ððð _t ri ¼ ε i σdV, ð3:16Þ V

_

where ε i represents the matrix of compatible strains due to a unit displacement in the direction of ri, and σ is the exact stress matrix due to the applied forces rm. The unit displacements can be used in turn for all the points where member forces are applied, and therefore,

106

3 Optimal Displacement Method of Structural Analysis r 2 ,u 2

Fig. 3.4 A general element with its nodal loads and nodal displacements

r 3 ,u 3

r ,u 4 4

...

r1 ,u1

r ,u i i

...

r n ,u n

rm ¼

ððð

_t

ε σdV,

ð3:17Þ

V

where: _

ε¼

n_ _ o _ t ε1 ε2 . . . εn :

ð3:18Þ

For a linear system the total strain,

can be expressed as,

& 't e ¼ exx eyy ezz exy eyz exz :

ð3:19Þ

e¼bu,

ð3:20Þ

where b is the exact strain due to the unit displacement u. The stress-strain relationship can be written as, σ ¼ χbu,

ð3:21Þ

where: 2

1$v v 6 v 1 $ v 6 6 v v 6 E 6 χ¼ 6 ð1 þ vÞð1 $ 2vÞ 6 6 6 4 Substituting in Eq. 3.17 leads to,

v v 1$v

3 1 $ 2v 2

0

0

1 $ 2v 2

0

0

1 $ 2v

0

0

2

7 7 7 7 7 7 7 7 7 5

ð3:22Þ

3.2 Formulation

107

rm ¼ or

ððð

V

_t

ε χbdVum ,

rm ¼ km um ,

where:

km ¼

ððð

ð3:23Þ

ð3:24Þ

_t

ε χbdV,

V

ð3:25Þ

represents the element stiffness matrix. The evaluation of the matrix b, representing the exact strain distributions can often be difficult, if not impossible. Hence in case there is no exact distribution, an approximate relationship may be used. Usually the matrix b is selected such that it will satisfy the equations of compatibility at least. Denoting this approximate _

_

_

matrix by ε and using ε ¼ b results in: ððð _ _ km ¼ b t χ b dV: V

ð3:26Þ

This equation will be used for the derivation of the stiffness matrices of a finite element in Sect. 3.5.1. As an example, consider a prismatic bar element shown in its local coordinate system, in Fig. 3.5. According to the definition of such an element, only axial forces are present. From the theory of elasticity, the axial strain is expressed as: εxx ¼ strain ¼

∂ux , ∂x

ð3:27Þ

The displacement ux along the longitudinal axis of the bar can be expressed as: ux ¼ A 1 x þ A 2 :

ð3:28Þ

From the boundary conditions: ux ¼ u1 at x ¼ 0,

ux ¼ u4 at x ¼ L:

ð3:29Þ

u4 $ u1 L

ð3:30Þ

Hence: A1 ¼ By substitution in Eq. 3.28:

and

A 2 ¼ u1 :

108

3 Optimal Displacement Method of Structural Analysis

y

Fig. 3.5 A bar element in its local coordinate system

r 4,u 4

r1 ,u1

j

i

x

z

ux ¼

u4 $ u1 x þ u1 : L

ð3:31Þ

Now axial strain can be evaluated as: εxx

∂ux 1 1 ¼ ðu2 $ u1 Þ ¼ ½ $1 ¼ L L ∂x

!

" u1 þ1 ( : u2

ð3:32Þ

The above strain distribution is exact, and 1 b^ ¼ b ¼ ½ $1 L

þ1 (:

ð3:33Þ

Since a bar element is one dimensional, χ is a 1 % 1 matrix defined as: χ ¼ E:

ð3:34Þ

Substituting in Eq. 3.26 leads to: km ¼

ðL 0

! " 1 $1 E ½ $1 L 1 L

1 (Adx,

ð3:35Þ

and ! EA 1 km ¼ L $1

" $1 : 1

ð3:36Þ

This method will also be used for the derivation of the finite element stiffness matrices in subsequent sections.

3.2 Formulation

3.2.3

109

Element Stiffness Matrix Using Castigliano’s Theorem

In this section, a different approach is described for the formation of element stiffness matrices, using Castigliano’s theorem. Consider a general element as shown in Fig. 3.4. Suppose that loads are applied at certain points (specified as nodes) 1, 2, . . ., n. Let vi be the displacement of node i along the applied load pi. The loads are applied in a pseudo-static manner increasing gradually from zero. Assuming a linear behaviour, the work done by an external force p ¼ {p1, p2, . . ., pn} through the displacement v ¼ {v1, v2, . . ., vn} can be written as: 1 W ¼ ðp1 v1 þ p2 v2 þ . . . þ pn vn Þ: 2

ð3:37Þ

According to the principle of the conservation of energy, W ¼ U,

ð3:38Þ

1 U ¼ ðp1 v1 þ p2 v2 þ . . . þ pn vn Þ: 2

ð3:39Þ

and therefore:

If a small variation is now given to vi while keeping the other displacement components constant, then the variation of v with respect to vi can be written as: ! " ∂U 1 ∂p1 ∂p2 ∂pn ¼ p þ v1 þ v2 þ . . . þ vn : ∂vi 2 i ∂vi ∂vi ∂vi

ð3:40Þ

According to Castigliano’s theorem: ∂U ¼ pi : ∂vi

ð3:41Þ

! " ∂p1 ∂p2 ∂pn pi ¼ v1 þ v2 þ . . . þ vn , ∂vi ∂vi ∂vi

ð3:42Þ

Thus,

or in a matrix form for all i ¼ 1, . . ., n, we have:

110

3 Optimal Displacement Method of Structural Analysis

2

∂p1 6 ∂ 2 3 6 v1 p1 6 ∂p1 6 p2 7 6 6 7 6 6 6 * 7 ¼ 6 ∂v2 6 7 6 * 4 * 5 6 6 * pn 6 ∂p 4 1 ∂vn

∂p2 ∂v1 ∂p2 ∂v2 * * ∂p2 ∂vn

* * * * * * * *

* *

3 ∂pn ∂v1 7 72 3 7 v1 ∂pn 76 v 7 76 2 7 ∂v2 7 7 76 6 * 7: * 7 4 7 * 5 * 7 v ∂pn 7 5 n ∂vn

ð3:43Þ

According to definition, the above coefficient matrix forms the stiffness matrix of the elastic body defined by its n nodes as illustrated in Fig. 3.4. A typical element of the stiffness matrix kij is given by: kij ¼

∂pj : ∂vi

ð3:44Þ

Using Castigliano’s first theorem: ( ) 2 ∂ ∂U ∂ U kij ¼ : ¼ ∂vi ∂vj ∂vi ∂vj

ð3:45Þ

Similarly: 2

kji ¼

∂pi ∂ U ¼ : ∂vj ∂vj ∂vi

ð3:46Þ

Since the order of differentiation should not affect the result for our problems, we have: kij ¼ kji ,

ð3:47Þ

which is a proof of the symmetry of the stiffness matrices both for a structure and for an element. As an example, consider a prismatic bar element as shown in its local coordinate system, Fig. 3.5. According to the definition of such an element, only axial forces are present. The strain energy of this bar can be calculated as: ððð ððð ð 1 E EA 2 U¼ σxx εxx dxdydz ¼ ε2xx dxdydz ¼ εxx dx: ð3:48Þ 2 2 2 On the other hand:

3.2 Formulation

111

εxx ¼

∂ux : ∂x

ð3:49Þ

Using Eq. 3.31, by substituting in Eq. 3.48, the strain energy of the bar is calculated to be: U¼ Hence

+ EA * 2 u4 $ 2u4 u1 þ u21 : 2L

ð3:50Þ

2

k11 ¼

∂ U EA , ¼ L ∂u21 2

k14 ¼ k41 ¼

∂ U EA , ¼$ ∂u1 ∂u4 L

ð3:51Þ

2

k44 ¼

∂ U EA , ¼ L ∂u24

kij ¼ 0 for all other components. Therefore, the stiffness matrix of a bar element in the selected local coordinate system is obtained, and: 2 3 1 r1 6 0 6 r2 7 6 6 7 6 r3 7 EA 6 0 6 6 7¼ 6 r4 7 L 6 6 $1 6 7 4 0 4 r5 5 0 r6 2

3.2.4

0 0 0 0 0 0

0 0 0 0 0 0

$1 0 0 1 0 0

0 0 0 0 0 0

32 3 0 u1 6 u2 7 07 76 7 6 7 07 76 u3 7: 6 7 07 76 u4 7 0 54 u5 5 0 u6

ð3:52Þ

The Stiffness Matrix of a Structure

Let p and v represent the joint loads and joint displacements of a structure. Then the force-displacement relationship for the structure can be expressed as, p ¼ Kv,

ð3:53Þ

where K is a αN % αN symmetric matrix, known as the stiffness matrix of the structure. Expanding the ith equation of the above system, the force pi can be expressed in terms of the displacements {v1,v2, . . .,vaN} as: pi ¼ Ki1 v1 þ Ki2 v2 þ . . . þKiαN vαN :

ð3:54Þ

112

3 Optimal Displacement Method of Structural Analysis

A typical coefficient Kij is the value of the force pi required to be applied at the ith component of the structure in order to produce a displacement vj ¼ 1 at j and zero displacements at all the other components. The member forces r can be related to nodal forces p by: p ¼ Br:

ð3:55Þ

Using the contragradient relationship, the joint displacements v can be related to member distortions u by: u ¼ Bt v:

ð3:56Þ

For each individual member of the structure, the member forces can be related to member distortions by an element stiffness matrix km. A block diagonal matrix containing these element stiffness matrices is known as the unassembled stiffness matrix of the structure, denoted by k. Obviously: r ¼ ku:

ð3:57Þ

This equation together with Eqs. 3.55 and 3.56 yields: p ¼ BkBt v:

ð3:58Þ

K ¼ BkBt ,

ð3:59Þ

Therefore,

is obtained. The matrix K is singular since the boundary conditions of the structure are not yet applied. For an appropriately supported structure, the deletion of the rows and columns of K corresponding to the support constraints results in a positive definite matrix, known as the reduced stiffness matrix of the structure. A symmetric matrix S is called positive definite if xtSx > 0 for every non-zero vector x. As shown before, the stiffness matrix K of a structure is symmetric. This matrix is also positive definite since, pt v ¼ ðKvÞt v ¼ vt Kt v ¼ vt Kv ¼ 2W,

ð3:60Þ

and W is always positive. Let us illustrate the stiffness method by means of a simple example. Consider a fixed end beam with a load P applied at its mid span. This beam is discretized as two beam elements, as shown in Fig. 3.6a with two degrees of freedom for each node (axial deformation is ignored for simplicity). The components of element forces and element distortions are depicted in Fig. 3.6b and those of the entire structure are illustrated in Fig. 3.6c. For each element such as element 1, the stiffness matrix can be written as:

3.2 Formulation

113

a

Fig. 3.6 Illustration of the analysis of a simple structure. (a) A fixed ended beam S. (b) Member forces and member distortions. (c) Nodal forces and nodal displacements of the entire structure

P

1

1

2

L

b r2 ,u 2

1 1

c

p ,v

2 2

1

3 2 k11 r1 6 r2 7 6 k21 6 7¼6 4 r3 5 4 k31 r4 k41

k12 k22 k32 k42

k13 k23 k33 k43

K13 K23 K33 K43 K53 K63

K14 K24 K34 K44 K54 K64

2

r7 ,u 7

4 4

p ,v 2

p ,v

1 1

r8 ,u8

r5 ,u

p ,v

p ,v

2

r6 ,u 6

r3 ,u 3

r ,u

3

L

r4 ,u4

1

2

3 3

32 3 u1 k14 6 7 k24 7 7 6 u2 7 , k34 54 u3 5 k44 u4

6 6

p ,v

5 5

ð3:61Þ

and for the entire structure we have: 3 2 K11 p1 6 p2 7 6 K21 6 7 6 6 p3 7 6 K31 6 7¼6 6 p4 7 6 K41 6 7 6 4 p5 5 4 K51 p6 K61 2

K12 K22 K32 K42 K52 K62

K15 K25 K35 K45 K55 K65

32 3 v1 K16 6 v2 7 K26 7 76 7 6 7 K36 7 7 6 v3 7 : 6 7 K46 7 7 6 v4 7 K56 54 v5 5 K66 v6

ð3:62Þ

Element stiffness matrices k1 and k2 can be easily constructed using the definition of kij. For a beam element, ignoring its axial deformation, these terms are shown in Fig. 3.7. The structure has a uniform cross section and both elements have the same length. Therefore, using the force displacement relationship from Chap. 1:

114

3 Optimal Displacement Method of Structural Analysis

Fig. 3.7 Stiffness coefficients of a beam element ignoring its axial deformation

k 21 u 1=1 k 11

k 41

k 22

k 31

k12

k 42

u2 =1

k 43 k 23

k 24 u3 =1

k 13

k33

2

6=L2 2EI 6 6 $3=L k1 ¼ k2 ¼ L 4 $6=L2 $3=L

$3=L 2 3=L 1

k44 u4 =1

k 14

$6=L2 3=L 6=L2 3=L

k 32

3 $3=L 1 7 7: 3=L 5 2

k 34

ð3:63Þ

The unassembled stiffness matrix is an 8 % 8 matrix of the form: !

k1 k¼ 0

" 0 : k2

ð3:64Þ

Now consider the equilibrium of the joints of the structure, resulting in, p1 ¼ r 1 , p 2 ¼ r 2 , p3 ¼ r 3 þ r 5 , p4 ¼ r 4 þ r 6 , p5 ¼ r 7 , p6 ¼ r 8 :

ð3:65Þ

or in a matrix form we have, 2

3

2

1 p1 6 p2 7 6 * 6 7 6 6 p3 7 6 * 6 7¼6 6 p4 7 6 * 6 7 6 4 p5 5 4 * * p6

* 1 * * * *

* * 1 * * *

* * * 1 * *

* * * * 1 * * 1 * * * *

* * * * 1 *

2 3 3 r1 7 * 6 6 r2 7 7 6 * 76 r3 7 7 6 7 *7 76 r4 7, 6 7 *7 76 r5 7 5 7 * 6 6 r6 7 4 1 r7 5 r8

ð3:66Þ

and more compactly, p ¼ Br, where:

ð3:67Þ

3.2 Formulation

115

2

1 * 6* 1 6 6* * B¼6 6* * 6 4* * * *

* * 1 * * *

* * * 1 * *

* * * * 1 * * 1 * * * *

* * * * 1 *

3 * *7 7 *7 7, *7 7 *5 1

is known as the equilibrium matrix. Consider now the compatibility of displacements: u 1 ¼ v1 , u2 ¼ v2 , u3 ¼ u 5 ¼ v3 , u4 ¼ u6 ¼ v 4 , u7 ¼ v5 , u 8 ¼ v6 :

ð3:68Þ

In a matrix form we have, 3 2 1 * u1 6 u2 7 6 * 1 6 7 6 6 u3 7 6 * * 6 7 6 6 u4 7 6 * * 6 7¼6 6 u5 7 6 * * 6 7 6 6 u6 7 6 * * 6 7 6 4 u7 5 4 * * * * u8 2

* * 1 * 1 * * *

* * * 1 * 1 * *

3 * * 2 3 * *7 7 v1 6 7 * *7 76 v2 7 6 7 * *7 76 v3 7, 6 7 * *7 76 v4 7 4 5 * *7 7 v5 1 * 5 v6 * 1

ð3:69Þ

and in compact form: u ¼ Ev ¼ Bt v:

ð3:70Þ

where: 2

1 6* 6 6* 6 6* E¼6 6* 6 6* 6 4* *

* 1 * * * * * *

* * 1 * 1 * * *

* * * * * * 1 * * * 1 * * 1 * *

3 * *7 7 *7 7 *7 7, *7 7 *7 7 *5 1

is known as the compatibility matrix. The reason for the matrix E being the transpose of the matrix B, has already been discussed in the previous chapter, however, by using the principle of virtual work, a simple proof can be obtained. Consider: W ¼ work done by external loads ¼ 12 vt p,

116

3 Optimal Displacement Method of Structural Analysis

U ¼ strain energy ¼ 12 ut r. Then equating W and U, leads to E ¼ Bt and completes the proof. It should be mentioned that this equality holds for a general structure, and it is the result of the contragradient relationship introduced in Chap. 1. The stiffness matrix of the entire structure is then obtained as: 2

6=L2 6 $3=L 6 2 2EI 6 6 $6=L K¼ L 6 6 $3=L 4 0 0

$6=L2 3=L 12=L2 0 $6=L2 $3=L

$3=L 2 3=L 1 0 0

$3=L 1 0 4 3=L 1

0 0 $6=L2 3=L 6=L2 3=L

3 0 0 7 7 $3=L 7 7: 1 7 7 3=L 5

ð3:71Þ

2

Applying the boundary conditions, v1 ¼ v2 ¼ v5 ¼ v6 ¼ 0, by deleting the rows and columns corresponding to zero displacements, leads to the formation of the following reduced stiffness matrix: !

p3 p4

"

¼

! 2EI 12=L2 0 L

0 4

"! 3

" v3 : v4

ð3:72Þ

3L ¼ $PL Since p4 ¼ 0 and p3 ¼ $ P, therefore v3 ¼ p24EI 24EI :

3.2.5

3

Stiffness Matrix of a Structure; an Algorithmic Approach

From the above simple example, it can be seen that the matrix B is a very sparse Boolean matrix and the direct formation of BkBt using matrix multiplication requires a considerable amount of storage. In the following, it is shown that one can form BkBt with an assembling process (known also as planting), as follows: Consider an element “a” of a structure, as shown in Fig. 3.8, for which the element stiffness matrix can be written as, !

kii ka ¼ kji

" kij , kjj

ð3:73Þ

i and j are the two end nodes of member a. Pre and post multiplication in the form of BkBt has the following effect on ka:

3.2 Formulation

117

Fig. 3.8 A structural model S

a 7 3

2

0 60 6 60 6 6I 6 60 6 60 6 40 0

6

8

3 0 07 7 07 7! "! 07 7 kii kij 0 0 0 I 0 07 7 kji kjj 0 0 0 0 0 I7 7 05 0 2 1 0 26 60 36 60 0 46 ¼ 6 56 60 66 60 740 8 0

2

0 60 6 6 " 60 6I 0 0 0 ¼6 60 I 0 0 6 60 6 40 0

0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 kii 0 0 0 kji 0 0 0 0

0 0 0 0 0 0 0 0

5

4

2

1

3 0 07 7 07 7! " 07 7 0 0 0 kii 0 kij 0 0 ð3:74Þ 07 7 0 0 0 kji 0 kjj 0 0 I7 7 05 0 3 0 0 0 0 0 07 7 0 0 07 7 kij 0 0 7 7 0 0 07 7 kjj 0 0 7 7 0 0 05 0 0 0

The adjacency matrix of S is also an 8 % 8 matrix, and the effect of node 4 being adjacent to node 6, is the existence of unit entries in the same locations as the submatrices of the element “a”. One can build up the adjacency matrix of a graph by the addition of the effect of one member at a time. In the same way, one can also form the overall stiffness matrix of the structure by the addition of the contribution of every member in succession. As an example, for the graph shown in Fig. 3.8, the overall stiffness matrix has the following pattern: 122 3 4 5 6 1 1 * * 1 * 26 6* 1 * * 1 36 6* * 1 * * 46 61 * * 1 1 56 6* 1 * 1 1 66 6* * * 1 1 74 * * 1 * 1 8 * * * * *

7 * * * 1 1 1 * 1

8 3 * * * *7 7 1 *7 7 * *7 7: 1 *7 7 * 17 7 1 15 1 1

ð3:75Þ

118

3 Optimal Displacement Method of Structural Analysis

Non-zero entries are shown by “1”. For a stiffness matrix each of these non-zero entries is an η % η submatrix, where η is the degrees of freedom of each node of the structure. As an example, for a planar truss η ¼ 2, and for a space frame η ¼ 6. The formation of the stiffness matrix by the above process is known as the assembling or planting of the stiffness matrix of a structure.

3.3

Transformation of Stiffness Matrices

Methods for the formation of element stiffness matrices have been presented in the previous section. In the following the stiffness matrices for bar and beam elements are transformed to global coordinate systems using the transformation described in Sect. 3.2.1. From Eq. 3.7, we have: r ¼ Tr,

ð3:76Þ

u ¼ Tu:

ð3:77Þ

From the definition of an element stiffness matrix in a local coordinate system: r ¼ ku:

ð3:78Þ

By substitution of Eqs. 3.76 and 3.77 into the above equation: r ¼ T$1 kTu ¼ Tt kTu:

ð3:79Þ

By definition of a stiffness matrix in a global coordinate system: r ¼ ku:

ð3:80Þ

Comparison of Eqs. 3.79 and 3.80 results in: k ¼ Tt kT:

3.3.1

ð3:81Þ

Stiffness Matrix of a Bar Element

Equation 3.52 provides the stiffness matrix of a bar element in its local coordinate system. This matrix in the global system, as shown in Fig. 3.9, can be written as: k¼ Denoting T in Eq. 3.32 by,

!

T T

"t

! * + T k

T

"

:

ð3:82Þ

3.3 Transformation of Stiffness Matrices

119 r5 ,u5

Fig. 3.9 A bar element of a space truss

j

y r2 ,u2 i O

r4 ,u4

r 6 ,u 6 r 1,u 1

r3 ,u3

x

z

2

T11 T ¼ 4 T21 T31

T12 T22 T32

3 T13 T23 5, T33

ð3:83Þ

km can be written as, 2

T211 6 T11 T12 6 EA 6 6 T11 T213 k¼ L 6 6 $T11 4 $T11 T12 $T11 T13

T212 T12 T13 $T11 T12 $T212 $T12 T13

3

sym: T213 $T11 T13 $T12 T13 $T213

T211 T11 T12 T11 T13

T212 T12 T13

T213

7 7 7 7, 7 7 5

ð3:84Þ

where “sym.” denotes the symmetry of the matrix. The entries of the above matrix can be found using the Tij from Eq. 3.32. As an example, the stiffness matrix of bar 1 in the planar truss shown in Fig. 3.10 can be obtained as: pﬃﬃﬃ 2 1 , T11 ¼ , -12 ¼ pﬃﬃﬃ ¼ 2 2 2 2 2 x12 þ y12 þ z12 pﬃﬃﬃ y21 1 2 p ﬃﬃ ﬃ T12 ¼ , : ¼ $ ¼ $ -12 2 2 2 2 2 x12 þ y12 þ z12 x21

Therefore:

120

3 Optimal Displacement Method of Structural Analysis

Fig. 3.10 A planar truss and the selected global coordinate system

1

3 2

1

L

2

0:5 EA 6 $0:5 k1 ¼ pﬃﬃﬃ 6 L 2 4 $0:5 0:5

3.3.2

$0:5 0:5 0:5 $0:5

$0:5 0:5 0:5 $0:5

y

2

L

x

3 0:5 $0:5 7 7: $0:5 5 0:5

Stiffness Matrix of a Beam Element

Consider a prismatic beam element as shown in Fig. 3.11. The element forces and element distortions are defined by the following vectors, r ¼ fr1 ; r2 ; r3 ; . . . ; r12 gt , and u ¼ fu1 ; u2 ; u3 ; . . . ; u12 gt , where r1 to r3 are the force components at end i and r4 to r6 are moment components at end i. Also r7 to r9 are the force and r10 to r12 are the moment components, respectively at the end j, and ui (i ¼ 1, . . ., 12) are correspondingly the translations and rotations at the ends i and j of the element. Using one of the methods presented in Sect. 3.2.2, the stiffness matrix of the beam element, in the local coordinate system defined in Fig. 3.11, can be obtained from Eq. 3.83 as: 2

3 A 0 0 0 0 0 $A 0 0 0 0 0 2 6 0 12Iz =L2 7 0 0 0 6I =L 0 0 $12I =L 0 0 6I =L z z z 6 7 6 0 0 $6Iy =L 0 0 0 $12Iy =L2 0 $6Iy =L 0 7 0 12Iy =L2 6 7 6 0 7 0 0 J=2 ð 1 þ v Þ 0 0 0 0 0 $J=2 ð 1 þ v Þ 0 0 6 7 6 0 7 0 $6I =L 0 4I 0 0 0 $6I =L 0 2I 0 y y y y 6 7 7 E6 =L 0 0 0 4I 0 $6I =L 0 0 0 2I 0 6I z z z z 7 k¼ 6 7 $A 0 0 0 0 0 A 0 0 0 0 0 L6 6 7 2 6 0 $12Iz =L2 7 0 0 0 $6I =L 0 12I =L 0 0 0 $6I =L z y z 6 7 2 2 6 0 7 0 $12I =L 0 6I =L 0 0 0 12I =L 0 6I =L 0 y y y y 6 7 6 0 7 0 0 $J=2 ð 1 þ v Þ 0 0 0 0 0 J=2 ð 1 þ v Þ 0 0 6 7 4 0 0 2Iy 0 0 0 6Iy =L 0 4Iy 0 5 0 $6Iy =L 0 6Iz =L 0 0 0 2Iz 0 $6Iz =L 0 0 0 4Iz ð3:85Þ

3.3 Transformation of Stiffness Matrices

121 y

Fig. 3.11 A beam element in the local coordinate system

j

i

x

z

In this matrix, Iy, Iz and J are the moments of inertia with respect to the y and z axes and J is the polar moment of inertia of the section. E specifies the elastic modulus and v is the Poisson ratio. L denotes the length of the beam. For the two-dimensional case, the columns and rows corresponding to the third dimension can easily be deleted, to obtain the stiffness matrix of an element of a planar frame. The stiffness matrix in a global coordinate system can be written as: 2

6 k¼6 4

3t

T T T T

2

7* +6 7 k 6 5 4

3

T T T T

7 7: 5

ð3:86Þ

For the two-dimensional case: k¼

!

T T

"t

! * + T k

T

"

:

ð3:87Þ

The entries of k are as follows: k11 ¼ T211 α1 þ T221 α4z

k21 ¼ T11 T12 α1 þ T21 T22 α4z k22 ¼ T212 α1 þ T222 α4z k31 ¼ T21 α2z k32 ¼ T22 α2z k33 ¼ α3z

k41 ¼ $T211 α1 þ T221 α4z k42 ¼ $T21 T22 α4z $ T12 T11 α1 k43 ¼ $T21 α2z k44 ¼ $T21 α2z k51 ¼ $T21 T22 α4z $ T12 T11 α1 k52 ¼ $T221 α4z $ T212 α1 k53 ¼ $T22 α2z k54 ¼ T21 T22 α4z þ T12 T11 α1 k55 ¼ T222 α4z þ T212 α1

k61 ¼ T21 α2z k62 ¼ T22 α2z k63 ¼ α6z k64 ¼ $T21 α2z k65 ¼ $T22 α2z k66 ¼ α3z :

ð3:88Þ

122

3 Optimal Displacement Method of Structural Analysis

in which α1 ¼

EA z 6EIz z 4EIz z 12EIz 2EIz , α ¼ 2 , α3 ¼ , α4 ¼ : , and α6z ¼ 3 L 2 L L L L

As an example, for element 1 of the planar frame, shown in Fig. 3.12, we have, T11 ¼ 0

T12 ¼ 1 T21 ¼ $1 T22 ¼ 0,

and the stiffness matrix of the element is obtained as: 2

1:25 6 0 6 6 6 6 $0:75 k1 ¼ 10 6 6 $1:25 4 0 $0:75

3.4

200 0 6 0 0:75 $200 0 0 3

3

sym: 1:25 0 0:75

200 0

6

7 7 7 7: 7 7 5

Displacement Method of Analysis

Once the stiffness matrix of an element is obtained in the selected global coordinate system, it can be planted in the specified and initialised overall stiffness matrix of the structure K, using the process described in Sect. 3.2.5. Example. Let S be a planar truss with an arbitrary nodal and element numbering, as shown in Fig. 3.13. The entries of the transformation matrices of the members are calculated using Eqs. 3.32 and 3.33 as follows: pﬃﬃ pﬃﬃ y2 $y1 3$0 3 1$0 1 1 For bar 1: T11 ¼ x2 $x ¼ ¼ and T ¼ ¼ ¼ 12 2 : 2 2 2 2 2 Similarly, pﬃﬃ for bar 2: T11 ¼ 12 T12 ¼ $ 23 , and for bar 3: T11 ¼1, T12 ¼ 0. Using the following relationship, 2

3 2 Fix T211 y 6 Fi 7 EA 6 T11 T12 6 x7¼ 6 4 Fj 5 L 4 $T211 Fjx $T11 T12

T11 T12 T212 $T11 T12 $T212

$T211 $T11 T12 T211 T11 T12

32 x 3 δi $T11 T12 6 δiy 7 $T212 7 76 x 7 T11 T12 54 δj 5 δjy T212

ð3:89Þ

the stiffness matrices of the members are computed directly in the selected global coordinate system.

3.4 Displacement Method of Analysis

123 4m

Fig. 3.12 A planar frame

4m

y

A = 4 ´ 10-3 m2

3

2 2

I = 30 ´ 10-6 m4

1

E = 2 ´ 1011 N/m2

1

x

Fig. 3.13 A planar truss and the selected global coordinate system

30 kN

20 kN 2 y

4

3

1

3m

2 x

1 1m

3 1m

1m

Now the stiffness matrices can be formed using Eq. 5.62: 2 3 0:25 sym: 6 0:433 7 0:75 6 7:. For bar 1: k1 ¼ EA 2 4 $0:25 5 $0:433 0:25 $0:433 $0:75 0:433 0:75 2 3 0:25 sym: 6 $0:433 0:75 7 6 7: For bar 2: k2 ¼ EA 2 4 $0:25 5 0:433 0:25 0:433 $0:75 $0:433 0:75 2 3 1 sym: 6 0 0 7 6 7: For bar 3: k3 ¼ EA 2 4 $1 0 1 5 0 0 0 0 The overall stiffness matrix of the structure is an 8 % 8 matrix, which can easily be formed by planting the three member stiffness matrices as follows: 2

3 0:250 0:433 $0:250 $0:433 0 0 0 0 6 0:433 0:750 $0:433 $0:750 0 0 0 07 6 7 6 $0:250 $0:433 1:500 0 $0:250 0:433 $1:00 0 7 7 6 EA 6 0 1:500 0:433 $0:750 0 07 7: 6 $0:433 $0:750 K¼ 0 $0:250 0:433 0:250 $0:433 0 07 2 6 7 6 0 6 0 0 0:433 $0:750 $0:433 0:750 0 07 7 6 4 0 0 $1:00 0 0 0 1:00 0 5 0 0 0 0 0 0 0 0

124

3 Optimal Displacement Method of Structural Analysis

Partitioning K into 2 % 2 submatrices, it can easily be seen that it is pattern equivalent to the node adjacency matrix of the graph model of the structure as follows: 2

∗ 6 ∗ C ) C)t ¼ 6 40 0

∗ ∗ ∗ ∗

0 ∗ ∗ 0

3 0 ∗7 7: 05 ∗

This pattern equivalence simplifies certain problems in structural mechanics, such as ordering the variables for bandwidth or profile reduction. Methods for increasing the sparsity, using special cutset bases, and improving the conditioning of structural matrices, are discussed in Refs. [6, 7].

3.4.1

Boundary Conditions

The matrix K is singular, since the boundary conditions have to be applied. Consider, p ¼ Kv, and partition it for free and constraint degrees of freedom as: !

pf pc

"

¼

!

Kff Kcf

Kfc Kcc

"!

" vf : vc

ð3:90Þ

This equation has a mixed nature; pf and vc have known values and pc and vf are unknowns. Kff is known as the reduced stiffness matrix of the structure, which is non-singular for a rigid structure. For boundary conditions such as vc ¼ 0, it is easy to delete the corresponding rows and columns to obtain, pf ¼ Kff vf ,

ð3:91Þ

from which vf can be obtained by solution of the above set of equations. In a computer this can be done by multiplying the diagonal entries of Kcc by a large number such as 1020. An alternative approach is possible by equating the diagonal entries of Kcc to unity and all the other entries of these rows and columns to zero. If vc contains some specified values, pc will have corresponding vc values. A third method, which is useful when a structure has more constraint degrees of freedom (such as many supports), consists of the formation of element stiffness matrices considering the corresponding constraints, i.e. to form the reduced stiffness matrices of the elements in place of their complete matrices. This leads to some reduction in storage, and is also at the expense of additional computational effort.

3.4 Displacement Method of Analysis

125

As an example, the reduced stiffness matrix of the structure shown in Fig. 3.13 can be obtained from K, by deleting the rows and columns corresponding to the three supports 1, 3 and 4: !

" ! EA 1:5 20 ¼ 30 0 2

0 1:5

"!

" u2x : u2y

Solving for the joint displacements, we have: u2x ¼

40 40 and u2y ¼ : 1:5EA EA

The member distortions can easily be extracted from the displacement vector, and multiplication by the stiffness matrix of each member results in its member forces in the global coordinate system. As an example, for member 3 we have: 2 3 r2x 1 6 r2y 7 EA 6 0 6 6 7¼ 4 r4x 5 2 4 $1 r4y 0 2

0 0 0

sym: 1 0

32

3 2 3 40=1:5EA 13:33 76 40=EA 7 6 0 7 76 7¼6 7 54 5 4 $13:33 5: 0 0 0 0

A transformation yields the member forces in the local coordinate systems, r1 ¼ f $23:99 23:99 gt , r2 ¼ f $10:659 10:65 gt and r3 ¼ f 13:33 $13:33 gt .

3.4.2

General Loading

The joint load vector of a structure can be computed in two parts. The first part comes from the external concentrated loads and/or moments, which are applied to the joints defined as the nodes of S. The components of such loads are most easily specified in a global coordinate system and can be entered into the joint load vector p. The second part comes from the loads, which are applied to the spans of the members. These loads are usually defined in the local coordinate system of a member. For each member the fixed end actions (FEA) can be calculated using existing classical formulae or tables. A simple computer program can be prepared for this purpose. The fixed end actions should then be expressed in the global coordinate system using the transformation matrix given by Eq. 3.11. The FEA should then be reversed and applied to the end nodes of the members. These components can be added to p to form the final joint load vector. After p has been prepared and the boundary conditions imposed, the corresponding equations should be solved to obtain the joint displacements of the structure. Member distortions can then be extracted for each member in the reverse order to that used in assembling the p vector.

126

3 Optimal Displacement Method of Structural Analysis

a

b

P=6kN

q=1.2kN/m

4m

2m

1.4kN.m

3kN.m

2m

Fig. 3.14 A continuous beam and its equivalent loading

Example 1. A two span continuous beam is considered as shown in Fig. 3.14a. EI is taken to be constant along the beam. For continuous beams, the transformation matrix T from local coordinate to global coordinate is identity, and therefore km ¼ km , i.e. no transformation is required. Ignoring the axial deformation and using Eq. 3.63, the stiffness matrices of the elements are obtained as: 2

$0:75 $1:5 0:75 $1:5

0:75 1:5 64 6 1:5 4 k1 ¼ k2 ¼ 6 4 4 $0:75 $1:5 1:5 2

3 1:5 2 7 7: $1:5 5 4

Assembling the overall stiffness matrix and imposing the boundary conditions, the reduced stiffness matrix of the entire beam is obtained and the forcedisplacement relationship for beam is written as: !

" ! $1:40 8 ¼ 16 3 2

2 4

"!

" θ2z : θ3z

Solving the equations leads to: !

θ2z θ3z

"

¼

! 1 4 448 $2

$2 8

"!

" ! " $1:4 $0:0259 ¼ : 3 0:0598

Member forces are calculated as: 2

3 2 V1 0:75 1:5 6 M1 7 6 4 6 7 ¼ 166 1:5 4 V2 5 4 $0:75 $1:5 1:5 2 M2 2 3 1:779 6 0:772 7 7 ¼6 4 3:021 5, $3:256 and

$0:75 $1:5 0:75 $1:5

32 3 2 3 1:5 0 2:4 6 7 6 1:6 7 2 7 0 76 7þ6 7 5 4 2:4 5 $1:5 54 0 4 $0:0259 $1:6

3.4 Displacement Method of Analysis

2

3 2 0:75 V2 6 M2 7 6 1:5 6 7 6 4 V3 5 ¼ 164 $0:75 1:5 M3 2 3 3:814 6 3:258 7 7 ¼6 4 2:186 5: 0

127

1:5 $0:75 4 $1:5 $1:5 0:75 2 $1:5

32 3 2 3 1:5 0 3 6 7 6 7 2 7 76 $0:0259 7 þ 6 3 7 5 4 3 5 $1:5 54 0 4 þ0:0598 $3

Example 2. A portal frame is considered as shown in Fig. 3.15. The members are made of sections with A ¼ 150 cm2 and Iz ¼ 2 % 104cm4 and E ¼ 2 % 104 kN/ cm2. Calculate the joint rotations and displacements. The equivalent joint loads are illustrated in Fig. 3.16. Employing Eq. 3.88, the stiffness matrices for the members are obtained as: For member 1: 2

0:008 6 0 6 6 4 6 $1:5 k1 ¼ 10 6 6 0:008 4 0 $1:5

0:75 0 400 0 1:5 $0:75 0 0 200

3

sym: 0:008 0 1:5

0:75 0

400

7 7 7 7, 7 7 5

and for member 2: 2

0:6 6 0 6 6 0 k2 ¼ 104 6 6 $0:6 6 4 0 0

0:004 0:96 0 $0:004 0:96

3

sym: 320 0 $0:96 160

0:6 0 0

0:004 $0:96

320

7 7 7 7 7 7 5

For member 3: 2

0:008 6 0 6 6 1:5 46 k3 ¼ 10 6 6 $0:008 4 0 1:5

0:75 0 0 $0:75 0

3

sym: 400 $1:5 0:008 0 0 200 $1:5

0:75 0

400

7 7 7 7 7 7 5

Assembling the stiffness matrices and imposing the boundary conditions results in the following equations:

128

3 Optimal Displacement Method of Structural Analysis

Fig. 3.15 A portal frame and its loading

2

5kN

3

2 1.2kN/m

1

3

1

4

4m

5m 160kN.m

Fig. 3.16 Equivalent joint loads 7.4kN

y x

3 2 7:4 0:608 6 0 7 6 0 7 6 6 6 160 7 6 7 ¼ 104 6 1:5 6 6 0 7 6 $0:6 7 6 6 4 0 5 4 0 0 0 2

0:754 0:96 720 0 0 $0:004 $0:96 0:96 160

32

0:608 0 1:5

3 δ2x y 76 δ2 7 sym: 76 z 7 76 θ2 7 76 x 7 76 δ 7 76 3y 7 54 δ 5 0:754 3 θ3z $0:96 720

Solving these equations leads to: δ2x ¼ 0:0659167,

δ3x ¼ 0:0653377,

δ2y ¼ 2:617764E $ 04,

θ2z ¼ $8:983453E $ 05,

δ3y ¼ $2:617704E $ 04 and θ3z ¼ $1:16855E $ 04:

The final member forces can be found using the stiffness of the members, superimposed by the fixed end actions.

3.5

Stiffness Matrix of a Finite Element

In this section, a simple element is introduced from finite element methods, in order to show the capability of the method presented in Sect. 3.2.2, for the formation of element stiffness matrices.

3.5 Stiffness Matrix of a Finite Element

3.5.1

129

Stiffness Matrix of a Triangular Element

For plane stress and plane strain problems, the displacements of a node can be specified by two components, and therefore for each node of the triangular element, two degrees of freedom is considered, as shown in Fig. 3.17. Element forces and displacements are defined by the following vectors: rm ¼

&

r1

r2

. . . r6 gt and um ¼

&

u1

u2

...

u6 gt :

ð3:92Þ

A triangular element has its boundary attached continuously to the surrounding medium, and therefore no exact stiffness matrix can be derived. Therefore an approximate solution should be sought. The following displacement functions can be considered for the variation of the displacements, u ¼ α1 x þ α2 y þ α3 and v ¼ α4 x þ α5 y þ α6 ,

ð3:93Þ

where α1, α2, . . ., α6 are arbitrary constants which can be found from the displacements of the three nodes of the element. From the boundary conditions, at node ið#xi ; yi Þ,$u ¼ ui and v ¼ vi , at node j xj ; yj , u ¼ uj and v ¼ vj , at node kðxk ; yk Þ, u ¼ uk and v ¼ vk ,

ð3:94Þ

the constants can be evaluated. Substituting in Eq. 3.93 yields: h i o # $i , ykj x $ xj $ xkj y $ yj u1 þ ½$yki ðx $ xk Þ $ xki ðy $ yk Þ(u3 þ yji ðx $ xi Þ $ xji ðy $ yi Þ u5 , h i o nh , # $i v ¼ 1=2A ykj x $ xj $ xkj y $ yj u2 þ ½$yki ðx $ xk Þ $ xki ðy $ yk Þ(u4 þ yji ðx $ xi Þ $ xji ðy $ yi Þ u6 :

u ¼ 1=2A

nh

ð3:95Þ

where: 2A ¼ 2ðarea of the triangleÞ ¼ xkj yji $ xji ykj ,

ð3:96Þ

xmn ¼ xm $ xn and ymn ¼ ym $ yn :

ð3:97Þ

and

From Eq. 3.95, it is obvious that both u and v vary linearly along each edge of the element, and they depend only on the displacements of the two nodes on a particular edge. Therefore, the compatibility of displacements on two adjacent elements with common boundary is satisfied.

130

3 Optimal Displacement Method of Structural Analysis

Fig. 3.17 A triangular element

v2

y

u2

2

v3

v1

u3 3

u1

1

x

O

From the theory of elasticity, the nodal displacements utm ¼ {u1,u2, . . .,u6} are related to total strains et ¼ {exx,eyy,ezz} by the following: 2

3

∂u ∂x

7 6 7 6 7 6 exx 6 ∂v 7 7 6 4 5 e ¼ eyy ¼ 6 ∂y 7 7 6 exy 7 6 6 ∂u ∂v 7 5 4 þ ∂y ∂x 2

¼

3

2

y 1 4 kj 0 2A $x

kj

0 $xkj ykj

$yki 0 xki

0 xki $yki

yji 0 $xji

3 u1 36 u2 7 7 0 6 6 u3 7 7 $xji 56 6 u4 7: 7 yji 6 4 u5 5 u6 2

ð3:98Þ

This relationship can be written in matrix notation as, _

e ¼ b u,

ð3:99Þ

where: 2

y 1 4 kj 0 b¼ 2A $x kj

_

0 $xkj ykj

$yki 0 xki

0 xki $yki

yji 0 $xji

3 0 $xji 5: yji

ð3:100Þ

The above equation, indicates that for linearly varying displacement field, the strains are constant, and by Hooke’s law it also leads to constant stresses. Substituting the total strain e in Eq. 3.96 gives the stress-displacement relationship,

3.5 Stiffness Matrix of a Finite Element

2

3 2 ykj σxx E 4 σyy 5 ¼ 4 vykj 2Að1 $ v2 Þ $Ψx σxy kj

131

$vykj $xkj Ψykj

$yki $vyki Ψxki

vxki xki $Ψyki

yji vyji $Ψxji

3 u1 36 u2 7 7 $vyji 6 6 u3 7 7 $xji 56 6 u4 7, 7 Ψyji 6 4 u5 5 u6 2

ð3:101Þ

where v is the Poisson ratio and Ψ¼

1$v : 2

The stiffness matrix is then calculated using Eq. 3.26, and for convenience is presented in two separate parts as: k ¼ kn þ k s ,

ð3:102Þ

where kn represents the stiffness due to normal stresses and ks represents the stiffnesses due to shearing stresses. Thus: 2

y232 6 $vy32 x32 6 6 $y y Et 32 31 6 kn ¼ 4Að1 $ v2 Þ 6 6 vy32 x31 4 y y 32 21 $vy32 x21

x232 vx32 y31 $x32 x31 $vx32 y21 x32 x21

3

sym: y231 $vy31 x31 $y31 y21 vy31 x21

x231 vx31 y21 $x31 x21

y221 $vy21 x21

x221

7 7 7 7, 7 7 5

and 2

x232 6 $x32 y32 6 6 $x32 x31 Et 6 ks ¼ 4Að1 þ vÞ 6 6 x32 y31 4 x32 x21 $x32 y21

y232 y32 x31 $y32 x21 $y32 x21 y32 y21

3

sym: x231 $x31 y31 $x31 x21 x31 y21

y231 y31 x21 $y31 y21

x221 $x21 y21

y221

7 7 7 7: 7 7 5

ð3:103Þ

Using the same method, the stiffness matrices for other elements can be derived. Since there are many excellent books on finite element methods, no further studies are made here, and the interested reader may refer to McGuire and Gallagher [3], Przemieniecki [4], and Zienkiewicz [5], among many others. For the formation of well-conditioned stiffness matrices the reader may refer to Kaveh [6, 7].

132

3.6

3 Optimal Displacement Method of Structural Analysis

Computational Aspects of the Matrix Displacement Method

The main advantage of the displacement method is its simplicity for computer programming. This is due to the existence of a simple kinematical basis formed on a special cutset basis known as cocycle basis of the graph model S of the structure. Such a basis does not correspond to the most sparse stiffness matrix; however, the sparsity is generally so good that there is usually no need to look further. However, if an optimal cutset basis of S is needed, the displacement method encounters all the problems met by the force method, described in Chap. 3. The algorithm for the displacement method is summarized below. Algorithm Step 1: Select a global coordinate system and number the nodes and members of the structure. An appropriate nodal ordering algorithm will be discussed in Chap. 5. Step 2: After initialization of all the vectors and matrices, read or generate the data for the structure and its members. Step 3: For each member of the structure: (a) (b) (c) (d)

Compute L, L*, sinα, sinβ, sinγ, cosα, cosβ, cosγ; Compute the rotation matrix T; Form the member stiffness matrix k in its local coordinate system; Form the member stiffness matrix k in the selected global coordinate system; (e) Plant k in the overall stiffness matrix K of the structure.

Step 4: For each loaded member: (a) Read the fixed end actions; (b) Transform the fixed end actions to the global coordinate system and reverse it to apply at joints; (c) Store these joint loads in the specified overall joint load vector. Step 5: For each loaded joint: (a) Read the joint number and the applied joint loads; (b) Store it in the overall joint load vector. Step 6: Apply boundary conditions to the structural stiffness matrix K, to obtain the reduced stiffness matrix Kff. Repeat the same for the overall joint load vector. Step 7: Solve the corresponding equations to obtain the joint displacements. Step 8: For each member: (a) Extract the member distortions from the joint displacements; (b) Rotate the member distortions to the local coordinate system;

3.6 Computational Aspects of the Matrix Displacement Method

133

(c) Compute the member stiffness matrix; (d) Compute the member forces and fixed end actions. Step 9: Compute the joint displacements and the member forces. The application of the above procedure is now illustrated by a simple example, so that the reader can use it to fully understand the computational steps. Example. Consider a planar truss, as shown in Fig. 3.18. Member 1 has a uniform load of intensity 0.6 kN/m and at joint 2 a concentrated load of magnitude 1.05 kN is applied. The cross-section areas for members are 2A and 1.8A, respectively. The selected global coordinate system and the equivalent nodal forces are illustrated in Fig. 3.19. The stiffness matrices are formed as: for member 1: 2

0:64 0:48 6 0:48 2 0:36 k1 ¼ EA6 4 $0:64 $0:48 5 $0:48 $0:36

$0:64 $0:48 0:64 0:48

3 $0:48 $0:36 7 7: 0:48 5 0:36

and for member 2: 2

0 60 1:8 6 k2 ¼ EA4 0 3 0

0 þ1 0 $1

0 0 0 0

3 0 $1 7 7: 0 5 þ1

The overall stiffness matrix is then obtained as: 2

0:256 6 0:192 6 6 $0:256 K ¼ EA6 6 $0:192 6 4 0 0

0:192 $0:256 0:144 $0:192 $0:192 0:256 $0:144 0:192 0 0 0 0

$0:192 $0:144 0:192 0:744 0 $0:6

0 0 0 0 0 0

3 0 0 7 7 0 7 7: $0:6 7 7 0 5 0:6

The fixed end actions are shown in Fig. 3.19b, and calculated for member 1 as: 2

3 0 6 1:5 7 7 FEA1 ¼ 6 4 0 5: 1:5 These forces are reversed and transformed into the global coordinate system as:

134

3 Optimal Displacement Method of Structural Analysis 2

Fig. 3.18 A planar truss with general loading

1.05kN

m N/

k

0.6

1

3m

2 3

1 4m

Fig. 3.19 The selected coordinate system and the equivalent nodal loads

y

x

2

0:8 6 0:6 T1t ð$FEA1 Þ ¼ 6 4 0 0

$0:6 0:8 0 0

0 0 0:8 0:6

32 3 2 3 0 0 0:9 6 7 6 7 0 7 76 $1:5 7 ¼ 6 $1:2 7: $0:6 54 0 5 4 0:9 5 0:8 $1:5 $1:2

Superimposing the concentrated force at node 2 yields the final vector of external forces as: p ¼ f 0:9

$1:2

$0:15

$1:2 0

0 gt :

Substituting a large number such as 1.E + 30 for the diagonal entries corresponding to the zero displacement boundary conditions, 3 3 2 1:E þ 30 0:192 $0:256 $0:192 0 0 0 6 0 7 7 6 0:192 1:E þ 30 $0:192 $0:256 0 0 7 6 7 6 6 $0:15 7 7 6 $0:256 $0:192 0:256 0:192 0 0 7 ¼ EA6 6 7½v(: 6 $1:2 7 6 $0:192 $0:256 0:192 0:714 0 $0:6 7 7 6 7 6 4 0 5 5 4 0 0 0 0 1:E þ 30 0 0 0 0 $0:6 0 1:E þ 30 0 2

References

135

Solving these equations results in: v¼

1 f0 EA

0

0:845

$1:907 0

0 gt :

The member forces are now computed as: 2

1 0 26 0 0 6 r1 ¼ 4 5 $1 0 0 0 2 3 0:179 6 1:5 7 7 ¼6 4 $0:179 5, 1:5

$1 0 1 0

32 0 0:8 6 07 76 $0:6 0 54 0 0 0

0:6 0:8 0 0

32 3 2 3 0 0 0 0 6 7 6 7 0 0 7 76 0 7 þ 6 1:5 7 5 4 5 4 0:8 0:6 0:845 0 5 $0:6 0:8 $1:907 1:5

and 2

1 0 $1 36 0 0 0 6 r2 ¼ 4 5 $1 0 1 0 0 0 2 3 1:091 6 0 7 7 ¼6 4 $1:091 5: 0

32 0 0 6 07 76 1 0 54 0 0 0

$1 0 0 0

32 3 2 3 0 0 0:845 0 6 7 6 7 0 0 7 76 $1:907 7 þ 6 0 7 0 $1 54 0 5 4 0 5 1 0 0 0

References 1. Argyris et al (1979) Finite element method: the natural approach. Comput Methods Appl Mech Eng 7:1–106 2. Livesley RK (1975) Matrix methods of structural analysis, 2nd edn. Pergamon Press, New York 3. McGuire W, Gallagher RH (1979) Matrix structural analysis. Wiley, New York 4. Przemieniecki JS (1963) Matrix structural analysis of substructures. AIAA J 1:138–147 5. Zienkiewicz OC (1977) The finite element method in engineering, 3rd edn. McGraw-Hill, Maidenhead 6. Kaveh A (2004) Structural mechanics: graph and matrix methods, 3rd edn. Research Studies Press, Somerset 7. Kaveh A (1977) Optimal structural analysis, 1st edn. Research Studies Press, Chichester

Chapter 4

Ordering for Optimal Patterns of Structural Matrices: Graph Theory Methods

4.1

Introduction

In this chapter, methods are presented for ordering to form special patterns for sparse structural matrices. Such transformation reduces the storage and the number of operations required for the solution, and leads to more accurate results. Graph theory methods are presented for different approaches to reordering equations to preserve their sparsity, leading to predefined patterns. Alternative, objective functions are considered and heuristic algorithms are presented to achieve these objectives. Three main methods for the solution of structural equations require the optimisation of bandwidth, profile and frontwidth, especially for those encountered in finite element analysis. Methods are presented for reducing the bandwidth of the flexibility matrices. Bandwidth optimisation of rectangular matrices is presented for its use in the formation of sparse flexibility matrices. In this chapter entries of the stiffness and flexibility matrices are provided with the most appropriate specified patterns for solution of the corresponding equations. Realization of these patterns (or not) affects the formulation of the mathematical models and efficiency of solution. Many patterns can be designed depending on the solution scheme being used. Figure 4.1 shows some of the popular ones encountered in practice. Pattern equivalence of the stiffness matrix of a structure and cutset basis adjacency matrix C*C*t of its graph model, and pattern equivalence of the flexibility matrix of a structure with that of a generalized cycle basis adjacency matrix CCt of its graph model, reduce the size of the problem β-fold, β being the degrees of freedom of the nodes of the model for the displacement method, and β ¼ 1 to 6 depending on the type of structure being studied by the force method.

A. Kaveh, Computational Structural Analysis and Finite Element Methods, 137 DOI 10.1007/978-3-319-02964-1_4, © Springer International Publishing Switzerland 2014

138

4 Ordering for Optimal Patterns of Structural Matrices: Graph Theory Methods

Banded form

Profile form

Partitioned form

Nested partitioned Block matrix form form

Fig. 4.1 Different matrix forms

4.2

Bandwidth Optimisation

The analysis of many problems in structural engineering involves the solution of a set of linear equations of the form, Ax ¼ b,

ð4:1Þ

where A is a symmetric, positive definite and usually very sparse matrix. For large structures encountered in practice, 30–50 % of the computer execution time may be devoted to solving these equations. This figure may rise to about 80 % in non-linear, dynamic or structural optimisation problems. Different methods can be used for the solution of the system of equations, of which the Gaussian elimination is the most popular among structural analysts, since it is simple, accurate and practical, producing some very satisfactory error bounds. In the forward course of elimination, new non-zero entries may be created, but the back substitution does not lead to any new non-zero elements. It is beneficial to minimize the total number of such non-zero elements created during the forward course of the Gaussian elimination in order to reduce the round off errors and the computer storage. Matrix A of Eq. 4.1 can be transformed by means of row and column operations to a form which leads to the creation of a minimum number of non-zero entries during the forward course of the elimination. This is equivalent to the “a priori” determination of permutation matrices P and Q, such that: PAQ ¼ G:

ð4:2Þ

When A is symmetric and positive definite, it is advantageous to have G also symmetric so that only the non-zero elements on and above the diagonal of G need to be stored, and only about half as many arithmetic operations are needed in the elimination. The diagonal elements of A and G are the same, only in different positions. In order to preserve symmetry, P is taken as Qt so that Eq. 4.2 becomes: Qt AQ ¼ G:

ð4:3Þ

For transforming a symmetric matrix A into the forms depicted in Fig. 4.1, various methods are available, some of which will be described in this chapter. However, due to the simplicity of the banded form, most of the material presented

4.2 Bandwidth Optimisation

139

will be confined to optimising the bandwidth of the structural matrices, and other forms will only be introduced briefly. In the Gaussian elimination method, the time required to solve the resulting equations by the banded matrix technique is directly proportional to the square of the bandwidth of A. As mentioned before, the solution of these equations forms a large percentage of the total computational effort needed for the structural analysis. Therefore it is not surprising that a lot of attention is being paid to the optimisation of the bandwidth of these sparse matrices. A suitable ordering of the elements of a kinematical basis for a structure reduces the bandwidth of A, hence decreasing the solution time, storage and round-off errors. Similarly, ordering the elements of a statical basis results in the reduction of the bandwidth of the corresponding flexibility matrix of the structure. Iterative methods using different criteria for the control of the process of interchanging rows and columns of A are described by many authors, e.g. see Rosen [1] and Grooms [2]. For these methods, in general, the required storage and CPU time can be high, making them uneconomical. The first direct method for bandwidth reduction was recognized by Harary [3] in 1967, who posed the following question: For a graph S with N(S) nodes, how can labels 1, 2, . . ., N(S) be assigned to nodes in order to minimize the maximum absolute value of the difference between the labels of all pairs of adjacent nodes?

For a graph labelled in such an optimum manner, the corresponding adjacency matrix will have unit entries concentrated as closely as possible to its main diagonal. In structural engineering, Cuthill and McKee [4] developed the first graphtheoretical approach for reducing the bandwidth of stiffness matrices. In their work, a level structure was used which was called a “spanning tree” of a structure. The author’s interest in bandwidth reduction was initially motivated by an interest in generating and ordering the elements of cycle bases and generalized cycle bases of a graph, as defined in Chap. 2, in order to reduce the bandwidth of the flexibility matrices, Refs. [5, 6]. For this purpose a shortest route tree (SRT) has been used. The application of this approach has been extended to the elements of a kinematical basis (cutset basis) in order to reduce the bandwidth of stiffness matrices. Subsequently it has been noticed that there is a close relation between Cuthill-McKee’s level structure and the author’s SRT. However, there is a difference between these trees in that an SRT contains additional information about the connectivity properties of the corresponding structure. Further improvements have been achieved by employing special types of SRTs such as the longest and narrowest ones, Ref. [7]. Generation of a suitable SRT depends on an appropriate choice of starting node. Kaveh [5] used an end node of an arbitrary SRT, which was chosen from its last contour (level) having the least valency. Gibbs et al. [8] employed a similar node and called it a pseudo-peripheral node. Cheng [9] used an algebraic approach to select a single node or a set of nodes as the root of an SRT. Kaveh employed two simultaneous SRTs for selecting a

140

4 Ordering for Optimal Patterns of Structural Matrices: Graph Theory Methods

pseudo-peripheral node. A comparison of six different algorithms was made in Ref. [10]. Algebraic graph theory has also been used for finding a starting node, Kaveh [11] and Grimes et al. [12]. Paulino et al. [13] used another type of algebraic graphtheoretical approach employing the Laplacian matrix of a graph for nodal ordering.

4.3

Preliminaries

A matrix A is called banded, when all its non-zero entries are confined within a band, formed by diagonals parallel to the main diagonal. Therefore, Aij ¼ 0 when |i$j| > b, and Ak, k$b 6¼ 0 or Ak, k+b 6¼ 0 for at least one value of k. b is the halfbandwidth and 2b + 1 is known as the bandwidth of A. As an example, for 2

1 66 6 A¼6 6: 4: :

6 : 2 7 7 3 8 9 : :

: 9 8 4 :

3 : :7 7 :7 7, :5 5

ð4:4Þ

the bandwidth of A is 2b + 1 ¼ 2%2 + 1 ¼ 4. A banded matrix can be stored in different ways. The diagonal storage of a symmetric banded n%n matrix A is an n%(b + 1) matrix AN. The main diagonals are stored in the last column, and lower co-diagonals are stored down-justified in the remaining columns. As an example, AN for the above matrix is: 2

& 6& 6 AN ¼ 6 60 49 0

& 6 7 8 0

3 1 27 7 37 7: 45 5

ð4:5Þ

When A is a sparse matrix, this storage scheme is very convenient, since it provides direct access, in the sense that there is a simple one-to-one correspondence between the position of an entry in the matrix A(i, j) and its position in AN(i, j $ i + b + 1). Obviously, the bandwidth depends on the order in which the rows and columns of A are arranged. This is why iterative techniques seek a permutation of the rows and a permutation of columns to make the resulting bandwidth small. For symmetric matrices, identical permutations are needed for both the rows and the columns. When a system of linear equations has a banded matrix of coefficients and the system is solved by Gaussian elimination, with pivots being taken from the diagonals, all the operations are confined to the band and no new non-zero entries are generated outside the band. Therefore, the Gaussian elimination can be carried out

4.3 Preliminaries

141

in place, since a memory location is already reserved for any new non-zeros that might be introduced within the band. For each row i of a symmetric matrix A define, bi ¼ i $ jmin ðiÞ,

ð4:6Þ

where jmin(i) is the minimum column index in row i for which Aij 6¼ 0. Therefore, the first non-zero of row i lies bi positions to the left of the diagonal, and b is defined as: b ¼ maxðbi Þ:

ð4:7Þ

In Chap. 4, it has been shown that the stiffness matrix K of a structure is pattern equivalent to the cutset basis adjacency matrix C*C*t, where C* is the cutset basismember incidence matrix of the structural model S. Similarly, the flexibility matrix G is pattern equivalent to the cycle basis adjacency matrix CCt, where C is the cycle basis-member incidence matrix of S. Reducing the bandwidths of C*C*t and CCt directly influences those of K and G, respectively. Notice that the dimensions of C*C*t and CCt, for general space structures, are sixfold smaller than those of K and G, and therefore simpler to optimise. For the displacement method of analysis, there exists a special cutset basis whose elements correspond to stars of its nodes except for the ground node (cocycle basis). The adjacency matrix of such a basis naturally is the same as that of the node adjacency matrix of S, with the row and column corresponding the datum node being omitted. In this chapter, such a special cutset basis will be considered, and the nodes of S will be ordered such that the bandwidth of its node adjacency matrix is reduced to the smallest possible amount. Let A be the adjacency matrix of a graph S. Let i and j be the nodal numbers of member k, and let αk ¼ |i $ j|. Then the bandwidth of A can be defined as: bðAÞ ¼ 2Maxfαk : k ¼ 1, 2, . . . , MðSÞg þ 1,

ð4:8Þ

where M(S) is the number of members of S. In order to minimize the bandwidth of A, the value of b(A) should be minimized. The bandwidth of the stiffness matrix K of a structure is related to that of A by: bðkÞ ¼ βbðAÞ,

ð4:9Þ

where β is the number of degrees of freedom of a typical node of the structure. Papademetrious [14] has shown that the bandwidth minimization problem is an NP-complete problem. Therefore any approach to it is of interest primarily because of its heuristic value.

142

4.4

4 Ordering for Optimal Patterns of Structural Matrices: Graph Theory Methods

A Shortest Route Tree and Its Properties

The main tool for most of the ordering algorithms using graph-theoretical approaches is the shortest route tree of its model or its associate model. A shortest route tree rooted at a node O, called the starting node (root) of the tree, is denoted by SRTO and has the following properties: The path from any node to the root through the tree is a shortest path. An algorithm for generating an SRT is given in Sect. 1.4.7 and therefore only its properties relevant to nodal number are discussed here. An SRT decomposes (partitions) the node set of S into subsets according to their distance from the root. Each subset is called a contour (level) of the SRT, denoted by Ci. The contours of an SRT have the following properties: Adj ðCi Þ ( Ci$1 [ Ciþ1 , Adj ðC1 Þ ( C2 , Adj ðCm Þ ( Cm$1 :

1 d1i

and

#

d1i < d1L

$$

then

d1i ¼ d2i ði ¼ 1, . . . , kÞ

ð6:15Þ

After using the above condition, the desired optimal list {d1i , sj}(i ¼ 1, . . ., k; j ¼ 1, . . ., t) will be obtained. After finding optimal lists corresponding to type II γ-cycles, using relevant equilibrium submatrix, numerical values for each null vector are calculated. The list corresponding to the remaining subgraph will have DSI equal to 3. Three null vectors corresponding to such cycles will be obtained directly from the equilibrium submatrix which leads to suboptimal basis. For type III γ-cycles, finding an optimal list is a time consuming process and considering the fact that the number of cut outs is low in the real structures, the use of this process is not economical for improvement of the final null basis. Thus for each cycle of this type, graph scI is decomposed and all members corresponding to Type I and Type II self stress systems and all the nodes of degree 2 are removed. Algorithm Step 1: Generate the associate graph of finite element model and use an efficient method for its node numbering. It is obvious that a good numbering of this graph corresponds to a good numbering of the elements of a finite element model. This numbering leads to a banded adjacency matrix of the graph and correspondingly to a banded flexibility matrix. Since numbering the members of the interface graphs correspond to the element numbering of the finite elements, therefore such a numbering is the only parameter for controlling the bandwidth of the flexibility matrix. Step 2: Setup the equilibrium matrix of the finite elements model. Step 3: Generate the interface graph and perform its numbering. The numbering of this graph should be performed according to the element numbering of the considered finite elements model. After this numbering, the interface graph can easily be formed and its members can be numbered. Step 4: Find the Type I self stress systems. All multiple members of interface graph are identified and the values "1 and 1 are assigned to appropriate rows (corresponding to the member numbers) and the corresponding null vectors are created. Step 5: Find the Type II self stress systems. Using the Type I and Type II minimal cycles of the associate graph, the subgraphs scI relevant subgraphs are identified and their corresponding optimal lists are found. Step 6: Calculate numerical values of the optimal lists. Using optimal lists selected in Step 5, null vectors corresponding to the Type I and Type II cycles are

6.2 Force Method for Finite Element Models: Rectangular and Triangular Plane. . .

227

calculated from the relevant equilibrium submatrix. For each generator, unit load is applied at a cut in the generator and the internal forces are calculated to form a null vector. Step 7: Order the null vectors. At this step the constructed null vectors should be ordered such that their generators form a list with an ascending order. In the following the efficiency of this algorithm is demonstrated using numerical examples and a comparison is made through the results of the present algorithm and the LU factorization method. The comparisons are confined to those of sparsity, condition number and computational time of the formation of the flexibility matrices. It should be noted that all the algebraic methods use LU decomposition approach for the formation of the null basis or controlling the independence of the columns of the equilibrium matrix.

6.2.6

Numerical Examples

In this section examples with different topological properties are studied. The models are assumed to be supported in a statically determinate fashion. The effect of the presence of additional supports can separately be included for each special case with no difficulty. Example 1. A beam with one opening which is supported in a statically determinate fashion is depicted in Fig. 6.9. This structure is also discretized using rectangular and triangular finite elements. The properties of the model are: Number of rectangular elements ¼ 16, E ¼ 2e + 8 kN/m2, ν ¼ 0.3 Number of Triangular elements ¼ 16, t ¼ 0.02 m, nc ¼ 1 Number of type I self stress systems ¼ 44 (76 %), Number of nodes ¼ 36 Number of type II self stress systems ¼ 12, DSIT ¼ 59 ¼ (44 + 12 + 3) Pattern of the equilibrium matrix, the null basis matrices and the corresponding flexibility matrices for the present algorithm are illustrated in Figs. 6.10, 6.11, and 6.12, respectively. Comparison of the results of the displacement method and the present force method can be found in Ref. [4]. Example 2. A finite element model which is supported in a statically determinate fashion is depicted in Fig. 6.13. This structure is also discretized using quadrilateral and triangular finite elements. The properties of the model are: Number of quadrilateral elements ¼ 66, E ¼ 2e + 8kN/m2, ν ¼ 0.3 Number of Triangular elements ¼ 44, t ¼ 0.1 m Number of type I self stress systems ¼ 172, Number of nodes ¼ 115 Number of type II self stress systems ¼ 63, DSIT ¼ 235 ¼ (172 + 63)

228

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.9 A beam and the discretization of the selected part 0 10 20 30 40 50 60 70

0

20

40

60

80

100

120

nz = 315

Fig. 6.10 Pattern of the equilibrium matrix

Pattern of the null basis matrix with 1,100 entries for the present method is shown in Fig. 6.14. Pattern of the flexibility matrix using the present algorithms are illustrated in Fig. 6.15. Example 3. A circular disk, shown in Fig. 6.16, is analyzed using plane stress triangular elements with the following properties: Diameter ¼ 4.4 m, thickness ¼ 0.05 m, E ¼ 2e + 8 kN/m2, ν ¼ 0.3, Number of triangular elements ¼ 312, Number of nodes ¼ 169, Number of members of the natural associate graph ¼ 456 (type-1 S.E.Ss), First Betti number of the natural associate graph ¼ 145 (type-2 S.E.Ss), DSI ¼ 601 ¼ (456 + 145).

6.2 Force Method for Finite Element Models: Rectangular and Triangular Plane. . .

229

0

Fig. 6.11 Pattern of the null bases matrix

20

40

60

80

100

120 0

Fig. 6.12 Pattern of the flexibility matrix

20 40 nz = 286

60

40

60

0

10

20

30

40

50

60

0

10

20

30 nz = 619

50

230

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.13 A finite element model with qudrilateral and triangular elements

Fig. 6.14 Pattern of the null bases matrix

0 50 100 150 200 250 300 350 400 450 0

50

100

150

200

nz = 1100

Patterns of the null basis matrices are shown in Fig. 6.17, and pattern of the flexibility matrix using the present algorithm is illustrated in Fig. 6.18. For LU factorization the null basis contains 11,014 entries, while the present method leads to only 2,623 entries.

6.3

Finite Element Analysis Force Method: Triangular and Rectangular Plate Bending Elements

In this section, an efficient algorithm is presented for the formation of null bases for finite element models consisting of triangular and rectangular plate bending elements [5]. The null bases obtained by this algorithm are highly sparse and narrowly

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

231

0

50

100

150

200

0

50

100

150

200

nz = 2597

Fig. 6.15 Pattern of the flexibility matrix

a

b

Fig. 6.16 (a) A circular disk with loading (b) its natural associate graph

banded and can be used for optimal finite element analysis by force method. In the present method, using topological transformations the non-zero patterns of null bases are identified and their numerical values are calculated by an algebraic process. For this purpose, associate digraph and interface graph are utilized.

232

6 Optimal Force Method for FEMs: Low Order Elements

a

b

0

100

0

100

200

200

300

300

400

400

500

500

600

600

700

700

800

800

900

900

0

200

400

600

0

200

nz = 2632

400

600

nz = 11014

Fig. 6.17 Patterns of the null basis matrices; (a) The present approach (b) LU factorization approach 0

100

200

300

400

500

600

0

100

200

300

400

500

nz = 7525

Fig. 6.18 Pattern of the flexibility matrix using the presented method

600

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

6.3.1

233

Graphs Associated with Finite Element Models

An Associate Digraph: In this graph one node is associated with each element of the FEM and two nodes are connected with a member if the elements have a common edge. A typical member of the graph is directed from the node with smaller number to the node with higher number. Except the node numbered as 1, all the other nodes have one or two negatively incident (one or two entering members) members defined as the negative incidence number of the node (if the nodes are badly numbered this number can be increased). Owing to the importance of these numbers in recognizing the types of SESs, the negative incidence numbers of the nodes of the graph should carefully be calculated. In Fig. 6.19, a rectangular and a triangular FEM with element numbering and their corresponding associate digraphs and negative incidence number of nodes are shown. An Interface Graph. This graph can easily be constructed for triangular FEM using the following two rules: 1. This graph contains all the nodes of the FEM. 2. With each edge of an element of FEM, two graph members are associated. Therefore, in the interface of two elements, four members are present incident with the two end nodes of the common edge. For rectangular FEM the following additional rule should be used: 3. For each element a diagonal member is added in the interface graph. This member can be added between the first and third nodes of the element. These graphs for a rectangular and triangular FEM are shown in Fig. 6.20. The member numbering of the interface graph should be performed according to the numbering of the FEM, taking into account the primary nodal numbering of considered element in the model. Thus for each triangular element six, and for rectangular element nine members of the interface graph will be numbered sequentially. In Fig. 6.20, such a numbering is shown for a typical element (a).

6.3.2

Subgraphs Corresponding to Self-Equilibrating Systems

6.3.2.1

Definitions of Independent Elements Forces

For the generation of equilibrium matrix A of a FEM, a system of independent force systems should be defined and also their relations with the element nodal forces should be established. The system of independent element forces for a rectangular finite element contains four symmetric moments (F1,F3,F5,F7) and four

234

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.19 A rectangular and a triangular FEM with their associate digraphs and nodes incidence numbers

Fig. 6.20 A rectangular and a triangular FEM with their interface graphs and their numbering for a typical element (a)

anti-symmetric moments (F2,F4,F6,F8) and a set of four forces (F9), which are applied at four corners of the element. These forces are related to the nodal forces (S1 ~ S12) by a 12 ( 9 transformation matrix. A comprehensive study of these forces and their corresponding transformation matrix can be found in [12].

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

235

Fig. 6.21 Nodal forces for a triangular finite element

The system of independent element forces for a triangular finite element can be defined as three symmetric moments (F1,F3,F5) and three anti-symmetric moments (F2,F4,F6). These forces can be related to element nodal forces (S1 ~ S9) using Eq. 6.16. The nodal forces are shown in Fig. 6.21, and the defined element forces for a triangular finite element are illustrated in Fig. 6.22. The interface graph defined in the preceding section is formed based on the way these element forces are considered and members of this graph have one-to-one correspondence with the element forces. 2

6 0 6 S1 6 m12 6 S2 7 6 6 7 6 "l 12 6 S3 7 6 6 7 6 6 S4 7 6 0 6 7 6 6 S5 7 ¼ 6 6 7 6 "m12 6 S6 7 6 6 7 6 l12 6 S7 7 6 6 7 6 4 S8 5 6 6 0 6 S9 6 4 0 2

3

0

S ¼ TF

2 L12 m12 "l12 2 " L12 m12 "l12

m23 "l23

0 0

"m23 l23

0

0

0

0

0 0

0 0 2 L23 m23 "l23 2 " L23 m23 "l23

"m31 l31

0

0

0 0 0 0 m31 "l31

3 2 " L13 7 7 7 m31 7 7 "l31 7 7 0 7 7 7 7 0 7 7 0 7 2 7 7 7 L13 7 7 m31 5

ð6:16Þ

3 F1 6 F2 7 6 7 6 F3 7 6 7 6 F4 7 6 7 4 F5 5 F6 2

"l31

In the transformation matrix T, Lij is the length and lij, mij are direction cosines of the edge ij, which have the following definitions according to nodal coordinates: lij ¼

yj " yi x j " xi mij ¼ Lij Lij

Considering the above definitions, the degree of statical indeterminacy (DSI) for a rectangular and triangular plate bending FEM with determinate support conditions is as follows:

236

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.22 Independent element forces for a triangular finite element

DSI ¼ 9m " 3n þ 3

DSI ¼ 6m " 3n þ 3

ðFor rectangular FEMÞ ðFor triangular FEMÞ

ð6:17Þ

ð6:18Þ

where m is the total number of finite elements and n is the total number of nodes of FEM.

6.3.2.2

Self-Equilibrating Systems of Type I

Every set consisting of four members of interface graph, corresponding to two elements of the FEM with common edges, is called a self-equilibrating system of Type I. The corresponding subgraph contains two SESs. Therefore, the set of four members corresponding to the common edges of the two elements i and j(i < j), has two members mi and ni(m < n), and rj and sj (r < s). The two SESs obtained from this set are (m,r) with ("1, 1) and (n,s) with (1, 1). On the other hand, a null vector with non-zero entries ("1, 1) in rows (m,r) and another null vector with non-zero entries (1, 1) at rows (n,s) are formed. Obviously, the number of such minimal SESs is twice the number of the members of associate digraph, since each member of this graph passes from interface of two elements. Nearly, two-third of null vectors for a rectangular or triangular FEM are of this type, corresponding to high sparsity for the null basis matrix.

6.3.2.3

Self-Equilibrating Systems of Type II

For each two adjacent finite element (two adjacent node in the associate digraph) such as r and j (r < j) in which j have negative incidence 1, another type of SES can be constructed which is called as the self-equilibrating system of Type II. In Fig. 6.23a, two adjacent rectangular and triangular finite element as well as their

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

a

b

237

c

Fig. 6.23 (a) Two adjacent elements, (b) Related interface graphs, (c) Subgraphs of Type II SESs

associate digraphs are shown. Their corresponding interface graphs are also shown in Fig. 6.23b. The DSI of interface graph is 3 and thus correspond to three null vectors. Two null vectors are previously formed using four members in the interface of two elements. Therefore, in order to form the third SES, the generator of two type I SESs, should be removed from interface subgraph. Thus the DSI of remaining subgraph equals one and an independent null vector can simply be extracted. It should be noted that the remaining subgraph corresponding to rectangular elements have still six ineffective members (hidden members in Fig. 6.23c) which can analytically be shown that always lead to zero entries in related null vector. Thus the subgraphs corresponding to Type II SESs, of rectangular and triangular finite elements have always ten members. In Fig. 6.23, the interface graph and the subgraphs corresponding to Type II SESs are shown for two rectangular and triangular adjacent elements. For each node with a negative incidence two, a self equilibrating system of Type II can also be extracted. For each element k with negative incidence two which is adjacent to two elements i and j with k > i, j, pairs (i, k) or (j, k) can be used for the formation of a SES. Though both choices are valid, for maximum reduction in bandwidth of null basis matrix, the pair (max(i,j), k) should be selected. 6.3.2.4

Self-Equilibrating Systems of Type III

There are two elements i and j with k > i, j in the adjacency of an element such as k with negative incidence two. Using these three elements and from their related interface graph, a subgraph corresponding to another minimal SES can be decomposed which is defined as the self equilibrating system of Type III. The interface graph related to these three elements has DSI ¼ 6 and corresponds to six null vectors. Therefore, in order to maintain the independency of null vectors, one

238

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.24 Triangular and rectangular FEM and the corresponding Type III SESs

independent SES should be extracted from this graph. In this subgraph, there exist four and one member corresponding to te generators of Type I and II SESs, respectively, which using them five null vectors were previously formed. Thus the remaining subgraph after removing these members will be one degree statically indeterminate (DSI ¼ 1) and corresponds to an independent null vector. This process can be used without any changes for rectangular and triangular FEM. However, in rectangular finite elements the remaining subgraphs have always some ineffective members. In Fig. 6.24, these subgraphs are shown for triangular and rectangular FEMs.

6.3.2.5

Self-Equilibrating Systems of Type IV

In the previous sections, three types of SESs were defined. These systems are sufficient for formation of null bases of finite element models without openings. However, if a FEM contains one or more openings, then another type of SESs can be identified which is called the self equilibrating system of Type IV, Fig. 6.25. In fact, from each opening in the FEM three independent SESs can be extracted. The subgraphs corresponding to these SESs have usually more members than the previous systems and also their related null vectors have more non-zero entries.

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

239

Fig. 6.25 A FEM with an opening and related cycle from the associate digraph

Every cycle of the associate digraph, if their related elements have no common in one node, corresponds to an opening. Since, every cycle has the same number of members as its nodes; therefore a cycle with m members passes through m finite elements of the FEM. As m triangular and rectangular finite elements surrounding an opening have m and 2m nodes respectively, therefore using Eqs. 6.17 and 6.18 the DSI of their related interface subgraphs will be 3m + 3. However, in these subgraphs 3m self equilibrating systems consisting 2m SESs of Type I and m SESs of Types II and III are previously selected. Then simply by removing the generators of these SESs from interface subgraph corresponding to an opening, a subgraph with DSI ¼ 3 will be remained which corresponds to three null vectors. These three null vectors can simply be calculated using the remaining members of the interface graph as the columns of the related equilibrium submatrix and by utilizing an algebraic procedure. The null vectors related to openings which are calculated by the above process are subminimal. Finally, using the present procedures all minimal and subminimal SESs are simply calculated and the null basis matrix is generated. Due to the nature of present method, the calculated null bases are highly sparse and narrowly banded. However, for further reduction in bandwidth of null basis matrix (without any changes in sparsity) for each SES, an optimal list should be selected. Algorithm This algorithm consists of the following steps: Step 1. In this step the associate digraph of the considered FEM is formed. In order to have a banded null basis, the nodes of this graph should be numbered by any efficient nodal ordering algorithms. Obviously, the effect of final numbering should be considered in FEM and rectangular equilibrium matrix. However, the ordering of the elements of FEM (nodes of associate digraph) is sufficient for formation of a banded null basis and there is no need for nodal ordering of FEM. Step 2. The rectangular equilibrium matrix of the FEM is formed in this step. Step 3. Formation of the interface graph of the FEM and the numbering their members according to nodal and element numbering of FEM is performed n this step. Step 4. In this step the SESs of Type I are formed and the corresponding null vectors are obtained.

240

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.26 A triangular FEM with numbering

Step 5. Formation of the SESs of Type II and calculation of the corresponding null vectors by using an algebraic process (such as LU factorization) on the related submatrices is performed in this step. Step 6. Formation of the SESs of Type III, Numerical values of null vectors are found by a similar process in Step 5. Step 7. The SESs of Type IV are formed and calculation of the numerical values of related null vectors is carried out if the model contains one or more openings, similar to Steps 5 and 6. Step 8. The calculated null vectors are combined and ordered in a matrix in such a way that their generators make an ascending ordered list.

6.3.3

Numerical Examples

In this section three examples from triangular and rectangular FEM are studied. All of the models are assumed to be supported in a statically determinate fashion. The effect of indeterminate support conditions can separately be included with no difficulty [13]. However, the null basis matrices for each model are calculated using the present algorithm and LU factorization methods and the results are compared through computational time, sparsity, pattern of matrices and accuracy. Example 1. In this example (Fig. 6.26), the null basis matrix (B1) for a triangular FEM with statically determinate support conditions is calculated and the sparsity, computational time and two norms of AB1 matrix, namely Frobenious and infinite

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

a0

b

0

100

100

200

200

300

300

400

400

500

500

0

100

200

300

400

241

0

100

200

300

400

nz = 1873

nz = 10187

Fig. 6.27 Pattern of the null basis matrix: (a) LU factorization, (b) present method

Table 6.1 Comparison of the sparsity, computational time and accuracy of the present algorithm versus the LU factorization

LU factorization Present algorithm

Number of non-zero entries (nz)

Time LU Time

kAB1kfro

kAB1k∞

10,187 1,873

1.0000 0.7885

7.39e"13 2.95e"14

5.57e"12 1.86e"14

norms are compared with LU factorization method, Fig. 6.27 (Table 6.1). The FEM properties are as follows: Number of triangular elements ¼ 98, Number of nodes ¼ 64, DSI ¼ 399, Thickness ¼ 0.1 m, E ¼ 2e + 8 kN/m2, ν ¼ 0.3. Example 2. In Fig. 6.28, a rectangular 1.6 m ( 0.8 m plate which is discretized as 120 rectangular finite elements is shown. Patterns of the calculated null basis matrix using two methods are shown in Fig. 6.29. Also the results of the comparison are presented in Table 6.2. The properties of the model are as follows: Number of rectangular finite elements ¼ 128, Number of nodes ¼ 153, DSI ¼ 696, Thickness ¼ 0.05 m, E ¼ 2e + 8 kN/m2, ν ¼ 0.3. Example 3. In this example, a circular plate (with diameter 4) which is clamped at its center (determinate support condition) is studied, Fig. 6.30. Pattern of the null basis matrix for two methods and the comparison of results are shown in Fig. 6.31 and Table 6.3, respectively. The properties of the model are as follows:

242

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.28 The FEM of a rectangular plate and its numbering

a

b

0

200

0

200

400

400

600

600

800

800

1000

1000

0

200

400

600

0

200

nz = 24663

400

600

nz = 1856

Fig. 6.29 Pattern of the null basis matrix: (a) LU factorization, (b) present method

Table 6.2 Comparison of the sparsity, computational time and accuracy of the present algorithm versus the LU factorization

LU factorization Present algorithm

Number of non-zero entries (nz)

Time LU Time

kAB1kfro

kAB1k∞

24,663 1,856

1.0000 0.6539

3.76e"12 1.74e"14

2.62e"11 7.10e"15

6.3 Finite Element Analysis Force Method: Triangular and Rectangular Plate. . .

243

Fig. 6.30 The FEM of a circular plate

a

b

0

0

500

500

1000

1000

1500

1500

2000

2000

2500

2500

3000

3000

3500

3500

4000

4000

4500

4500

0

1000

2000

3000

0

1000

nz = 398389

2000

3000

nz = 17350

Fig. 6.31 Pattern of the null basis matrix: (a) LU method, (b) the present method Table 6.3 Comparison of the sparsity, computational time and accuracy of the present algorithm versus the LU factorization

LU factorization Present algorithm

Number of non-zero entries (nz)

Time LU Time

kAB1kfro

kAB1k∞

398,389 17,350

1.0000 0.1852

4.15e"11 7.51e"14

6.65e"10 1.95e"14

244

6 Optimal Force Method for FEMs: Low Order Elements

Number of triangular finite elements ¼ 800, Number of nodes ¼ 441, DSI ¼ 3,480, Thickness ¼ 0.2 m, E ¼ 2e + 7 kN/m2, ν ¼ 0.2. The results of examples clearly reveal the efficiency of the present method in reduction of non-zero entries and bandwidth. In Example 3, the difference of the computational time for two methods has been considerable which means that the complexity of present method is lower than LU method, and this difference becomes even more when the DSI is increased. The values of norms also indicate the higher accuracy of the present algorithm. Finally, the results show that, the present method can be used as an efficient tool for null basis calculation of plate bending FEM and optimal finite element force method because in all aspects of comparisons, (sparsity, computational time and accuracy) the present algorithm has considerable priority versus the LU method and thus versus other algebraic algorithms which LU factorization is one of the primary steps of those methods.

6.4

Force Method for Three Dimensional Finite Element Analysis

In this section an efficient method is presented for the formation of null bases of finite element models comprised of tetrahedron elements, corresponding to highly sparse and banded flexibility matrices [7]. This is achieved by associating special graphs to the finite element model and selecting appropriate subgraphs and forming the self stress systems on these subgraphs.

6.4.1

Graphs Associated with Finite Element Model

Here, the natural associate graph and the interface graph are utilized as defined in the following: The interface graph SI. This graph can be constructed using the following two rules: a. There is 1–1 correspondence between the nodes of the interface graph and the nodes of the FEM. b. For each edge of the tetrahedron, one independent member is associated. Therefore, if k tetrahedrons have a common edge, then the corresponding member of the interface graph will consists of k members (multiple members). A FEM and the corresponding interface graph are shown in Fig. 6.32a, b, respectively. The members of the interface graph should be numbered according to the FEM. For each tetrahedron element like a, six members of the interface graph should be numbered consequently. The numbering is performed according to the direction of the independent element forces. A typical numbering is shown in Fig. 6.32c.

6.4 Force Method for Three Dimensional Finite Element Analysis

a

b

245

c

Fig. 6.32 (a) A 3D finite element model; (b) The interface graph; (c) Numbering for the skeleton of a typical tetrahedron a

a

b

Fig. 6.33 The nodal and element forces of a tetrahedron element

6.4.2

The Pattern Corresponding to the Self Stress Systems

The nodal forces and independent element forces of a tetrahedron is defined as shown in Fig. 6.33. This is the same convention as that of the Przemieniecki [10]. Considering Fig. 6.32, in order to find the patterns corresponding to the self stress systems, the skeleton of tetrahedra are simulated as a space truss. This is possible since the independent element forces F1 to F6 are applied in the nodes and are along the edges of the tetrahedron, Fig. 6.33. The statical indeterminacy of a space truss with m members and n nodes is given as γ(S) ¼ m " 3n + 6, therefore the Degree of Statical Indeteminacy (DSI) of the entire FEM, supported in a statically determinate fashion, can be calculated with same relationship as: DSIT ¼ 6M " 3N þ 6

ð6:19Þ

where M is the number of tetrahedron elements and N is the total number of nodes of the FEM.

246

6 Optimal Force Method for FEMs: Low Order Elements

With the above simulation, the patterns of the self stress systems can be identified as follows:

6.4.2.1

Type I Self Stress Systems

Each k-multiple member of the interface graph is a subgraph on which k " 1 self stress systems can be generated. In other words, on a k-multiple member numbered as (i1,j2,l3, . . .,mk " 1,nk) with the condition (i < j < l < . . . < m < n), k " 1 self stress systems each formed on two single members can be constructed. Each k(k " 1)/2 combination of double members from the above list is valid for a self stress system but obviously, for maximum reduction in bandwidth of the final null basis, k " 1 pairs of duplicate members should be selected as (i,j), (j,l), . . ., (m,n). Each pair (i,j) with (i < j), corresponds to a null vector with their nonzero entries are located in rows i and j, and their numeric values are "1, 1, respectively. The member with bigger member number (j) is called the generator. Each pairs forms the underlying subgraph of a Type I self stress system. For finite elements models with tetrahedron elements, more than 85 % of total self stress systems are of Type I. Thus a large percent of the minimal null vectors can be formed only by the determination of member numbers of these pairs. It should be noted that in the process of the formation of the interface graph, these pairs and their numbers can simply be identified.

6.4.2.2

Type II Self Stress Systems

There are other types of self stress systems which are topologically identical to the minimal self stress systems of the corresponding space truss. In the other words, if a k-multiple member from the interface graph is substituted by a member, or if the generators of the Type I self stress systems are removed from SI, then the remaining subgraph is a graph, denoted by S. In general the self stress systems built on S are called Type II self stress systems. In general, the self stress systems which can be selected from subgraph S are called Type II systems. In fact these systems are γ-cycles, which correspond to the cycles of minimal lengths of the associate graph of the finite element model. A connected rigid subgraph Ck of S with γ(Ck) ¼ 1, which has no removable subgraph, is termed a γ-cycle of S, Ref. [11]. A removable subgraph Sj of a graph Si, is the elementary subgraph for which γ(Si " Sj) ¼ γ(Si). The associate graph of tetrahedron finite element models, denoted by A(S), is a graph in which to each tetrahedron element one node is associated and two such nodes are connected together by a member if their corresponding elements having a common face (3 common nodes). A finite element model with 24 tetrahedron elements is shown in Fig. 6.34a, its associate graph, which is the 1-skeleton of a polyhedron, is depicted in Fig. 6.34b.

6.4 Force Method for Three Dimensional Finite Element Analysis Fig. 6.34 A finite element model with 24 tetrahedron elements and the corresponding associate graph

247

a

b

Corresponding to the regional cycle basis of a planer graph (the set of cycles which are the boundaries of the internal regions [11]), in general, two types of minimal cycles can be extracted from the associate graph of a finite element model. These cycles are as follows.

6.4.2.3

Type I Minimal Cycles

In these cycles all the corresponding finite elements have two common nodes. Each cycle in this type passes through M finite elements for which its corresponding interface graph has N ¼ M + 2 nodes, and (3M " 1) Type I self stress systems can be extracted. Therefore, by using Eq. 6.8, the degree of statical indeterminacy of equivalent γ-cycle is 1. Thus each Type I cycle corresponds to one null vector.

6.4.2.4

Type II Minimal Cycles

A minimal cycle which surrounds an opening (in the form of a hole through a structure), is called Type II minimal cycle. Such a cycle passes through M finite elements and its corresponding interface graph has N ¼ M nodes, and 3M Type I self stress systems can be extracted. Again by using Eq. 6.8, the DSI of equivalent γ-cycle is 6. Thus each Type II cycle corresponds to six null vectors.

248

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.35 A typical subgraph S0

6.4.3

Relationship Between γ(S) and b1(A(S))

The goal of this section, which has theoretical importance, is to derive a relationship between the degree of statical indeterminacy of the 1-skeleton of a FE model S and the first Betti number of the associate graph of the model without openings (analogous to 2-dimensional fully triangulated trusses [11]). For this purpose an expansion is employed. 6.4.3.1

Number of Nodes of the Associate Graph

Consider a tetrahedron element. The 1-skeleton S corresponding to this element has six members and its associate graph is only a single node. Second tetrahedron element results in the addition of a typical subgraph S0 as shown in Fig. 6.35. Each time by adding this subgraph to the previous graph leads to addition of three members to the main graph S and one node to its associate graph. The associate graph which is formed using this process is a tree and therefore its number of nodes can simply be calculated as: 0

N ¼

MðSÞ " 3 3

ð6:20Þ

where M(S) and N0 are the number of members and nodes of the 1-skeleton of the model S and its associate graph A(S), respectively. Obviously N0 is also equal to the number of tetrahedron elements. Addition of one node and three members in each stage of expansion is the basic condition for validity of Eq. 6.20. Obviously, in this case the 1-skeleton can be viewed as a space truss having DSI equal 0. Some stages of the expansion process is shown in Figs. 6.36a–c. If the subgraph S0 is joined to the previous 1-skeleton in a manner that only one new member is added to S without addition of a new node (equivalent to the addition of a new tetrahedron element, one new node to A(S) and formation of a cycle in A(S)), then the DSI of corresponding space truss will be increased by unity, Fig. 6.36d. In such a case, Eq. 6.9 is not valid and must be modified using a new parameter. Clearly, the DSI of the space truss, γ(S), should be considered as this new parameter. Considering the above mentioned point, Eq. 6.20 is modified as

6.4 Force Method for Three Dimensional Finite Element Analysis

a

b

S M (S) = 6 N' = 1, M ' = 0 DSI = 0

c

249

d

S1 = S U S0

S2 = S1 U S0

S3 = S2 U S0

M (S) = 9

M(S) = 12

M (S) = 13

N' = 2, M ' = 1 DSI = 0

N' = 3, M' = 2 DSI = 0

N' = 4, M ' = 4 DSI = 1

Fig. 6.36 The process of expansion for the formation of a γ-cycle with DSI equal to unity

0

N ¼

MðSÞ þ 2γðSÞ " 3 3

ð6:21Þ

Substitution γ(S) ¼ M(S) " 3N(S) + 6 in Eq. 6.21 leads to 0

N ¼ MðSÞ " 2NðSÞ þ 3

ð6:22Þ

In fact, the number of nodes of the associate graph is equal to the DSI of the 1-skeleton S, when S is viewed as a two dimensional truss!

6.4.3.2

The Number of Members of the Associate Graph

Similar to the previous section, the number of members of the associate graph can also be determined. However, if the expansion process is in a manner that leads to an associate graph which is a tree, its number of members can be simply calculated using the property of a tree, i.e. 0

0

M ¼N "1

ð6:23Þ

or 0

M ¼

MðSÞ " 6 3

ð6:24Þ

Here again, if joining a subgraph S0 leads to the addition of only one member to the graph S (this case corresponds to the addition of two members and one node to A(S), and one unit increase in the DSI of space truss), then Eq. 6.24 can simply be modified using γ(S) as

250

6 Optimal Force Method for FEMs: Low Order Elements

0

M ¼

MðSÞ þ 5γðSÞ " 6 3

ð6:25Þ

or 0

M ¼ 2MðSÞ " 5NðSÞ þ 8

ð6:26Þ

Now it is possible to relate the number of independent cycles of the associate graph to the number of nodes and members of the 1-skeleton S. The dimension of the cycle space, or the first Betti number of the associate graph, can be calculated using 0

0

b1 ðAðSÞÞ ¼ M " N þ 1

ð6:27Þ

By substitution of M0 and N0 from Eqs. 6.11 and 6.26, the relation for b1(A(S)) is obtained in terms of M(S) and N(S) as b1 ðAðSÞÞ ¼ MðSÞ " 3NðSÞ þ 6

ð6:28Þ

and this is the relationship for the DSI of a three dimensional truss. The right hand of the above formula is identical to the DSI of a space truss. Examining further models with tetrahedron finite elements and their corresponding associate graphs, it becomes obvious that the relationship presented in earlier sections are valid for all the cases where A(S) is not the 1-skeleton of a polyhedron. If A(S) is the 1-skeleton of a polyhedron (Fig. 6.34b), then internal nodes will be created in the finite element model or graph S (a node is called internal if it is not positioned on the surface of the FEM). This case corresponds to situations where in the process of expansion, adding one tetrahedron element leads to the addition of three members and one node for the graph A(S). For such cases, the present relationship must be modified considering the contribution of the number of internal nodes as: 0

N ¼ MðSÞ " 2NðSÞ þ 3 þ Ni ðSÞ

ð6:29Þ

M ¼ 2MðSÞ " 5NðSÞ þ 8 þ 3Ni ðSÞ

ð6:30Þ

b1 ðAðSÞÞ ¼ γðSÞ þ 2Ni ðSÞ

ð6:32Þ

0

b1 ðAðSÞÞ ¼ MðSÞ " 3NðSÞ þ 6 þ 2Ni ðSÞ

ð6:31Þ

or

In which Ni(S) is the total number of internal nodes in the 1-skeleton S of the finite element model. The above equations are general relationships for finding the number of nodes, members and the dimension of the cycle space of an associate graph. Equation 6.32

6.4 Force Method for Three Dimensional Finite Element Analysis

a

b

251

c

Fig. 6.37 (a) Finite element model with its associate graph; (b) corresponding scI ; (c) corresponding γ-cycle

shows that, if no internal node is created, then the dimension of the cycle space of the associate graph is equal to the DSI of the corresponding space truss, and therefore for each cycle, an independent null vector can be formed. If there is one or more internal nodes in the model, then the dimension of the cycle space of A (S) is greater than of the DSI of S and thus all the null vectors corresponding to cycles cannot be used in formation of the final null basis. In this case, 2Ni(S) of vectors must be selected and ignored. Some of the null vectors will have the same generators. In Sect. 6.4.2 a method is presented for the selection of these vectors. It should be noted that, in 2-dimensional trusses and plane stress and strain finite elements, the dimension of the cycle space of A(S) is always equal to the DSI of the 1-skeleton S.

6.4.4

Selection of Optimal γ-Cycles Corresponding to Type II Self Stress Systems

Thus far, it is found out that each γ-cycle corresponds to a minimal cycle of the associate graph. Also each minimal cycle with n nodes from A(S) such as c passes through n tetrahedron elements. The subgraph scI (scI ' SI) which is relevant to these n elements and cycle c, is a base for the selection of an optimal γ-cycle. Such a subgraph may contain simple and multiple members, where each multiple member with k members corresponds to the overlap of k tetrahedron elements, and each simple member corresponds to the edge of a element in the boundary of the model. By applying a special condition to such subgraphs, lists corresponding to optimal γ-cycles can be obtained. A finite element model with four tetrahedron elements and its associate graph are shown in Fig. 6.37a. The corresponding scI which contains multiple and simple members is illustrated in Fig. 6.37b, and the corresponding γ-cycle is depicted in Fig. 6.37c. In general, from each scI many γ-cycles (self stress systems) can be extracted, since each simple member of scI is included in a γ-cycle, and all the members of a multiple member can be used in the formation of final self stress system. Thus for obtaining an optimal self stress system, on each scI , two basic selections should be performed:

252

6 Optimal Force Method for FEMs: Low Order Elements

Table 6.4 Lists corresponding to graph scI

d11

d12

d21

d22 d32

... ...

s1

s2

... ...

d1i

sj

d2i

... ...

d1k d2k

dni

1. Selection of the generator or the last member of a self stress system, which is required for the null vectors to be independent. 2. Selection of a list of members from graph scI with maximum possible number for the first member. Such a selection can reduce the bandwidth of the null basis matrix considerably. The mathematical representation of this selection can be written as Minimize ðj " iÞ

ð6:33Þ

where j is the generator’s member number and i is the least member number of the current γ-cycle. In the following a simple and fast method is presented for these selections.

6.4.5

Selection of Optimal Lists

In Table 6.4, members of a graph scI which are relevant to a Type I minimal cycle of A(S) are shown. In this table di, (i ¼ 1, . . ., k) are the member numbers of multiple members, where d1i < d2i < . . . < dni , and sj; (j ¼ 1, . . ., t) are the member numbers of simple members. All dm i with (m 6¼ 1) are already used as the generators of Type I self stress systems. Therefore, it is obvious that the max {d1i ,sj} ¼ d1L , (i ¼ 1, . . ., k, j ¼ 1, . . ., t) must be selected as the generator of the current γ-cycle. For maximizing the difference between the first member number and the generator of the current γ-cycle, the following condition can be used: ( ( ))* ( ) * find max dij *dij < d1L then d 1i ¼ max d ij ði ¼ 1, . . . , k, j ¼ 2, . . . , nÞ ð6:34Þ

Equation 6.34 means that, in each multiple member the largest dji , (j ¼ 2, . . ., n) which is also less than the generator’s member number (d1L ), should be substituted with d1i for all the indices of i. After using the above process, the remained list, {d1i ,sj}, (i ¼ 1, . . ., k, j ¼ 1, . . ., t) is the desired optimal list which corresponds to subgraph of the current optimal self stress system. After finding optimal lists corresponding to Type I minimal cycles, using relevant equilibrium submatrix, numerical values for each null vector are calculated. For Type II minimal cycles, finding an optimal list is a time consuming process and considering the fact that the number of openings is low in the real

6.4 Force Method for Three Dimensional Finite Element Analysis Table 6.5 Schematic view of three null vectors with identical generator

v1 0 v1

v2 0 0

0 v3 v3

0 v4 0

253 0 0 v5

v6 0 0

v7 v7 0

v8 v8 v8

structures, the use of this process is not economical for improvement of the final null basis. Thus for each cycle of this type, graph scI is decomposed and all members corresponding to Type I self stress systems are removed. The remaining subgraph has the DSI equal to 6, and then 6 null vectors can be calculated from relevant equilibrium submatrix. Obviously, these null vectors will be suboptimal. At this stage, considering the number of Type I self stress systems previously selected, decision for performing the rest of the process should be taken. Suppose t1 Type I self stress systems are identified previously. Then t2 ¼ DSIT " t1 Type II self stress systems should be selected. Therefore, if we have t2 ¼ b1(A(S)) + 5nc (with nc being the total number of openings) meaning that there is no internal node in the model, then all γ " cycles corresponding to cycles of A(S) should be involved in the formation of final null basis. Otherwise, t2 < b1(A(S)) + 5nc means that there are one or more internal nodes. This case corresponds to the generation of null vectors with identical generator numbers. These vectors can usually be grouped in triplex sets and some of them should be deleted. It should be noted that all vectors which have unique generators are valid and independent. In the following, an algebraic procedure is presented for the formation of a desired list of vectors. In Table 6.5, a schematic view for the patterns of three null vectors with identical generator is illustrated. These vectors correspond to three minimal cycles of A (S) which according to the process presented in Sect. 4.1, their corresponding optimal lists have identical generator as v8. According to Eq. 6.22, it is obvious that if v8 6¼ 0, then for the generator v8, the second vector will be the desired vector from these three sets (rows). In such a case, after the selection of one optimal vector, one cannot simply delete the remaining vectors. For this purpose the following two controls should be performed. a. Numerical cancellation control Each vector for which the numerical value of its generator is equal to zero has in fact another generator (closer nonzero entry to the generator). If there is no such a vector with identical generator among all the previously selected vectors, then this vector should be selected as a new and independent null vector. b. New generator control All combinations of m vectors (m is usually equal 3) for possibility of the formation of vectors with new generator should be calculated. These combinations should be found in a manner that the common generator member of vectors is removed. Here again, from newly created vectors, those with new generators should be selected as valid and independent null vectors. As an example, in Table 6.5, the combination of the first and third rows leads to a new vector with new generator v7. If there is no such a vector with this new generator among all

254

6 Optimal Force Method for FEMs: Low Order Elements

the previously selected null vectors, then this new vector should be selected as a valid and independent null vector with the generator as v7. Finally, using the above mentioned two conditions, t2 valid and independent vectors are identified and totally DSIT null vectors will be left. Since this process is performed on all vectors with identical generators, therefore all the desired vectors are obtained automatically and there is no need to additional information about the number of internal nodes. In the following an efficient algorithm is presented for finding the null basis of tetrahedron finite element models. Algorithm Step 1: Generate the associate graph of finite element model and use an efficient method for its node numbering, Kaveh [11]. It is obvious that a suitable numbering of this graph corresponds to good numbering of elements of finite element model. This numbering leads to a banded adjacency matrix of the graph and correspondingly to a banded flexibility matrix. Step 2: Setup the equilibrium matrix of finite elements model. Step.3 Generate the interface graph and perform its numbering. The numbering of this graph should be performed according to the element numbering of the considered finite elements model. Step 4: Find the Type I self stress systems. All multiple members of interface graph are identified and the values "1 and 1 are assigned to appropriate rows (corresponding to the member numbers). At the end of this step t1 minimal null vectors are created. Step 5: Find the Type II self stress systems. Using the Type I and Type II minimal cycles of the associate graph, relevant subgraphs are identified and their corresponding optimal lists are constructed. Step 6: Calculate numerical values of the optimal lists. Using optimal lists selected in Step 5, null vectors corresponding to the Type I and Type II minimal cycles are calculated from the relevant equilibrium submatrix. Step 7: Order the null vectors. At this step the constructed null vectors should be ordered such that their generators form a list with an ascending order. In the following the efficiency of this algorithm is demonstrated using two numerical examples and a comparison is made through the results of the present algorithm and the LU factorization method. The comparisons are confined to those of sparsity, condition number and computational time of the formation of the flexibility matrices.

6.4.6

Numerical Examples

In this section two examples with different topological properties are studied. The models are assumed to be supported in a statically determinate fashion. The effect of the presence of additional supports can separately be included for each special case with no difficulty, Kaveh and Fazli [13]. The patterns of the null basis matrix

6.4 Force Method for Three Dimensional Finite Element Analysis

255

Fig. 6.38 A thick beam-type structure and the associate graph of the selected part

B1 and the flexibility matrix G are formed for two examples, and the number nonzero entries of these matrices are denoted by nz. Example 1. A thick beam-type structure supported in a statically determinate fashion is depicted in Fig. 6.38. This structure is discretized using tetrahedron finite elements. The properties of the model are as follows: Number of tetrahedron elements ¼ 480, Number of nodes ¼ 205 Elastic modulus E ¼ 2e + 7 kN/m2, Poisson’s ratio ν ¼ 0.2 Number of Type I self stress systems ¼ 2,032 (89.5 %) First Betti number of the associate graph ¼ 317 (independent cycles) Number of Type II self stress systems ¼ 239 Number of internal nodes (Ni) ¼ 39, DSIT ¼ 2,271 ¼ (2,032 + 239). The sparsity of the final null basis obtained by the present algorithm is approximately 12 % of LU method, as shown in Fig. 6.39. The conditioning numbers, the ∞ norms and the Frobenius norms of AB1 are given for the present method and LU factorization approach, where A is the equilibrium matrix. The computational time is lower than 50 % for the present algorithm. The flexibility matrix shown in Fig. 6.40 is quite banded (Table 6.6). In the above table, λmax/λmin is the condition number, and kk∞, kkfro are the ∞ norm and Frobenius norm of AB1, respectively. Example 2. A thick flat plate with 3D tetrahedra in a single layer is considered which is supported in a statically determinate fashion as depicted in Fig. 6.41. The 1-skeleton

256

6 Optimal Force Method for FEMs: Low Order Elements

a

b

0

0

500

500

1000

1000

1500

1500

2000

2000

2500

2500

0

500

1000

1500

2000

0

500

nz = 6576

1000

1500

2000

nz = 54822

Fig. 6.39 Patterns of B1(2,880 ( 2,271) and the number of nonzero entries, nz, of null basis; (a) Present algorithm; (b) LU factorization 0

500

1000

1500

2000

0

500

1000

1500

2000

nz = 50759

Fig. 6.40 Pattern of the flexibility matrix G(2,271 ( 2,271) and the number of its nonzero entries obtained by the present algorithm Table 6.6 Comparison of the condition number of G, the norms and the computational time LU Present Algorithm

Time/LU time 1.00 0.48

λmax/λmin 1.67e+5 3.74e+5

kAB1k∞ 5.73e"12 0.00

kAB1kfro 1.29e"12 0.00

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

257

Fig. 6.41 A double layer grid, and the associate graph of the selected part of the grid

of this model is similar to a double layer space structure. The associate graph of this model is also shown in Fig. 6.41. The properties of the model are as follows: Number of tetrahedron elements ¼ 904, Number of nodes ¼ 402 Elastic modulus E ¼ 2e + 7 kN/m2, Poisson’s ratio ν ¼ 0.2 Number of Type I self stress systems ¼ 3,719 (88.0 %) First Betti number of the associate graph ¼ 505 (independent cycles) Number of Type II self stress systems ¼ 505 Number of internal nodes (Ni) ¼ 0, DSIT ¼ 4,224 ¼ (3,719 + 505) Here again, the sparsity of final null basis obtained by the present algorithm is approximately 10.5 % of LU method, as depicted in Fig. 6.42, while its computational time is nearly 11 % and also the condition number of G is improved, Table 6.7. The flexibility matrix G is also well structured as shown in Fig. 6.43. In this chapter low order elements were presented. Higher order element will be discussed in subsequent chapter.

6.5

Efficient Finite Element Analysis Using Graph-Theoretical Force Method: Brick Element

In this section, an efficient graph theoretical method is presented for FEA of models composed of 3D brick elements. For this purpose first independent force systems and flexibility matrix of the element are presented, followed by the formation of the

258

6 Optimal Force Method for FEMs: Low Order Elements

a

b

0

0

500

500

1000

1000

1500

1500

2000

2000

2500

2500

3000

3000

3500

3500

4000

4000

4500

4500

5000

5000 0

1000

2000

3000

0

4000

1000

2000

3000

4000

nz = 129976

nz = 13746

Fig. 6.42 Patterns of B1(5,424 ( 4,224) and the number of nonzero entries, nz, of null basis; (a) present algorithm; (b) LU factorization Table 6.7 Comparison of the condition number of G, the norms and the computational time LU Present Algorithm

Time/LU time 1.00 0.11

λmax/λmin 1.34e+6 1.68e+5

kAB1k∞ 1.64e"10 1.94e"15

kAB1kfro 1.83e"11 1.18e"14

minimal subgraphs of the graph models of the considered FEMs. Then the selfequilibrating systems are constructed on these subgraphs forming a statical basis of the FEM corresponding to highly sparse and banded flexibility matrix.

6.5.1

Definition of the Independent Element Forces

In displacement method we use three forces at each node of the element, while in the force method, as shown in Fig. 6.44, it is preferable to select twelve edge force systems plus six diagonal force systems on six faces of the brick element between the second and third nodes of the current face. These element forces can be related to nodal forces using Eq. 6.35 as S ¼ TF

ð6:35Þ

where lij is the length and mij, nij, pij are the direction cosines of the line between nodes i and j.

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . . Fig. 6.43 Pattern of the flexibility matrix G (4,224 ( 4,224) and the number of its nonzero entries obtained by the present algorithm

259

0 500 1000 1500 2000 2500 3000 3500 4000 0

1000

2000

3000

4000

nz = 98238

Fig. 6.44 Nodal and element force systems of a brick element

6.5.2

Flexibility Matrix of an Element

Formulation of a discrete element equivalent to the actual continuous structure is the first step in matrix structural analysis. For a linear system it can be assumed that the stresses σ.are related to the forces F by linear equation as σ ¼ cF

ð6:36Þ

The matrix c represents statically equivalent stresses system due to the unit force F. The flexibility matrix of an element can be written as

3 2 m13 S1 6 S2 7 6 n13 7 6 6 6 S3 7 6 p13 7 6 6 6 S4 7 6 0 7 6 6 6 S5 7 6 0 7 6 6 6 S6 7 6 0 7 6 6 6 S7 7 6 m31 7 6 6 6 S8 7 6 n31 7 6 6 6 S9 7 6 p31 7 6 6 6 S10 7 6 0 7 6 6 6 S11 7 6 0 7 6 6 6 S12 7 6 0 7 6 6 6 S13 7 ¼ 6 0 7 6 6 6 S14 7 6 0 7 6 6 6 S15 7 6 0 7 6 6 6 S16 7 6 0 7 6 6 6 S17 7 6 0 7 6 6 6 S18 7 6 0 7 6 6 6 S19 7 6 0 7 6 6 6 S20 7 6 0 7 6 6 6 S21 7 6 0 7 6 6 6S 7 6 0 6 22 7 6 4S 5 4 0 23 0 S24

2

0 0 0 0 0 0 m34 n34 p34 m43 n43 p43 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 m24 n24 p24 0 0 0 m42 n42 p42 0 0 0 0 0 0 0 0 0 0 0 0

m12 n12 p12 m21 n21 p21 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m78 n78 p78 m87 n87 p87

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m68 n68 p68 0 0 0 m86 n86 p86

0 0 0 0 0 0 0 0 0 0 0 0 m56 n56 p56 m65 n65 p65 0 0 0 0 0 0

m15 n15 p15 0 0 0 0 0 0 0 0 0 m51 n51 p51 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m37 n37 p37 0 0 0 0 0 0 0 0 0 m73 n73 p73 0 0 0

0 0 0 0 0 0 0 0 0 m48 n48 p48 0 0 0 0 0 0 0 0 0 m84 n84 p84

0 0 0 m26 n26 p26 0 0 0 0 0 0 0 0 0 m62 n62 p62 0 0 0 0 0 0

0 0 0 0 0 0 m35 n35 p35 0 0 0 m53 n53 p53 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 m47 n47 p47 0 0 0 0 0 0 m74 n74 p74 0 0 0

( ) # $ # $ mij ¼ xi " xj =lij , nij ¼ yi " yj =lij , and pij ¼ zi " zj =lij :

0 0 0 0 0 0 0 0 0 0 0 0 m57 n57 p57 0 0 0 m75 n75 p75 0 0 0

0 0 0 0 0 0 0 0 0 m46 n46 p46 0 0 0 m64 n64 p64 0 0 0 0 0 0

0 0 0 m25 n25 p25 0 0 0 0 0 0 m52 n52 p52 0 0 0 0 0 0 0 0 0

0 0 0 m23 n23 p23 m32 n32 p32 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

3 0 0 7 7 0 7 72 3 0 7 7 F1 7 6 0 7 7 6 F2 7 6 F3 7 0 7 76 7 7 6 0 7 7 6 F4 7 6 F5 7 0 7 76 7 7 6 0 7 7 6 F6 7 6 F7 7 0 7 76 7 7 6 0 7 7 6 F8 7 6 F9 7 0 7 76 7 7 6 0 7 76 F10 7 6 F11 7 0 7 76 7 7 6 0 7 76 F12 7 6 F13 7 m67 7 76 7 7 6 n67 7 76 F14 7 6 F15 7 p67 7 76 7 7 6 m76 7 76 F16 7 7 4 n76 7 F17 5 p76 7 7 F18 0 7 7 0 5 0

260 6 Optimal Force Method for FEMs: Low Order Elements

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

fm ¼

ð

V

ct φcdV

261

ð6:37Þ

The integration is taken over the volume of the element, where φ is the matrix relating the stresses to strains ε = φ σ in three dimensional problems. The primary step in achieving the flexibility matrix of an element is determining the matrix c. It is obvious that the ith column of c represents the resultant stresses due to unit element force Fi in force method and also stresses due to nodal forces S is equal to the ith column of T utilizing displacement method. Hence we can form matrix c using stiffness properties of the brick element using the displacement method. Now the flexibility matrix of the element in the force method is formed from Eq. 6.37 using Gauss numerical integration method with eight Gauss points.

6.5.3

Graphs Associated with Finite Element Model

Here, topological properties of the FEM are transferred into the connectivity of its interface graph and natural associate graph.

6.5.3.1

Interface Graph

Interface graph of a FEM, denoted by IG(FEM), is constructed by the following rules: 1. Nodes of the IG(FEM) correspond to the nodes of FEM. 2. For each edge of a break element, one new member is added to the IG(FEM). 3. For each face of a break element, one new diagonal member is added to the IG (FEM). This member is located between second and third nodes of the current face of the element. In fact there is one to one to one correspondence between element forces and member of the IG(FEM). The members of the interface graph are numbered according to the element numbers of the FEM. In this way for each element, corresponding members in interface graph are numbered consequently and then members of the next element are numbered. A FEM and the corresponding interface graph and the schematic numbering of the members corresponding to nth element in the interface graph are illustrated in Fig. 6.45.

6.5.3.2

Natural Associate Graph

The natural associate graph represented by NAG(FEM) is constructed by the following rules:

262

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.45 (a) Finite element model (b) Interface graph of the FEM (c) Schematic numbering of the nth element

1. Nodes of the NAG(FEM) correspond to the elements of FEM. 2. For each pair of elements in FEM having four common nodes, one member is added between the corresponding two nodes in NAG(FEM). NAG(FEM) can be constructed using the following procedure: One of the preliminary steps in FEA is defining the elements with their connected nodes. In this way the element connectivity matrix is constructed which contains the elementnode incidence relationships. In the process of constructing the element connectivity matrix, another matrix which contains node-element incidence properties can be formed. This matrix is named the node connectivity matrix. Now using the element connectivity and the node connectivity matrices leads to an algorithm with complexity O(n) for an efficient generation of NAG. In order to recognize the adjacent elements to the nth element which have common four nodes or one common face, first the connected nodes to the nth element are identified from the element connectivity matrix. In the subsequent step using the node connectivity matrix, elements which have at least one common node with the nth element are identified. Now it is convenient to seek for the adjacent elements in this reduced search space. A FEM and its corresponding NAG are illustrated in Fig. 6.46.

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

263

Fig. 6.46 Finite element model and Finite element model with natural associate graph Fig. 6.47 Space truss model equivalent to a brick element

6.5.4

Topological Interpretation of Static Indeterminacy

6.5.4.1

Degree of Static Indeterminacy of the FEM

Each bi-action element force in a brick element can be considered as bi-action element forces in a bar element. In this way, the force system of the brick element will be equivalent to the force system of the corresponding space truss as indicated in Fig. 6.47. Thus calculating the degree of static indeterminacy (DSI) and forming the self equilibrating systems of the FEM are replaced by the DSI and self equilibrating systems of the equivalent truss model. In this way using the DSI of a space truss with n nodes and m members as DSI ¼ m " 3n + 6, the degree of indeterminacy of a FEM is obtained as. DSI ¼ 18E " 3N þ 6

ð6:38Þ

where E is the number of brick elements and N is the total number of the nodes of the FEM.

6.5.4.2

Pattern of Type I Self-Equilibrating Systems

For each k multiple member in equivalent truss model of FEM, there are k unknown forces and one equilibrium equation in the member’s direction. Thus DSI of the

264

6 Optimal Force Method for FEMs: Low Order Elements

substructure is equal to k " 1 and k " 1 self equilibrating systems can be generated on each k multiple member of interface graph of the FEM. In this way, first each k multiple members are arranged in ascending order as (m1, m2, m3,. . ., mk"1, mk). where (m1 < m2 < m3 < . . . < mk"1 < mk). Each selection of two members from this list is valid to construct a type I self-equilibrating system, but in order to achieve a better bandwidth reduction; selection of adjacent members from the defined list is preferable. Therefore k " 1 duplicate members are selected as (m1, m2), (m2, m3),. . ., (mk"1, mk). Each pair (mi, mj) with i < j represents the numbers of corresponding self-equilibrating system. The member with bigger number is selected as the generator of the current SES and also as a redundant force. The null vectors corresponding to the type I SESs have two non-zero entries in rows i and j equal to "1 and 1, respectively. Therefore by generating type I SESs, about three fourths of null basis is formed with maximum sparsity. These SESs are generated easily in the process of constructing natural associate graph of the FEM.

6.5.4.3

Relationship Between γ(S) and NAG(FEM)

By reducing the generators of the type I SESs from IG(FEM), the remaining subgraph is called graph S, with its associate graph A(S) being equivalent to NAG(FEM). In order to generate other types of the SESs, a relationship between the DSI of the equivalent truss of graph S and the natural associate graph of the FEM should be established. For achieving this aim an expansion process is employed. Consider a brick element, as illustrated in Fig. 6.48a. The corresponding graph S is denoted by S1 and NAG(FEM) is a single node. The equivalent structure is determinate. The graph S corresponding to two brick elements, denoted by S2, is constructed by adding the subgraph S01 (Fig. 6.49a) to the graph S1 as indicated in Fig. 6.48b, and also one node and one member is added to NAG(FEM) with the DSI becoming one. Consequently by adding subgraph S01 to the previous graph, it adds one node and member to the NAG(FEM) and it is growing as a tree and the DSI increases by unity. In some stages of the expansion process adding subgraph S02 (Fig. 6.49b) to the previous graph S from two faces, as shown in Fig. 6.48d, is equivalent to adding one node and two members to the NAG(FEM) and a cycle is formed in the NAG(FEM). In this case, the DSI of corresponding truss is increased by three. Considering the above points, the number of the nodes and members of the NAG (FEM) can be calculated as 0

N ¼

MðSÞ þ 2γðSÞ " 3 15

ð6:39Þ

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

S1

S2

S3

S4

M(S 1) = 18

M(S 2) = 31

M(S 3) = 44

M(S 4) = 53

N' = 1, M' = 0

N' = 2, M' = 1

N' = 3, M' = 2

N' = 4, M' = 4

DSI = 0

DSI = 1

DSI = 2

DSI = 5

a

b

c

265

d

Fig. 6.48 The expansion process for the formation of a γ-cycle without internal node

Fig. 6.49 A typical S0 subgraphs. (a) S01, (b) S02 0

M ¼

MðSÞ þ 17γðSÞ " 18 30

ð6:40Þ

Now, the relation between the DSI of the equivalent truss of the graph S and the independent cycles of the natural associate graph of FEM can be established. The

266

6 Optimal Force Method for FEMs: Low Order Elements

a

b

S7 M(S7) = 84 N' = 7, M' = 9 DSI = 12

S8 M(S8) = 90 N' = 8, M' = 12 DSI = 15

Fig. 6.50 The expansion process for the formation of a γ-cycle with internal node

first Betti number of the natural associate graph of the FEM, states the number of independent cycles of this graph which is expressed as 0

0

b1 ðNAGðFEMÞÞ ¼ M " N þ 1

ð6:41Þ

Adding M0 to both sides of the Eq. 6.41 leads to 0

0

0

b1 ðNAGðFEMÞÞ þ M ¼ 2M " N þ 1

ð6:42Þ

Substituting Eqs. 6.39, 6.40, 6.41, and 6.42 results in γðSÞ ¼ b1 ðNAGðFEMÞÞ þ M

0

ð6:43Þ

According to this equation, the DSI of the equivalent truss of the graph S can be expressed as the sum of the number of members and the first Betti number of NAG (FEM) that corresponds to type II and type III self-equilibrating systems. The graph S corresponding to the eight brick elements denoted by S8 is constructed by adding one node and six members to the graph S7 as illustrated in Fig. 6.50. This process adds one node, three members and three minimal cycles to the NAG(FEM), and also the DSI of the equivalent graph increases by three. When the FEM or corresponding graph S has an internal node and the NAG (FEM) becomes a polyhedral, then Eqs. 6.39 and 6.40 will be modified as 0

N ¼

MðSÞ þ 2γðSÞ þ 3Ni ðSÞ " 3 15

ð6:44Þ

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

267

Fig. 6.51 Pattern of a Type II self-equilibrating system 0

M ¼

MðSÞ þ 17γðSÞ þ 33Ni ðSÞ " 18 30

ð6:45Þ

Therefore the relation between the DSI of equivalent truss of graph S and the first Betti number of the NAG(FEM) is modified as 0

γðSÞ ¼ b1 ðNAGðFEMÞÞ þ M " 2Ni ðFEMÞ

ð6:46Þ

Comparing Eqs. 6.43 and 6.46 demonstrates that the FEM has no internal node the DSI of the equivalent truss of graph S is equal to the sum of the number of members and first Betti number of NAG(FEM), however, when the FEM has one or more internal nodes, 2Ni(FEM) self-equilibrating systems are not independent and must be ignored.

6.5.4.4

Pattern of Type II Self-Equilibrating Systems

As mentioned, type II self-equilibrating systems as indicated in Fig. 6.51 are topologically identical to the subgraph of graph S which corresponds to the two connected nodes of the natural associate graph of the FEM. The most important point in type II self-equilibrating systems is to select an appropriate generator. Because by eliminating these generators from graph S, the sub-structure of type III SESs and primary structure of the structure S must be stable. To achieve this, the following rule for appropriate selection of generators of type II SESs is suggested. In this way avoiding instability of the subsequent type of the SESs, the following procedure is applied, as indicated in Fig. 6.52. For a type II SESs (in any coordinate system such as Cartesian, cylindrical or spherical) generators of the type II SESs in directions 1, 2 and 3 are the chosen members which are numbered as 8, 11, and 23.

6.5.4.5

Pattern of Type III Self-Equilibrating Systems

According to the expansion process in models without opening, sub-structures which are topologically identical to the minimal cycles of the natural associate

268

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.52 Selected generators of the type II SES

Fig. 6.53 Minimal cycles of the natural associate graph of the FEM

graph of FEM contains some type II and one type III self-equilibrating systems as indicated in Figs. 6.53a and 6.54. 6.5.4.6

Type I Minimal Cycles

These minimal cycles of the natural associate graph of the FEM pass through E elements which have two common nodes and one edge. Corresponding interface graph of these elements have N = 4E + 2 nodes. Therefore using Eq. 6.37 the DSI of the related sub-structure is equal to 6E. Obviously 5E " 1 and E type I and type II self-equilibrating systems can be extracted from the mentioned sub-structure. The DSI of the remaining sub-structure is 1. Thus each type I minimal cycle of the natural associate graph of the FEM contains a type III self-equilibrating system and one null vector. Avoiding instability of the primary structure S, the procedure indicated in Fig. 6.55 is applied to selection of the generators of the type III SESs. For a type III SESs (in any coordinate system such as Cartesian, cylindrical or spherical) generators of SESs perpendicular to the directions 1, 2 and 3 are chosen members which are numbered as 70, 51 and 17, respectively.

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

269

Fig. 6.54 Selected generators of the type III SESs

6.5.4.7

Type II Minimal Cycles

For models with openings, each independent cycle of the natural associate graph which surrounds an opening of the FEM is called type II minimal cycle of the natural associate graph. Considering that this cycle passes through E elements and its corresponding interface graph has N = 4E nodes. Using Eq. 6.37, the DSI of the related sub-structure is equal to 6E + 6. Obviously 5E and E type I and type II selfequilibrating systems can be extracted from the mentioned sub-structure. Therefore the DSI of the remaining sub-structure is 6. Thus each type II minimal cycle of the natural associate graph of the FEM contains six self-equilibrating systems of type III and six corresponding null vectors. These null vectors can easily be generated on the corresponding sub-structure utilizing an algebraic method.

270

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.55 Eight elements and the corresponding hexahedron natural associate graph

6.5.5

Models Including Internal Node

For Ni internal node in the FEM, 2Ni self-equilibrating systems are not independent from others and then should be selected and ignored. Since each type III SESs include some type I and type II SESs, therefore ignoring each type I and type II SESs causes the corresponding type III SESs not to be valid. Therefore for any internal node, two type III SESs should be selected and ignored. The following procedure should be applied to select dependent SESs. Considering graph S7 its equivalent truss is twelve times statically indeterminate. As it can be seen from Fig. 6.50a, b adding one node and six members consisting of three edge members and three diagonal ones, to graph S7 forms graph S8. Then the DSI of the equivalent truss is increased by three. Considering the equivalent truss of graph S7 and by eliminating the restraints corresponding to the generators of SESs, the primary structure which is determinate and stable is obtained. Also corresponding primary structure of graph S8 is obtained by eliminating the generators of SESs of graph S7 plus the above mentioned three diagonal members, form graph S8. Eight elements correspond to each internal node and the natural associate graph corresponding to these elements is a hexahedron. At the beginning, from each hexahedron as illustrated in Fig. 6.55, node 8 is considered as the last node which makes up the hexahedron. These nodes must be distinct from each others. From each hexahedron three type III SESs corresponding to three minimal cycles of NAG(FEM) which pass through the selected node, are ignored and also the mentioned three diagonal members used as generator of three type II SESs corresponding to three members of NAG(FEM) which pass through the selected node. In Figs. 6.55 and 6.56, the red members represent the modified type II SESs.

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

271

Fig. 6.56 A finite element model and the corresponding natural associate graph

6.5.6

Selection of an Optimal List Corresponding to Minimal Self-Equilibrating Stress Systems

The main goal of this section is the selection of a member list for each selfequilibrating systems which has the nearest member numbers to the generator of the system. Minimize ðAbsði " jÞÞ

ð6:47Þ

where j is the number of the generator member, and i are the number of members of the self-equilibrating system. Consider dm i ; (i ¼ 1, 2, . . ., k; m ¼ 1, 2, . . ., n), representing the member numbers of the multiple members where d1i < d2i < . . . < dni , and sj; (j ¼ 1, 2, . . ., t) representing the member numbers of the simple members. Since dm i with m 6¼ 1 is used as generator of type I of SESs, as illustrated in Figs. 6.52 and 6.53, the generator of type II and type III of SESs must be selected from {d1i ,sj}. For maximum bandwidth reduction, from each multiple member one member is selected which has the nearest number to the generator’s number of the selfequilibrating system. In order to achieve this goal, Eq. 6.48 should be applied. ( ) ( ) ( ( )) find dij jAbs dij " d1G ¼ Min abs dij " d1G

ð6:48Þ

d1i ¼ dij ði ¼ 1, 2, . . . , k; j ¼ 1, 2, . . . , nÞ

ð6:49Þ

then

where, d1G is the generator of the self-equilibrating system. Algorithm. Step 1: Number the nodes of the FEM. Nodal numbering does not affect the pattern of the flexibility matrix of the FEM.

272

6 Optimal Force Method for FEMs: Low Order Elements

Step 2: Define the brick elements through its eight nodes. Use an efficient method for element numbering, for having small bandwidth for the null basis matrix B1 and the flexibility matrix. Step 3: Generate the natural associate graph of the FEM. Step 4: Generate the interface graph of the FEM in manner that its member numbering is according to the element numbering of the FEM. In this way for each element, corresponding members in interface graph are numbered consequently. Step 5: Construct the equilibrium matrix of the FEM. Step 6: Set up the type I self-equilibrating systems and calculate the corresponding null vectors which have two nonzero entries in the rows corresponding to the member numbers. Step 7: Set up the type II self-equilibrating systems and calculate the corresponding null vectors form the relevant equilibrium sub-matrix. Step 8: Set up the type III self-equilibrating systems and calculate the corresponding null vectors form the relevant equilibrium sub-matrix. Step 9: Construct the statical basis (null basis) of the FEM by arranging the null vectors in the null basis in the ascending manner utilizing the highest member number of the corresponding self-equilibrating systems. The efficiency of this algorithm is shown through two examples by comparing the required computational times for the construction of the null basis matrices, also non-zero patterns and condition numbers of the flexibility matrices. In this comparison (a) Present method (b) Turn-back method (c) Gauss-Jordan elimination method are considered.

6.5.7

Numerical Examples

In this section two FEMs are considered, one of these models is assumed to be supported in statically indeterminate fashion and the other supported in a determinate fashion. Null basis and flexibility matrices are formed and the required computational times, and the condition numbers are calculated. In the following examples, nz represents the number of non-zero entries and λmax/λmin is the ratio of the extreme eigenvalues taken as the condition number of a matrix. Example 1. A thick arch type structure, having internal radius of 8 m, discretized by brick elements. The corresponding FEM is supported in a statically indeterminate fashion as illustrated in Fig. 6.57. The mechanical and topological properties of the model are as follow: Poisson’s ratio ¼ 0.2; Elastic modulus E ¼ 2E + 10 N/m2; Density ρ ¼ 2,400 kg/m3; Number of nodes ¼ 165; Number of internal nodes (Ni) ¼ 27; Number of elements ¼ 80; Number of members of the natural associate graph ¼ 172;

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

273

Fig. 6.57 A thick arch type structure discretized by brick elements

Fig. 6.58 Pattern of the equilibrium matrix of the FEM for Example 1

First Betti number of the natural associate graph (independent cycles) ¼ 93; Number of Type I self-stress systems ¼ 740 (77.8 %); Number of Type II self-stress systems ¼ 172 (18.1 %); Number of Type III self-stress systems ¼ 93 " 2 ( 27 ¼ 39 (4.1 %); DSIInternal ¼ 951; DSIExternal ¼ 21; DSITotal ¼ 972; The pattern of the equilibrium matrix of the FEM is displayed in Fig. 6.58. The nodes and elements of the FEM are numbered in a way to produce a banded equilibrium matrix. This characteristic facilitates the Turn-back method to form a null basis with less required computational time and more banded form. The null basis of the FEM can be constructed using a mixed algebraic-graph theoretical and pure algebraic methods. In mixed methods, graph theoretical

274

6 Optimal Force Method for FEMs: Low Order Elements

Table 6.8 Definition of the element forces Force system 1 2 3 4 5 6 7 8 9

Location (nodes) 1,3 3,4 2,4 1,2 5,7 7,8 6,8 5,6 1,5

Force system 10 11 12 13 14 15 16 17 18

Location (nodes) 3,7 4,8 2,6 3,5 4,7 4,6 2,5 2,3 6,7

Table 6.9 Member list corresponding to the type II or type III self-equilibrating systems d11 d11

d12 d22 d32

...

S1

S2

...

d1G d2G ⋮ dnG

...

d1i d2i ⋮ dni

...

Sj

...

d1k d2k

approach is utilized to form columns of the null basis which are related to the internal indeterminacies where algebraic procedures form the columns corresponding to the external indeterminacies. In this case, graph theoretical approach is employed together with the Turn-back method (Tables 6.8 and 6.9). As displayed in Fig. 6.59 and Table 6.10, applying the mixed graph theoretical with Turn-back method lead to a highly sparse and banded null basis; however it requires some additional computational time than using the mixed graph theoretical with QR decomposition method. Pure algebraic methods are also used to form the null basis of the equilibrium matrix of the FEM. From Fig. 6.59 and Table 7.3 it can be observed that each pure algebraic method has better performance when they are used together with the present graph theoretical method (Fig. 6.60). Example 2. An arch type structure with an internal radius of 8 m, discretized by brick elements. As shown in Fig. 6.61, this structure has two openings and is supported in a statically determinate fashion. Properties of the model are as follow: Mechanical Properties Poisson’s ratio ¼ 0.2; Elastic modulus E ¼ 2E + 10 N/m2; Density ρ ¼ 2,400 kg/m3; Topological Properties Number of nodes ¼ 108; Number of internal nodes (Ni) ¼ 0; Number of elements ¼ 38; Number of members of the natural associate graph ¼ 59 First Betti number of the natural associate graph (independent cycles) ¼ 22;

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

275

Fig. 6.59 Pattern of the null basis B1 matrices corresponding to Example 1 utilizing: (a) Graph theoretical-Turn back method (b) Graph theoretical-QR method (c) Turn-back method (d) Gauss Jordan elimination method

Number of Type I self-stress systems ¼ 275 (75.1 %); Number of Type II self-stress systems ¼ 59 (16.1 %); Number of Type III self-stress systems ¼ 20 + 2 ( 6 ¼ 32 (8.8 %); DSIInternal ¼ 366; DSIExternal ¼ 0; DSITotal ¼ 366;

276

6 Optimal Force Method for FEMs: Low Order Elements

Table 6.10 Comparison of the optimality characteristics of the null basis matrices B1 and the flexibility matrices G for the FEM of Example 1 Flexibility matrix G

Null basis B1 Time (sec) Graph theoreticalTurn back method Graph theoreticalQR method Turn back method Gauss Jordan elimination method

Time

λmax Time λmin Present method

nz

nz entries

2.464

1.000

2.475e +07

1.000

0.824

0.334

1.692

70.358

28.554

25.921

10.520

4.699e +08 2.776e +07 5.655e +06

entries Present

method

1.622 7.279

Fig. 6.60 Pattern of the flexibility matrices G corresponding to Example 1 utilizing: (a) Graph theoretical-Turn back method (b) Graph theoretical-QR method (c) Turn-back method (d) Gauss Jordan elimination method

6.5 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

277

Fig. 6.61 An arch type structure containing two openings, discretized by brick elements

Fig. 6.62 Pattern of the null basis B1 matrices corresponding to Example 2 utilizing: (a) Graph theoretical method (b) Turn-back method (c) Gauss Jordan elimination method

278

6 Optimal Force Method for FEMs: Low Order Elements

Fig. 6.63 Pattern of the flexibility matrices G corresponding to Example 2 utilizing (a) Graph theoretic method (b) Turn-back method (c) Gauss Jordan elimination method

Table 6.11 Comparison of the optimality characteristics of the null basis matrices B1 and the flexibility matrices G for the FEM of Example 2 Null basis B1 Time (sec) Graph theoretical method Turn back method Gauss-Jordan elimination method

Time

Flexibility matrix G λmax Time λmin Present method

0.171

1.000

16.816

98.339

6.394

37.392

nz

nz entries

entries Present

method

2.860e 1.000 +4 1.033e 1.103 +4 1.814e 3.296 +5

The pattern of the null basis and flexibility matrices are illustrated in Figs. 6.62 and 6.63 and it can be easily seen that present graph theoretical method and Turnback method lead to banded null basis and flexibility matrices. Table 6.11 containing the optimality characteristics of the applied methods reveals that the presented method requires acceptable computational time for constructing the null basis of the FEM. As mentioned, the present method leads to a highly sparse and banded flexibility matrix requiring a low computational time with an acceptable condition number. These examples are also analyzed by the standard displacement method and the integrated force method. Figure 6.64 illustrates the pattern of the reduced stiffness matrix, and its optimality characteristics are provided in Table 6.12. It can be seen that for these examples, the displacement method analyzes the model with less unknowns than the presented graph theoretical force method, and the reduced stiffness matrix Kr has less non-zero entries than the flexibility matrix G.

References

279

Fig. 6.64 Pattern of the reduced stiffness matrix Kr corresponding to (a) Example 1 (b) Example 2

Table 6.12 Optimality characteristics of the stiffness matrix Kr

Example 1 Example 2

λmax λmin

nz

1.793e+4 2.667e+5

0.428 0.797

nz entries

entries Present

method

References 1. Kaveh A, Koohestani K, Taghizadeh N (2007) Efficient finite element analysis by graphtheoretical force method. Finite Elem Anal Des 43(6–7):543–554 2. Kaveh A, Koohestani K (2009) Efficient graph-theoretical force method for two dimensional rectangular finite element analysis. Commun Numer Methods Eng 25(9):951–971 3. Kaveh A, Ebrahimi E (2012) Graph-theoretical force method of finite element models with triangular and rectangular elements. Asian J Civil Eng 13(5):597–616 4. Kaveh A (1974) Applications of topology and matroid theory to the analysis of structures. Ph. D. thesis, Imperial College, London University, London, UK 5. Kaveh A, Koohestani K (2008) Efficient finite element analysis by graph-theoretical force method; triangular and rectangular plate bending elements. Finite Elem Anal Des 44:646–654 6. Kaveh A, Koohestani K (2007) An efficient graph theoretical method for plate bending finite element analysis via force method. Eng Comput 24(7):679–698 7. Kaveh A, Koohestani K (2008) Efficient graph-theoretical force method for three dimensional finite element analysis. Commun Numer Methods Eng 24(11):1533–1551 8. Kaveh A, Tolou Kian SJ (2012) Efficient finite element analysis using graph-theoretical force method; brick element. Finite Elem Anal Des 54:1–15 9. Kaveh A, Roosta GR (1988) Comparative study of finite element nodal ordering methods. Eng Struct 20(1–2):86–96 10. Przemieniecki JS (1968) Theory of matrix structural analysis. McGraw-Hill, New York

280

6 Optimal Force Method for FEMs: Low Order Elements

11. Kaveh A (2004) Structural mechanics: graph and matrix methods, 3rd edn. Research Studies Press (Wiley), Somerset 12. Watkins DS (2002) Fundamentals of matrix computations, 2nd edn. Wiley, New York 13. Kaveh A, Fazli H (2008) Analysis of frames by substructuring technique based on using algebraic and graph methods. Commun Numer Methods Eng 24(10):867–874

Chapter 7

Optimal Force Method for FEMS: Higher Order Elements

7.1

Introduction

In this chapter force method for the analysis of finite element models comprising of higher order elements are studied. In the first part, an efficient graph theoretical force method is presented for the analysis of FEMs comprising of higher order triangular elements, corresponding to highly sparse and banded flexibility matrices [1]. This is achieved by associating special graphs to a finite element model, and selecting subgraphs for the formation of localized self stress systems. In second part, a method is described for the formation of null bases for FEMs comprised of higher order rectangular plane stress and plane strain elements (serendipity family elements) leading to highly sparse and banded flexibility matrices for optimal finite element analysis by force method [2]. In the third part, an competent method is described for the formation of null bases of finite element models (FEMs) consisting of hexahedron elements, corresponding to highly sparse and banded flexibility matrices. This is achieved by associating special graphs with the FEM and selecting appropriate subgraphs and forming the self-equilibrating systems on these subgraphs [3].

7.2

Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

This part introduces an efficient method for the finite element analysis of models comprised of higher order triangular elements. The presented method is based on the force method and benefits graph theoretical transformations. For this purpose, minimal subgraphs of predefined special patterns are selected. Self-equilibrating systems (S.E.Ss) are then constructed on these subgraphs leading to sparse and

A. Kaveh, Computational Structural Analysis and Finite Element Methods, 281 DOI 10.1007/978-3-319-02964-1_7, © Springer International Publishing Switzerland 2014

282

7 Optimal Force Method for FEMS: Higher Order Elements

banded null basis. Finally, well-structured flexibility matrices are formed for efficient finite element analysis.

7.2.1

Definition of the Element Force System

Defining appropriate structural elements is the first step of structural analysis. Based on the analysis approaches, structural elements are formulated in different manners. In case of higher order triangular elements (in-plane forces), in displacement method two forces are employed at each node of the element, while in force method the following force system is utilized. Considering an O(n) element first, 3n sets of edge bi-action forces are described between adjacent side nodes. Then n(n ! 1)/2 bi-action forces are added between adjacent nodes parallel to side 23. The same forces are added parallel to side 13. Finally n ! 1 bi-action forces are added in the same manner, parallel and in the closest position to side 12. Force systems corresponding to the second, third and fourth order elements are shown in Fig. 7.1a–c. These independent element forces denoted by F are related to nodal forces S using Eq. 7.1. S ¼ TF

7.2.2

ð7:1Þ

Flexibility Matrix of the Element

The flexibility matrices of higher order triangular elements can simply be formed using the stiffness matrices of such elements. f m ¼ ðTr Þt ðKr Þ!1 Tr

ð7:2Þ

where the subscript r indicates that, corresponding orders of matrices to dependent forces are reduced.

7.2.3

Graphs Associated with Finite Element Model

In order to benefit topology in finite element analysis, first some topological transformations of FEM are needed. In this relation ten different graphs are presented in Ref. [4]. Here natural associate graph and interface graph are used that are defined in the following:

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

a

283

b

c

Fig. 7.1 Element force systems of higher order triangular elements

7.2.3.1

Natural Associate Graph

The natural associate graph of a FEM is represented by NAG(FEM). This graph reveals elements adjacency properties and as illustrated in Fig. 7.2a is constructed by following rules: 1. Each node of NAG(FEM) corresponds to each element of the FEM. 2. Two nodes of NAG(FEM) are connected with a member if two corresponding O(n) elements have n + 1 common nodes on a common edge. Natural associate graph can easily be generated using the following procedure: Connected nodes with a considered element are identified using element connectivity matrix. 1. Connected elements with these nodes are identified using node connectivity matrix.

284

7 Optimal Force Method for FEMS: Higher Order Elements

a

b

Fig. 7.2 Natural associate graph and interface graph of the corresponding FEM comprised of fourth order elements

2. All identified elements in step 2, at least have one common node with the considered element in step 1. Now among these identified elements the one which has n + 1 common nodes with the considered element, is desirable. 7.2.3.2

Interface Graph

The interface graph of a FEM is represented by IG(FEM). This graph corresponds to the force system of the FEM, and as indicated in Fig. 7.2b it is constructed by the following rules: 1. Each node of IG(FEM) corresponds to the each node of the FEM. 2. Members of the IG(FEM) correspond to the force system of FEM between their adjacent nodes. 3. Each support condition is considered as a member of IG(FEM). Members of the interface graph corresponding to the element forces are numbered according to element numbering. Meantime corresponding members to support conditions are numbered before members of their connected elements.

7.2.4

Topological Interpretation of Static Indeterminacies

7.2.4.1

Degree of Static Indeterminacy of the FEM

As mentioned in Sect. 3.1, the introduced element force system is comprised of a number of bi-action forces. In accordance with Przemieniecki [5] each bi-action force can be considered as force system of a bar element, hence force system of the equivalent truss element can be employed instead of the force system of the original element. Therefore, the DSI of the FEM and self-equilibrating systems can be conveniently explored.

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

285

Each higher order triangular element has n2 + 3n ! 1 bi-action forces, and hence the DSI of a FEM which is comprised of E elements and contains N nodes is obtained as ! " DSI ¼ n2 þ 3n ! 1 E ! 2N þ 3

ð7:3Þ

Following self-equilibrating systems are found. Then null vectors which express the equilibrium conditions of the self-equilibrating systems are generated.

7.2.4.2

Pattern of the Type I Self-Equilibrating Systems

The interface graph of a FEM contains double members at the interface of two elements (Fig. 7.2b). Each double member of the interface graph correspond to a Type I self-equilibrating system. The self-equilibrating system consisting of two members numbered as i and j (j > i). The member with bigger number is selected as the generator of the self-equilibrating system and is considered as the redundant force of the FEM. Typical null vector corresponding to a Type I self-equilibrating system contains two nonzero entries in ith and jth rows equal to !1 and 1, respectively. The above mentioned double members can conveniently be identified while natural associate graph is being generated.

7.2.4.3

Identification of Other Self-Equilibrating Systems Using an Expansion Process

This section adopts a method to identify other self-equilibrating systems and locate other redundant forces. Consider a graph S the same as the interface graph of the FEM with generators of Type I self-equilibrating systems being removed.

7.2.4.4

Models Excluding Openings

Consider a general triangular element as shown in Fig. 7.3a. The corresponding graph S contains NE ¼ (n + 1)(n + 2)/2 nodes and ME ¼ (n + 1)(n + 2) ! 3 members, thus the equivalent truss is determinate. The NAG(FEM) is an isolated node. When another element is added (Fig. 7.3b, c), each time NE ! n ! 1 nodes and ME ! n members are added to the corresponding graph S, thus the indeterminacy is increased by n ! 1. The NAG(FEM) grows with a node and member. This is true while The NAG(FEM) is growing like a tree (with no cycle). In some steps of the expansion process adding an element grows NAG(FEM) by a node and two members, and a cycle is formed in the natural associate graph, as illustrated in Fig. 7.3d. In this situation (NE ! 2n ! 1) nodes and (ME ! 2n) members are added to the corresponding graph S, hence the indeterminacy is increased by 2(n ! 1) + 1.

286

7 Optimal Force Method for FEMS: Higher Order Elements

a

c

b

d

Sc1

Sc2

Sc3

Sc4

N(Sc1) = NE

N(Sc2) = 2NE-n-1

N(Sc3) = 3NE-2n-2

N(Sc4) = 4NE-4n-3

M(Sc1) = ME

M(Sc2) = 2ME-n

M(Sc3) = 3ME-2n

M(Sc4) = 4ME-4n

N' = 1, M' = 0 γ (Sc1) = 0

N' = 2, M' = 1

N' = 3, M' = 2

N' = 4, M' = 4

γ (Sc2) = n-1

γ (Sc3) = 2(n-1)

γ (Sc4) = 4(n-1)+1

Fig. 7.3 Expansion process in FEMs comprising of O(n) elements and without opening

Based on the above mentioned remarks it can clearly be seen that each member and cycle of natural associate graph corresponds to n ! 1 and one degrees of indeterminacy, respectively. This conclusion is theoretically validated as follows. Considering the above points, the number of nodes and members of the natural associate graph are derived as 0

M ¼ 0

MðSÞ þ ME γðSÞ ! ME nð M E ! 1Þ

N ¼

MðSÞ þ γðSÞ ! 1 ðME ! 1Þ

ð7:4Þ ð7:5Þ

Now the first Betti number is epmloyed to calculate the number of independent cycles of the natural associate graph of the FEM 0

0

b1 ðNAGðFEMÞÞ ¼ M ! N þ 1

ð7:6Þ

By substituting Eqs. 7.4 and 7.5 in Eq. 7.6, the degree of static indeterminacy of the equivalent truuss is obtained using the natural associate graph of the FEM: γðSÞ ¼ b1 ðNAGðFEMÞÞ þ ðn ! 1ÞM0

ð7:7Þ

This equation shows that the subgraphs of S which correspond to memebers of NAG(FEM), represent n ! 1 degree of indeterminacy and n ! 1 self-equilibrating systems can be constructed which are called Type II self-equilibrating systems. Meantime subgraphs of S which correspond to independent cycles of NAG(FEM), represent one degree of indeterminacy and one self-equilibrating system, called Type III self-equilibrating systems, can be formed.

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

7.2.5

287

Models Including Opening

In this section, an expansion process is employed in the process of expanding a cycle which surounds an opening, to identiy new degrees of indeterminacy and corresponding the new self-equilibrating systems. Consider the finite element model of Fig. 7.4a. Adding an element in manner shown in Fig. 7.4b, adds NE ! n ! 2 nodes and ME ! n members to the corresponding graph S, thus increasing the indeterminacy of equivalent truss by (n ! 1) + 2. When the final element is added as shown in Fig. 7.4c, NE ! 2n ! 1 nodes and ME ! 2n members are added to the corresponding graph S leading to an increase of DSI by 2(n ! 1) + 1. In the step which is shown in Fig. 7.4b if the Type II self-equilibrating systems are ignored, two new self-equilibrating systems can be recognized and considering Fig. 7.4c there is one new self-equilibrating system. As pointed out, the truss corresponding to a mininal cycle of NAG(FEM) that surrounds an opening, contains three self-equilibrating systems. These self-equilibrating systems are classified in Type III self-equilibrating systems. Here, Eq. 7.7 is modified by adding the term 2nc. Each cycle of NAG(FEM) which surrounds an opening is considered as an independent cycle by the first Betti number as 0

γðSÞ ¼ b1 ðNAGðFEMÞÞ þ ðn ! 1ÞM þ 2nc

ð7:8Þ

where nc is the number of openings in FEM.

7.2.5.1

Pattern of Type II Self-Equilibrating Systems

Subgraphs of the graph S which correspond to members of the NAG(FEM) are the underlying subgraphs of Type II self-equilibrating systems. If n is considered as the order of elements, n ! 1 Type II self-equilibrating systems can be constructed on each subgraph. Consider a triangular element; the second element can be attached from each three sides. Depending on the side to which the second element is attached, generators of Type II self-equilibrating systems are selected in different ways. Figure 7.5a shows a second order element indicated by bold nodes which is connected to three elements from three sides. In each case the corresponding generator is identified by dashed red line. The same is shown in Fig. 7.5b, c considering third and fourth order elements. Meantime as it can be noticed from Fig. 7.5, the pattern of the generators can conveniently be expanded for elements with higher orders.

288

7 Optimal Force Method for FEMS: Higher Order Elements

a

b

c

So1

So2

So3

N(So1) = No1

N(So2) = No1+ NE-n-2

N(So3) = No1+ 2NE-3n-3

M(So1) = M o1

M(So2) = M o1+ ME-n

M(So3) = M o1+ 2ME-3n

N' = N'o1, M' = M'o1

N' = N'o1+1, M' = M'o1+1

DSI = DSI o1

DSI = DSI o1+(n-1)+2

N' = N'o1+2, M' = M'o1+3 DSI = DSI o1+3(n-1)+3

Fig. 7.4 Expansion process in FEMs comprising of O(n) elements, containing an opening

a

b

c

Fig. 7.5 Appropriate generators of Type II self-equilibrating systems from models comprised of second, third and fourth order elements

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

7.2.5.2

289

Pattern of Type III Self-Equilibrating Systems

Subgraphs of the graph S which correspond to minimal cycles of the NAG(FEM) are underlying subgraphs of Type III self-equilibrating systems. These minimal cycles can be categorized into two classes.

7.2.5.3

Type I Minimal Cycles

These cycles pass through elements which all have one certain node in common. As discussed in expansion process, subgraphs of graph S corresponding to Type I minimal cycles lead to one Type III self-equilibrating system. Figure 7.6a–c represent the underlying subgraphs and generators of Type III self-equilibrating systems corresponding to second, third and fourth order elements. The generators are indicated by dashed red lines. Meantime Fig. 7.6 implies that the pattern of the generators can conveniently be expanded for the elements with higher orders.

7.2.5.4

Type II Minimal Cycles

These cycles pass through elements which surround an opening. According to the expansion process, subgraphs of the graph S corresponding to Type II minimal cycles contain three self-equilibrating systems of Type III. Consider Fig. 7.4b, in this situation, based on expansion process two Type III self-equilibrating systems are formed hence the two corresponding generators can simply be selected from members of the last added element. The last Type III self-equilibrating systems is formed when the Type II minimal cycle is completed (Fig. 7.4c). Here again the corresponding generator is simply selected from the members of the added element.

7.2.5.5

Self-Equilibrating Systems Corresponding to the External Indeterminacies

These self-equilibrating systems are formed in relation with the external degrees of indeterminacy. For this purpose, each indeterminate restraint forces is considered as a redundant force. Here unlike the internal redundant forces, the external ones are not bi-action forces. Thus the corresponding self-equilibrating systems will require simple support conditions. A typical self-equilibrating system regarding to an external indeterminacy shown in Fig. 7.7b is formed based on a tree of NAG (FEM) which connects the external redundant force to a close simple support configurations. However, it is essential that these self-equilibrating systems remain independent from each other. After selecting two ends of the tree, it is desirable the above mentioned tree to pass through elements with close numbers.

290

7 Optimal Force Method for FEMS: Higher Order Elements

a

b

c

Fig. 7.6 Appropriate generators of Type III self-equilibrating systems from models comprised of second, third and fourth order elements a

b

Fig. 7.7 A self-equilibrating system corresponding to an indeterminate support condition

7.2.6

Selection of an Optimal List Corresponding to Minimal Self-Equilibrating Stress Systems

Consider a general self-equilibrating system of Type II or Type III. According to the above procedure SESs consist of single members like sj and double members d1i

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

291

Table 7.1 Member list of a typical Type II or Type III self-equilibrating system d11 d21

d12 d22

...

s1

s2

...

g

...

d1i d2i

...

sj

...

d1k d2k

which have a twin member as d2i . It is clear that replacing each d1i member with its twin member does not affect the topology of the corresponding SES, Table 7.1. Thus except the generators which are unique and necessary for the independency of the SESs, other double members can be replaced by their twin members. Here, considering the bandwidth reduction of the null basis, the following procedure is utilized to select members with closer numbers to the generator number. ! " ! " If abs g ! d2i < abs g ! d1i then

d2i ! d1i

ð7:9Þ

where, g is the generator number.

Algorithm Step 1: Define and number the nodes of the FEM. Step 2: Define the triangular elements and use an efficient numbering method to reduce the bandwidth of the null basis and flexibility matrice. Step 3: Generate the natural associate graph of the FEM based on the adjacency of the elements. Step 4: Generate the interface graph of the FEM in a manner that its members are numbered according to the element numbering of the FEM. Step 5: Select the Type I self-equilibrating systems and form the corresponding null vectors. Step 6: Set up the Type II and Type III self-equilibrating systems, and form the corresponding null vectors consisting of the members’ forces when the generator’s force is equal to unity. Step 7: Finally assemble the null basis (static basis) of the FEM by arranging the null vectors in the ascending order of the highest member number of the selfequilibrating systems. The above algorithm is implemented in MATLAB and is used to analyze three structures, and the efficiency of the present method is illustrated through these examples.

7.2.7

Numerical Examples

In this section three examples are studied. In each case, first the structure is idealized using second order triangular elements. The null basis matrices are constructed utilizing the present method and two algebraic procedures, namely

292

7 Optimal Force Method for FEMS: Higher Order Elements

the Gauss-Jordan elimination method and QR factorization. Then the results are contrasted through normalized computational time for the formation of the null basis matrices and nonzero pattern and condition numbers of the flexibility matrices. In the following examples, nz represents the number of non-zero entries and, the ratio of the extreme eigenvalues, λmax/λmin, is taken as the condition number of the matrices. In the second step, each example is idealized using second, third and fourth order elements. Then the properties of the flexibility matrices obtained from the present method are compared to those of the corresponding stiffness matrices. For this purpose condition s and the number of unknowns, namely the DSI for the force method and the DKI degree of kinematic indeterminacy for the displacement method are utilized. Example 1. Consider a beam structure with determinate support conditions. The beam is bent under a uniformly distributed load of intensity q ¼ 10 kN/m. The structure is idealized using plain stress triangular elements. As indicated in Fig. 7.8, three types of elements are generated using second, to fourth order elements. All models have the same nodes with the following mechanical properties: Thickness ¼ 0.01 m, E ¼ 2e + 8 kN/m2, ν ¼ 0.3. Topological properties of the models are collected in Table 7.2. Figure 7.9 displays pattern of the null basis matrices employing the present method and two other algebraic procedures for the model comprising of second order elements. In this relation Table 7.3 contains other optimality characteristics of the force method procedures. It is clear that the graph theoretical method forms the most well-structured null basis in smallest computational time. Figure 7.10 shows the pattern of the flexibility matrices for the models comprising of the second to fourth order elements. It is noticeable that as the order of elements increases the DSI of the model decreases. Meanwhile, with identical number of nodes for the models, the DKI stays the same, as illustrated in Fig. 7.11. Table 7.4 contains the ratio of the DKI/DSI. Finally, for a model with second order elements, the average σxx stresses at nodes of Path 1 are compared through the results of the present force method and displacement method in [1]. Example 2. A beam structure which depicted in Fig. 7.12 is bent under a uniformly distributed load of intensity q ¼ 10 kN/m. The structure is analyzed three times using second, third and fourth order elements. Plane stress elements are considered with the following mechanical properties: thickness ¼ 0.01 m, E ¼ 2 e + 8 kN/m2, ν ¼ 0.3. Topological properties of the models are collected in Table 7.5. Figure 7.13 and Table 7.6 reveal the optimality characteristics of the present graph theoretical and the two algebraic force methods. The present method leads to a sparse and banded null basis using the smallest computational time. QR factorization method forms a null basis in reasonable computational time and leads to a well-conditioned flexibility matrix but the flexibility matrix is not well-structured at all.

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

293

a

b

c

Fig. 7.8 FEM of the beam of Example 1 Table 7.2 Topological properties for the FEM of Example 1 FEM Element type Second order Third order Fourth order

Nodes 481 481 481

Elements 216 96 54

Self-equilibrating systems

Force system

Type I 600 384 276

Int. 1,944 1,632 1,458

Type II 300 256 207

Type III 85 33 16

DSI 985 673 499

Comparing Figs. 7.14 and 7.15 shows that, by using higher order triangular elements there will be fewer compatibility conditions than equilibrium equations. Thus utilizing the graph theoretical force method can be attractive and economical.

294

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.9 Pattern of the null basis B1 corresponding to Example 1 with second order elements and utilizing (a) Graph theoretical method (b) Gauss Jordan elimination method (c) QR decomposition method Table 7.3 Comparison of the optimality characteristics of the null basis matrices B1 and flexibility matrices G for the FEM of Example 1

Graph theoretical method Gauss Jordan elimination method QR factorization method

Null basis B1 Time Time Present method

Flexibility matrix G

1.00 106.48

1.00 17.01

1.21e+4 3.08e+4

3.06

43.27

8.22e+3

nz

nz entries

entries Present

method

λmax λmin

Table 7.7 contains the ratios of the DKI/DSI and also condition numbers of the stiffness matrices. Example 3. A cross section of a retaining wall is idealized using triangular elements, as illustrated in Fig. 7.16. The model is analyzed three times using second, third and fourth order elements. Here plane strain elements are utilized with the following mechanical properties: Thickness ¼ 1 m, E ¼ 2e + 7 kN/m2, ν ¼ 0.2. Topological properties of the models are collected in Table 7.8. Optimality characteristics of the employed force methods can be seen in Fig. 7.17 and Table 7.16. The patterns of the flexibility matrices utilizing present method are illustrated in Fig. 7.18. The difference between nonzero numbers of the

7.2 Finite Element Analysis of Models Comprised of Higher Order Triangular Elements

a0

b

200

150

400

c

0

200 300

450

800

400

600

0

200

400

600

800

0

100

300

600

295

0

150

nz = 22419

300

450

600

0

100

nz = 20431

200

300

400

nz = 19123

Fig. 7.10 Pattern of the flexibility matrices G corresponding to Example 1, considering (a) second order (b) third order (c) fourth order elements

a0

b

200

200

200

400

400

400

600

600

600

800

800

800

0

200

400

600

800

c

0

0

200

nz = 13251

400

600

800

0

0

nz = 29987

200

400

600

800

nz = 41403

Fig. 7.11 Pattern of the reduced stiffness matrices Kr corresponding to Example 1 considering (a) second order (b) third order (c) fourth order elements Table 7.4 Optimality characteristics of the reduced stiffness matrices Kr for Example 1.

Element type

λmax λmin

DKI DSI

Second order Third order Fourth order

6.22e+4 9.45e+4 1.59e+5

0.97 1.42 1.92

Fig. 7.12 Beam structure of Example 2

296

7 Optimal Force Method for FEMS: Higher Order Elements

Table 7.5 Topological properties for the FEM of Example 2 FEM Element type Second order Third order Fourth order

Nodes 350 742 1,278

Elements 144 144 144

Self-equilibrating systems

Force system

Type I 368 552 736

Int. 1,296 2,448 3,888

Type II 184 368 552

Type III 47 47 47

DSI 599 967 1,335

Fig. 7.13 Pattern of the null basis B1 corresponding to Example 2 with second order elements and utilizing (a) Graph theoretical method (b) Gauss Jordan elimination method (c) QR decomposition method Table 7.6 Comparison of the optimality characteristics of the null basis matrices B1 and flexibility matrices G for the FEM of Example 2 Null basis B1

Time Graph theoretical method Gauss Jordan elimination method QR factorization method

Time Present method

Flexibility matrix G

nz

nz entries

entries Present

method

λmax λmin

1.00 60.33

1.00 11.50

4.24e+4 1.27e+5

1.67

26.21

1.66e+3

matrices is due to the use of different SESs for external redundant forces (Table 7.9). The patterns of reduced stiffness matrices and the corresponding condition numbers are provided in Fig. 7.19 and Table 7.10, respectively.

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

a

b

0

c

0

200

150

400

600

600

599

900

800 0

150

300

450

1200

0

599

0

300

300

450

297

200

400

600

800

0

300

600

900

1200

nz = 54869

nz = 30579

nz = 13685

Fig. 7.14 Pattern of the flexibility matrices G corresponding to Example 2 considering (a) second order (b) third order (c) fourth order elements

a

b

0

c

0

150

300

300

600

450

900

600

1200

0

600 1200 1800 2400

0

150

300

450

600

0

300

nz = 9319

600

0

900 1200

600

1200

1800

2400

nz = 110811

nz = 45527

Fig. 7.15 Pattern of the reduced stiffness matrices Kr corresponding to Example 2 considering (a) second order (b) third order (c) fourth order elements

Table 7.7 Optimality characteristics of the reduced stiffness matrices Kr for Example 2

7.3

Second order Third order Fourth order

λmax λmin

DKI DSI

5.94e+4 2.08e+5 6.25e+5

1.16 1.53 1.91

Finite Element Analysis of Models Comprised of Higher Order Rectangular Elements

In this section, an efficient method is developed for the formation of null bases of finite element models (FEMs) consisting of rectangular plane stress and plane strain serendipity family elements, corresponding to highly sparse and banded flexibility matrices. This is achieved by associating special graphs with the FEM and selecting appropriate subgraphs and forming the self-equilibrating systems (SESs) on these subgraphs. The efficiency of the present method is illustrated through three examples.

298

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.16 A section of the retaining wall of Example 3

Table 7.8 Topological properties for the FEM of Example 3 FEM Element type Second order Third order Fourth order

7.3.1

Nodes 150 319 551

Elements 63 63 63

Self-equilibrating systems

Force system

Type I 166 249 332

Int. 567 1,071 1,701

Type II 83 166 249

Type III 21 21 21

Ext. 34 50 66

DSI 301 483 665

Definition of Element Force System

For the generation of the equilibrium matrix A of a FEM, a set of independent forces system should be defined and also their relations with the element nodal forces should be established. In displacement method we have two forces at each node of the element. For an element with N nodes, 2 & N nodal forces can be defined. Using three equilibrium equations, 2N ! 3 independent forces will remain. In other words, there are 2N ! 3 independent element forces in an element with N nodes. The nodal forces and element forces systems are shown in Fig. 7.20 for rectangular plane stress and plane strain serendipity family elements with various numbers of boundary nodes. For the higher order elements, the element forces system can be obtained with the same procedure. These element forces can be related to the nodal forces for a rectangular element by a (2N) & (2N ! 3) transformation matrix using Eq. 7.10 as

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

299

Fig. 7.17 Pattern of the null basis B1 corresponding to Example 3 with second order elements and utilizing (a) Graph theoretical method (b) Gauss Jordan elimination method (c) QR decomposition method

a0

b

c

0

100

75

150

200

150

300

300 225 300

450

400 0

75

150

225

300

0

600 0

100

nz = 6927

200

300

400

0

150

nz = 16443

300

450

600

nz = 42261

Fig. 7.18 Pattern of the flexibility matrices G corresponding to Example 3 considering (a) second order (b) third order (c) fourth order elements Table 7.9 Comparison of the optimality characteristics of the null basis matrices B1 and flexibility matrices G for the FEM of Example 3 Null basis B1

Time Graph theoretical method Gauss Jordan elimination method QR factorization method

Time Present method

Flexibility matrix G

nz

nz entries

entries Present method

λmax λmin

1.00 48.17

1.00 5.54

5.02e+5 1.19e+5

1.62

13.07

4.80e+3

300

a

7 Optimal Force Method for FEMS: Higher Order Elements

b

0

75

150

225 0

75

150

225

c

0

0

150

250

300

500

450

750

0

150

300

450

1000 0

250

nz = 17784

nz = 5276

500

750

1000

nz = 44176

Fig. 7.19 Pattern of the reduced stiffness matrices Kr corresponding to Example 3 considering (a) second order (b) third order (c) fourth order elements Table 7.10 Optimality characteristics of the reduced stiffness matrices Kr for Example 3 Second order Third order Fourth order

λmax λmin

DKI DSI

4.35e+3 1.31e+4 3.61e+4

0.88 1.21 1.55

S ¼ TF

ð7:10Þ

Transformation matrix can be formed simply as ðn1 ; n2 Þ ¼ end nodes of element force Fj For i ¼ 1 : N For j ¼ 1 : 2N ! 3 If i ¼¼ n1 Tð2i ! 1, jÞ ¼ mn1 n2 and Tð2i, jÞ ¼ nn1 n2 If i ¼¼ n2 Tð2i ! 1, jÞ ¼ mn2 n1 and Tð2i, jÞ ¼ nn2 n1 End End Where xi and yi are the Cartesian coordinates of node i, mij ¼

xi !xj lij

,

nij ¼

yi !yj lij

are

the direction cosines and lij is the length of the line between nodes i and j.

7.3.2

Flexibility Matrix of the Element

In this case flexibility matrices of higher order triangular elements can simply be formed using the stiffness matrices of such elements.

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

301

Fig. 7.20 A set of rectangular serendipity family elements

f m ¼ ðTr Þt ðKr Þ!1 Tr

ð7:11Þ

where the subscript r indicates that, corresponding orders of matrices to dependent forces are reduced.

7.3.3

Graphs Associated with Finite Element Model

In order to transfer the topological property of a finite element model to the connectivity of a graph ten different graphs are previously introduced in Chap. 4.

302

7 Optimal Force Method for FEMS: Higher Order Elements

Here natural associate graph and interface graph are used that are defined in the following:

7.3.3.1

Natural Associate Graph

The natural associate graph represented by NAG(FEM) is constructed by the following rules: 1. Nodes of the NAG(FEM) correspond to the elements of FEM. 2. For each pair of elements in FEM having (N + 4)/4 common nodes, one member is added between the corresponding two nodes in NAG(FEM). NAG can be constructed using the following procedure: One of the preliminary steps in FEM is defining the elements with their connected nodes. In this way the element connectivity matrix is constructed which contains the element-node incidence relationships. In the process of constructing the element connectivity matrix, another matrix which contains node-element incidence properties can be formed. This matrix is named the node connectivity matrix. Now using the element connectivity and the node connectivity matrices leads to an algorithm with complexity O(n) for an efficient generation of NAG. In order to recognize the adjacent elements to the nth element which have (N + 4)/4 common nodes or one common face, first the connected nodes to the nth element are identified from the element connectivity matrix. In the subsequent step using the node connectivity matrix, elements which have at least one common node with the nth element are identified. Now it is convenient to seek for the adjacent elements in this reduced search space. A FEM and its corresponding NAG are illustrated in Fig. 7.21.

7.3.3.2

An Interface Graph

The interface graph of a finite element model denoted by IG (FEM) can easily be constructed for rectangular FEM using the following rules: 1. This graph contains all the nodes of the FEM. 2. With the all edges of an element of FEM, N graph members are associated. Therefore, in the interface of two elements, 2-multiple members are presented. 3. For each element with N nodes, 2N ! 3 members should be considered in the interface graph. Thus, N ! 3 ¼ (2N ! 3) ! N) diagonal members should be added. This graph for a quadratic and cubic FEM is shown in Fig. 7.22. The member numbering of the interface graph should be performed according to the numbering of the FEM, taking into account the primary nodal numbering of a considered element in the model. Thus, for each rectangular element 2N ! 3 members of the interface graph will be numbered sequentially according to the patterns which were illustrated in Fig. 7.20.

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

303

Fig. 7.21 A quadratic rectangular FEM with its natural associate graph (bold lines) for a circular plate

Fig. 7.22 A quadratic, cubic and quartic rectangular FEM with their interface graphs

7.3.4

Topological Interpretation of Static Indeterminacies

7.3.4.1

Degree of Static Indeterminacy of the FEM

Considering Fig. 7.20, in order to find the patterns corresponding to the selfequilibrating systems, a rectangular element is simulated as a planar truss formed as the 1-skeleton of the rectangular element together with some diagonal members. This is possible since the independent element forces applied at in the nodes and are

304

7 Optimal Force Method for FEMS: Higher Order Elements

along the edges of the rectangular element. The statical indeterminacy of a planar truss with m members and n nodes is given as γ(S) ¼ m ! 2n + 3; therefore, the degree of statical indeterminacy (DSI) of the entire model supported in a statically determinate manner can be calculated with the same relationship as DSI ¼ ð2N ! 3Þ & M ! 2n þ 3

ð7:12Þ

Where M is the total number of elements, N is the number of nodes within an element and n is the total number of nodes of the FEM. Following self-equilibrating systems are found. Then null vectors which express the equilibrium conditions of the self-equilibrating systems are generated.

7.3.4.2

Type I Self-Equilibrating Systems

Each multiple member of the interface graph is a subgraph on which one selfequilibrating system can be generated. In other words, on a 2-multiple member numbered as (i, j) with the condition (i < j), one self-equilibrating system can be constructed (extracted). Each pair such as (i, j) for which (i < j) corresponds to a null vector with their non-zero entries being located in rows i and j, and their numeric values are (!1, 1), respectively. The member with bigger member number (j) is called the generator. These pairs are called Type I self-equilibrating systems. For a FEM we have N4 & M0 self-equilibrating systems of Type I, where M0 is the number of members of the associate graph of the model.

7.3.4.3

Type II Self-Equilibrating Systems

There are other types of self-equilibrating systems which are extracted from two adjacent elements of FEM. In other words, for two adjacent elements with N nodes, the DSI can be calculated as: DSI ¼ ð2N ! 3Þ & M ! 2n þ 3 #

) DSI ¼ ð2N ! 3Þ & 2 ! 2 & 2N !

Nþ4 4

$

þ3¼

N!7

ð7:13Þ

N 4

self-equilibrating systems were generated as Type I systems. Thus the number of remaining self-equilibrating systems is Type II ¼

N N N !1! ¼ !1 2 4 4

ð7:14Þ

In other words, N4 ! 1 SESs should be extracted from two adjacent elements. This number is equal to the number of internal nodes of the remaining subgraph after

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

305

Fig. 7.23 (a) Subgraph corresponding to SESs of Type II, (b) Pattern of Type II self-equilibrating systems in horizontal direction

deleting the generators of SESs of Type I. For example, the remaining subgraphs for two adjacent cubic elements are shown in Fig. 7.23a in two directions. In this figure, the diagonal members are curved for better illustration. After deleting the generators corresponding to Type I SESs, the null vectors should be calculated from the remaining subgraph. These null vectors can easily be generated on the corresponding sub-structure utilizing an algebraic method. For instance, results SESs in horizontal direction are shown in Fig. 7.23b. In a FEM, the total number of Type II SESs can be calculated as: # $ N Type II ¼ M & !1 4 0

ð7:15Þ

where M0 is the number of members of the associate graph of the model and N is the number of nodes of an element. The most important point in Type II self-equilibrating systems is to select an appropriate generator. In fact by eliminating these generators from graph S, the

306

7 Optimal Force Method for FEMS: Higher Order Elements

sub-structure of Type III SESs and the primary structure of the structure S must remain stable.

7.3.4.4

Type III Self-Equilibrating Systems

Sub-structures which are topologically identical to the minimal cycles of the natural associate graph of FEM contains some Type I, II and one Type III self-equilibrating systems.

Type I Minimal Cycles of NAG(S) These minimal cycles of the natural associate graph of the FEM pass through four elements which have one common node. Corresponding interface graph of these elements have n nodes and m edges for a FEM with N-node elements. m ¼ 4 & ð2N ! 3Þ # $ Nþ4 n ¼ 4N ! 4 & þ 1 ¼ 3 & ð N ! 1Þ 4

ð7:16Þ ð7:17Þ

Subsequently, the DSI of the interface graph is DSI ¼ m ! 2n þ 3

) DSI ¼ 4 & ð2N ! 3Þ ! 2 & ð3 & ðN ! 1ÞÞ þ 3 ¼ 2N ! 3

ð7:18Þ

! " 0 The N, N ¼ N4 & M ¼ N4 & 4 , SESs are Type I and there are N ! 4, ! ! " ! "" 0 N ! 4 ¼ M & N4 ! 1 ¼ 4 & N4 ! 1 , SESs of Type II. DSI ! ðTypeI&IIÞ ¼ ð2N ! 3Þ ! ðN þ ðN ! 4ÞÞ ¼ 1

ð7:19Þ

Therefore, one independent SES should be extracted. This SES with eight members can be formed for any types of rectangular elements around the common node as is indicated bold in Fig. 7.24.

Type II Minimal Cycles of NAG(S) Each minimal cycle that surrounds an opening is called the Type II minimal cycle. Such a cycle passes through M0 , (M0 ' 8), finite elements and its corresponding ! " 0 0 interface graph has 3N 4 ! 1 & M nodes and M & (2N ! 3) members. The DSI of subgraph is

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

307

Fig. 7.24 The SES of Type III corresponding to the common node of four rectangular elements

#

$ 3N 0 !1 &M þ3 DSI ¼ M & ð2N ! 3Þ ! 2 & 4 # $ N 0 !1 þ3 ) DSI ¼ M & 2 0

0

0

that N4 & M SESs of Type I and Type II ¼ M & extracted.

!N 4

ð7:20Þ

" ! 1 SESs of Type II can be

# $ % # $& N N N 0 0 !1 ! M & þM & !1 þ3 ¼ 3 ð7:21Þ DSI!Type I & Type II ¼ M & 2 4 4 0

Therefore, each Type II minimal cycle corresponds to three null vectors which are calculated utilizing an algebraic method.

7.3.5

Selection of Generators for SESs of Type II and Type III

The most important point in Type II and Type III self-equilibrating systems is to select appropriate generators. This is by eliminating these generators from graph S, the sub-structure of primary structure of the structure S must remain stable. To achieve this, the following rule for appropriate selection of generators of Type II SESs is suggested. For quadratic and rectangular element the generators of SESs Type II and Type III are illustrated in Figs. 7.25 and 7.26, respectively. It should be noted that the generators corresponding to Type I were chosen previously. In addition, the generators corresponding to an opening are the last non-zero entries of its columns which are not common with the previously selected generators.

308

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.25 Selected generators of the Type II SES

Fig. 7.26 Selected generators of the Type III SES

7.3.6

Algorithm

Step 1: Generate the associate graph of the FEM and use an efficient method for its nodal numbering see Chaps. 4 and 5. It is obvious that good numbering of this graph corresponds to good numbering of elements of the FEM. This numbering leads to a banded adjacency matrix of the graph and correspondingly to a banded flexibility matrix. Since the numbering of the members of the interface graphs corresponds to the element numbering of the finite elements, such a numbering is the only parameter for controlling the bandwidth of the flexibility matrix. Step 2: Set up the equilibrium matrix of the FEM. Step 3: Generate the interface graph and perform its numbering. The numbering of this graph should be performed according to the element numbering of the

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

309

considered FEM. After this numbering the interface graph can easily be formed and its members can be numbered. Step 4: Find the Type I self-equilibrating systems. All multiple members of the interface graph are identified and the values !1 and 1 are assigned to appropriate rows (corresponding to the member numbers) and the corresponding minimal null vectors are created. Step 5: Find the Type II self-equilibrating systems. The N4 ! 1 SESs of Type II should be extracted from two adjacent elements. Step 6: Find the Type III self-equilibrating systems. For each minimal cycle of natural associate graph of FEM with four members (one common node), one SES and with eight or more members (Opening), three SESs should be extracted. Step 7: Order the null vectors. At this step the constructed null vectors should be ordered such that their last non-zero entries form a list with an ascending order.

7.3.7

Numerical Examples

In this section three FEMs are considered, one of these models is assumed to be supported in statically indeterminate fashion and the other two supported in a determinate fashion. The effect of the presence of additional supports can separately be included for each special case with no difficulty. The equilibrium matrices are formed. Null bases and the flexibility matrices are constructed and the required computational times, and the condition numbers are calculated. In all the following examples, nnz represents the number of non-zero entries and λmax/λmin is the ratio of the extreme eigenvalues taken as the condition number of a matrix. The comparison between present algorithm and algebraic force method is shown in Table 7.11 for all three examples. Finally the present method is validated through comparison of resulting stresses using the present graph-theoretical force method and the displacement method. Example 1. The lining of a tunnel is considered supported in a statically determinate manner, and its applied load is depicted in Fig. 7.27. This structure is discretized using rectangular 8-node finite elements. The properties of the model are as follows: Poisson’s ratio ¼ 0.3; Elastic modulus E ¼ 2e + 7 kN/m2; Thickness t ¼ 1.00 m Number of rectangular 8-node elements ¼ 100 Number of nodes ¼ 405 DSI ¼ 100 & 13 ! 2 & 405 + 3 ¼ 493 Number of Type I self-equilibrating systems ¼ 296 (60 %) Number of Type II self-equilibrating systems ¼ 148 (30 %) Number of Type III self-equilibrating systems ¼ 49 (10 %)

310

7 Optimal Force Method for FEMS: Higher Order Elements

Table 7.11 The comparison between present algorithm and algebraic force method for all three examples Condition number (flexibility matrices)

λmax λmin

Norms max|A & B1|

Example

Computational time LU time

Present method

LU Present factorization method

LU factorization

Tunnel lining Circulate beam Retaining wall (8-node) Retaining wall (12-node)

1.21 0.45 0.84

47.65 9.38e+5 2.68e+4

1.63e+5 8.73e+7 4.28e+7

1.08e!15 4.04e!14 5.51e!14 1.76e!13 7.43e!12 2.67e!13

0.78

3.59e+5

8.01e+7

1.22e!14 1.90e!13

Fig. 7.27 A lining of a tunnel, the discretization and loading of the structure

The interface and natural associate graphs of the FEM model are illustrated in Fig. 7.28. The pattern of the equilibrium matrix is shown in Fig. 7.29. The sparsity of the final null basis obtained by the present method is approximately 6.7 % of that of QR method and 6.07 % of the LU method as depicted in Fig. 7.30. The flexibility matrix, G, is also well-structured as shown in Fig. 7.31. The results are verified by standard displacement method in Table 7.12. Example 2. A circular plate and its applied load are shown in Fig. 7.32. The internal and external diameters are 1.00 and 5.00 m, respectively. This structure is discretized using 12-node rectangular finite elements. The properties of the model are as follows:

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

311

Fig. 7.28 Interface and natural associate graphs of Example 1. (a) Interface graph, (b) natural associate graph

Fig. 7.29 Pattern of the equilibrium matrix for Example 1

312

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.30 Patterns and number of non-zero entries of the null bases of Example 1: (a) present algorithm, (b) QR factorization and (c) LU factorization

Fig. 7.31 Patterns of the flexibility matrix G ¼ Bt1 FmB1 for Example 1 using the proposed method

Poisson’s ratio ¼ 0.3; Elastic modulus E ¼ 2e + 7 kN/m2; Thickness t ¼ 1.00 m Number of rectangular 12-node elements ¼ 384 Number of nodes ¼ 2,064 DSI ¼ 384 & 21 ! 2 & 2064 + 3 ¼ 3939 Number of Type I self-equilibrating systems ¼ 2,160 ((55 %) Number of Type II self-equilibrating systems ¼ 1,440 ((36 %) Number of Type III self-equilibrating systems ¼ 336 (internal nodes) + 3 (an opening) ¼ 339 ((8.5 %)

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

313

Table 7.12 Comparison of the displacement method and the present force method for Example 1 Method

Element stresses Displacement method

Element

σ xx

σ yy

The present force method σ xy

2

kN/cm 1 10 20 30 40 50 60 70 80 90 100

!0.1806 !0.0918 !0.3097 !0.6168 !0.8943 !1.0196 !0.9346 !0.6790 !0.3672 !0.1244 !0.1361

σ xx kN/cm

!0.7815 !0.7379 !0.6082 !0.3721 !0.1416 !0.0346 !0.1073 !0.3214 !0.5666 !0.7240 !0.7739

0.2763 !0.2186 !0.4040 !0.4470 !0.3060 !0.0306 0.2586 0.4333 0.4265 0.2627 0.3446

σ yy

σ xy

!0.7815 !0.7379 !0.6082 !0.3721 !0.1416 !0.0346 !0.1073 !0.3214 !0.5666 !0.7240 !0.7739

0.2763 !0.2186 !0.4040 !0.4470 !0.3060 !0.0306 0.2586 0.4333 0.4265 0.2627 0.3446

2

!0.1806 !0.0918 !0.3097 !0.6168 !0.8943 !1.0196 !0.9346 !0.6790 !0.3672 !0.1244 !0.1361

Fig. 7.32 A circulate plate with an opening

The interface and natural associate graph of the FEM model is illustrated in Figs. 7.33 and 7.21. The pattern of the equilibrium matrix is shown in Fig. 7.34. The sparsity of the final null basis obtained by the present method is approximately 0.46 % of the QR method and 1.9 % of the LU approach as depicted in Fig. 7.35.

314

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.33 The interface graph of Example 2

Fig. 7.34 Pattern of the equilibrium matrix for Example 2

The flexibility matrix is also well-structured as shown in Fig. 7.36. The results are verified by the standard displacement method in Table 7.13. Example 3. The FEM of a dam which is supported in a statically indeterminate fashion is depicted in Fig. 7.37. This structure is discretized using 8-node and 12-node rectangular finite elements separately. It should be noted that the number of support elements depends on the choice of 8 or 12 nodes per finite element. The properties of the models are:

7.3 Finite Element Analysis of Models Comprised of Higher Order. . .

315

Fig. 7.35 Patterns and the number of non-zero entries of the null bases of Example 2: (a) present algorithm, (b) QR factorization and (c) LU factorization

Fig. 7.36 Patterns of flexibility matrix G ¼ Bt1 FmB1 for Example 2 using the proposed method Poisson’s ratio ¼ 0.3; Elastic modulus E ¼ 2e + 7 kN/m2; Thickness t ¼ 1.00 m Case 1: Number of rectangular 8-node elements ¼ 192, Number of nodes ¼ 681 Case 2: Number of rectangular 12-node elements ¼ 192, Number of nodes ¼ 1,117 DSI8 ! node ¼ 192 & 13 ! 2 & 681 + 82 ¼ 1, 216 DSI12 ! node ¼ 192 & 21 ! 2 & 1117 + 122 ¼ 1, 920 Number of Type I self-equilibrating systems, Case 1 ¼ 664 (58.5 %) Number of Type II self-equilibrating systems, Case 1 ¼ 332 (29 %) Number of Type III self-equilibrating systems, Case 1 ¼ 141 (12.5 %) Number of Type I self-equilibrating systems, Case 2 ¼ 996 (55 %) Number of Type II self-equilibrating systems, Case 2 ¼ 664 (36.8 %) Number of Type III self-equilibrating systems, Case 2 ¼ 141 (8.2 %)

316

7 Optimal Force Method for FEMS: Higher Order Elements

Table 7.13 Comparison of the displacement method and the present force method for Example 2 Method

Element stresses Displacement method

Element

σ xx

σ yy

The present force method σ xy

2

kN/cm 337 340 343 346 349 352 255 358 361 364 367 370 373 376 379 382

!2.6962 !2.7314 !2.7879 !2.8348 !2.8483 !2.8220 !2.7686 !2.7158 !2.6939 !2.7288 !0.8483 !2.4803 !2.5319 !2.3806 !2.6693 !2.7092

σ xx kN/cm

!2.8935 !2.8329 !2.7646 !2.7121 !2.6972 !2.7262 !2.7870 !2.8547 !2.9112 !2.9929 !2.4765 !2.7025 !2.7105 !2.6839 !2.7841 !2.9526

0.0188 0.0701 0.0793 0.0464 !0.0098 !0.0612 !0.0813 !0.0572 0.0059 0.1074 !0.0155 !0.0552 0.0084 0.0927 !0.4591 !0.0679

σ yy

σ xy

!2.8935 !2.8329 !2.7646 !2.7121 !2.6972 !2.7262 !2.7870 !2.8547 !2.9112 !2.9929 !2.4765 !2.7025 !2.7105 !2.6839 !2.7841 !2.9526

0.0188 0.0701 0.0793 0.0464 !0.0098 !0.0612 !0.0813 !0.0572 0.0059 0.1074 !0.0155 !0.0552 0.0084 0.0927 !0.4591 !0.0679

2

!2.6962 !2.7314 !2.7879 !2.8348 !2.8483 !2.8220 !2.7686 !2.7158 !2.6939 !2.7288 !0.8483 !2.4803 !2.5319 !2.3806 !2.6693 !2.7092

The interface and natural associate graphs of the FEM model are illustrated in Fig. 7.38 for FEM with 8-node elements. The interface graph for other cases can simply be obtained. The final null basis obtained for both cases by the present method are depicted in Figs. 7.39 and 7.40. The flexibility matrix is also wellstructured as shown in Fig. 7.41. The results are verified by the standard displacement method in Table 7.14.

7.4

Efficient Finite Element Analysis Using GraphTheoretical Force Method: Hexa-Hedron Elements

Formation of a suitable null basis for equilibrium matrix is the main problem of finite elements analysis via force method. For an optimal analysis, the selected null basis matrices should be sparse and banded corresponding to sparse, banded and well-conditioned flexibility matrices. In this section, an efficient method is developed for the formation of null bases of finite element models (FEMs) consisting of hexahedron elements, corresponding to highly sparse and banded flexibility matrices. This is achieved by associating special graphs with the FEM and selecting appropriate subgraphs and forming the self-equilibrating systems (SESs) on these subgraphs.

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

317

Fig. 7.37 A retaining wall and the corresponding rectangular meshes

7.4.1

Independent Element Forces and Flexibility Matrix of Hexahedron Elements

In the force method the efficiency of this analysis depends on the required time for the formation of the matrix. G ¼ Bt1 FmB1 and its characteristics, i.e. sparsity and bandedness together with its conditioning. For the formation of a well-structured matrix G, one should select a well-structured B1 matrix.

318

7 Optimal Force Method for FEMS: Higher Order Elements

For the generation of the equilibrium matrix A of a FEM, a system of independent force systems should be defined and also their relations with the element nodal forces should be established. In displacement method we have three forces at each node of the element. For an element with N nodes, 3 & N nodal forces can be defined. Using six equilibrium equations, 3N ! 6 independent forces will be remained. In other words, there are 3N ! 6 independent element forces in an element with N nodes. The nodal forces and element forces systems are shown in Fig. 7.42 for hexahedron elements with various numbers of boundary nodes. For the higher order elements, the element forces system can be obtained with the same procedure. These element forces F can be related to the nodal forces S for a N-node element by a (3N) & (3N ! 6) transformation matrix using Eq. 7.22 as S ¼ TF

ð7:22Þ

Transformation matrix can be formed simply as where xi, yi and zi are the Cartesian coordinates of node i, mij ¼ (xi ! xj)/lij, nij ¼ (yi ! yj)/lij, and pij ¼ (zi ! zj)/lij, are the direction cosines and lij is the length of the line between nodes i and j. Formulation of a discrete element equivalent to the actual continuous structure is the first step in matrix structural analysis. For a linear system it can be assumed that the stresses σ .are related to the forces F by linear equation as σ ¼ cF

ð7:24Þ

The matrix c represents a statically equivalent stress system due to the unit force F. The flexibility matrix of an element can be written as ð f m ¼ ct φcdV ð7:25Þ V

The integration is taken over the volume of the element, where φ is the matrix relating the stresses to strains ε ¼ φσ in three dimensional problems. The primary step in the formation of the flexibility matrix of an element is determining the matrix c. It is obvious that the ith column of c represents the resultant stresses due to unit element force Fi in the force method and also stresses due to nodal forces S is equal to the ith column of T utilizing the displacement method. Hence, we can form matrix c using the stiffness properties of the hexahedron element using the displacement method. Now the flexibility matrix of the element in the force method is formed from Eq. 7.25 using Gauss numerical integration method with sixty four Gauss points (4 & 4 & 4 Gauss Points Integration).

3 2 m13 S1 6 S2 7 6 n13 7 6 6 6 S3 7 6 p13 7 6 6 6 S4 7 6 0 7 6 6 6 S5 7 6 0 7 6 6 6 S6 7 6 0 7 6 6 6 S7 7 6 m31 7 6 6 6 S8 7 6 n31 7 6 6 6 S9 7 6 p31 7 6 6 6 S10 7 6 0 7 6 6 6 S11 7 6 0 7 6 6 6 S12 7 6 0 7 6 6 6 S13 7 ¼ 6 0 7 6 6 6 S14 7 6 0 7 6 6 6 S15 7 6 0 7 6 6 6 S16 7 6 0 7 6 6 6 S17 7 6 0 7 6 6 6 S18 7 6 0 7 6 6 6 S19 7 6 0 7 6 6 6 S20 7 6 0 7 6 6 6 S21 7 6 0 7 6 6 6S 7 6 0 6 22 7 6 4S 5 4 0 23 0 S24

2

0 0 0 0 0 0 m34 n34 p34 m43 n43 p43 0 0 0 0 0 0 0 0 0 0 0 0

0 0 0 m24 n24 p24 0 0 0 m42 n42 p42 0 0 0 0 0 0 0 0 0 0 0 0

m12 n12 p12 m21 n21 p21 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m57 n57 p57 0 0 0 m75 n75 p75 0 0 0

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m78 n78 p78 m87 n87 p87

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m68 n68 p68 0 0 0 m86 n86 p86

0 0 0 0 0 0 0 0 0 0 0 0 m56 n56 p56 m65 n65 p65 0 0 0 0 0 0

m15 n15 p15 0 0 0 0 0 0 0 0 0 m51 n51 p51 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 m37 n37 p37 0 0 0 0 0 0 0 0 0 m73 n73 p73 0 0 0

0 0 0 0 0 0 0 0 0 m48 n48 p48 0 0 0 0 0 0 0 0 0 m84 n84 p84

0 0 0 m26 n26 p26 0 0 0 0 0 0 0 0 0 m62 n62 p62 0 0 0 0 0 0

0 0 0 0 0 0 m35 n35 p35 0 0 0 m53 n53 p53 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 m47 n47 p47 0 0 0 0 0 0 m74 n74 p74 0 0 0

0 0 0 0 0 0 0 0 0 m46 n46 p46 0 0 0 m64 n64 p64 0 0 0 0 0 0

0 0 0 m25 n25 p25 0 0 0 0 0 0 m52 n52 p52 0 0 0 0 0 0 0 0 0

0 0 0 m23 n23 p23 m32 n32 p32 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

ð7:23Þ

3 0 0 7 7 0 7 72 3 0 7 7 F1 7 6 0 7 7 6 F2 7 7 6 0 7 6 F3 7 7 7 6 0 7 7 6 F4 7 7 6 0 7 6 F5 7 7 7 6 0 7 7 6 F6 7 7 6 0 7 6 F7 7 7 7 6 0 7 7 6 F8 7 7 6 0 7 6 F9 7 7 7 6 0 7 76 F10 7 7 6 0 76 F11 7 7 7 6 0 7 76 F12 7 7 6 m67 76 F13 7 7 7 6 n67 7 76 F14 7 7 6 p67 76 F15 7 7 7 6 m76 7 76 F16 7 7 4 n76 7 F17 5 p76 7 7 F18 0 7 7 0 5

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . . 319

320

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.38 Interface graph and natural associate graph for both cases of Example 3, (a) Interface graph for 8-node element, (b) Interface graph for 12-node element and (c) Associate graph for both cases

Fig. 7.39 Patterns and number of non-zero entries of null bases of Example 3 (8-node element): (a) present algorithm, (b) QR factorization and (c) LU factorization

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

321

Fig. 7.40 Patterns and number of non-zero entries of null bases of Example 3 (12-node element): (a) present algorithm, (b) QR factorization and (c) LU factorization

Fig. 7.41 Patterns of flexibility matrix G ¼ Bt1 FmB1 of Example 3, (a) 8-node element, (b) 12-node element

7.4.2

Graphs Associated with Finite Element Models

7.4.2.1

An Interface Graph

The interface graph of a finite element model denoted by IG (FEM) can easily be constructed for hexahedron FEM using the following rules: 1. This graph contains all the nodes of the FEM.

8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 136 144 152 160 168

Element

Method

0.0560 !0.0430 0.2060 0.0655 !1.0339 0.0694 !0.1939 !0.1546 !0.1114 !0.0979 !0.0891 !0.0809 !0.0729 !0.0650 !0.0570 !0.0491 !0.0412 !0.0333 !0.0254 !0.0176 !0.0097

0.9029 !1.0513 0.0529 1.4381 !1.4361 0.0060 1.2847 1.1539 1.0027 0.8474 0.6999 0.5645 0.4431 0.3363 0.2446 0.1683 0.1073 0.0612 0.0294 0.0104 0.0019

0.0436 !0.2987 !0.0512 0.1071 !0.9329 0.0006 0.0540 !0.0439 !0.0788 !0.0779 !0.0703 !0.0616 !0.0527 !0.0441 !0.0359 !0.0281 !0.0210 !0.0145 !0.0089 !0.0044 !0.0012

0.0560 !0.0430 0.2060 0.0655 !1.0339 0.0694 !0.1939 !0.1546 !0.1114 !0.0979 !0.0891 !0.0809 !0.0729 !0.0650 !0.0570 !0.0491 !0.0412 !0.0333 !0.0254 !0.0176 !0.0097

0.9029 !1.0513 0.0529 1.4381 !1.4361 0.0060 1.2847 1.1539 1.0027 0.8474 0.6999 0.5645 0.4431 0.3363 0.2446 0.1683 0.1073 0.0612 0.0294 0.0104 0.0019

σ yy 0.0436 !0.2987 !0.0512 0.1071 !0.9329 0.0006 0.0540 !0.0439 !0.0788 !0.0779 !0.0703 !0.0616 !0.0527 !0.0441 !0.0359 !0.0281 !0.0210 !0.0145 !0.0089 !0.0044 !0.0012

σ xy 0.0322 0.0346 0.2399 0.1210 !0.9501 0.0736 !0.1916 !0.1572 !0.1124 !0.0975 !0.0884 !0.0802 !0.0722 !0.0643 !0.0564 !0.0485 !0.0407 !0.0329 !0.0251 !0.0173 !0.0096

kN/cm2

σ xx 0.8158 !0.6760 0.1925 1.6845 !1.1784 0.0340 1.2772 1.1521 1.0021 0.8475 0.7002 0.5648 0.4434 0.3366 0.2449 0.1686 0.1075 0.0614 0.0296 0.0105 0.0020

σ yy

σ xx

σ xy

σ xx

σ yy

Displacement method

Displacement method

kN/cm2

12-node The present force method

Element stresses 8-node

Table 7.14 The comparison of the displacement method and the present force method for Example 3

0.0426 !0.3087 !0.0526 0.0629 !0.9567 0.0035 0.0490 !0.0411 !0.0764 !0.0767 !0.0695 !0.0608 !0.0521 !0.0436 !0.0354 !0.0278 !0.0207 !0.0143 !0.0088 !0.0044 !0.0012

σ xy 0.0322 0.0346 0.2399 0.1210 !0.9501 0.0736 !0.1916 !0.1572 !0.1124 !0.0975 !0.0884 !0.0802 !0.0722 !0.0643 !0.0564 !0.0485 !0.0407 !0.0329 !0.0251 !0.0173 !0.0096

σ xx

0.8158 !0.6760 0.1925 1.6845 !1.1784 0.0340 1.2772 1.1521 1.0021 0.8475 0.7002 0.5648 0.4434 0.3366 0.2449 0.1686 0.1075 0.0614 0.0296 0.0105 0.0020

σ yy

0.0426 !0.3087 !0.0526 0.0629 !0.9567 0.0035 0.0490 !0.0411 !0.0764 !0.0767 !0.0695 !0.0608 !0.0521 !0.0436 !0.0354 !0.0278 !0.0207 !0.0143 !0.0088 !0.0044 !0.0012

σ xy

The present force method

322 7 Optimal Force Method for FEMS: Higher Order Elements

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

323

Fig. 7.42 A set of hexahedron elements

2. With each edge of an element of FEM, (N + 4)/12 graph elements are associated. 3. For each element with N nodes, 3N ! 6 members should be considered in the interface graph. Thus, 2N ! 10 ¼ (3 N ! 6) ! (N + 4) diagonal members should be added. Þ ð2N!10Þ Therefore, in the interface of two elements (common side), 4 & ðNþ4 12 þ 6 multiple members are present. The member numbering of the interface graph should be performed according to the numbering of the FEM, taking into account the primary nodal numbering of a consider element in the model. Thus, for each hexahedron element 3N ! 6 edges of the interface graph will be numbered sequentially according to the patterns which were illustrated in Fig. 7.43. In this figure, quartic element numbering was neglected and just the element forces are displayed. Numbering of this type can be easily obtained according to the pattern of other elements.

7.4.2.2

Natural Associate Graph

The natural associate graph represented by NAG(FEM) is constructed by the following rules: 1. Nodes of the NAG(FEM) correspond to the elements of FEM. Þ 2. For each pair of elements in FEM having ð2Nþ8 common nodes (N ¼ the number 6 of nodes of an element), one member is added between the corresponding two nodes in NAG(FEM).

324

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.43 Nodal numbering and element forces for hexahedron elements; (a) linear, (b) quadratic, (c) cubic and (d) quartic

NAG(FEM) can be constructed using the following procedure: One of the preliminary steps in FEA is defining the elements with their connected nodes. In this way the element connectivity matrix is constructed which contains the elementnode incidence relationships. In the process of constructing the element connectivity matrix, another matrix which contains node-element incidence properties can be formed. This matrix is named the node connectivity matrix. Now using the element connectivity and the node connectivity matrices leads to an algorithm with complexity O(n) for an efficient generation of NAG. In order to recognize the adjacent elements to the nth element which have Þ common ð2Nþ8 nodes or one common face, first the connected nodes to the nth 6 element are identified from the element connectivity matrix. In the subsequent step using the node connectivity matrix, elements which have at least one common node with the nth element are identified. Now it is convenient to seek for the adjacent elements in this reduced search space. A FEM and its corresponding NAG are illustrated in Fig. 7.44.

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

325

Fig. 7.44 Finite element model (black part) with natural associate graph (blue part)

7.4.3

Negative Incidence Number

Negative incidence number (NIN) is necessary for each node of NAG(FEM). This number can be found as following: After generation of natural associate graph of the FEM, use an efficient method for its nodal numbering. A typical edge of the graph connects smaller number to the node with higher number. Negative incidence number of each node is the number of its adjacent nodes with smaller nodal number. Except the node numbered as 1, all the other nodes have one, two or three negatively incident edges defined as the negative incidence number of the node. Owing to the importance of these numbers in recognizing the types of SESs, the negative incidence numbers of the nodes of the graph should carefully be calculated. In Fig. 7.45, a hexahedron FEM with element numbering, its corresponding associate graph and negative incidence number of nodes are shown. The nodes should be numbered such that the incidence numbers do not become large. Any simple nodal ordering will lead to a logical ordering.

7.4.4

Pattern Corresponding to Self-Equilibrating Systems

Considering Fig. 7.43, in order to find the patterns corresponding to the selfequilibrating systems, a hexahedron element is simulated as a spatial truss formed as the 1-skeleton of the hexahedron element together with some diagonal members. This is possible since the independent element forces are applied in the nodes and are along the edges of the element. In Fig. 7.46, an IG(FEM) with four quadratic elements is shown which is simulated as a spatial truss containing some multiple members. The statical indeterminacy of a spatial truss with m members and n nodes is given as γ(S) ¼ m ! 3n + 6; therefore, the degree of statical indeterminacy (DSI)

326

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.45 Finite element model (black part) with natural associate graph (blue part); (a) nodal numbering of NAG; (b) negative incidence numbers of NAG

Fig. 7.46 An IG(FEM) with four quadratic elements is shown which is simulated as a spatial truss

of the entire model supported in a statically determinate manner can be calculated with the same relationship as DSI ¼ ð3N ! 6Þ & M ! 3n þ 6

ð7:26Þ

where M is the total number of finite elements, N is the number of nodes of one element and n is the total number of nodes of the FEM. With the above simulation, the patterns of the self-equilibrating systems can be identified as follows:

7.4.4.1

Type I Self-Equilibrating Systems

For each k multiple member in equivalent truss model of FEM, there are k unknown forces and one equilibrium equation in the member’s direction. Thus DSI of the

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

327

substructure is equal to k ! 1 and k ! 1 self-equilibrating systems can be generated on each k multiple member of interface graph of the FEM. In this way, first each k multiple members are arranged in ascending order as (m1, m2, m3,. . ., mk!1, mk). where (m1 < m2 < m3 < . . . < mk!1 < mk). Each selection of two members from this list is valid to construct a Type I self-equilibrating system, but in order to achieve a better bandwidth reduction; selection of adjacent members from the defined list is preferable. Therefore k ! 1 duplicate members are selected as (m1, m2), (m2, m3),. . ., (mk!1, mk). Each pair (mi, mj) with i < j represents the numbers of corresponding self-equilibrating system. The member with bigger number is selected as the generator of the current SES and also as a redundant force. The null vectors corresponding to the Type I SESs have two non-zero entries in rows i and j equal to !1 and 1, respectively. For FEMs with hexahedron elements, more than 75 % of the total self-stress systems are of Type I. Thus, a large percent of the minimal null vectors can be formed only by the determination of member numbers of these pairs. It should be noted that in the process of the formation of the interface graph, these pairs and their numbers can simply be identified.

7.4.4.2

Type II Self-Equilibrating Systems

There are other types of self-equilibrating systems which are extracted from two adjacent elements of FEM. In other words, for two adjacent elements with N nodes, the DSI can be calculated as: DSI ¼ ð3N ! 6Þ & M ! 3n þ 6 ! 2N þ 8 " þ6¼ N!2 ) DSI ¼ ð3N ! 6Þ & 2 ! 3 & 2N ! 6 2N ! 10 6 |ﬄﬄﬄﬄ{zﬄﬄﬄﬄ}

diagonal members of one side

þ

2N þ 8 6 |ﬄﬄﬄ{zﬄﬄﬄ}

4N ! 2 6 |ﬄﬄﬄ{zﬄﬄﬄ}

¼

other members of one side

ð7:27Þ self-equilibrating

number of members of common side of two adjacent elements systems were generated as Type I systems. Thus the number of remaining selfequilibrating systems is Type II ¼ ðN ! 2Þ !

#

4N ! 2 6

$

¼

2N ! 10 6

ð7:28Þ

In other words, 2N!10 SESs should be extracted from two adjacent elements. For 6 example, the remaining subgraphs for two adjacent quadratic elements are shown in

328

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.47 Subgraph corresponding to SESs of Type II

Fig. 7.47. After deleting the generators corresponding to Type I SESs, the null vectors should be calculated from the remaining subgraph. These null vectors can easily be generated on the corresponding sub-structure utilizing an algebraic method. Apart from the aforementioned about generating the SESs of Type II, if there is at least a negative incidence number higher than one in a FEM, another important point should be considered which is explained below: Some of the calculated SESs of Type II are not independent of the others. For example, for a FEM with four quadratic elements M0 is equal to four, where M0 is the number of members of the associate graph of the model. The number of SESs of Type II is 18 instead of 20 ¼ 4 & 5. In other words, two SESs are dependent and should not be selected. For determining the independent SESs, an appropriate approach is proposed. In this approach, independent SESs will be recognized utilizing negative incidence number of elements. The SESs of Type II are extracted from two adjacent elements in a FEM which are the same as members of NAG(FEM). If a member of NAG(FEM) connects two elements Mi and Mj where i < j, the number of independent SESs of Type II which can be extracted from the subgraph corresponding to these two adjacent elements is equal to: Type II ¼

# $ modðN; 8Þ α ! NINj & 4

ð7:29Þ

Where NINj is the negative incidence number of jth element and α is 1,6,8 and 13 for linear, quadratic, cubic and quartic elements, respectively. For linear element, a SES Type II can be generated on each two adjacent elements on a FEM [6]. For other types of element, after deleting the generators corresponding to Type I SESs, the main diagonal member (longer diagonal member) of the jth element which is located in the common side with other adjacent elements with smaller number than j, the null vectors should be calculated from the remaining subgraph. The most important point in Type II self-equilibrating systems is to select an appropriate generator. In fact by eliminating these generators and the generators

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

329

corresponding to Type I SESs from IG(FEM), the sub-structure of Type III SESs and the primary structure of the IG(FEM) must remain stable.

7.4.4.3

Type III Self-Equilibrating Systems

Sub-structures which are topologically identical to the minimal cycles of the natural associate graph of FEM contain some Type I, Type II and one or six Type III selfequilibrating systems. (a) Type I minimal cycles of NAG(FEM) These minimal cycles of the natural associate graph of the FEM pass through 2Nþ8

Nþ4 four elements which have 46 þ 1 ¼ Nþ16 common nodes or Nþ16 12 12 ! 1 ¼ 12 common edges. Corresponding interface graph of these elements have n nodes and m edges for a FEM with N node elements.

m ¼ 4 & ð3N ! 6Þ # $ 2N þ 8 N þ 16 11N ! 16 ¼ n ¼ 4N ! 4 & þ 6 12 4

ð7:30Þ ð7:31Þ

Subsequently, the DSI of the interface graph is DSI ¼ m ! 3n þ 6 ) DSI ¼ 4 & ð3N ! 6Þ ! 3 & ¼

15N !6 4

11N ! 16 þ6 4 ð7:32Þ

2Nþ8 31N!20 The 4 & 4N!2 SESs are Type I and there are 4, 18, 32 and 46 SESs 6 ! 24 ¼ 12 of Type II for linear, quadratic, cubic and quartic elements, respectively. 8 N ¼ 8 ) DSI ! ðType I & IIÞ ¼ 1 > > < N ¼ 20 ) DSI ! ðType I & IIÞ ¼ 1 ð7:33Þ DSI ! ðTypeI&IIÞ ¼ N ¼ 32 ) DSI ! ðType I & IIÞ ¼ 1 > > : N ¼ 44 ) DSI ! ðType I & IIÞ ¼ 1

Therefore, one independent SES should be extracted. This SES with thirteen members can be formed for any types of hexahedron elements around the common edge as is indicated bold in Fig. 7.48. It should be noted that in a FEM, all of the SESs of Type III which are extracted from any four elements around one common edge, are not independent with all previous selected SESs. Independent ones should be selected utilizing NINs of elements. For this purpose, NIN of four elements with common edge should not be more than 2. In Fig. 7.45, a FEM with eight elements is shown. The independent SESs of Type III should be selected utilizing these three sets of elements:

330

7 Optimal Force Method for FEMS: Higher Order Elements

Fig. 7.48 The SES of Type III corresponding to the common edge of four elements

Fig. 7.49 A FEM with an opening and its NAG

f E1 E2 E3 E4 g, f E1 E2 E5 E6 g and f E1 E3 E5 E7 g: In other words, E8 should not be in selections, because the NIN of E8 is 3. (b) Type II minimal cycles of NAG(FEM) Each minimal cycle that surrounds an opening is called the Type II minimal cycle (Fig. 7.49). Such a cycle passes through M0 (M0 ' 8) finite elements and its ! " 0 corresponding interface graph has N ! 2Nþ8 & M nodes and M0 & (3N ! 6) 6 members. The DSI of subgraph is # $ 2N þ 8 0 0 DSI ¼ M & ð3N ! 6Þ ! 3 & N ! & M þ 6 ) DSI 6 0

¼ M & ð N ! 2Þ þ 6 0

and M &

4N ! 2 6 |ﬄﬄﬄ{zﬄﬄﬄ}

number of members of common side of two adjacent elements (Eq. 7.28) can be extracted.

ð7:34Þ 0

SESs of Type I and M &

!2N!10" 6

SESs of Type II

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

331

0

DSI ! ðType I&IIÞ ¼ M & ðN ! 2Þ þ 6 % # $& 4N ! 2 2N ! 10 0 0 þM & ! M & 6 6 ¼6

ð7:35Þ

Therefore, each Type II minimal cycle corresponds to six null vectors which are calculated utilizing an algebraic method.

7.4.5

Selection of Generators for SESs of Type II and Type III

The most important point in Type II and Type III self-equilibrating systems is to select appropriate generators, because by eliminating these generators from IG (FEM), the sub-structure of primary structure of the IG(FEM) must remain stable. To achieve this, the following rule for appropriate selection of generators of Type II SESs is suggested. For quadratic hexahedron element the generators of SESs Type II and III are illustrated in Tables 7.15 and 7.16, respectively. Directions 1, 2 and 3 are shown in Fig. 7.50. In these Tables, Nα,β indicates the βth node of element α and NIN(d )j is the negative incidence number of element j in direction d. In other words, NIN(d)j is one if j has an adjacent element i where i < j in direction d. It should be noted that the generators corresponding to Type I were chosen previously. In addition, the generators corresponding to an opening are the last six non-zero entries of its columns which are not common with the previously selected generators. For other element types, generators corresponding to Type II and Type III can be obtained following aforementioned patterns. Algorithm. Step 1: Generate the associate graph of the FEM and use an efficient method for its nodal numbering [4]. It is obvious that good numbering of this graph corresponds to good numbering of elements of the FEM. This numbering leads to a banded adjacency matrix of the graph and correspondingly to a banded flexibility matrix. Since numbering the members of the interface graphs corresponds to the element numbering of the finite elements, such a numbering is the only parameter for controlling the bandwidth of the flexibility matrix. Negative incidence number of the NAG(FEM) should be calculated in this step. Step 2: Set up the equilibrium matrix of the FEM. Step 3: Generate the interface graph and perform its numbering. The numbering of this graph should be performed according to the element numbering of the considered FEM. After this numbering the interface graph can easily be formed and its members can be numbered. Step 4: Find the Type I self-equilibrating systems. All multiple members of the interface graph are identified and the values !1 and 1 are assigned to appropriate

332

7 Optimal Force Method for FEMS: Higher Order Elements

Table 7.15 Generators of Type II SESs in directions 1, 2 and 3 (k < i < j)

rows (corresponding to the member numbers) and the corresponding minimal null vectors are created. Þ Step 5: Find the Type II self-equilibrating systems. The ð2N!10 SESs of Type II should 6 be extracted from two adjacent elements and independent ones should be selected among these SESs utilizing the approach which is explained in Sect. 5.2. Calculate

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

333

Table 7.16 Generators of Type III SESs in planes 1-2, 2-3 and 1-3 (i < j < k < l )

Fig. 7.50 Typical view of an element with corner nodes and determining directions 1, 2 and 3

the corresponding null vectors from the relevant equilibrium sub-matrix in this step. Step 6: Find the Type III self-equilibrating systems. For each minimal cycle of Nþ4 natural associate graph of FEM with four members (Nþ16 12 common nodes or 12

334

7 Optimal Force Method for FEMS: Higher Order Elements

common edges and NINj < 3 for j ¼ 1, 2, 3, 4), one SES and with eight or more members (opening) six SESs should be extracted. Calculate the corresponding null vectors from the relevant equilibrium sub-matrix. Step 7: Order the null vectors. At this step the constructed null vectors should be ordered such that their last non-zero entries form a list with an ascending order.

7.4.6

Numerical Examples

In this section two FEMs are considered, which are assumed to be supported in statically indeterminate fashion. The translations of each support node are fixed in all three directions. The equilibrium matrices are formed. Null bases and the flexibility matrices are constructed and the required computational times, and the condition numbers are calculated. In all the following examples, nz represents the number of non-zero entries and λmax/λmin is the ratio of the extreme eigenvalues taken as the condition number of a matrix. The comparison between present algorithm and algebraic force method will be shown in the conclusion section. Example 1. An arch wall structure which is supported in a statically indeterminate fashion is illustrated in Fig. 7.51. This structure is discretized using 20-node hexahedron elements. The properties of the model are as follows: Poisson’s ratio ¼ 0.2; Elastic modulus E ¼ 2E + 10 N/m2; Density ρ ¼ 2,400 kg/m3; Internal radius ¼ 8.0 m; Number of 20-node hexahedron elements ¼ 80 Number of nodes ¼ 557 DSIInternal ¼ 80 & 54 ! 3 & 557 + 6 ¼ 2655; DSIExternal ¼ 15 & 3 ! 6 ¼ 39 Number of Type I self-equilibrating systems ¼ 1,996 (75.0 %) Number of Type II self-equilibrating systems ¼ 620 (23.3 %) Number of Type III self-equilibrating systems ¼ 39 (1.7 %)

The pattern of the equilibrium matrix is shown in Fig. 7.52. The sparsity of the final null basis obtained by the present method is approximately 22.41 % of the LU method as depicted in Fig. 7.53. The flexibility matrix G is also well-structured as shown in Figs. 7.54 and 7.55. It should be added that the total DSI for the force method of this structure is 2,655 + 39 ¼ 2,694, while the DOFs for the displacement method is 557 & 3 ¼ 1,671, indicating less number of equations for the latter approach. However, since in the force method nearly 75 % of the null vectors are found by simple graph theoretical approach, one should compare 1,671 with approximately 25 % of 2,694 ¼ 673, showing the superiority of the force method. Example 2. A dome with an opening which is supported in a statically indeterminate fashion is illustrated in Fig. 7.56. The internal and external diameters are 5.00

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

335

Fig. 7.51 An arch wall structure which is supported in a statically indeterminate fashion 0 200 400 600 800 1000 1200 1400 1600 0

500

1000

1500

2000

2500

3000

3500

4000

nz = 17925

Fig. 7.52 Pattern of the equilibrium matrix for Example 1

and 5.50 m, respectively. This structure is discretized using 20-node hexahedron elements. The properties of the model are as follows: Poisson’s ratio ¼ 0.2; Elastic modulus E ¼ 2E + 10 N/m2; Density ρ ¼ 2,400 kg/m3; Number of 20-node hexahedron elements ¼ 84, Number of nodes ¼ 648 DSIInternal ¼ 84 & 54 ! 3 & 648 + 6 ¼ 2598; DSIExternal ¼ 60 & 3 ! 6 ¼ 174 Number of Type II self-equilibrating systems ¼ 636 (24.4 %) Number of Type III self-equilibrating systems ¼ 72 (four elements with common edges) + 6 (an opening) ¼ 78 (3.0 %)

336 Fig. 7.53 Patterns and number of non-zero entries of the null bases of Example 1: (a) present algorithm, (b) LU factorization

7 Optimal Force Method for FEMS: Higher Order Elements

a0

b0

500

500

1000

1000

1500

1500

2000

2000

2500

2500

3000

3000

3500

3500

4000

4000

0

500

1000

1500

2000

0

2500

Fig. 7.54 Patterns of the flexibility matrix G ¼ Bt1 FmB1 for Example 1: (a) present algorithm, (b) LU factorization

a

b

0

500 1000 1500 2000 2500 0

500

1000

1500

2000

2500

nz = 398466

Fig. 7.55 Patterns of the flexibility matrix G ¼ Bt1 FmB1 for Example 1: (a) present algorithm, (b) LU factorization

a

b

0

500 1000 1500 2000 2500 0

500

1000

1500

500

1000

1500

2000

nz = 159318

nz = 35709

2000

nz = 398466

2500

2500

7.4 Efficient Finite Element Analysis Using Graph-Theoretical Force Method:. . .

337

Fig. 7.56 A dome with an opening: 3D view, bottom view and a section 0 200 400 600 800 1000 1200 1400 1600 1800 0

500

1000

1500

2000

2500

3000

3500

4000

4500

nz = 24544

Fig. 7.57 Pattern of the equilibrium matrix for Example 2

The pattern of the equilibrium matrix is shown in Fig. 7.57. The sparsity of the final null basis obtained by the present method is approximately 28.1 % of the LU approach as depicted in Fig. 7.16. The flexibility matrix is also well-structured as shown in Figs. 7.58 and 7.59.

338

7 Optimal Force Method for FEMS: Higher Order Elements

a

b

0

500

1000

1500

2000

2500

3000

3500

4000

4500 0

500

1000

1500

2000

2500

nz = 315547

Fig. 7.58 Patterns and the number of non-zero entries of the null bases of Example 2: (a) present algorithm, (b) LU factorization

a

b

500

500

1000

1000

1500

1500

2000

2000

2500

2500

0

0

0

500 1000 1500 2000 2500 nz = 1151626

0

500 1000 1500 2000 2500 nz = 3743075

Fig. 7.59 Patterns of flexibility matrix G ¼ Bt1 FmB1 for Example 2: (a) present algorithm, (b) LU factorization

Finally, it is hoped that the extension of elements for the force method continues similar to those of the displacement method, to enable these dual approaches to be efficiently utilized in the analysis of large-scale finite element models.

References 1. Kaveh A, Tolou Kian MJ (2013) Efficient finite element analysis of models comprised of higher order triangular elements. Acta Mech 224(9):1957–1975

References

339

2. Kaveh A, Massoudi MS, Massoudi MJ (2014) Efficient finite element analysis using graphtheoretical force method; rectangular plane stress and plane strain serendipity family elements. Periodica Polytechnica in print 3. Kaveh A, Massoudi MS, Massoudi MJ (2013) Efficient finite element analysis using graphtheoretical force method; hexahedron elements. Comput Struct 128:175–188 4. Kaveh A, Roosta GR (1998) Comparative study of finite element nodal ordering methods. Eng Struct 20(1–2):86–96 5. Przemieniecki JS (1968) Theory of matrix structural analysis. McGraw-Hill, New York 6. Kaveh A, Koohestani K (2007) An efficient graph theoretical method for plate bending finite element analysis via force method. Eng Comput 24(7):679–698

Chapter 8

Decomposition for Parallel Computing: Graph Theory Methods

8.1

Introduction

In the last decade, parallel processing has come to be widely used in the analysis of large-scale structures. This chapter is devoted to the optimal decomposition of structural models using graph theory approaches. First, efficient graph theory methods are presented for the optimal decomposition of space structures. The subdomaining approaches are then presented for partitioning of finite element models. A substructuring technique for the force method of structural analysis is discussed. Several partitioning algorithms are developed for solution of multi-member systems, which can be categorised as graph theory methods and algebraic graph theory approaches. For the graph theory method, Farhat [1] proposed an automatic finite element domain decomposer, which is based on a Greedy type algorithm and seeks to decompose an FEM into balanced domains, sharing a minimum number of common nodal points. In order to avoid domain splitting, Al-Nasra and Nguyen [2] incorporated geometrical information of the FEM into an automatic decomposition algorithm similar to the one proposed by Farhat [1]. The Sparpak uses nested dissection due to George and Liu [3], which uses a level tree for dissecting a model. Kaveh and Roosta [4] employed different expansion processes for decomposing space structures and finite element meshes. Applications of the methods of this chapter are by no means confined to structural systems; these methods can equally be applied to other large-scale problems like the analysis of hydraulic systems and electrical networks.

A. Kaveh, Computational Structural Analysis and Finite Element Methods, 341 DOI 10.1007/978-3-319-02964-1_8, © Springer International Publishing Switzerland 2014

342

8.2 8.2.1

8 Decomposition for Parallel Computing: Graph Theory Methods

Earlier Works on Partitioning Nested Dissection

The term “nested dissection” was introduced by George [5], following a suggestion of Birkhoff. Its roots lie in finite element substructuring, and it is closely related to the tearing and interconnecting method of Kron [6]. The central concept for nested dissection is the removal of a set of nodes from the graph (separator) of a symmetric matrix (or the model of a structure) that leaves the remaining graph in two or more disconnected parts. In nested dissection, these parts are themselves further divided by the removal of sets of nodes, with the dissection nested to any depth. If the variables of each subgraph are grouped together, by ordering the nodes of their nodes contiguously followed by numbering the nodes, in the separator, then the following block form will be obtained: 2

A11 4 0 A31

0 A22 A32

3 A13 A23 5: A33

ð8:1Þ

The blocks A11 and A22 may themselves be ordered to such a form by using dissection sets. This way every level defines a nested dissection order. The significance of the above partitioning of the matrix is twofold: first, the zero blocks are preserved in the factorisation, thereby limiting fill; second, factorisation of the matrices A11 and A22 can proceed independently, thereby enabling parallel execution on separate processors. When a complicated design is assembled from simpler substructures, it makes sense to exploit these natural substructures. The resulting ordering is likely to be good, simply because, when each variable is eliminated, only the other variables of its substructures are involved.

8.2.2

A Modified Level-Tree Separator Algorithm

The separator routine in Sparspak, FNDSEP, finds a pseudo-peripheral node in the graph and generates a level structure from it. It then chooses the median level in the level structure as the node separator. However, this choice may separate the graph into widely disparate parts. In a modification made by Pothen et al. [7], the node separator is selected to the smallest level k, such that the first k levels together contain more than half of the nodes. A node separator is obtained by removing from the nodes in level k those nodes that are not adjacent to any node in level k # 1, and therefore these are added to the part containing the nodes in the first k # 1 levels. The other part has nodes in levels k + 1 and higher. Although such a method is

8.3 Substructuring for Parallel Analysis of Skeletal Structures

343

simple; however, the spectral bisection method computes a smaller node separator than the Sparspak algorithm.

8.3

8.3.1

Substructuring for Parallel Analysis of Skeletal Structures Introduction

In many engineering applications, particularly in the analysis and design of large systems, it is convenient to allocate the design of certain components (substructures) to individual design groups. The study of each substructure is carried out more or less independently, and the dependencies between the substructures are resolved after the study of individual substructures is completed. The dependencies among the components may of course require redesign of some of the substructures, so the above procedure may be iterated several times. As an example, suppose for a structural model, we choose a set of nodes I and their incident members which, if removed, disconnect it into two substructures. If the variables associated with each substructure are numbered consecutively, followed by the variables associated with I, then the partitioning of the stiffness matrix A will be as that of Eq. 8.1. The Cholesky factor L of A, correspondingly, will be partitioned as, 2

L11 4 0 L¼ t W13

0 L22 t W23

3 0 0 5, L33

ð8:2Þ

where t t t t A11 ¼ L11 L11 , A22 ¼ L22 L22 , W13 ¼ L11 A13 , W23 ¼ L22 L23 ,

and t t t #1 L33 ¼ A33 # A13 A#1 L33 11 A23 # A23 A22 A23 :

ð8:3Þ

Therefore, A11 and A22 correspond to each substructure, and the matrices A13 and A23 represent the “glue” which relates the substructures through the nodes of I. Since the factors of A11 and A22 are independent, they can be computed in either order, or in parallel if two processors are available. Finally, in some design applications, several substructures may be identical, for example, have the same configuration and properties, and each substructure may be regarded as a superelement, which is constructed once and used repeatedly in the design of several structures. In the above example, A11 and A22 could be identical.

344

8.3.2

8 Decomposition for Parallel Computing: Graph Theory Methods

Substructuring Displacement Method

For the analysis of skeletal structures and for the finite element method, using the displacement approach, an appropriate formulation such as the Galerkian method reduces to solving the following matrix equation, Kv ¼ p,

ð8:4Þ

where K is the global stiffness matrix, and v and p are the nodal displacement and nodal force vectors, respectively. To distribute the computation after decomposing the model into q subdomains, each subdomain can be treated as a super element and mapped onto the processors. Various methods for decomposition will be presented in this chapter. The global stiffness matrix and nodal force vector are equivalent to the assembly of its components for q subdomains: K¼

q X j¼1

kj and p ¼

q X

pj :

j¼1

ð8:5Þ

Equation 8.4 can be written in the following partitioned form: !

Kii Kbi

Kib Kbb

"!

vi vb

"

¼

!

" pi : pb

ð8:6Þ

In the above equation, a boundary node is defined as a node which is part of more than one subdomain and degrees of freedom at the boundary nodes are treated as boundary degrees of freedom. The vectors vi and vb are displacements, and pi and pb are forces, corresponding to internal and boundary nodes, respectively. Each subdomain requires solution of an equation, similar to Eq. 8.4: ½k&j ½d&j ¼ ½p&j :

ð8:7Þ

For the full domain, Eq. 8.7 can be written in partitioned form as: !

kii kbi

kib kbb

"!

vi vb

"

¼

!

" pi : pb

ð8:8Þ

Using static condensation for eliminating the interior degrees of freedom of each subdomain, the effective stiffnesses and load vectors on the interface boundaries are obtained. For internal nodes we have, ½kii &½vi & þ ½kib &½vb & ¼ ½pi &, or

ð8:9Þ

8.3 Substructuring for Parallel Analysis of Skeletal Structures

½vi & ¼ ½kii f½pi & # ½kib &½vb &g:

345

ð8:10Þ

Substituting in Eq. 8.8 leads to, ½kbi &½kii f½pi & # ½kib &½vb &g þ ½kbb &½vb & ¼ ½pb &,

ð8:11Þ

½k( &½vb & ¼ ½pb & # ½kib ½kii &½pi &,

ð8:12Þ

n o ½k( & ¼ ½kbb & # ½kbi &½kii ½kib & ,

ð8:13Þ

or

where

is the condensed super element stiffness matrix and ½p( & ¼ ½pb & # ½kbi &½kii ½pi &,

ð8:14Þ

is the modified load vector. A summation of the interface conditions for the subdomains leads to the formation of the global interface stiffness matrix K* and the global interface load vector p* as follows: K( ¼

q X j¼1

k(j and p( ¼

q X j¼1

p(j :

ð8:15Þ

K is symmetric and positive definite, and K* has the same properties. The following interface system can now be solved: ½K( &½vb & ¼ ½p( &:

ð8:16Þ

Once vb is found, the internal degrees of freedom for a subdomain can be evaluated employing Eq. 8.10. A natural route to parallelism now is to provide it through domain decomposition by distributing the substructures onto the processors available. Several approaches can be used to solve Eq. 8.4. In the following, three broad classifications are briefly discussed:

346

8 Decomposition for Parallel Computing: Graph Theory Methods

8.3.3

Methods of Substructuring

8.3.3.1

Direct Methods

A substructuring method can be used to obtain the condensed stiffness matrix on each subdomain in parallel on the different processors. In order to create matrix K*, it is necessary to condense the stiffness matrix of each substructure (subdomain), i.e. from Eq. 8.13 the product [kbi][kii]# 1[kib] should be calculated. The explicit formation of [kii]# 1[kib] requires NBDOF triangular system resolutions, where MbDOF is the number of subdomain boundary degrees of freedom (DOF). This step can be considered as follows: Each internal DOF makes its contribution to the stiffness of each boundary DOF, such that the behaviour of the condensed boundary is equivalent to the behaviour of the entire domain. This step can be executed step by step, so that only the internal DOF connected to the boundary DOF updates the boundary stiffness matrix. This requires the internal DOF to appear at the bottom of the internal stiffness matrix kii, so that they are modified by the elimination of all other internal DOF. A frontal method can be used, which has the advantage of allowing very flexible strategies concerning the sequence of elimination of equations. When this method is applied to subdomain condensation, it is necessary to assemble the boundary DOF in the frontal matrix, and to retain them until all the internal DOF have been eliminated. At the end of the frontal elimination process, the frontal matrix is exactly the condensed matrix [kbi][kii]# 1[kib]. The interface system of equations is then solved employing a direct approach (e.g. skyline method) on a single machine. Although the direct methods are simple and terminate in a fixed number of steps, the interface solution dominates the overall computational cost when the interface system is large, thus limiting the overall efficiency. In such a case, however, a distributed algorithm can be used for factorisation of the direct method to overcome this difficulty.

8.3.3.2

Iterative Methods

A different method to avoid the explicit inverse of kii in Eq. 8.13 is the use of an iterative approach. Among the iterative solutions, the conjugate gradient method is a promising candidate, because of its inherent parallelism and its rate of convergence. The theory of the conjugate gradient method is well known [8]. One iteration of this method for solving a system of equations Kv ¼ p is given as: fug ¼ ½K&ff g, t

t

α ¼ frg frg=ff g fug,

ð8:17aÞ ð8:17bÞ

8.3 Substructuring for Parallel Analysis of Skeletal Structures

fvnew g ¼ fvg þ αff g,

347

ð8:17cÞ

frnew g ¼ frg þ αfug,

ð8:17dÞ

ff new g ¼ frnew g þ λff g:

ð8:17fÞ

t

t

λ ¼ frnew g frnew g=frg frg,

ð8:17eÞ

Before each iteration, the vectors {v}, {f} and {r} are set to {vnew}, {fnew} and {rnew}, respectively. The vectors are initialised as, frg ¼ fpg # ½K&fv0 g,

ð8:18aÞ

ff g ¼ frg,

ð8:18bÞ

And

where {v0} is usually taken as null, unless some approximation to the solution is known. Iteration is terminated when the residual is small. One criterion for handling the iteration is, krk=kpk < ε,

ð8:19Þ

where ε is the tolerance specified for the problem. In structural analysis, the vector r is the potential gradient and is identical to the residual force vector, (p # Kv) in the linear case. The vector f is the gradient direction to generate the displacement vector v. For discussion and further details, the reader may refer to Law [9]. Preconditioned Conjugate Gradient (PCG) methods form a large class of the many iterative methods that have been suggested to reduce the cost of forming condensed stiffness matrices. A saving in total time may be achieved, since the predominant matrix-vector product at each iteration is computed in parallel. For further detailed discussion, the interested reader may refer to Keyes and Gropp [10].

8.3.3.3

Hybrid Methods

These methods use a combination of the direct and iterative methods. For instance, the components of the condensed matrix k* may be obtained for the substructures using the direct method, and the resulting interface can be solved using an iterative approach. A comparative study of direct, iterative and hybrid methods is made by Chadha and Baugh [11]. In the following sections, algorithms are presented for partitioning of the nodes of structural graph models, which can be incorporated in any program available for

348

8 Decomposition for Parallel Computing: Graph Theory Methods

the analysis of skeletal structures. Domain decomposition algorithms are presented in Chap. 8.

8.3.4

Main Algorithm for Substructuring

Let S be the graph model of a structure. The following algorithm decomposes S into q subgraphs with equal or near equal number of nodes (support nodes are not counted) having the least number of interface nodes: Step 1: Delete all the support nodes with their incident members, and denote the remaining subgraphs by Sr. Step 2: Determine the distance between each pair of nodes of Sr, and evaluate the eccentricities of its nodes. Step 3: Sort the remaining nodes (RN) in ascending order of their eccentricities. Step 4: Select the first node of RN as the representative node of the subgraph S1 to be determined and find a second node as the representative node of subgraph S2 with a maximum distance from S1. Step 5: Find the third representative node with the maximum least distance from S1 and S2, and denote it with S3. Step 6: Subsequently, select a representative node of subgraph Sk for which the least distance from S1, S2, . . . ,Sk # 1 is maximum. Repeat this process until q representative nodes of the subgraphs to be selected are found. Step 7: For each subgraph Sj (j ¼ 1, . . . ,q), add an unselected node ni of RN, if it is adjacent only to Sj and its least distance from all nodes of other subgraphs is maximum. Step 8: Continue the process of Step 7, without the restriction of transforming one node to each subgraph Sj, until no further node can be transferred. The remaining nodes in RN are interface nodes. Step 9: Transfer the support nodes to the nearest subgraph. Once the nodes for each subgraph Sj are found, the incidence members can easily be specified. The algorithm is recursively applied to the selected substructures, decomposing each substructure into smaller ones, resulting in a further refinement.

8.3.5

Examples

Example 1. A double-layer grid supported at four corner nodes is considered and partitioned into q ¼ 2, 4 substructures, Fig. 8.1. The corresponding node adjacency matrices (pattern of their stiffness matrices) are illustrated in Fig. 8.2a, b. For the case q ¼ 2, the selected substructures are further refined with q0 ¼ 2 and 3, and the corresponding matrices are shown in Fig. 8.3a, b.

8.3 Substructuring for Parallel Analysis of Skeletal Structures

349

Fig. 8.1 A double-layer grid S

Fig. 8.2 Patterns of the adjacency matrices for different values of q. (a) q ¼ 2. (b) q ¼ 4

Example 2. A dome-type space structure supported at six nodes is considered and partitioned into q ¼ 2, 3, 4 and 5 substructures, Fig. 8.4. The corresponding node adjacency matrices are illustrated in Fig. 8.5a–d. For the case q ¼ 2, the selected substructures are further refined with q0 ¼ 2 and 3, and the corresponding matrices are shown in Fig. 8.6a, b. Once the subgraphs and the interface nodes are specified, ordering the nodes of each subgraph reduces the bandwidth of each block, and appropriate numbering of the interface nodes, results in banded bordered for the entire matrix.

350

8 Decomposition for Parallel Computing: Graph Theory Methods

Fig. 8.3 Patterns of the adjacency matrices for q ¼ 2 and q0 ¼ 2 and 3. (a) q ¼ 2 and q0 ¼ 2. (b) q ¼ 2 and q0 ¼ 3

Fig. 8.4 A dome-type space structure

8.3.6

Simplified Algorithm for Substructuring

In the following, a simplified algorithm is presented which requires less storage and computer time than the main algorithm, at the expense of selecting subgraphs with a slightly higher number of interface nodes for some structural models. In this approach, the number of distances to be considered and compared for finding the nodes of substructures is far less than when the main algorithm is used, where the distances between each pair of nodes of S are required. This simplified algorithm consists of the following steps: Step 1: Form an SRT rooted from an arbitrary node, in order to find a representative node of S1 with maximum distance from the root. The selected node is also denoted by S1.

8.3 Substructuring for Parallel Analysis of Skeletal Structures

351

Fig. 8.5 Patterns of the adjacency matrices for different values of q. (a) q ¼ 2. (b) q ¼ 3. (c) q ¼ 4. (d) q ¼ 5

Fig. 8.6 Patterns of the adjacency matrices for q ¼ 2 and different values of q0 . (a) q ¼ 2 and q0 ¼ 2. (b) q ¼ 2 and q0 ¼ 3

352

8 Decomposition for Parallel Computing: Graph Theory Methods

Step 2: Form an SRT rooted from S1, to calculate the distance between each node of S and S1, and find the representative node S2 in a maximum distance from S1. Step 3: Form an SRT rooted from S2, to calculate the distance between each node of S and S2 and find the representative node S3 in a maximum least distance from the selected nodes. Repeat this process until q representative nodes S1, S2, . . . , Sq, forming a transversal , are selected. Step 4: For each subgraph Si, find a node adjacent to the previously formed Si only, with maximum least distance from other representative nodes, in turn. Step 5: Continue the process of Step 4, without the restriction of transforming one node to each subgraph Si, until no further node can be transferred.

8.3.7

Greedy Type Algorithm

In this algorithm, the weight of a node is taken as the number of elements incident with that node. The interior boundary of a subdomain Di is defined as the subset of its boundary that will interface with another subdomain Di. The total number of elements in a given mesh is denoted by M(FEM). Step 1: Start with a node and add incident elements one by one having the least current weight. The current weight is taken as the number of unselected elements at that stage incident with that node. Continue this process until M(FEM)/q elements are selected as D1. Step 2: Select an interior node of D1, and repeat Step 1 to form D2. Step k: Repeat Step 2 for k ¼ 3, 4, . . . , q with an interior node of Dk # 1 and form subdomain Dk. This process is a Greedy type algorithm, which selects one element of minimal current weight at a time and completes a domain when N(FEM)/q (+1 if remainder 6¼ 0) elements are selected for the formation of that subdomain. The current weight of an element is updated when an incident element is joined to the expanding subdomain.

8.4

Domain Decomposition for Finite Element Analysis

In this section, efficient algorithms are developed for automatic partitioning of unstructured meshes for the parallel solution of problems in the finite element method. These algorithms partitions a domain into subdomains with approximately equal loads and good aspect ratios, while the interface nodes are confined to the smallest possible. Examples are included to illustrate the performance and efficiency of the presented algorithms.

8.4 Domain Decomposition for Finite Element Analysis

8.4.1

353

Introduction

Domain decomposition is attractive in finite element computations on parallel architectures, because it allows individual subdomain operations to be performed concurrently on separate processors and serial solutions on a sequential computer to overcome limitation of core storage capacity. Given a number of available processors q, an arbitrary finite element model (FEM) is decomposed into q subdomains, where formation of element matrices, assembly of global matrices, partial factorisation of the stiffness matrix and state determination or evaluation of generalised stresses can be carried out independently of similar computations for the other subdomains, and hence can be performed in parallel. In parallel processing of subdomains, the time to complete a task will be the time to compute the longest subtask. An algorithm for domain decomposition will be efficient if it yields subdomains that require an equal amount of execution time. In other words, the algorithm has to achieve a load balance among the processors. In general, this will be particularly ensured if each subdomain contains an equal number of elements or an equal total number of degrees of freedom. However, for some numerical techniques based on domain decomposition, a balanced number of elements or total degrees of freedom among the subdomains does not imply balancing of the subdomain calculations themselves. The use of a frontal subdomain solver provides a relevant example. In this case, the computing load within a domain is not only a function of the number of elements within the subdomain, but also the element numbering. Thus, the optimal number of elements is a priori unknown and can vary significantly from one subdomain to another. In order to reduce the cost of synchronisation and message passing between the processors in a parallel architecture, the amount of interface nodes should be minimised, because the parallel solution for the generalised displacements usually requires explicit synchronisation on a shared-memory multiprocessor and message passing on local-memory ones. In a domain decomposition method, another significant mesh partitioning factor which should be considered is the subdomain aspect ratio. This ratio has a vital impact on the convergence rate of the iterative approaches for the finite element tearing and interconnecting method. The above features suggest that an automatic finite element domain decomposer should meet four basic requirements in order to be efficient: 1. It should be able to handle irregular geometry and arbitrary discretisation in order to be general purpose. 2. It must yield a set of balanced subdomains in order to ensure that the overall computational load be as evenly distributed as possible among the processor. 3. It should minimise the amount of interface nodes in order to reduce the cost of synchronisation and/or message passing between the processors. 4. It must result in subdomains with proper aspect ratios, in order to improve the convergence rate of the domain decomposition based iterative method.

354

8 Decomposition for Parallel Computing: Graph Theory Methods

Methods of subdomaining are well documented in the literature, see for example Farhat and Wilson [12], Farhat [1], Dorr [13] Malone [14], Farhat and Roux [15], Farhat and Lesoinne [16], Topping and Khan [17], Topping and Sziveri [18], Vanderstraeten and Keunings [19], and Kaveh and Roosta [4]. Several automatic domain decomposition methods that address the load balance and minimum interprocessor computation problems have already been reported in the literature. In general, these algorithms can be grouped into two categories: engineering based and graph theory based methods. For engineering based approaches, one can refer to those of Ref. [20], and for graph theory based methods the algorithms of Ref. [21] can be referred to. In this section, two efficient algorithms are presented to decompose one- to three-dimensional finite element models of arbitrary shapes. The first method is a graph based method and uses a general expansion process. The second is an engineering based approach. In these algorithms the resulted subdomains generally have good aspect ratios, especially when the elements have this property originally.

8.4.2

A Graph Based Method for Subdomaining

In this algorithm, first the associate or incidence graph model G of the FEM is generated. Then a good starting node R1 of G is selected. R1 is taken as the first node of the first subgraph G1. Next G1 is expanded from R1. The process of expansion is continued such that the equality of the total degrees of freedoms of subdomains is provided. G2 is formed similar to G1, but it is expanded from R2, which is an unselected node in a maximum distance from R1. R2 should contain no node of G1. The process of expansion is executed in a manner that provides the connectedness of the subgraph being formed (if it is possible). A similar approach is employed and G3, . . . , Gq are generated, and the subdomains of the FEM corresponding to the selected subgraphs of G are identified. The steps of the algorithm are as follows: Step 1: Use the associate or incidence graph G of the considered FEM and form an SRT rooted from an arbitrary node of G, in order to find a node R1 with maximum distance from the root. Step 2: Generate subgraph Gi (i ¼ 1 to q) as follows: (a) Form an SRT rooted from Ri in order to calculate the distance between each node of G and Ri (Ri is taken as the first selected node of Gi), and find an unselected node Ri + 1 with maximum distance from Ri. (b) Find all the unselected boundary nodes of Gi, and denote them by UBN. (c) Associate an integer with each node ni of UBN which is the same as its distance from Ri plus the number of unselected nodes adjacent to ni minus the number of selected nodes adjacent to ni. Then detect the node with minimum integer and add it to Gi. (d) If the total degrees of freedom of the corresponding subdomain is less than [TDOF + W0(q # 1)]/q, then repeat the above steps from Step (b);

8.4 Domain Decomposition for Finite Element Analysis

355

otherwise, execute Step 2 to generate subgraph Gi + 1. TDOF is the total degrees of freedom of the FEM and W0 is the total degrees of freedom for the nodes of the corresponding subdomain which are also contained in unselected elements. In the above algorithm, only the connectivity of the nodes of G is considered, and no labels for edges of G, list or matrices of edges are needed. Therefore, the formation of SRTs of G and data keeping will be more simple and efficient. Since valencies of the nodes of an associate or incidence graph of an FEM are not generally very different, the adjacency list is an efficient means of keeping the connectivity data of G. The adjacency list of a graph G is a matrix containing N(G) rows and Δ columns, where Δ is the maximum degree of the nodes of G. The ith row contains the labels of the nodes adjacent to the node i. Step 1 is carried out to select a good starting node in the generated associate or incidence graph G. Using the adjacency list of G, Step (a) can be performed as follows; however, any other type of list may also be used: 1. Select all the nodes of the Rith row of the adjacency list of G. The distance between these nodes and the root is equal to unity. 2. Select all the unselected nodes of the rows j (j is an element of the set of the selected nodes in the previous step). The distance of these nodes from the root is one more. 3. Repeat Step 2 until all nodes are selected. The last instruction of Step (a) is carried out to select the first node of the next subgraph. This node should not be included in the previously generated subgraphs (i.e. it should be an unselected node). In Step (b), UBN contains unselected nodes which are adjacent to selected nodes of Gi. In order to extend Gi, a node of UBN will be added to Gi in every execution of Step (c). In this step an integer will be associated with each node of UBN which defines the best possible node, having the following properties: 1. It is near to the root. 2. It does not make the next UBN very large. 3. It is connected to Gi with more nodes, which leads to a desirable configuration for Gi. This integer is equal to the distance from Ri plus the valency of the node minus the number of selected adjacent nodes multiplied by 2. The value of [TDOF + W0(q # 1)]/q is not needed to be calculated in every execution of Step (d). Since every subdomain should have at least TDOF/q degrees of freedom, W0 can be calculated when the degrees of freedom of a subdomain becomes more than TDOF/ q. Additional value, W0(q # 1)/q, is considered, since the degrees of freedom of the interface nodes of subdomains are calculated in two or more subdomains and the degrees of freedom of the subdomains should be equal or nearly equal. In this algorithm, a disconnected subdomain may be generated. This happens when no node can be found in Step (b). In such a case, an unselected node with

356

8 Decomposition for Parallel Computing: Graph Theory Methods

minimum distance should be added to the considered subgraph. In order to avoid such situations, one should avoid decomposing a small FEM into many subdomains. However, the following modifications can always be used: 1. Formation of a single SRT from an arbitrary node to find a good starting node, may not lead to the best node; however, the existing good starting node algorithms can be used to select a better node. 0 00 2. If a subgraph Gi contains two components Gi and Gi , one can exchange nodes of 0 00 Gi or Gi with the adjacent subgraphs to provide connectedness for Gi. 3. Use a non-deterministic heuristic of combinatorial optimisation such as Simulated Annealing to improve the initial partitioning to avoid the formation of multiconnected subdomains.

8.4.3

Renumbering of Decomposed Finite Element Models

Once the subdomains and interface nodes are specified, the nodes and/or elements of each subdomain and the interface nodes can be renumbered for bandwidth, profile or frontwidth reduction, depending on whether a band, profile or frontal solver is exploited, respectively. The process of renumbering includes the following steps: (I) Renumber the internal nodes/elements of the subdomains M1, . . . , Mq using an available algorithm. (II) Select an interface node connected to M1 which is contained in a minimum number of elements as the starting node, and number the interface nodes using a nodal ordering algorithm. In the process of renumbering, when possible, priority is given to the nodes connected to lower numbered subdomains. It should be noted that, for a specified solver such as a frontal solver, the resulted subdomains and interface nodes should also satisfy additional conditions. For example in a frontal solver, a necessary condition for the applied domain decomposition approach to be feasible is that the number of degrees of freedom lying on the interface of any subdomain be smaller than the frontwidth associated with the direct (one domain) approach. However, such conditions cannot always be satisfied using the existing decomposition heuristics, because they generally depend on the shape and the connectivity of FEMs, see Lesoinne et al. [22].

8.4.4

Computational Results of the Graph Based Method

Example 1. A finite element model is considered with λ ¼ 606, α ¼ 1961; each element has 4 corner nodes and 4 mid-side nodes, and each node has 2 degrees of

8.4 Domain Decomposition for Finite Element Analysis

357

Fig. 8.7 A finite element model and its decompositions. (a) q ¼ 4 using the associate graph. (b) q ¼ 6 using the associate graph. (c) q ¼ 4 using the incidence graph. (d) q ¼ 6 using the incidence graph

freedom and is decomposed into 2, . . . ,6 subdomains, as shown in Fig. 8.7a–d for q ¼ 4 and 6, where λ and α denote the numbers of elements and nodes, respectively. The degrees of freedom of the selected subdomains and interface nodes for q ¼ 2, . . . ,6 are illustrated in Table 8.1, when associate and incidence graphs are used. Example 2. An L-shaped finite element model is considered with λ ¼ 2,400, α ¼ 1,281, and each node has degrees of freedom equal to 2. The model is decomposed into 6 and 12 subdomains, as shown in Fig. 8.8a–d. The degrees of freedom of the subdomains and interface nodes using associate and incidence graphs are illustrated in Table 8.2. Example 3. A finite element model is considered with λ ¼ 528, α ¼ 307, and each node has 2 degrees of freedom. The model is decomposed into 2, 3 and 4 subdomains, and the decomposed models for q ¼ 4 are shown in Fig. 8.9a–b. The degrees of freedom of the subdomains and interface nodes using associate and incidence graphs are illustrated in Table 8.3. The patterns of the node adjacency matrices employing the associate graph for the model, after ordering, are shown in Fig. 8.10a–c.

358 Table 8.1 Results of Example 1

8 Decomposition for Parallel Computing: Graph Theory Methods q 2

Type of graph Associate Incidence

DOFs of subdomains; interface nodes 2002, 1994; 74 2016, 2008; 102

3

Associate Incidence

1370, 1360, 1352; 160 1352, 1370, 1352; 150

4

Associate Incidence

1048, 1052, 1060, 1044; 280 1022, 1030, 1030, 1022; 182

5

Associate Incidence

856, 860, 868, 852, 842; 352 828, 844, 848, 826, 816; 240

6

Associate Incidence

724, 730, 744, 748, 672, 706; 394 700, 728, 714, 692, 692, 694; 296

Fig. 8.8 An L-shaped finite element model and its decompositions. (a) q ¼ 6 using the associate graph. (b) q ¼ 12 using the associate graph. (c) q ¼ 6 using the incidence graph. (d) q ¼ 12 using the incidence graph

8.4 Domain Decomposition for Finite Element Analysis

359

Table 8.2 Results of Example 2 q 8

Type of graph Associate Incidence

DOFs of subdomains; interface nodes 462,462,462,462,462,462; 210 462,462,462,462,462,462; 216

12

Associate Incidence

244,244,248,248,246,246,242,252,246,232,236,232; 342 252,252,250,250,268,268,246,268,260,232,234,242; 440

Fig. 8.9 A finite element model and its decompositions. (a) q ¼ 4 using the associate graph. (b) q ¼ 4 using the incidence graph

Table 8.3 Results of Example 3

8.4.5

q 2

Type of graph Associate Incidence

DOFs of subdomains; interface nodes 324,324;34 322,322;30

3

Associate Incidence

224,216,218;42 222,222,220;50

4

Associate Incidence

170,164,170,168;58 176,164,170,170;66

Discussions on the Graph Based Method

This algorithm has low time complexity and is simple to program and leads to efficient partitioning of a finite element model into subdomains with the required properties; therefore it can also be considered as a good educational approach. The finite element model that should be partitioned can contain meshes with different dimensions, types and sizes. Although the problem of aspect ratios of the subdomains is not dealt with explicitly in this section, the algorithm has the feature of expansion in all directions, leading to good aspect ratios.

360

8 Decomposition for Parallel Computing: Graph Theory Methods

Fig. 8.10 The patterns of the ordered node adjacency matrices

8.4.6

Engineering Based Method for Subdomaining

Definitions. A level structure L(r) of a finite element model rooted from an element r (as the root), is defined as a partitioning of the set of elements into levels l1(r), l2(r), . . . , ld(r) such that: 1. l1(r) ¼ {r}. 2. All elements adjacent to elements in level li(r) (1 < i < d) are in levels li # 1(r), li(r) and li + 1(r). 3. All elements adjacent to elements in level ldi(r) are in levels ld # 1(r) and ld(r). The overall level structure may be expressed as the set L(r) ¼ {l1(r), l2(r), . . ., ld(r)}, where d is the depth of the level structure and is simply the total number of levels, and two elements are adjacent if they share a common node. The element adjacency list of a finite element mesh contains the lists of elements adjacent to each element. The element-node list of an FEM contains the lists of nodes of each element and is generally employed as an input for data connectivity of finite element models. Following Webb and Froncioni [23], the node-element list contains the lists of elements containing each node of the finite element mesh.

8.4 Domain Decomposition for Finite Element Analysis

361

A genre structure is a level structure in which each level is divided into one or more genres, and the index of each genre, as defined below, simply shows the pseudo-distance between the root and its elements. The overall genre structure rooted from an element r may be expressed as the set G(r) ¼ {g0(r), g1(r), g2(r), . . . , gs(r)}, in which the pseudo-distance between r and the elements of genre gi(r) is equal to i. The index vector IVr(i) of a genre structure rooted from an element r is an (n + 1)-dimensional vector whose ith array (i ¼ 0, . . ., n) defines the total number of elements of gi (j ¼ 0, . . . , i), i.e. IVr ðiÞ ¼

i X j¼0

j gi ðrÞj:

ð8:20Þ

Thus, the cardinality of genre i (0 < i ) n) is simply equal to IVr(i) # IVr(i # 1), and the cardinality of g0(r) is equal to 1. The following scheme (in pseudo code) should be used to form a genre structure from an arbitrary starting element r, to generate its index vector and to find the pseudo-distances pd(r,ei) between the root r and all elements ei(i ¼ 1, . . . , λ, where λ denotes the number of elements) of the considered finite element model. In this scheme, D ∈ {1,2,3} denotes the highest dimension of the elements in the model, and CCN(gi(r),e) denotes the set of common corner nodes between the elements of genre gi(r), and the element e. 1. Set g0(r) ¼ {r},IVr(0) ¼ 1, pd(r,r) ¼ 0 and mask r. 2. Set i ¼ 1, a ¼ 0 and b ¼ 0. 3. for j ¼ D to 1 step 1 for k ¼ a to b (I) put each unmasked element e with |CCN(gk(r), e)| * j into gi(r). (II) if |g1 (r) 6¼ 0| then set IVr(i) ¼ IVr(i # 1) + |g1 (r)|, pd(r,e) ¼ i(e ∈ gi(r)), i ¼ i + 1 and mask the elements of gi(r). end for end for 4. If Ivr(i # 1) < λ then set a ¼ b + 1, b ¼ i and repeat Step 3.

8.4.7

Genre Structure Algorithm

Step 1: Form a genre structure rooted from an arbitrary element, and select an element e1s from its last genre. Step 2: Calculate the pseudo-distance between e1s and each element, and select an element e2s with maximum pseudo-distance from e1s .

362

8 Decomposition for Parallel Computing: Graph Theory Methods

Step 3: Calculate the pseudo-distance between e2s and each element. If q ¼ 2, then go to Step 5. Step 4: Find an unselected element eis (i ¼ 3,4, . . . ,q) contained in genres gj1(e1s ), gj2(e2s ), gj3(e3s ), . . ., gji # 1(ei#1 s ), such that the least value of IVr(jk # 1) be maximum, where jk > 0, k ¼ 1, . . ., i # 1, then calculate the pseudo-distances between eis and the elements. Step 5: For each selected element ejs (j ¼ 1, . . . , q) and each element ek (k ¼ 1, . . . , λ), assign an integer in (ejs ,ek) as follows,

where

# $ # $ # $ in esj ; ek ¼ λ þ mpd esj ; ek # pd esj ; ek ,

ð8:21Þ

# $ % # $ & mpd esj ; ek ¼ min pd esi ; ek j1 ) 1 ) q, 1 6¼ j :

Step 6: Let ejs be the first element of the subdomain Mi, calculate the weight of Mi and mask eis , where i ¼ 1, . . ., q. Step 7: Find an expandable subdomain Mi with minimum weight, add an unmasked element ek with maximum non-zero priority number Pi ¼ CN + in (eis ,ek) to Mi, update the weight Mi and mask ek, where if |CCN(Mj,ek)| ) 3 then CN ¼ |CCN (Mi,ek)|, else CN ¼ 3. If there is no element to be added to Mi, this subdomain is not expandable and should be masked. If there are several elements with the maximum priority number Pi, then select the one with the minimum sum of integers corresponding to eis (i ¼ 1, . . ., q). Repeat this step until all the elements are masked. In this algorithm, the weight of a subdomain Mi can be taken as an arbitrary single number such as the number of the elements of Mi, the total degrees of freedom of the nodes of Mj, a function of the number and labels of the elements of Mi, and so on. However, here the total degrees of freedom of the nodes of a subdomain, is considered as the weight of the subdomain. Obviously, in this method only the corner nodes of a finite element mesh should be provided; i.e. mid-side nodes and interior nodes are not needed. This increases the efficiency of the algorithm and results in saving computer storage space for finite element models with high order elements. An important problem which should be contemplated in a domain decomposition method is the connectedness of the elements of a single subdomain. In this algorithm, multicomponent subdomains are avoided. Since the integers which are calculated in Step 5 are more than zero, hence the priority number Pi of an element ei corresponding to the subdomain Mi will be zero if |CCN(Mi,ei)| ¼ 0, i.e. the element is not connected to Mi with a corner node. As stated in Step 7, an element with priority number Pi ¼ 0 cannot be added to Mi. This provides the connectedness of Mi (i ¼ 1, . . . , q); however, it leads to differences between the weights of the subdomains, because when a subdomain cannot be expanded and is masked, the

8.4 Domain Decomposition for Finite Element Analysis

363

other unmasked subdomains are still expanded. This problem has been nearly remedied in the present algorithm by Steps 2 and 4. In these steps, the first elements of the subdomains are selected in such a manner that there are enough elements to be added to them for further expansion of the subdomains. For complete balanced loads for subdomains, one can let elements with zero priority numbers be also added to a subdomain, in which case multicomponent subdomains will be generated. However, there are several non-deterministic heuristics used in combinatorial optimisation such as Simulated Annealing, Stochastic Evolution and Tabu Search which can be used for better load balancing of subdomains and reduction in the number of interface nodes, see for example Reference 13. These combinatorial optimisation methods are normally included in an FEM decomposition algorithm as follows: Step I: Invoke a direct partitioning scheme to produce an initial decomposition of reasonable quality. Step II: Use an optimisation procedure to improve the initial partitioning. The second step generally needs high computer time, hence this algorithm is designed for careful partitioning of the finite element meshes in order to avoid (as far as possible) the use of optimisation procedures for general cases. However, this method can be applied as a direct method in Step I. This will be efficient, since the more the load balancing of subdomains and the less the number of interface nodes produced by a direct partitioning scheme, the less cost for the applied optimisation method. The Step 1 of the algorithm presented in this section is carried out to find a good starting element e1s for the first subdomains M1. Step 2 is executed in order to calculate the pseudo-distance between e1s and each element, and to find an element e2s as the good starting element of the second subdomain M2. It should be noted that, when q > 2, it is needed to know the index of genres containing a specified element because it is needed for the selection of the starting elements of subdomains Mi (i ¼ 3, . . . , q). Step 3 should be carried out to calculate the pseudo-distance between e2s and each element of the considered finite element mesh. Also in this step, the index of genres containing a specified element should be defined for q > 2. Step 4 is executed in order to find good starting elements for subdomains Mi (i ¼ 3, . . . , q). The condition contained in this step is included in order to provide the starting elements of subdomains to be unobtrusive when the process of expansion is performed in Step 7. This condition increases the probability that a subdomain will remain expandable while the other subdomains are being expanded. Step 5 is carried out in order to calculate an integer for each selected (starting) element and each element of the finite element mesh. This integer is always more than zero since λ is always more than or equal to a pseudo-distance between two elements, and a pseudo-distance is always equal to or more than zero. The integers calculated in this step affect the priority number of elements in two ways when the process of expansion is performed: (1) the elements which are added to a subdomain have lower priority numbers for other subdomains, (2) the elements of a subdomain do

364

8 Decomposition for Parallel Computing: Graph Theory Methods

not have flange positions in relation to the region of the subdomain (loosely speaking). These effects make the number of boundary interior nodes of a subdomain low and its aspect ratio a desired value. The less the differences between the geometrical dimensions of a subdomain with a given area/volume, the smaller the boundary and the better aspect ratio of the subdomain. However, this remark is true when the elements have originally good aspect ratios. For more details about the aspect ratio of a subdomain, see the recent paper of Farhat et al. [24] in which their final choice has been to compute the aspect ratio AR of a subdomain Mi as follows: Surface ðMi Þ ðtwo dimensional problemsÞ Surface of circumscribed circle Volume ðMi Þ ðthree dimensional problemsÞ ARðMi Þ ¼ c3 + Volume of circumscribed sphere ð8:22Þ ARðMi Þ ¼ c2 +

where c2 and c3 are scaling constants designed such that 0 < AR ) 1. Step 6 is executed in order to initialise the subdomains Mi (i ¼ 1, . . ., q) and their weights and to mask their first (starting) elements. The elements of a subdomain are masked only in order to forbid their repeated selection. Step 7 contains the expansion process of the algorithm. In every execution of this step, an element with maximum priority number corresponding to a subdomain Mi is added to Mi, where Mi is the subdomain with current minimum weight. This way of expansion leads to equal loads for subdomains such that the subdomains remain expandable, and this condition is provided in the process of selecting eis (i ¼ 1, . . . , q) and giving a priority number to an element corresponding to the subdomain being formed. The priority number defined in this step is simply designed to give more priority to an element connected to a subdomain Mk with more corner nodes in comparison with an element connected to Mk with less corner nodes having the same integers.

8.4.8

Example

Consider the simple finite element mesh, as shown in Fig. 8.11a, with each node having 2 degrees of freedom, and suppose it to be decomposed into three subdomains. The steps of the present algorithm are performed as follows: Step 1: A genre structure is rooted from an arbitrary element such as the element 15. The elements of each genre are recognised with the index of the genre as illustrated in Fig. 8.11b. The last genre, g8(15), contains the element 6; hence e1s ¼ 6.

8.4 Domain Decomposition for Finite Element Analysis

a

365

b 3

4

8

9

10

13

14

15

16

19

20

21

22

1

2

7

5

6

5

4

3

4

5

8

11

12

4

2

1

2

4

7

18

3

1

0

1

3

6

24

4

2

1

2

4

7

17 23

c

d 14

10

6

3

1

0

6

7

8

9

13

17

15

11

7

4

2

1

3

4

5

8

12

16

16

12

8

5

4

3

1

2

4

7

11

15

17

13

9

8

7

6

0

1

3

6

10

14

e

f 13

9

8

7

6

7

16.31 17

33.26 22

26.21 22

28.22 29.12 19 20

31.7 17

12

8

5

4

3

4

12.33 17.28 20 15

22.24 24

26.20 24

25.14 23

27.9 21

11

7

4

2

1

2

9.34 14

14.29 19

20.24 24

21.19 21.14 27 37

23.11 28

10

6

3

1

0

1

7.34 14

12.28 19

18.21 21

17.19 29

19.11 29

17.14 33

g

Fig. 8.11 Illustration of the steps for the example. (a) A simple two-dimensional FEM. (b) Genres of G(15). (c) Genres of G(6). (d) Genres of G(19). (e) Genres of G(23). (f) Integers of the elements. (g) Decomposition of the FEM for q ¼ 3

Step 2: G(6) is formed to calculate the pseudo-distance between the element 6 and other elements. The elements of each genre are assigned with the index of the genre; this index is same as the pseudo-distances between the root and the elements of the genre. In Fig. 8.11c the pseudo-distance between the root (element 6) and other elements are depicted; the element 19 belongs to the last genre of G(6), having the highest pseudo-distance from the root, and thus e2s ¼ 19.

366

8 Decomposition for Parallel Computing: Graph Theory Methods

Step 3: G(19) is generated, and the pseudo-distances between the root and the elements are shown in Fig. 8.11d. Since q > 2, therefore Step 4 should be executed. Step 4: Two elements 2 and 23 satisfy the condition of this step, since 2∈g10 ð6Þ and g7 ð19Þ

IV6 ð9Þ ¼ 16, IV19 ð6Þ ¼ 11 23∈g7 ð6Þ and g10 ð19Þ

IV6 ð6Þ ¼ 11, IV19 ð9Þ ¼ 16, and minfIV6 ðiÞ, IV19 ðjÞg < 11, where 0 ) i, j ) 16 and ði; jÞ 6¼ ð9; 6Þ and ð6; 9Þ: Element 2 or 23 can be selected for e3s arbitrarily; suppose e3s ¼ 23. Figure 8.11e shows the pseudo-distances between e3s and the other elements. Step 5. For each element, three integers are assigned corresponding to e3s , e2s and e3s . These integers are respectively illustrated in Fig. 8.11f for each element. Step 6: Execution of this step leads to M1 ¼ {6}, M2 ¼ {19} and M3 ¼ {23}. The weights of M1, M2 and M3 are the same and equal to 8, and their elements are masked. Step 7. This step is carried out λ # q ¼ 21 times, and in each execution one element with maximum priority number is added to a subdomain with current minimum weight as follows: All subdomains have the same weight; hence the subdomain M1 is selected arbitrarily to be expanded. The elements with non-zero priority numbers which are connected to M1 are 5, 11 and 12, and their priority numbers are 2 + 29, 1 + 25 and 2 + 27, respectively. Thus element 5 is added to M1 and is masked. The weight of M1 is now equal to 12. The subdomains M2 and M3 have minimum current weight. The subdomain M2 is selected arbitrarily to be expanded. The elements 13, 14 and 20 are connected to M2, and their priority numbers are 2 + 34, 1 + 29 and 2 + 29, respectively. Hence the element 13 is added to M2 and is masked. The current weight of M2 is now equal to 12. The subdomain M3 has the least current weight. The elements 16, 17, 18, 22 and 24 are connected to M3, and their priority numbers are 1 + 27, 2 + 27, 1 + 25, 2 + 29 and 2 + 29, respectively. The priority numbers of the elements 22 and 24 are maximum; however, element 24 is added to M3 because the sum of its integers is less than that of the element 22. The element 24 is masked. The weight of the subdomain M3 is now equal to 12. The repetitions of this step lead to the decomposition as illustrated in Fig. 8.11g.

8.4 Domain Decomposition for Finite Element Analysis

8.4.9

367

Computational Results of the Engineering Based Method

Two examples are studied in this section, using the direct method for the formation of their element adjacency list. Example 1. A multiconnected finite element mesh is shown in Fig. 8.12a, and decomposed into 2, 3, 4, 8 and 16 subdomains as illustrated in Fig. 8.12b–f. In this example, each node has 2 degrees of freedom. Computational time is provided in Table 8.4. Example 2. A multiconnected H-shaped finite element mesh with each node having 2 degrees of freedom is shown in Fig. 8.13a, and decomposed into 2, 4, 5, 8, 16 and 32 subdomains as illustrated in Fig. 8.13b–g. Computational time is provided in Table 8.5.

8.4.10 Discussions The algorithm developed in this section is designed as a pre-processor for concurrent finite element computations. It may also serve as an automatic decomposer for serial solutions on a sequential computer, to overcome limited core storage capacity. This algorithm has low time complexity and leads to efficient partitioning of a finite element mesh into subdomains with required properties. A finite element mesh to be partitioned, may contain various meshes with different dimensions, types and sizes. The algorithm uses a simultaneous expansion process which is an improved version of the algorithm presented in the previous section for substructuring. In this algorithm the method for selecting the first (representative) element for each subdomain is improved, and the better priority numbers for elements to be added to the expanding subdomains are defined in order to form subdomains with more appropriate properties. This algorithm is designed to have properties required for an efficient decomposition and leads to subdomains with the following properties: 1. Low computer space and time requirements. In the present algorithm only the corner nodes are needed to be given, and this leads to a large space saving in FEMs with high order elements. The time complexity of the algorithm is independent of the number of nodes for the considered FEM, and the critical step of the algorithm takes O(λ2θ) operations in worst-case. 2. General in use. The algorithm can be employed to decompose unstructured FEMs without any restriction, and an arbitrary parameter can be considered as the loads of the subdomains. 3. Balance loads for subdomains. Selection of the starting elements of subdomains and the expansion process are performed in a manner which leads to an efficient

368

8 Decomposition for Parallel Computing: Graph Theory Methods

Fig. 8.12 Decompositions of the multiconnected finite element mesh. (a) A multiconnected FEM with 1,152 elements and 1,248 nodes. (b) q ¼ 2. (c) q ¼ 3. (d) q ¼ 4. (e) q ¼ 8. (f) q ¼ 16 Table 8.4 Computational time q Time (sec.)

2 29.00

3 29.28

4 29.44

8 30.59

16 33.39

8.4 Domain Decomposition for Finite Element Analysis Fig. 8.13 Decompositions of a multiconnected H-shaped finite element mesh. (a) A multiconnected H-shaped FEM with 1,340 elements and 1,042 nodes. (b) q ¼ 2. (c) q ¼ 4. (d) q ¼ 5. (e) q ¼ 8. (f) q ¼ 16. (g) q ¼ 32

369

370

8 Decomposition for Parallel Computing: Graph Theory Methods

Table 8.5 Computational time q Time (sec.)

2 33.95

4 31.75

5 32.90

8 37.14

16 40.70

32 52.79

balancing of loads. However, in order to decrease the differences between the loads of subdomains, the following steps are included which should be executed in place of Steps 1–4 of the original algorithm: (a) Find the pseudo-distance between each element and all the element of the finite element mesh. (b) Find q elements e1s , e2s , . . ., eqs , provided that eis (i ¼ 1, . . ., q), which is contained in genres gj1(e1s ), gj1(e2s ), . . ., gj1(eqs ), is selected in such a way that the least value of IV(jk # 1) is maximum where jk 6¼ 0 (k ¼ 1, . . ., q). However, this takes more operations than those of Steps 1–4. 4. Close to minimum number of interface nodes. In this algorithm, the number of interface nodes is kept to the least possible by selecting the elements to be added to a subdomain which have not high priority numbers for the other subdomains, and have a proper position in relation to the previously selected elements of the subdomains. 5. Good aspect ratios for subdomains. When the elements of the considered finite element mesh have aspect ratios with proper values, the algorithm leads to a decomposition with subdomains having reasonable aspect ratios. This is because the subdomains are expanded in all directions, which makes the denominators of the equations introduced by Farhat et al. [24] to be increased.

8.5

Substructuring: Force Method

The force method can be employed in parallel analysis of structures. In this section, the formulation of substructuring is provided, and an algorithm is presented for such analysis. The computational process is illustrated using simple examples. In this section, the notations and formulations presented in Chap. 3 will be used.

8.5.1

Algorithm for the Force Method Substructuring

Once a structural model has been decomposed using any of the methods presented in the previous sections, the following approach can be used for the analysis employing the force method:

8.5 Substructuring: Force Method

371

In order to support a substructure in a statically determinate fashion, cuts are introduced at members incident with the interface nodes contained in the corresponding substructure, except at one arbitrary node where the substructure is connected to the previous one. For a given substructure Si, let the external forces be denoted by pi, and redundant forces by qi. Then the substructure Si can be analysed for the internal forces in the substructure (not coupling redundants) qi in the aforementioned manner, i.e. !

v0i v1i

"

!

D00 ¼ D10

D01 D11

" ! i

" pi : qi

ð8:23Þ

For continuity within the substructure: # $ qi ¼ # D#1 11 D10 i pi:

ð8:24Þ

# $ t v0i ¼ D00 # D10 D#1 11 D10 i pi ,

ð8:25Þ

v0i ¼ Fi pi ,

ð8:26Þ

Deflections corresponding to the nodal force are,

that is

where Fi is the flexibility transformation matrix for the ith substructure. Internal forces are obtained as,

or

# $ ri ¼ B0 # B1 D#1 11 D10 i pi ,

ð8:27Þ

ri ¼ Bi pi

ð8:28Þ

# $ Bi ¼ B0 # B1 D#1 11 D10 i ,

ð8:29Þ

and

where Bi is the force transformation matrix in the redundant substructure. The matrices Fi and Bi are formed for each substructure, in turn. For the complete structure S composed of q substructures (S1,S2, . . ., Sq), the force vector pi acting on a substructure “s” is given by,

372

8 Decomposition for Parallel Computing: Graph Theory Methods

psi ¼ ½ ae

!

" pe be & , qc

ð8:30Þ

where qc are the coupling redundants. On a particular substructure, there will be three different types of forces: pee is the external force vector, pec is the coupling redundant forces vector, and peb contains the statically determinate connection forces. For the entire structure, the following matrices Aee and Bec are defined:

and

Then:

% & At ¼ aeð1Þ ; aeð2Þ ; . . . ; aeðqÞ ,

ð8:31Þ

% & Bt ¼ beð1Þ ; beð2Þ ; . . . ; beðqÞ :

ð8:32Þ

2

3 psð1Þ 6 psð2Þ 7 6 7 7 ps ¼ 6 6 . . . 7 ¼ ½ Aee 4 ... 5 psðqÞ

Bec &

!

" pc : qc

ð8:33Þ

The forces ps can be partitioned according to three types of forces pei, peb, and pec as mentioned before. Then: 2

3 2 pei aei 4 peb 5 ¼ 4 aeb pec aec

3 0 ! " p beb 5 e : qc bec

ð8:34Þ

It is obvious that, whereas qc may produce peb and pec forces, it does not produce pei forces. The flexibility matrix of the entire structure corresponding to pe and qc can be formed using Eqs. 8.25 and 8.33 as: ! and

f ee f ce

f ec f cc

"

2 " Feð1Þ t Aee 4 ¼ t Aec !

... FeðqÞ

3

5½ Aee

Bec &,

ð8:35Þ

8.5 Substructuring: Force Method Fig. 8.14 A singlebay fourstorey frame with geometric and connectivity properties

373

a

b 44.4kN

9 10

44.4kN

[email protected]

7

44.4kN 44.4kN

7 5 4 3 1 1

12 9 6 3

10 11 8 8 6 5 4 2 2

7.64m

!

ve vc

"

¼

!

f ee f ce

f ec f cc

"!

" pe : pc

ð8:36Þ

For continuity across the cut sections of the structure, vc ¼ 0,

ð8:37Þ

qc ¼ #f #1 cc f ce pe

ð8:38Þ

hence:

Deflections of the structure are then given as, # $ ve ¼ f ee # f ec f #1 cc f ce pe ,

ð8:39Þ

making the complete analysis of the structure feasible.

8.5.2

Examples

Example 1. A single-bay four-storey frame is considered, as shown in Fig. 8.14. The forces are depicted in Fig. 8.14a, and the nodal and element orderings are given in Fig. 8.14.b. For this frame, I ¼ 41,623.14 cm4 (for all members) and E ¼ 2.1 + 105 N/m2. The model is decomposed into two substructures as illustrated in Fig. 8.15. The analysis is performed and the bending moments are obtained as provided in Table 8.6. Example 2. A three-bay pitched-roof frame together with material properties and dimensions are shown in Fig. 8.16.

374

8 Decomposition for Parallel Computing: Graph Theory Methods

Fig. 8.15 Decomposition of the structural model

a

44.4kN

Fb

b Fc

44.4kN 44.4kN

44.4kN Fb

Table 8.6 Bending moments of Example 1

Nodes 1 3

5

7 9 10

8

6

4 2

End nodes of members 1–3 3–1 3–4 3–5 5–3 5–6 5–7 7–5 7–8 7–9 9–7 9–10 10–9 10–8 8–10 8–7 8–6 6–8 6–5 6–4 4–6 4–3 4–2 2–4

Fc

Bending moments(kN.m) #219.17 78.93 185.07 #106.14 0.34 2.94 7.34 96.2 112.7 #16.48 58.04 58.04 58.04 #58.04 #16.48 112.7 #96.2 #52.84 170.28 #117.43 #78.93 185.07 #106.14 #219.17

This model is partitioned into two substructures, as illustrated in Fig. 8.17, where different groups of loads on each substructure are shown. For all the members, I ¼ 0.2 m4 and E ¼ 2.1 + 105 N/m2. The bending moments for members of this frame are presented in Table 8.7. The substructuring analysis, using the force method for frame structures, can be generalised to the analysis of other types of structures when the algebraic force method is employed, Plemmons and White [25]. In this method, appropriate partitioning of the incidence matrices of the structural graph models is performed,

8.5 Substructuring: Force Method 10kN

375 10kN

10kN 4m

10kN

8m

[email protected]

Fig. 8.16 A three bay pitched-roof frame 10kN

10kN Ft

10kN

Substructure I

Ft

Substructure II

Fig. 8.17 Decomposition of the structural model Table 8.7 Bending moments of Example 2

Nodes 1 5 6

7 2 8

9 3 10 11

End nodes of members 1–5 5–1 5–6 6–5 6–7 7–6 7–8 7–2 2–7 8–7 8–8 9–8 9–10 9–3 3–9 10–9 10–11 11–10

Bending moments(kN.m) #72 #30 30 3 3 26 65 #91 #116 45 45 28 3.03 60 40.02 52 52 79

leading to well structured equilibrium equations. It is proved that sparse null bases can then be constructed in parallel, using the proposed decomposition. The performance of the method is illustrated by some examples from skeletal structures.

376

8 Decomposition for Parallel Computing: Graph Theory Methods

References 1. Farhat C (1988) A simple and efficient automatic FEM domain decomposer. Comput Struct 28:579–602 2. Al-Nasra M, Nguyen DT (1991) An algorithm for domain decomposition in finite element analysis. Comput Struct 39:277–289 3. George A, Liu JWH (1978) Algorithms for partitioning and numerical solution of finite element systems. SIAM J Numer Anal 15:297–327 4. Kaveh A, Roosta GR (1995) Graph-theoretical methods for substructuring, subdomaining and ordering. Int J Space Struct 10(2):121–131 5. George A (1971) Computer implementation of the finite elements. Report STAN-CS-71-208, Ph.D. thesis, Computer Science Department, Stanford University, CA 6. Kron G (1959) Diakoptics, piecewise solution of large-scale systems. A series of 20 chapters in the Electrical Journal, London 7. Pothen A, Simon H, Liou KP (1990) Partitioning sparse matrices with eigenvectors of graphs. SIAM J Matrix Anal Appl 11:430–452 8. Jennings A, McKeown JJ (1992) Matrix computation. Wiley, Chichester 9. Law KH (1986) A parallel finite element solution method. Comput Struct 23:845–858 10. Keyes DE, Gropp WD (1987) A comparison of domain decomposition techniques for elliptic partial differential equations and their parallel implementation. SIAM J Sci Statist Comput 8:166–202 11. Chadha HS, Baugh JW Jr (1996) Network-distributed finite element analysis. Adv Eng Softw 25:267–280 12. Farhat C, Wilson E (1987) A new finite element concurrent computer program architecture. Int J Numer Methods Eng 24:1771–1792 13. Dorr MR (1988) Domain decomposition via Lagrange multipliers. Report No. UCRL-98532, Lawrence Livermore National Laboratory 14. Malone JG (1988) Automated mesh decomposition and concurrent finite element analysis for hypercube multiprocessor computers. Comput Methods Appl Mech Eng 70:27–58 15. Farhat C, Roux FX (1991) A method of finite element tearing and interconnecting and its parallel solution algorithm. Int J Numer Methods Eng 32:1205–1227 16. Farhat C, Lesoinne M (1993) Automatic partitioning of unstructured meshes for the parallel solution of problems in computational mechanics. Int J Numer Methods Eng 36:745–764 17. Topping BHV, Khan AI (1995) Parallel finite element computations. Saxe-Coburg Publications, Edinburgh 18. Topping BHV, Sziveri J (1995) Parallel sub-domain generation method. In: Proceedings of Civil Comp 95, Edinburgh, pp 449–457 19. Vanderstraeten D, Zone O, Keunings R (1993) Non-deterministic heuristic for automatic domain decomposition in direct parallel finite element calculation. In: Proceedings of 6th SIAM conference on parallel processing, SIAM, pp 929–932 20. Felippa CA (1975) Solution of linear equations with skyline-stored symmetric matrix. Comput Struct 5:13–29 21. Kaveh A, Roosta GR (1994) A graph theoretical method for decomposition in finite element analysis. In: Proceedings of Civil-Comp 94, Edinburgh, pp 35–42 22. Lesoinne M, Farhat C, Geradin M (1991) Parallel/vector improvements of the frontal method. Int J Numer Methods Eng 32:1267–1281 23. Webb JP, Froncioni A (1986) A time-memory trade-off frontwidth reduction algorithm for finite element analysis. Int J Numer Methods Eng 23:1905–1914 24. Farhat C, Maman N, Brown GW (1995) Mesh partitioning for implicit computations via iterative domain decomposition: impact and optimization of the subdomain aspect ratio. Int J Numer Methods Eng 38:989–1000 25. Plemmons RJ, White RE (1990) Substructuring methods for computing the null space of equilibrium matrices. SIAM J Matrix Anal 11:1–22

Chapter 9

Analysis of Regular Structures Using Graph Products

9.1

Introduction

In this chapter, an efficient method is presented for the analysis of non-regular structures which are obtained by addition or removal of some members to regular structural models. Here a near-regular structure is divided into two sets, namely “the regular part of the structure” and “the excessive members”. Regular part refers to the structure for which the inverse of the stiffness matrix can be obtained by the previously developed simplified methods, and excessive members refer to those which cause the non-regularity of the regular structure [1].

9.2

Definitions of Different Graph Products

Many structures have regular patterns and can be viewed as the Cartesian product, strong Cartesian product, or direct product of a number of simple graphs. These subgraphs, used in the formation of the entire model, are called the generators of that model. Graph products were developed in the past 50 years (see e.g. Imrich and Klavzar [2]) for mathematical aspects, and Kaveh [3] for extensive applications.

9.2.1

Boolean Operation on Graphs

In order to explain the products of graphs, let us consider a graph S as a subset of all unordered pairs of its nodes. The node set and member set of S are denoted by N(S) and M(S), respectively. The nodes of S are labelled as v1,v2, . . . , vM, and the resulting graph is a labelled graph. Two distinct adjacent nodes, vi and vj, form a member, denoted by vivj ∈ M(S). A. Kaveh, Computational Structural Analysis and Finite Element Methods, 377 DOI 10.1007/978-3-319-02964-1_9, © Springer International Publishing Switzerland 2014

378

9 Analysis of Regular Structures Using Graph Products

A Boolean operation on an ordered pair of disjoint labelled graphs K and H results in a labelled graph S, which has N(K) ! N(H) as its nodes. The set M(S) of members of S is expressed in terms of the members in M(K) and M(H), differently for each Boolean operations. Three different operations are discussed in this chapter, corresponding to Cartesian product, strong Cartesian product and direct product of two graphs.

9.2.2

Cartesian Product of Two Graphs

Many structures have regular patterns and can be viewed as the Cartesian product of a number of simple graphs. These subgraphs, which are used in the formation of a model, are called the generators of that model. The simplest Boolean operation on a graph is the Cartesian product K ! H introduced by Sabidussi [4]. The Cartesian product is a Boolean operation S ¼ K ! H, in which, for any two nodes u ¼ (u1,u2) and v ¼ (v1,v2) in N(K) ! N(H), the member uv is in M(S) whenever u1 ¼ v1 and u2 v2 ∈MðHÞ,

ð9:1aÞ

u2 ¼ v2 and u1 v1 ∈MðKÞ:

ð9:1bÞ

or

As an example, the Cartesian product of K ¼ P2 and H ¼ P3 is shown in Fig. 9.1. In this product, the two nodes (u1,v2) and (v1,v2) are joined by a member, since the condition (9.1b) is satisfied. The Cartesian product of two graphs K and H can be constructed by taking one copy of H for each node of K and joining copies of H corresponding to adjacent nodes of K by matching of size N(H). The graphs K and H will be referred to as the generators of S. The Cartesian product operation is symmetric, i.e. K ! H ﬃ H ! K. For other useful graph operations, the reader may refer to the work by Gross and Yellen [5]. Examples. In the first example, the Cartesian product C7 ! P5 of the path graph with five nodes denoted by P5 and a cycle graph shown by C7 is illustrated in Fig. 9.2. Two representations of the Cartesian product C3 ! P4 are illustrated in Fig. 9.3. The Cartesian product Pm1 ! Pm2 ! Pm3 of three paths forms a threedimensional mesh. As the second example, the Cartesian product of P6 ! P4 ! P5, resulting in a 5 ! 3 ! 4 mesh, is shown in Fig. 9.4. A graph can be the product of more than two specific graphs, such as paths and cycles. As the third example, the product of three graphs, P2 ! K3 ! P4, is shown

9.2 Definitions of Different Graph Products Fig. 9.1 The Cartesian product of two simple graphs

379

a

b

v1 u2

u1 K=P2

v2

w2

(v1 ,u 2 )

(u1 ,u2 )

H=P3

(u 1 ,v2 )

(u 1 ,w2 )

S

a b

Fig. 9.4 Representation of a 5 ! 3 ! 4 mesh

(v 1 ,w2 )

=

Fig. 9.2 Representation of C7 ! P5

Fig. 9.3 Two different representations of C3 ! P4

(v1 ,v 2 )

380

9 Analysis of Regular Structures Using Graph Products

Fig. 9.5 The Cartesian product of three graphs P2 ! K3 ! P4. (a) Generators. (b) Product

a

b P4

P2

Fig. 9.6 The Cartesian product of S by P4. (a) Generators. (b) Product

K3

a

b

S

P4

in Fig. 9.5. The product of a general graph and a path, S ! P4, is illustrated in Fig. 9.6.

9.2.3

Strong Cartesian Product of Two Graphs

This is another Boolean operation, known as the strong Cartesian product. The strong Cartesian product is a Boolean operation S ¼ K⊠H in which, for any two distinct nodes u ¼ (u1,u2) and v ¼ (v1,v2) in N(K) ! N(H), the member uv is in M (S) if: u1 ¼ v1 and u2 v2 ∈MðHÞ,

ð9:2aÞ

u2 ¼ v2 and u1 v1 ∈MðKÞ,

ð9:2bÞ

or

or

9.2 Definitions of Different Graph Products Fig. 9.7 The strong Cartesian product of two simple graphs. (a) Generators. (b) S ¼ K⊠H

381

a

b (v

v1 u1 K=P2

u2

v2

w2

H=P3

u1 v1 ∈MðKÞ and u2 v2 ∈MðHÞ:

1 ,u 2 )

(v1 ,v 2 )

(v 1 ,w2 )

(u1 ,u2 )

(u 1 ,v2 )

(u 1 ,w2 )

=

S

ð9:2cÞ

As an example, the strong Cartesian product of K ¼ P2 and H ¼ P3 is shown in Fig. 9.7. In this example, the nodes (u1,u2) and (v1,v2) are joined, since the condition (9.2c) is satisfied. Examples. In the first example, the strong Cartesian product P7⊠P5 of a path graph with seven nodes, denoted by P7 and the path graph P5 is illustrated in Fig. 9.8. As the second example, the strong Cartesian product C7⊠P4 is shown in Fig. 9.9.

9.2.4

Direct Product of Two Graphs

This is another Boolean operation, known as the direct product, introduced by Weichsel [6], who called it the Kronecker Product. The direct product is a Boolean operation S ¼ K*H, in which, for any two nodes u ¼ (u1,u2) and v ¼ (v1,v2) in N (K) ! N(H), the member uv is in M(S) if: u1 v1 ∈MðKÞ and u2 v2 ∈MðHÞ:

ð9:3Þ

As an example, the direct product of K ¼ P2 and H ¼ P3 is shown in Fig. 9.10. Here, the two nodes (u1,u2) and (v1,v2) are joined, since the condition (9.3) is satisfied. Examples. The direct product P7*P5 of the path graph P7 and path graph P5 is illustrated in Fig. 9.11. As the second example, the direct product C7*P4 is shown in Fig. 9.12.

382

9 Analysis of Regular Structures Using Graph Products

Fig. 9.8 Strong product representation of P7⊠P5

Fig. 9.9 Strong product representation of C7⊠P4

Fig. 9.10 The direct product of two simple graphs

a

b (v1 ,u 2 ) (v1 ,v 2 ) (v 1 ,w2 )

v1 u1

*

K=P2

Fig. 9.11 Direct product representation of P7*P5

u2

v2 H=P3

w2

= (u1 ,u2 )

(u 1 ,v2 ) S

(u 1 ,w2 )

9.3 Analysis of Near-Regular Structures Using Force Method

383

Fig. 9.12 Direct product representation of C7*P4

9.3

Analysis of Near-Regular Structures Using Force Method

Different simple and efficient methods for the analysis of structures are provided in Kaveh [3]. In the analysis of some near-regular structures one can solve the regular part independently and then superimpose the effect of the additional part. For such models, the matrices corresponding to regular part have canonical forms and their eigensolution or inversion can easily be performed [1]. The effect of member changing the regular to a near-regular structure can then be added. In this method, linear behaviour is assumed for the structures. Here we use the force method, and instead of selecting a statically determinate basic structure (standard method) we employ the regular part of the structure as the basic structure [7]. A new algebraic method is introduced for the force method of analysis for efficient analysis of large near-regular structures. In this part, we use the force method, however, instead of selecting a statically determinate basic structure we employ the regular part of the structure as the basic structure. Those additional elements are considered as redundant elements. This method is applied to truss and frame structures. In the present approach we can have missing elements instead of additional elements. In order to demonstrate this problem, consider the truss shown in Fig. 9.13a. This structure consists of a regular part P4⊠P10 as shown in Fig. 9.13b and has become a near-regular because of having additional 10 bars. The main aim is to decompose these two parts in order to arrive at the analysis of the near-regular structure using the results of the analysis of the regular part. In Fig. 9.13c the positions of the excessive members are highlighted, where the regular part is shown in broken lines. It should be mentioned that for some regular structures the stiffness matrices can be formed in special block forms, known as the canonical forms, Kaveh [3]. Here we assume that only the members cause irregularity and no additional nodes are

384

9 Analysis of Regular Structures Using Graph Products

Fig. 9.13 (a) An irregular truss. (b) The regular part as the strong Cartesian product P4⊠P10. (c) The excessive members being highlighted [1]

present except those of the regular part, i.e. the nodes of the two ends of each excessive members are in the regular part of the structure. At the beginning, methods suggested are presented for the formation of the matrices required in the force method. Obviously one can also obtain these matrices by other approaches. The present method consists of two groups of structures as described in the following: The first group is related to the analysis of those structures in which the excessive members have caused the irregularity. The second group is about those structures which require addition of some members to alter the near-regular structure to a regular one. In this case, by assuming pairs of members with two identical

9.3 Analysis of Near-Regular Structures Using Force Method

385

modulus of elasticity having positive and negative signs are added to those places where we need to have members to make the near-regular structure into regular one. In this case, the members with negative sign will be treated as the excessive members. In the above force method, the internal forces of the excessive members will be considered as redundants, and the corresponding forces will be applied at the regular part of the structure as external loads to incorporate the effect of such members. Thus the regular part will be the main structure to be analyzed. This means we analyze the near-regular structure by considering the regular part and adding the effect of the internal forces of the complementary members as external loads. Here we assume that the removal of the excessive members will leave the structure geometrically stable, and considering the topology of the regular structures this assumption is quite logical. In each remaining section first the formulation will be presented, and then through a simple example the process of analysis will be described in a step by step manner. Then by some practical examples, the efficiency of the method will be demonstrated. First the formation of the flexibility matrix is described. It should be mentioned that this matrix can be formed using any other available method.

9.3.1

Formulation of the Flexibility Matrix

In this section a method is presented for the formation of the matrix B in the following form: B ¼ ½ B0

B1 '

ð9:4Þ

where the ith column of B0 is a vector of internal force of the structure under a unit value of a load applied at the ith DOF of the structure (Pi ¼ 1), and the ith column of B1 is a vector containing the internal forces of the structure under the unit load applied at the position of the ith redundant of the (Xi ¼ 1) structure. According to the above definitions for the formation a matrix B we are looking for a method by means of which having the externally applied loads of the structure we find the internal forces of the members. In the following a method is presented for this problem using the equilibrium matrix, though one can also find this employing the existing traditional method. In order to calculate the internal forces of the regular structure under the external loading we proceed as the following: In the global coordinate system we have

386

9 Analysis of Regular Structures Using Graph Products

SΔ ¼ P ) Δ ¼ S(1 P

ð9:5Þ

where S( 1 is the inverse of the stiffness matrix of the DOFs of the regular part of the structure. Using the theorems previously developed for the block matrices, S( 1 can be formed using the blocks constituting S. This matrix can be obtained using some concepts of graph products or employing concepts from group theory. According to the definition of equilibrium matrices of the members of the structure and Eq. 9.5, in general, the following form can be written for the deformation of the local coordinate systems of the members of the regular structure: δ ¼ At Δ ¼ At S(1 P

ð9:6Þ

where A is the equilibrium matrix of the regular structure. Considering the equilibrium equations in the local coordinate system and Eq. 9.6, in general the vector of internal forces of the members of the regular structure under the action of an imaginary external unit load can be obtained as: " ! " ! Q0 ¼ sδ ¼ s At S(1 P ¼ sAt S(1 P

ð9:7Þ

Here s is the block diagonal matrix containing the stiffness of the members of the regular part of the structure. Therefore the vector of internal forces of a regular structure can be obtained having the external forces in the following form: Q0 ¼ R P

R ¼ sAt S(1

;

ð9:8Þ

If X contains the internal forces of the excessive members in the global coordinate system and P is the external force vector of the structure, then the internal forces of the members of the regular structure when part of it is nearregular, can be obtained as: Q1 ¼ RP þ R X

ð9:9Þ

Thus for the analysis of near-regular structure discussed in here, Q1 is the vector of internal forces of the regular structure. The vectors P and X can be expressed as: X ¼ NX

;

P ¼ IP

ð9:10Þ

I is a unit matrix and N is a matrix for transforming the local coordinate system to the global coordinate system. X is the internal force vector of excessive members. Here the method for the formation of A and N is explained. If AT is the equilibrium matrix of a near-regular structure, then by partitioning according to the numbers of internal forces of the excessive members, the matrices A and N can be formed as:

9.3 Analysis of Near-Regular Structures Using Force Method

387

$ # % A T ¼ A #N

ð9:11Þ

where A is the equilibrium matrix of the regular structure. In Eqs. 9.9 and 9.10 by taking the common factor and extracting the vector of the assumed forces we have: Q1 ¼ R ½ I

&

P N' X

'

ð9:12Þ

In general case the internal forces of the near-regular structure will be as follows: &

Q1 Q¼ X

'

ð9:13Þ

Therefore adding X to Eq. 9.12 the matrix Q can be written as éQ ù éR RN ù é P ù Q = ê 1ú = ê I úû êë Xúû ë X û ëZ

ð9:14Þ

Now the matrices B0 and B1 can be formed by partitioning the above matrix according to the numbering of the internal forces of the excessive members of the structure in the following form: B0 ¼

&

R Z

'

;

B1 ¼

&

RN I

'

ð9:15Þ

In the above equations Z is a matrix of zeros with dimension t ! k and I is a unit matrix of dimension t ! t. The matrices B0 and B1 have dimensions (e + t) ! k and (e + t) ! t, respectively. t is the total number of internal forces of the excessive members, k is the DOFs of the near-regular structure in global coordinate system and e is the number of internal forces of the regular structure. In this method we need to form the matrix AT and in the subsequent section a simple method will be presented for this formation. The formation of the matrix B can be summarized as follows: Step 1: Form the matrices S( 1 and s for the regular structure. Step 2: Form the matrix AT for all the members of the structure consisting of regular and excessive members. Step 3: Partition AT using Eq. 9.11 and form the matrices A and N. Step 4: Calculate the matrix R using Eq. 9.8. Step 5: Calculate the matrices B0 and B1 using Eq. 9.15.

388

9.3.2

9 Analysis of Regular Structures Using Graph Products

A Simple Method for the Formation of the Matrix AT

A general method for the formation of the equilibrium matrix consists of writing equilibrium of the forces at the nodes of the structure. For a quick calculation of the matrix AT one can assemble the rotation matrices of the members of the nearregular structure. Then it can be partitioned using the relationship presented in the previous section. In the following the approach for positioning the rotation matrices of the members in each column of the equilibrium matrix is illustrated. For the formation of the equilibrium matrix AT of the near-regular matrix we perform the following process: If we consider i as the nodal DOFs of the assumed member j in the local coordinate system, and r are the nodal DOFs of the assumed member j in the global coordinate system, then the columns corresponding to i in the matrix AT will be as follows: AT ðr; iÞ ¼ Tjt

ð9:16Þ

The remaining rows of these columns are zero. We repeat this process for all the members of the near-regular structure. Tj is the modified rotation matrix of the jth member. This matrix can be represented as follows: Space truss member

Tj = [T1 - T1 ] , T1 = [Cosa Cosb Cosg ] Planar frame member

Tj = [T1 T3 ]

,

és T3 = s1-1s 2 T2 , s j = ê 1 ës 2

s2 ù s1 úû - Sina

ð9:17Þ

0ù é Cosa Sina 0ù é- Cosa T1 = êê- Sina Cosa 0úú , T2 = êê Sina - Cosa 0úú êë 0 êë 0 0 1úû 0 1úû

sj is the jth block of the stiffness matrix s. Here, α, β and γ are the angles with the x, y and z axis, respectively. Similar to Eq. 9.11 the above matrix can be transformed to A and N by partitioning and numbering the internal forces of the excessive members which are separated from the structure.

9.4 Analysis of Regular Structures with Excessive Members

389

Algorithm The algorithm can be summarized as: Step 1: Formation of the matrices Tj for members of the near-regular structure. Step 2: Formation of the equilibrium matrix AT by assembling the rotation matrices of the near-regular matrix using Eq. 9.16. Step 3: Formation of the matrices A and N by partitioning of the matrix AT.

9.4

Analysis of Regular Structures with Excessive Members

In this section the analysis of those structures for which the irregularity is produced by excessive members is studied. Here the force method is used for the analysis, with the only difference that instead of removing member to obtain a statically determinate structure, members are removed to transform the structure into a regular one. Here the relationships required for the force method are presented. Base on the concepts of the force method, the internal forces of the members of the near-regular structure can be expressed as: &

P ½Q' ¼ ½B' X

'

ð9:18Þ

After the formation of the matrix B which was described in Sect. 9.2, one can calculate the internal forces of the excessive members using the following relationships: D2 ¼ B1 t F B1

;

X ¼ (D2

(1

D1 ¼ B1 t F B0

ð9:19Þ

D1 P

ð9:20Þ

Here F is a block matrix of dimension (e + t) ! (e + t) and contains all the flexibility matrices of the members of the near-regular structure. The matrix D1 is of dimension t ! k and the matrix D2 is of dimension t ! t. This means that for calculating the internal forces of the excessive members, only the inverse of a matrix of dimension t is needed. At the end, the forces of X are added to the external force vector P denoted by P∗ which is defined as the equivalent external load of the regular structure. According to this, the displacements of the structure can be obtained by the inverse of the stiffness matrix of the regular structure as follows: P∗ ¼ P þ NX

ð9:21Þ

390

9 Analysis of Regular Structures Using Graph Products

Δ ¼ S(1 P∗

ð9:22Þ

where S( 1 is the inverse of the stiffness matrix of the DOFs of the regular structure and can be obtained using the existing methods, Kaveh [3]. The vector Δ contains the displacements of the near-regular structure. The matrix N is the transformation matrix of the internal forces in excessive members from local to global coordinate systems.

9.4.1

Summary of the Algorithm

Step 1: Numbering the DOFs, nodes and members of the near-regular structure and formation of the external force vector P. Step 2: Formation of the matrices S( 1 and s . Step 3: Formation of the equilibrium matrix of the near-regular structure using Eq. 9.16. Step 4: Calculation of the B0 and B1 matrices using Eq. 9.15. Step 5: Formation of the flexibility matrix F in a block diagonal form for all the members of the near-regular structure. Step 6: Calculation of the matrices D1 and D2 using Eq. 9.19. Step 7: Calculation of the vector X using Eq. 9.20. Step 8: Calculation of the equivalent external load of the regular structure using Eq. 9.21. Step 9: Calculation of the nodal displacements of the near-regular structure using Eq. 9.21. The above explanations are further explained through the following simple example.

9.4.2

Investigation of a Simple Example

For the 10-bar truss shown in Fig. 9.14, deleting member 10, the structures become regular. Here using the force method, the internal force of the member 10 is calculated and as an additional force it is added to the external forces. Then the regular structure is analyzed with the new loads. It should be noted that in the standard force method the basic structure is selected for a redundant structure is often statically determinate. For the structure of this example we have four statical indeterminacy and four redundants should be chosen. However, in our approach the basic structure is selected as a regular structure which is not necessarily statically determinate. In this example, EA is assumed to be unit for all the members and the external load vector is as follows:

9.4 Analysis of Regular Structures with Excessive Members

391

Fig. 9.14 A 10-bar truss transformable to a regular structure

P ¼ ½ 10 0

0

0 0 't

20

One of the methods for the formation of the equilibrium matrix is to use the equilibrium equations of forces at the nodes and formation of the matrix of the coefficients of the forces. In this example the equilibrium matrix of the near-regular structure is calculated using this approach. The obtained equilibrium matrix is partitioned into A and N using Eq. 9.11. The force equilibrium equations will be as follows: pﬃﬃﬃ pﬃﬃﬃ P1 ¼ (Q2 ( pﬃﬃﬃ 2=2Q4 pﬃﬃﬃ , P2 ¼ Q1 þ p2ﬃﬃ=2Q ﬃ 4 pﬃﬃﬃ P3 ¼ Q2 þ p2 8 ﬃ ﬃﬃﬃ=2Q3 ( Q6p(ﬃﬃﬃ 2=2Q8 , P4 ¼ pﬃﬃﬃ 2=2Q3 þ Q5 þ p2ﬃﬃ=2Q P5 ¼ Q6 þ 2=2Q7 þ 2= 5Q10 , P6 ¼ 2=2Q7 þ Q9 þ 1= 5Q10

The relation between the equilibrium matrix A and the vector of external and internal forces of the structure can be written as: P ¼ AQ

ð9:23Þ

In this way the matrix A and the partitioning considering the excessive member 10 will be as follows: - 0.7071 0 é0 - 1 ê1 0 0 0.7071 ê ê0 1 0.7071 0 A T = [A N ]= ê 0 0 0 . 7071 0 ê ê0 0 0 0 ê 0 0 ëê0 0

0

0

0

0

0

0

0

0

0

0

0 -1

0

- 0.7071 0

1 0

0 1

0 0.7071

0.7071 0

0 0

0

0

0.7071

0

1

0

ù ú ú 0 ú ú 0 ú 0.9844ú ú 0.4472ûú 0

Thus the matrix N will be as N ¼ ½0

0 0

0

0:9844

0:4472 't

The stiffness matrix of the regular structure by elimination of the member 10 will become:

392

9 Analysis of Regular Structures Using Graph Products

2

0:6767 6 (0:1767 6 6 (0:5 S¼6 6 0 6 4 0 0

(0:1767 (0:5 0:6767 0 0 1:3535 0 0 0 (0:5 0 0

0 0 0 0 0 (0:5 0:8535 0 0 0:6767 0 0:1767

3 0 0 7 7 0 7 7 0 7 7 0:1767 5 0:6767

Replacing two columns 5 and 6 with columns 3 and 4, and also their corresponding rows, This matrix will get Form III pattern and calculating its eigenvalues leads to the formation of S(1. Here, s is a block diagonal matrix having the stiffness of the members of the regular structure. Since the structure is a truss, therefore this matrix becomes a diagonal one. s ¼ diagf0:5; 0:5; 0:3535; 0:3535; 0:5; 0:5; 0:3535; 0:3535; 0:5g In this relation diag represents a block diagonal matrix. Substituting the above matrices in Eq. 9.15 leads to the formation of B0 and B1 matrices. 2

0:3535 6 (0:6464 6 6 0:3535 6 6 (0:5 6 6 0 B0 ¼ 6 6 (0:1464 6 6 0:2071 6 6 (0:3535 6 4 (0:1464 0

0:8311 (0:1688 0:0923 0:2387 0 (0:0382 0:0540 0:0923 (0:0382 0

0:1846 0 0:1846 0 0:4459 0:2928 (0:2612 0 0 0:5857 (0:1847 0 0:2612 0 (0:4459 0:2928 (0:1847 0 0 0

0:1464 0:1464 0:3535 (0:2071 0 0:6464 0:5 (0:3535 (0:3535 0

3 (0:0382 (0:0382 7 7 (0:0923 7 7 0:0540 7 7 7 0 7 (0:1688 7 7 0:2387 7 7 0:0923 7 7 0:8311 5 0

B1 ¼ ½ (0:1138 (0:1138 (0:2749 0:1610 0 (0:5026 (0:5540 0:2749 (0:0557 1 't

The flexibility matrix of the near-regular structure F in general is a block diagonal matrix. Since the considered structure is a truss, thus this matrix has numerical values in its diagonal. n pﬃﬃﬃo pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃ pﬃﬃﬃ F ¼ diag 2, 2, 2 2, 2 2, 2, 2, 2 2, 2 2, 2, 2 5

Using Eq. 9.19 the matrices D1 and D2 are formed as follows: D1 ¼ ½ (0:8719

D2 ¼ ½6:4045'

(0:2277

(1:0997

0 (2:1050

(0:1109 ';

Now employing Eq. 9.20, the internal forces of the excessive members are calculated as:

9.5 Analysis of Regular Structures with Some Missing Members

393

X ¼ (D2 (1 D1 P ¼(

h i 1 1 4 ! (0:8719 . . . 0 . . . :½ 10 0 0 20 0 0 't ¼ (1:3624 6:4045

Substituting the values of X in Eq. 9.21, the vector P∗ can be obtained. Now according to Eq. 9.22, multiplying this vector with S( 1, the displacement vector of the nodal forces of the near-regular structure can be obtained.

Δ ¼ f 25:8839

P∗ ¼ f10; 0; 0; 20; 1:2177; 0:6089gt 6:7609

12:6449

23:4314

8:3473 (3:0400 gt

It should be noted that the aim of this example was the explanation of the method by means of a simple example and for showing the capabilities of the presented method is not sufficient. The reduction in dimensions of the matrices achieved by the present method and the speed of calculation will be illustrated in Sect. 9.5.

9.5

Analysis of Regular Structures with Some Missing Members

In this section we consider those structures which need addition of some members to become a regular one. Obviously the method presented in the previous section can not be applied directly for these structures. However, for transforming this case to the previous one, a pair of members with equal modulus of elasticity having different signs, are added where we have lack of members for regularity. In the next step the members with negative modulus of elasticity are considered as excessive members and separated from the structure. The remaining process of the analysis is the same as the previous case. The internal forces of the members with negative modulus of elasticity are calculated and added to the external forces. Then the regular structure with the external loads together with the internal forces of the excessive members which are applied as the additional external forces, is analyzed. In the following a simple example is considered for further explanation.

9.5.1

Investigation of an Illustrative Simple Example

In this section a simple example is used to describe the process of the algorithm. In a 4-bar structure shown in Fig. 9.15a, it is obvious that if we add a member between the nodes 2 and 3, the structure will be transformed into Form II and one can easily calculate its inverse. Now we add two members 5 and 6 of identical properties have modulus of elasticity of different signs between the two nodes 2 and 3. This is

394

9 Analysis of Regular Structures Using Graph Products

Fig. 9.15 (a) An irregular structure. (b) The irregular structure with a pair of members being added. (c) Representation of the internal forces in the regular structure. (d) The added member with negative modulus

logical assumption because the property of one member can be nullified by the other member. Member 6 has negative modulus of elasticity and we consider it as a member separation of which transforms the structure into a regular one as shown in Fig. 9.15c. The structure obtained in this way is equivalent to the basic structure of the force method. From here onward all the previous steps can be employed. Figure 9.15d shows the excessive bar with negative modulus of elasticity which is separated from the truss shown in Fig. 9.15b. The external force vector will be as follows: P ¼ ½0 0

0

(10 't

For the formation of the equilibrium matrix of the structure shown in Fig. 9.15b, one can either use the equations corresponding to the equilibrium of the forces at the nodes, or alternatively use Eq. 9.16.

9.5 Analysis of Regular Structures with Some Missing Members

é1 ê0 AT = ê ê0 ê ë0

0

0

0.8944

0

395

0

ù 0 - 1 - 0.4472 0 0 úú 1 0 0 0.8944 0.8944ú ú 0 1 0 0.4472 0.4472û

The matrices A and N can be obtained by partitioning the above matrix according to Eq. 9.11 and numbering the internal force of the separated member. 2

1 60 A¼6 40 0

0 0 1 0

0 (1 0 1

0:8944 (0:4472 0 0

3 0 0 7 7 0:8944 5 0:4472

2

3 0 6 0 7 7 N¼6 4 0:8944 5 0:4472

;

The matrix S(1 corresponding to the regular structure shown in Fig. 9.15c, and the matrix s are as follows:

S(1

2

1:5935 6 2:0508 ¼6 4 (0:4065 1:9491

s ¼ diagf 0:5

2:0508 9:8338 (1:9491 9:3465 0:5

(0:4065 (1:9491 1:5935 (2:0508

1 0:4472

3 1:9491 9:3465 7 7 (2:0508 5 9:8338

0:4472 g

It can be seen that the matrix S can be transformed into Form II by multiplying the row and column 2 by (1. Using the above matrices and Eq. 9.8, the matrix R is obtained as: 2

0:7967 1:0254 6 (0:2032 (0:9745 6 R¼6 6 (0:1016 (0:4873 4 0:2272 (1:1464 0:2272 1:0896

(0:2032 0:7967 (0:1016 0:2272 0:2272

3 0:9745 (1:0254 7 7 0:4872 7 7 (1:0896 5 1:1464

Using Eq. 9.15, the matrices B0 and B1 are obtained as: 2

0:7967 6 (0:2032 6 6 (1:1016 B0 ¼ 6 6 0:2272 6 4 0:2272 0

1:0254 (0:2032 (0:9745 0:7967 (0:4873 (0:1016 (1:1464 0:2272 1:0896 0:2272 0 0

3 3 2 0:9745 0:2540 6 0:2540 7 (1:0254 7 7 7 6 7 6 0:4872 7 7; B1 ¼ 6 0:1270 7 6 (0:2840 7 (1:0896 7 7 7 6 4 0:7159 5 1:1464 5 0 1

The flexibility matrix F for the truss shown in Fig. 9.15b will be as follows:

396

9 Analysis of Regular Structures Using Graph Products

F ¼ diagf 2

2

2:2360 (2:2360 g

1 2:2360

Using Eq. 9.19 the following matrices are obtained: D1 ¼ ½ 0:5081

2:4364 0:5081

2:5635 '; D2 ¼ ½(0:6351'

The matrix X is calculated from Eq. 9.20 as X ¼ (D2 (1 :D1 P ¼ 1 ! ½ 0:5081 2:4364 0:5081 2:5635 ':½ 0 0 0 (10 't ¼ (40:3607 0:6351 Adding X to the vector of external loads according to Eq. 9.21 and multiplying the matrix S(1 employing Eq. 9.22 we will have: P∗ ¼ ½ 0 0 (36:0997 (28:0498 't Δ ¼ ½ (40 (191:803 0 (201:803 't Finally, using X and Eq. 9.14 one can find the internal forces of the structure shown in Fig. 9.15b as follows: Q ¼ ½ (20

0 (10

22:3607 (40:3607

(40:3607 't

One can recognize the equality of the internal forces of the entries 5 and 6.

9.6

Practical Examples

Here four examples are presented. The first two examples correspond to Sect. 9.3 and the third example belongs to Sect. 9.4. The fourth example corresponds to the combination of the methods presented in Sects. 9.3 and 9.4. The latter example is chosen as a frame structure to showing the applicability of the presented method to other skeletal structures other than trusses. Example 1. A truss with 47 members is considered in the form of a single layer rotational dome, having two members 26 and 27 making the truss a near-regular one, Fig. 9.16. If we remove these two members then the remaining regular structure can easily be solved using the method of Kaveh and Rahami [8]. The value of EA ¼ 1 N is assumed to be identical for all the members and the force P2z ¼ 10 N and P8y ¼ 20 N are applied at nodes 2 and 8, in z direction and y direction, respectively. For solution of this problem the members 26 and 27 are considered as excessive members. For the above near-regular structure the equilibrium matrix AT has dimension 30 ! 47 and by partitioning using Eq. 9.11, the matrices A and N with

9.6 Practical Examples

397

Fig. 9.16 The space dome of the Example 1 with 47 members

Fig. 9.17 (a) The regular structure obtained by deleting the excessive members. (b) The deformed shape of the structure of Example 1

dimensions 30 ! 45 and 30 ! 2 are obtained. The matrix S of the regular structure with dimension 30 ! 30 is formed by deleting the excessive members, shown in Fig. 9.17a, as follows: S¼

5 X i¼1

ðPi * Ai Þ

In this relation the matrices Ai and Pi are the submatrices constituting the matrix S. Using the method presented in Kaveh and Rahami [8] the inverse of the matrix S is formed by using the eigenvalues and eigenvectors of five 6 ! 6 matrices.

398

9 Analysis of Regular Structures Using Graph Products

s is a diagonal matrix of dimension 45 ! 45 containing the stiffness matrices of the members of the regular structure and F is a diagonal matrix of dimension 47 ! 47 consisting of the flexibility matrices of the near-regular structure . Therefore the matrices B0 and B1 of dimension 47 ! 30 and 47 ! 2 can be obtained using the Eq. 9.15. The matrices D1 and D2 of dimension 2 ! 30 and 2 ! 2 obtained by using Eq. 9.19 as follows: D1 ¼

"

1

3

14

... 5:6163 . . . (30:9438 . . . (62:3208 . . . (29:6975 & ' 426:7122 (156:126 D2 ¼ (156:126 267:2256

30

... ...

#

The vector of the internal forces of the excessive members X can be calculated from Eq. 9.20 as: X¼(inv h

)&

426:7122 (156:126

(156:126 267:2256 i 1 3 14 30 t 0 ... 10 ... 20 ... 0

'* " 1 ...

3

5:6163

14

30

... (30:9438 ...

... (62:3208 ... (29:6975 ...

#

X¼ f(1:9283,0:7148gt

The matrix of the internal forces of the near-regular structure can be obtained by substituting X in Eq. 9.14 as follows: h it 1 3 44 2 4 45 46 47 Q ¼ 8:7095 0:1002 (0:6626 (2:1497 . . . (0:6706 0:4741 (1:9283 0:7148

We substitute X in Eq. 9.21, and substitute the vector of the equivalent external forces of the regular structure in Eq. 9.22. The vector of the displacements for the near-regular structure can then be obtained using the inverse of the stiffness matrix of the regular structure. h it 3 4 27 28 29 30 1 2 Δ ¼ (70:6421 (47:0104 399:9523 117:5682 . . . (9:5185 110:8598 56:1357 90:5555

The deformed shape of the structure is shown in Fig. 9.17b. If we solve the structure by a conventional method we have to find the inverse of a matrix of dimension 30 ! 30, while the present approach requires the inverse of 5 matrices of dimension 6 ! 6 and the inverse of the matrix D1 of dimension 2 ! 2. Example 2. A communication space tower studied in Kaveh and Rahami [8], is considered in here. This model can be expressed as the product two graphs. In practice the towers have horizontal belts employed in different heights. These belts make their model irregular. In here we want to analyze those towers which have

9.6 Practical Examples

399

Fig. 9.18 A communication space tower with five pairs of excessive members being shown at the central core of the structure

belts at the central core of the structure (Fig. 9.18). If we consider these members as additional ones, we will obtain a regular structure. Such a regular structure can be generated by rotation of one of its faces. Using the force method the internal forces of the excessive members will be calculated and together with other external loads will be applied to the regular structure. This tower has 84 nodes and 330 members. The load applied to the structure is P ¼ 1 kN applied in all DOFs of the 4 upper nodes of the tower. The value of EA ¼ 100 N for all members is considered to be identical. The structure has 80 free nodes and 330 members and 10 members belong to the belt of the structure.

400

9 Analysis of Regular Structures Using Graph Products

Forming the equilibrium matrix of the near-regular structure which is of dimension 240 ! 330, and its partitioning by employing Eq. 9.11, the matrices A and N of dimensions 240 ! 320 and 240 ! 10 can be obtained. The stiffness matrix S of the regular structure is of 240 ! 240 and has the following form: S¼

4 X i¼1

ðPi * Ai Þ

Using the method presented in Kaveh and Rahami [8], the inverse of the stiffness matrix of the regular structure can be obtained using the eigenvalues and eigenvectors of four 60 ! 60 matrices. In this example, since the structure is truss, s is a diagonal matrix of dimension 320 ! 320 containing the stiffness matrices of the members of the regular structure. Having the above mentioned matrices and using Eq. 9.15 one can easily form the B0 and B1 matrices which are of dimension 330 ! 240 and 330 ! 10. For this structure, the flexibility matrix F of the nearregular structure has dimension 330 ! 330. Having the matrices B0, B1 and F, the matrices D1 and D2 of dimensions 10 ! 320 and 10 ! 10 are obtained from Eq. 9.19. Here D2 is as follows: 2

0:2999 6 (0:1033 6 6 0:2226 6 6 (0:2225 6 6 0:3121 D2 ¼ 6 6 (0:3121 6 6 0:4017 6 6 (0:4017 6 4 0:4912 (0:4912

(0:1033 0:2999 (0:2225 0:2226 (0:3121 0:3121 (0:4017 0:4017 (0:4912 0:4912

0:2226 (0:2225 2:9468 (2:7704 5:5850 (5:5850 8:3715 (8:3715 11:1580 (11:158

(0:2225 0:2226 (2:7704 2:9468 (5:5850 5:5850 (8:3715 8:3715 (11:158 11:1580

0:3121 (0:3121 5:5850 (5:5850 15:3371 (15:183 25:2968 (25:296 35:3849 (35:385

(0:3121 0:3121 (5:5850 5:5850 (15:1836 15:3371 (25:2968 25:2968 (35:385 35:3849

0:4017 (0:4017 8:3715 (8:3715 25:2968 (25:296 48:2162 (48:0854 71:7608 (71:760

(0:4017 0:4017 (8:3715 8:3715 (25:296 25:2968 (48:085 48:2162 (71:760 71:7608

0:4912 (04912 11:1580 (11:158 35:3849 (35:385 71:7608 (71:760 116:622 (116:514

3 (0:4912 0:4912 7 7 (11:1581 7 7 11:1581 7 7 (35:385 7 7 35:3849 7 7 (71:760 7 7 71:7608 7 7 (116:514 5 116:622

In this way and using Eq. 9.20, the vector of internal forces of the excessive members can be obtained as X ¼ f(84:7178, ( 84:7178, ( 85:0153, ( 85:0153, ( 106:172, ( 106:172, ( 133:071, ( 133:071, ( 163:915, ( 163:915gt

Adding X to the external load vector using Eq. 9.21 and applying the load to the regular structure in Eq. 9.22, the vector displacements for the near-regular structure is obtained. The nodal displacements in some DOFs are as follows: h 61 Δ ¼ . 1. . 9513:052

121

. . . 31468:4

181

. . . 65318:71

The deformed shape of the structure is shown in Fig. 9.19.

240 ...

it

9.6 Practical Examples

401

Fig. 9.19 A communication transmission tower together with its deformation

For solution of this structure using a conventional stiffness method we have to find the inverse of 240 ! 240, while the present approach requires the inverse of four matrices of dimension 60 ! 60 and the inverse of the matrix D1 of dimension 10 ! 10 to complete the analysis of the near-regular structure.

402

9 Analysis of Regular Structures Using Graph Products

Fig. 9.20 (a) Two and three dimensional representations of the 43-bar structure. (b) The corresponding regular structure [1]

Example 3. Consider a 43-bar truss structure shown in Fig. 9.20a. This structure becomes a cyclically symmetric structure by addition of two members between the nodes 12 and 14, and nodes 11 and 13. A pair of members with identical geometry and equal modulus of elasticity having different signs, are added where we have lack of members for regularity. In the next step the members with negative modulus of elasticity are considered as excessive members are separated from the structure. The external forces consist of P2z ¼ P8y ¼ 10 N. For all the member we consider EA ¼ 1 N. Forming the equilibrium matrix AT of the near-regular structure according to Eq. 9.16, which is of dimension 30 ! 47, and its partitioning by employing Eq. 9.11, the matrices A and N of dimensions 30 ! 45 and 30 ! 2 are obtained. It should be noted that the matrix AT corresponds to the near-regular structure which has both members of positive and negative modulus of elasticity.

9.6 Practical Examples

N¼

"

1

403

19

27

30

0 . . . 0:0248 0:8650 0:5010 0 0 0 0 0 0 ... 0 0 ... 0 0 0 (0:1375 0:8347 0:5332 0:1375 (0:8347 (0:5332 . . . 0

#t

The matrix S corresponding to Fig. 9.20b can be expressed as S¼

5 X i¼1

ðPi * Ai Þ

Thus as we observed in Example 1, the inverse of the stiffness matrix which is of dimension 30 ! 30 can be calculated by evaluating the eigenvalues and eigenvectors of five matrices of dimension 6 ! 6. The diagonal matrix s is of dimension 45 ! 45 containing the stiffness matrices of the members of the structure shown in Fig. 9.20b. The diagonal matrix F is of dimension 47 ! 47 containing the flexibility matrices of the members of the regular structure together with the excessive members. Using Eq. 9.15 one can easily form the B0 and B1 matrices which are of dimension 47 ! 30 and 47 ! 2. Having the matrices B0, B1 and F, the matrices D1 and D2 of dimensions 2 ! 30 and 2 ! 2 are obtained from Eq. 9.19. Here D2 is as follows: D2 ¼

&

(19:4620 1:8186

1:8186 (10:7629

'

Equation 9.20 can be employed to find the vector X as: X ¼ (D2 (1 D1 P # & '(1 " 14 3 h i (19:4620 1:8186 1 3 14 30 t . . . 0:3207 . . . 1:3854 . . . ¼( 0 . . . 10 . . . 10 . . . 0 1:8186 (10:7629 . . . (0:0640 . . . (0:3018 . . .

X¼

&

0:8584 (0:1949

'

Using Eq. 9.21 the equivalent external forces of the regular structure are obtained, and using the inverse of S one can easily find the nodal displacements of the near-regular structure by Eq. 9.22. h i 1 3 14 19 27 30 t P∗ ¼ 0 . . . 10 . . . 10 . . . 0:0213 0:7426 0:4301 0:0268 (0:1627 (0:1039 (0:0268 0:1627 0:1039 . . . 0

h it 1 29 30 2 Δ ¼ (105:512 (134:943 380:134 51:383 . . . (18:489 50:4185 (75:205 116:529

The internal forces of the regular structure and the excessive members can be found using Eq. 9.14 as

404

9 Analysis of Regular Structures Using Graph Products

Fig. 9.21 (a) A 24-story irregular frame with bracing. (b) The regular part of the irregular frame with fictitious columns of positive modulus of elasticity being added. (c) The bracing part consisting of 32 bracing elements and 49 fictitious bending elements with negative modulus of elasticity being added as shown at the top of the structure [1] h it 5 6 44 46 1 2 3 43 45 47 Q ¼ 6:8906 (0:9281 1:7904 (0:950 (5:60 . . . (0:0003 0:8584 (0:1949 0:8584 (0:1949

As it can be seen, the internal forces in members 44, 46 and 45, 47 which are the added pairs of members are the same as the entries of X. For solution of this near-regular structure using a conventional stiffness method we have to find the inverse of 30 ! 30, while the present approach requires the inverse of 5 matrices of dimension 6 ! 6 and the inverse a matrix of dimension 2 ! 2 to complete the analysis of the near-regular structure . Example 4. A 24-story 3D frame is shown in Fig. 9.21a, with 49 columns and 84 beam in each story. The dimensions of all the beams and columns are assumed to be identical in all the stories. In each face of the building 8 bracing elements are added to increase the stiffness of the structure. Naturally these elements make the model irregular. Here using the presented method, the bracing elements are decomposed from the structure the analysis is performed for two separate parts, namely the regular bending frame and the excessive bracing elements.

9.6 Practical Examples

405

If the top part of the structure is fixed similar to the bottom part, then the bending frame structure can be easily analyzed using the method presented in [3]. Therefore here we consider pairs of bending elements with positive and negative modulus of elasticity at end part of the structure, similar to the columns of the other stories, as illustrated in Fig. 9.21b. The columns are connected to the top part of the structure and fixed at the other ends. The elements with negative modulus of elasticity are considered as excessive members and are separated from the structure. Therefore the excessive members consist of 32 bracing elements and 49 bending elements with negative modulus of elasticity. These elements are highlighted in Fig. 9.21c. In this way, the regular structure consists of a 24 story frame together with additional bending elements with positive modulus of elasticity as illustrated in Fig. 9.21b. The required parameters for the analysis are as k ¼ 7, 056, t ¼ 326, and e ¼ 19, 446. Using Eqs. 9.16 and 9.17 and employing the rotation and stiffness matrices of the elements in their local coordinate systems, the matrix AT of dimension 7056 ! 19772 can be constructed. For the formation of the stiffness matrices of the elements, the local coordinate systems should be selected such that the form given in Eq. 9.17 is formed. For each bending elements, six internal forces, and for bracing elements only one axial force are assumed. It should be noted that the fictitious elements with + modulus of elasticity contribute in the formation of this matrix. By partitioning the matrix AT we obtain two matrices A and N having dimensions 7056 ! 19446 and 7056 ! 326, respectively. Since the regular part contains 3,241 bending elements, thus the unassembled stiffness matrix s is of dimension 19446 ! 19446. The assembled matrix S of the regular part has dimension 7056 ! 7056. Utilizing the method of Ref. [3], the inverse of this matrix can easily be obtained calculating the eigenvalues of 24 matrices of dimension 294 ! 294. In this way forming the inverse of the stiffness matrix and using Eq. 9.8, the matrix R of dimension 19446 ! 7056 can be obtained. Having this matrix the matrices B0 and B1 of dimensions 19772 ! 7056 and 19772 ! 326 will be formed using Eq. 9.15. The flexibility matrix F contains the flexibility of all the elements of the near-regular structure (bending elements, bracing elements, and pair of fictitious elements with + and ( signs). This matrix is a block matrix such that for the bending members blocks are 6 ! 6 and for the bracing members the blocks are 1 ! 1. Thus the dimension of F is 19772 ! 19772. With help of Eq. 9.19 the matrices D1 and D2 of dimensions 326 ! 7056 and 326 ! 326 are obtained, respectively. Using Eq. 9.20 and finding the inverse of D2 leads to the vector of unknown X of dimension 326 ! 1. Substituting this in Eq. 9.21, the equivalent external force vector of the regular part of the structure is obtained. Multiplying the inverse of the stiffness matrix of the regular part, the displacement vector of the near-regular structure of dimension 7056 ! 1 is obtained. It can be observed that the analysis of the problem with the help of this method for frame structures is the same as that of the trusses which were discussed in the previous examples, with the only difference that the rotation and stiffness matrices

406

9 Analysis of Regular Structures Using Graph Products

for the bending elements in the local coordinate systems should be defined according to the Eq. 9.17. In this problem instead of inverting the stiffness matrix of dimension 7056 ! 7056 in direct analysis of the near-regular structure, one needs to find the inverse of the matrix D2 of dimension 326 ! 326, and calculate the eigenvalues of 24 matrices of dimension 294 ! 294. This shows the efficiency of the present method. Obviously increasing the number of stories this efficiency will become more apparent. In other words in this method a matrix of dimension 7056 is decomposed into 24 matrices of dimension 294.

References 1. Kaveh A, Rahami H, Mirghaderi SR, Ardalan Asl M (2013) Analysis of near-regular structures using the force method. Eng Comput 30:21–48 2. Imrich W, Klavzˇar S (2009) Product graphs; structure and recognition. Wiley, New York 3. Kaveh A (2013) Optimal analysis of structures by concepts of symmetry and regularity. Springer/GmbH, Wien/New York 4. Sabidussi G (1960) Graph multiplication. Math Z 72:446–457 5. Gross J, Yellen J (1998) Graph theory and its applications. CRC Press, New York 6. Weichsel PM (1962) The Kronecker product of graphs. Proc Am Math Soc 13:47–52 7. Argyris JH, Kelsey S (1960) Energy theorems and structural analysis. Butterworth, London 8. Kaveh A, Rahami H (2011) Block circulant matrices and applications in free vibration analysis of cyclically repetitive structures. Acta Mech 217:51–62

Chapter 10

Simultaneous Analysis, Design and Optimization of Structures Using Force Method and Supervised Charged System Search

10.1

Introduction

Developing methods with higher computation efficiency is a crucial subject in advanced engineering problems of multi-physics nature. For instance, analyzing structures with larger number of members requires larger memory size and longer computation time. In addition, this costly computation has to be repeated many times, typically over 5,000 times, because the cross section size of the members is not determined in the early stages of designing such structures. Therefore, reducing the size of structural matrices and eliminating the unduly repetitions in the design and analysis procedures can lead to a considerable reduction in the computation efficiency [1, 2]. In this chapter, this goal is achieved utilizing meta-heuristics algorithms which minimize the energy function indirectly. Besides, design procedure and minimizing the weight of the structure is added to the analysis procedure. One of the most reliable meta-heuristic methods recently developed is Charged System Search (CSS) [3, 4], that is used in here. In this chapter, supervisor agents are considered to increase the exploration ability of the CSS algorithm. This method is called supervised CSS abbreviated as SCSS. Also a new formulation of the penalty function is made to improve the performance of the supervised CSS. Designing structures with minimum weight can be achieved by using minimum energy methods, and members with pre-defined stress ratios [5], instead of the direct solution of classic equations. This results in avoiding not only the repetitive computations in the design and analysis, but also avoiding the computation of the solution of equations with large matrices. For this purposed, one needs to formulate the equations based on the minimum energy principle, and employ them in an efficient optimization algorithm. Combining the SCSS algorithm and the force method provides a suitable means for this purpose. The former is a suitable optimization algorithm and the latter can be used to derive the energy equations. In the first part of this chapter, supervisor agents are introduced. In the second part energy formulation based on the force method is derived and the supervised SCSS algorithm is applied to the analysis procedure. In the third part, using the A. Kaveh, Computational Structural Analysis and Finite Element Methods, 407 DOI 10.1007/978-3-319-02964-1_10, © Springer International Publishing Switzerland 2014

408

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

SCSS and prescribed stress ratios, structures are analyzed and designed, and finally in the last part weight minimization is performed by imposing the analysis procedure as a constraint to the SCSS. In recent years the CSS has been applied successfully to many engineering optimization problems. For optimal design of structures, CSS has performed very well and improved all of the resulted design parameters and weights achieved by the other algorithms. Large-scale structures are analyzed and designed in this chapter in order to show the accuracy of the method when applied to different kinds of structures.

10.2

Supervised Charged System Search Algorithm

In the CSS algorithm, each vector of variables is an agent that moves through the search space and finds the minimal solutions [3, 4]. Throughout the search process, an agent might go to a coordinate in the search space that already has been searched by the same agent or another. If this coordinates have a good fitness, it will be saved in the Charged Memory [3] but if this coordinate does not have a good fitness, it will not be saved anywhere. Therefore, this step of the search process becomes redundant. This unnecessary step adversely affects the exploration ability of the algorithm. In this chapter, the supervisor agents are introduced to improve the exploration ability of the CSS algorithm. The supervisor agent is an independent agent of constant values that repels the agent if its coordinate has a bad fitness or attracts the agents if its coordinate has a good fitness. This procedure is repeated in all of the iterations and gives an overall view of the search space. The number of supervisor agents is selected at the beginning of the algorithm, and then their constant coordinates in the search space are determined as follows: xsj, i

! " ði " 1Þ xmax, j " xmin, j þ xmin, j ¼ NOSA " 1

ð10:1Þ

where NOSA is the number of supervisor agents, and xsj,i is the jth variable of the ith supervisor agent; xmin,j and xmax,j are the minimum and the maximum limits of the jth variable. The kind of the force for these agents is determined as p ¼ log

#

fit fiti

$

ð10:2Þ

where p is the same as the parameter in the original version of the CSS [3], fiti is equal to the fitness value of the ith supervisor agent and fit is the average value of the fitness of the normal agents. Calculating other properties of the supervisor agents such as force and radius are similar to the standard CSS algorithm [3]. Supervisor agents do not move from their coordinate determined from Eq. 10.1, yet they apply additional forces on the normal agents. By doing so, they determine the fitness

10.3

Analysis by Force Method and Charged System Search

409

values of their fixed coordinate and its neighborhood, resulting in a better exploration ability of the CSS algorithm.

10.3

Analysis by Force Method and Charged System Search

In the presented approach, force method is applied to analyze structures. Since this method leads to less number of unknowns, it is preferred to displacement method. In the force method, the redundant forces are unknowns, whereas in the displacement method, the nodal displacements are unknowns. In this method [1, 2, 5], the energy relationships of the structure that satisfies the compatibility, forcedisplacement and equilibrium conditions are derived, and then, minimized using the SCSS. Suppose {p} ¼ {p1,p2,. . .,pn}t is the vector of nodal forces, {q} ¼ {q1, q2,. . .,qn}t is the vector of redundant forces, and {r} ¼ {s1,s2,. . .,sm}t comprises of the internal forces of the members. Equilibrium condition results in the following equation [1, 2]: r ¼ B0 p þ B1 q ¼ ½ B0

% & p B1 ' q

ð10:3Þ

In addition, the complementary energy function is: Uc ¼

1 t r Fm r 2

ð10:4Þ

where [Fm] is the unassembled flexibility matrix of the structure. According to the Castigliano’s principle, a group of the redundant forces that minimize the complementary energy function is the exact solution that satisfies compatibility condition. By substituting {r} from Eq. 10.3 in Eq. 10.4, the following equation obtained: 1 U ¼ ½ pt 2 c

where ½H' ¼ ½ B0 B1 't ½Fm '½ B0 submatrices leads to: Uc ¼

% & p q '½H' q t

ð10:5Þ

B1 '. Decomposing matrix [H] into four

! " ! " ! " ! " ( 1' fpgt Hpp fpg þ fpgt Hpq fqg þ fqgt Hqp fpg þ fqgt Hqq fpg 2 ð10:6Þ

In the classical method, the derivative of Uc in terms of {q} is calculated and is equated to zero leading to:

410

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

! ""1 ! " Hqp fpg fqg ¼ " Hqq

ð10:7Þ

! " Fu ¼ fqgt Hqp fpg

ð10:8Þ

Since [H] is symmetric, [Hqp]t ¼ [Hpq], Ref. [5]. Accordingly, in the classical method the inverse of [Hqq] needs to be calculated. This is a difficult task, and requires extensive computer memory, especially in the case of large scale structures. Therefore, finding {q} that minimizes the complementary energy without calculating the inverse of [Hqp] reduces the computation time and computer memory. The first term of Eq. 10.6 is constant and the second and third terms are equal. It can be shown that the third and fourth terms of Uc are symmetric. Therefore

is the equation that should be minimized [5]. Enhanced Charged System Search [4] is used to minimize Eq. 10.8. In this part, the force method analysis is applied to different types of structures to illustrate the performance of the method. Case Study 1. The first example is an 11-member truss with three degrees of statical indeterminacy, as shown in Fig. 10.1. Consequently, the energy function includes three variables. The classical method that calculates the exact and minimum amount of Uc leads to 419.8475, whereas, using the present approach with CSS, Uc ¼ 419.8476 is obtained and {q} is calculated as: fqg ¼ f4:6394 " 3:7629 8:1900gt The optimization history is shown in Fig. 10.2. The number of agents is selected as 20. Case Study 2. The second example is an unbraced planar frame with constant EI having 36( of statical indeterminacy, as shown in Fig. 10.3. In this example, the axial force, shear and moment in the first node of the beams are considered as the redundant forces. As a result, the energy function includes 36 variables. Note that only the bending energy is considered as the energy of the frame. Loading condition is considered as: 1. A load "10 kN in the y-direction at nodes 8–11, 2. A load 10 kN in the x-direction at nodes 8–11, 3. A bending moment 10 kN.m in the x-y surface at nodes 8–11. The exact calculation of Uc leads to 1,234.8; while it is Uc ¼ 1,249.2 utilizing the CSS algorithm. Figure 10.4 shows the variation of FU versus the number of iterations. As shown above, there is a very close agreement between the exact and the calculated value for the energy function, verifying the accuracy of the algorithm. In this case, the redundant forces are obtained as follows:

10.3

Analysis by Force Method and Charged System Search

411

Fig. 10.1 A simple truss and the selected basic structure (Case Study 1): (a) A planar truss. (b) The selected basic structure

Fig. 10.2 Variation of FU versus the number of iterations in the 11-member truss (Case Study 1)

{q} ¼ {1.1275,5.3155,14.0096,2.4854,4.8316,12.0549,4.0405,4.2845, 10.7913, "3.0551,1.2459,2.9740,"4.0016,1.3874,3.2303,5.5762,1.4122,1.3221,0.0660, 0.2315,0.4707,0.1680,0.2155,0.4678,0.4265,0.1987,0.2503,"0.1444,0.0425, "0.0728, 0.0540,0.0052,0.0351,0.0373,0.0847, 0.0901}t Case Study 3. In the third example, a 40-element grilling system is considered to illustrate the accuracy of the force method and CSS in analyzing space frames. Geometry, nodal loads and basic structure are shown in Fig. 10.5. Torsion and shear in z direction, and moment around the axis with a greater moment of inertia in each member are considered as redundant forces. Both the torsion and bending energies are considered as energy function in this structure. G, I and E are constant for members and the Poisson’s ratio (υ) is considered 0.3. The cross-sections of members are considered to be 272 W-section as given in LRFD-AISC. Using the least square regression, the polar moment of inertia (J) is expressed as a function of the moment of inertia (I):

412

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

Fig. 10.3 An unbraced planar frame (Case Study 2)

Fig. 10.4 Variation of FU versus the number of iterations in the unbraced planar frame analysis (Case Study 2)

10.3

Analysis by Force Method and Charged System Search

413

Fig. 10.5 A 40-element grillage (Case Study 3). (a) Geometry. (b) Node and element ordering. (c) Basic structure

J ¼ 1:04I

ð10:9Þ

E ¼ 2Gð1 þ υÞ

ð10:10Þ

Also

By substituting Eqs. 10.9 and 10.10 in [Fm], the energy function is derived. The exact calculation of energy using the classical method leads to 170,840, whereas, using the present approach Uc ¼ 177,460 is obtained. The redundant forces, {q}, are shown in Table 10.1. Case Study 4. The Last example of this part is a 26-story tower with 246( of statical indeterminacy selected from Ref. [6], as shown in Fig. 10.6a, b. The energy function has 246 unknowns. The cross section and module of elasticity for all of the elements are considered constant and equal. Geometry and basic structure is shown in Fig. 10.6c. The loading on the structure consists of: 1. The vertical load at each node in the first section is equal to "3 kips ("13.344 kN) 2. The vertical load at each node in the second section is equal to "6 kips ("26.688 kN) 3. The vertical load at each node in the third section is equal to "9 kips ("40.032 kN)

414

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

Table 10.1 The calculated redundant forces of 40-element grilling system (Case Study 3) ) 104

q1 q2 q3 q4 q5 q6 q7 q8 q9 q10 q11 q12 q13 q14 q15 q16 q17 q18

"0.914 0.2167 "3.9005 "0.6323 0.314 "0.3381 0.1307 "0.0469 2.8322 0.4806 "0.3335 2.1219 "0.7939 0.2277 3.3177 0.1725 "1.4645 "0.8168

q19 q20 q21 q22 q23 q24 q25 q26 q27 q28 q29 q30 q31 q32 q33 q34 q35 q36

"0.0084 "1.8335 0.7346 "1.0314 3.6083 0.0769 "0.0497 0.0678 5.0685 1.0572 "0.1714 5.5207 0.4753 "4.0345 0.0442 "0.3564 "3.7443 0.055

q37 q38 q39 q40 q41 q42 q43 q44 q45 q46 q47 q48 q49 q50 q51 q52 q53 q54

"0.0312 "5.1336 "0.0287 0.5316 "1.9493 0.0136 "0.0397 0.0061 4.5725 "0.2432 "1.6436 0.296 1.2002 "5.6626 0.1194 1.1286 "5.547 "0.17

q55 q56 q57 q58 q59 q60 q61 q62 q63 q64 q65 q66 q67 q68 q69 q70 q71 q72

0.7119 "4.0377 "0.2541 0.0398 "6.1707 2.1362 0.1051 "3.0445 1.9832 "0.0718 0.2401 1.3579 0.0941 "2.4965 "0.2361 "0.8848 "3.9475 0.2642

4. The horizontal load at each node on the right side in the x direction is equal to "1 kips ("4.448 kN) 5. The horizontal load at each node on the left side in the x direction is equal to 1.5 kips (6.672 kN) 6. The horizontal load at each node on the front side in the y direction is equal to "1 kips ("4.448 kN) 7. The horizontal load at each node on the back side in the y direction is equal to 1 kips (4.448 kN) In this example, the exact calculation of the energy function leads to 1.8008 ) 107, and it is obtained as 1.8252 ) 107 using the force method and CSS that is very close to the exact value.

10.4

Procedure of Structural Design Using Force Method and the CSS

In this section, design and optimization procedures are added to the analysis presented in the previous section. There are two major approaches to formulate the objective function in the simultaneous analysis and design of an optimal structure: 1. Using the pre-selected stress ratio. 2. Minimizing the structure weight.

10.4

Procedure of Structural Design Using Force Method and the CSS

415

Fig. 10.6 A 26-story tower. (a) Geometry and grouping. (b) Top view. (c) Basic structure (Case Studies 4 and 10)

10.4.1 Pre-selected Stress Ratio In this approach [5], a preselected stress ratio is assumed for each member, and then the complementary energy is minimized as the objective function. If the cross sections Ai (i ¼ 1,. . .,m) are known, then the analysis can be performed using a meta-heuristics method such as CSS, described in the Sect. 3.

416

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

However, usually the cross sectional areas are not determined at the beginning of the design procedure. This problem leads to a new formulation of the complementary energy that eliminates Ai (i ¼ 1,. . .,m) from the energy function [5]. Each agent in the CSS is a vector of redundant forces. Moreover, according to Eq. 10.3, the internal forces of members, {r}, is obtained from the selected agents. The ratio between the stress in each member (σ i) and its corresponding allowable stress (σ a) is defined as C: C¼

σi σa

ð10:11Þ

where σ i ¼ Arii . By substituting σ i in Eq. 10.11, the cross section area of each member is obtained in terms of the internal force ri, stress ratio C, and the allowable stress σ a Ai ¼

ri Cσ a

ð10:12Þ

Consequently, one can express the unassembled flexibility matrix of each member as a function of L, E, q and C as follows: Fm ¼

L 1 ¼ ¼ gðq; C; L; EÞ EA Ef ðr; L; CÞ

ð10:13Þ

Substituting Fm in Eq. 10.4, leads to the elimination of Ai from the formulation of the complimentary energy: MinUc ¼

1 ½p 2E

q 't ½ B0

B1 't ½gðq; C; LÞ'½ B0

B1 '½ p

q'

ð10:14Þ

Pre-selected stress ratio is a parameter controlling the weight of the structure and stress constraint, simultaneously. Therefore, by minimizing the energy function in the analysis procedure, weight optimization and stress constraints satisfaction are fulfilled. Case Study 5. As an example consider the truss shown in Fig. 10.7. This truss is designed with the constraints explained in Table 10.2 and using Eq. 10.14 as the objective function. In this example, two cases are considered. In case I, the stress ratios of the members is different, whereas in case II, it is assumed to be constant for all the members. For the sake of simplicity, the cross-sections are selected as hollow squares, as shown in Fig. 10.8. In this example, a population of 20 agents is considered in the CSS algorithm. The magnitude of Ai is determined considering the selected values of Ci. Enhanced CSS with supervisor agent is utilized in the simultaneous analysis and design of this structure and the results are shown in Tables 10.3 and 10.4. The convergence history is shown in Fig. 10.9. To verify the efficiency of the present method and combining the CSS algorithm and force

10.4

Procedure of Structural Design Using Force Method and the CSS

417

Fig. 10.7 A simple truss with pre-selected stress ratios (Case Study 5). (a) Geometry. (b) Basic structure

method in minimizing the structural weight, the design parameters and redundant forces obtained from CSS, are compared to those computed using the Genetic Algorithm (GA), reported by Kaveh and Rahami [5]. The comparison results are shown in Tables 10.3 and 10.4 for Case I and Case II, respectively. In this example, the exact calculation of the energy function leads to 6.5989 ) 105, and it is obtained as 6.6056 ) 105 using the force method and CSS for case I. Besides, the exact calculation of the energy function leads to 7.5368140 ) 105, and it is obtained as 7.5368147 ) 105 using the force method and CSS for case II. The close agreement between these values verifies the accuracy of the calculated redundant forces shown in Tables 10.3 and 10.4 for case I and case II, respectively. Also variation of FU versus the iteration is shown in Fig. 10.9.

10.4.1.1

Fully Stress Design (FSD) for Statically Indeterminate Structures

In this part, the presented CSS and force method is applied to an Optimally Criteria Method (OCM), namely Fully Stress Design (FSD). FSD leads to a correct optimal weight for statically determinate structures under a single load condition. In the FSD all the members are supposed to be subjected to their maximal allowable stresses [5]. Achieving such a design for an indeterminate structure with fixed geometry is not always possible. Even by changing the geometry, a FSD may not be achieved. Here a formulation presented by Kaveh and Rahami [5] is used to indirect analysis in the process of optimization. This formulation can be applied to all types of structures, however, a truss with the following strain energy is considered: Uc ¼

X P2 L EA

¼

X γP2 LA γEA2

¼

1 X 2 σi wi γE

ð10:15Þ

It should noted that for constant E and γ, the minimum weight can be achieved only when the stresses in all the members are identical. Therefore, in Eq. 10.15, the term corresponding with the stresses, i.e. σ 2i , may be moved out of the summation. On the other hand, in the design procedure, one can consider the fully stress

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

418

Table 10.2 Design data for the 11-bar planar truss (Case Study 5) Design variables Redundant and size variables q1; q2; q 3; A1; A2; A3; A4; A5; A6; A7; A8; A9; A10; A11 Material and section property Young’s modulus is assumed to be constant Density of the material: ρ ¼ 0.00277 kg/cm3 ¼ 0.1 lb/in3 pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ A ¼ 0:4h2 , r ¼ 0:4A, thicknesst ¼ 0:1h: Constraint data Stress ratios Case 1: C ¼ {0.9, 0.8, 0.85, 0.8, 0.9, 0.85, 0.95, 0.9, 0.8, 0.9, 0.95} Case 2: ci ¼ 1; i ¼ 1, . . ., 11 For tensile members Fa * 0.6 Fy and λi * 300 For compressive members λi * 200 h* + i Fa ¼ *

1"

λi 2C2 c

Fy

λ3 5 3λi i 3þ8Cc "8C3 c 2

E Fa ¼ 12π 23λ2 i

+

for

for λi * Cc

λi * Cc

Stress constraints σ i < 234.43 MPa; i ¼1, . . ., 11 Fig. 10.8 A hollow square cross-section (Case Study 5)

Table 10.3 Optimal design comparison for the 11-bar truss (Case Study 5) (case 1) Weight (N) 2,136.25 Size variable(cm2) A2 A3 A4 A1 11.55 13.36 41.20 4.44 Redundant variables )103 (N) q2 q1 123.04 "5.04

A5 4.44 q3 244.69

A6 42.51

A7 6.94

A8 9.15

A9 61.02

A10 9.71

A11 17.51

10.4

Procedure of Structural Design Using Force Method and the CSS

419

Table 10.4 Optimal design comparison for the 11-bar truss (Case Study 5) (case 2) Weight (N) 1,914.84 Size variable (cm2) A1 A2 A3 A4 11.55 13.36 41.20 4.44 Redundant variables )103 (N) q1 q2 94.04 "0.0000541

A5 4.44

A6 42.51

A7 6.94

A8 9.15

A9 61.02

A10 9.71

A11 17.51

q3 198.66

Fig. 10.9 Variation of FU versus the iteration in the design procedure for the 11-member truss (Case Study 5)

constraint instead of minimum weight. This is because the minimum weight corresponds to a structure for that all the members are subjected to their maximum allowable stress. Case Study 6. As an example, consider the structure shown in the Fig. 10.10, selected from Ref. [7]. The design and member size constraints are reported in Table 10.5. Redundant forces in this example are selected as internal forces in members 1 and 9. Twenty agents are selected in the CSS algorithm.

420

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

Fig. 10.10 A 10-bar truss example (Case Studies 6 and 7, Ref. [7])

Table 10.5 Design data for the 10-bar planar truss (Case Study 6) Loading Node Px: kips (kN) Py: kips (kN) 2 0 "100("444.8) 4 0 100("444.8) Design variables Variables: q1; q2 (and A1;A2;A3;A4;A5;A6;A7;A8;A9;A10 in case 3) Material property and constraint data Young’s modulus: E ¼ 1e7 psi ¼ 6.895e7MPa Density of the material: ρ ¼ 0.1 lb/in3 ¼ 0.00277 kg/cm3 For all members: Ai + 0.1 in2; i ¼ 1, . . ., 10 Stress constraints (a) FSD Case 1: |σ i| * 25 ksi(172.375 MPa); i ¼1, . . ., 10 Case 2: |σ i| * 25 ksi; i ¼ 1, . . ., 8, 10 and |σ 9 | * 50 ksi (344.75 MPa) (b) Weight minimization Case 3: |σi| * 25 ksi; i ¼ 1, . . ., 8, 10 and |σ9| * 50 ksi (344.75 MPa)

10.5

Pz: kips (kN) 0 0

Minimum Weight

In the second approach of simultaneous design and analysis of structures, the objective function is the weight of the structure, and the equilibrium, compatibility, and force/displacement conditions are the constraints. In summary, all these three conditions are called analysis criteria for simplicity. Other constraint such as stress, displacement, dynamical properties, and etc. can also be imposed to the fitness function. Penalty function is the most common approach to satisfying the constraints. The penalty function imposes a penalty to the fitness value of the solution, if the constraint is not satisfied: f ¼ A þ αB

ð10:16Þ

In Eq. 10.16, f is the fitness value, A is the objective function and B is the penalty

10.5

Minimum Weight

421

function and α is often selected as a big number. According to this equation, when B goes to zero and A goes to its minimum value, f goes to the minimum value of the fitness. However, since the minimum complementary energy is not zero, this form of penalty function cannot be used. In this case, W is minimum while the corresponding Uc is not minimum, i.e. the structure is not analyzed yet. Also a small value of α does not guarantee the minimum value of the B. On the other hand, in a structure that is in equilibrium and compatibility state, sum of the complementary energy Uc and the strain energy U is zero. Therefore, instead of the complementary energy, the sum of the complementary energy and the strain energy is used as the analysis criteria and is imposed to CSS as a constraint. The strain energy is a function of nodal displacements as follows: fdg ¼ ½B0 't ½Fm 'ð½B0 'fpg þ ½B1 'fqgÞ

ð10:17Þ

1 U ¼ fdgt ½K'fdg " fdgt fFg 2

ð10:18Þ

and

where [K] is the stiffness matrix and {F} is the nodal force vector. For equilibrium, U is negative and U + Uc is equal to zero. This formulation is used for the 10-bar truss example (Case Study 6) of Case III. Table 10.6 shows the results. Twenty agents are selected in the CSS algorithm. Also the resulting minimum weight is compared to the one obtained by Kaveh and Rahami in [5], and Kaveh and Hassani in [8] for the same example. The result of comparison is shown in Table 10.7. Similar to the other cases, CSS with supervisor agents have shown a better performance. Kaveh and Rahami in [5] used a different formulation to impose the analysis criteria as a constraint. In this method, using the derivative of Uc in Eq. 10.6 with respect to {q} leads to: " ! " ∂Uc ! ¼ Hqp fpg þ Hqq fqg ¼ 0 ∂q

ð10:19Þ

Equation 10.19 indicates that the complementary energy of the structure is equal to its minimum value in the compatibility condition. Thus {q} should be selected such that Eq. 10.19 holds. The left hand of this equation is a zero vector and it should be changed to a scalar. The best way is calculation of the norm, because the norm of a vector is equal to zero when all the entries is equal to zero. Here, we use the equilibrium itself. For this purpose we can write ' '! " ! " (( Fðq; AÞ ¼ WðAÞ 1 þ αnorm Hqp fpg þ Hqq fqg

ð10:20Þ

Having {q} and {A}, the magnitude of F can be calculated from Eq. 10.20 and its minimum for a large value of α corresponds to minimum W. Other constraints such as stress constraints, displacement constraints or dynamical properties constraints

422

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

Table 10.6 Results of the 10-bar planar truss (Case Study 6) (case 1–3) Case 1 (FSD) A ¼ {7.94 0.10 8.05 3.91 Case 2 (FSD) A ¼ {7.77 0.24 8.25 3.79 Case 3 (weight minimization) A ¼ {7.77 0.24 8.25 3.79 Table 10.7 Optimal design comparision for the 10-bar truss (Case Study 6)

0.10

0.10

5.73

5.57 5.54

0.11}in2, W ¼ 1,591.8 lb

0.1011 0.22

5.97

5.41

3.67

0.31}in2, W ¼ 1,591.8 lb

0.1011 0.22

5.97

5.41

3.67

0.31}in2, W ¼ 1,516.2 lb

Method

(GA) [5]

(ACO) [8]

Kaveh and Ahmadi [25]

Best weight (case 1) lb Best weight (case 2) lb Best weight (case 3) lb

1,593.5 1,723.5 1,519.2

1,593.5 1,723.5 1,519.2

1,591.8 1,724.6 1,516.2

can be applied to Eq. 10.20 after normalizing and selecting a penalty coefficient. Therefore, the final formulation will be as follow:

MinFðq;AÞ¼

Find!q,A;A∈f Sd or Sc g nc '! " ! " (( X A l ρ 1þαnorm H maxð0,gm ðAÞÞ p H q þ f gþ f g i i qp qq i i¼1

X ne

'

m¼1

ð10:21Þ

where Sd and Sc are the discrete and continuous sections, respectively. gm(A) corresponds to violation of the constraints. Because of indirect analysis, internal forces in earlier iterations are not reliable. In other words, since the redundant forces are not exact, the calculated constraints are not exact either, and cannot be relied on. Reliability criteria can be norm([Hqp]{p} + [Hqq]{q}). Accordingly, the design constraints penalty function can be altered to: Fðq;AÞ¼

X ne

nc ' '! " ! " (( X RðnormÞ f p gþ H f q g þ A l ρ 1þαnorm H maxð0,gm ðAÞÞ i i qp qq i i¼1 m¼1

ð10:22Þ

where R(norm) is a function of norm([Hqp]{p} + [Hqq]{q}). This function can be considered as follows: RðnormÞ ¼ logð10 þ NORMÞ

ð10:23Þ

where NORM is equal to norm([Hqp]{p} + [Hqq]{q}). In all of the examples studied in the following, Eq. 10.22 has been used in the CSS algorithm.

10.5

Minimum Weight

423

Table 10.8 Design data for the 10-bar planar truss (Case Study 7) Material property and constraint data Young’s modulus: E ¼ 1e7 psi ¼ 6.895e7MPa Density of the material: ρ ¼ 0.1 lb/in3 ¼ 0.00277 kg/cm3 Stress constraints |σ i| * 25 ksi(172.375 MPa); i ¼1, . . ., 10 Nodal displacement constraint in all directions of the co-ordinate system |Δi| * 2 in (5.08 cm); i ¼ 1, . . ., 4 List of the available profiles Case 1: (Discrete sections) Ai ¼ {1.62, 1.80, 1.99, 2.13, 2.38, 2.62, 2.63, 2.88, 2.93, 3.09, 3.13, 3.38, 3.47, 3.55, 3.63, 3.84, 3.87, 3.88, 4.18, 4.22, 4.49, 4.59, 4.80, 4.97, 5.12, 5.74, 7.22, 7.97, 11.5, 13.5, 13.9, 14.2, 15.5, 16.0, 16.9, 18.8, 19.9, 22.0, 22.9, 26.5, 30.0, 33.5} in2 Ai ¼ {10.4516, 11.6129, 12.8387, 13.7419, 15.3548, 16.9032, 16.9677, 18.5806, 18.9032, 19.9354, 20.1935, 21.8064, 22.3871, 22.9032, 23.4193, 24.7741, 24.9677, 25.0322, 26.9677, 27.2258, 28.9677, 29.6128, 30.9677, 32.0645, 33.0322, 37.0322, 46.5806, 51.4193, 74.1934, 87.0966, 89.6772, 91.6127, 99.9998, 103.2256, 109.0320, 121.2901, 128.3868, 141.9352, 147.7416, 170.9674, 193.5480, 216.1286} cm2 Case 2: (Continuous sections) 0.1 * Ai * 35 in2 (225.8960) cm2; i ¼ 1, . . ., 10

Case Study 7: A 10-bar Planar Truss. The 10-bar truss as shown in Fig. 10.10 is considered for optimal design. Table 10.8 contains the necessary data. As seen displacement constraint is added to the design procedure. Two cases are considered, the first is optimal design using discrete sections and the second corresponds to continuous sections. Equation 10.22 is used as the objective function in the CSS, where a population of 20 CPs is used. In both cases, A and q are variables. In discrete case a code is utilized that moves the section between two available sections to one of them based of a probabilistic function. Results are obtained in Tables 10.9 and 10.10 for discrete and continuous sections, respectively. Case Study 8: A 25-bar Space Truss. Geometry, nodal ordering and grouping of members are sown in Fig. 10.11 and Table 10.11, respectively. Table 10.12 contains the necessary data for design. Table 10.13 contains the results and shows the efficiency of this method and combining the CSS and force method compared to the other algorithms. In this example, the calculated maximum displacement in case 1 and case 2, using exact displacement method, are equal to 0.3482 in and 0.3503 in. and those of the present method are 0.3496 in and 0.3498 in. respectively. There is another set of areas for case 2 as A ¼ {0.10, 0.10, 3.7598, 0.10, 1.8932, 0.7755, 0.1408, 3.8460} and the corresponding weight is equal to 468.1998. Maximum displacement of this set of areas leads to 0.3497 in.

Weight lb(kN) 5,490.738 (24.4228) 5,491.71 (24.4271) 5,613.84 (24.9704) 5,517.72 (24.5702) 5,475.40 (24.3817)

A1 33.50 33.50 33.50 33.50 30.00 A2 1.62 1.62 1.62 1.62 1.62

A3 22.90 22.90 22.90 22.90 22.90

A4 14.20 15.50 15.50 14.20 16.00

Weight lb (kN)

5,061.90 (22.5153) 5,089.0 (22.6359) 5,076.85 (22.5818) 5,107.3 (22.7173) 5,084.9 (22.6176) 5,112.0 (22.7382) 5,080.0 (22.5958) 5,076.66 (22.5810) 5,066.98 (22.5379) 5,095.46 (22.6899) 5,059.39 (22.5041)

Method

Kaveh and Rahami [5] Schmit and Farshi [11] Schmit and Miura [12] Schmit and Miura [12] Venkayya [13] Gellatly and Berke [14] Dobbs and Nelson [15] Rizzi [16] Khan and Willmert [17] Kaveh and Hassani [8] Kaveh and Ahmadi [25]

30.67 33.43 30.67 30.57 30.42 31.35 30.50 30.73 30.98 30.86 30.5

A1 0.1 0.1 0.1 0.37 0.13 0.1 0.1 0.1 0.1 0.1 0.1

A2 22.87 24.26 23.76 23.97 23.41 20.03 23.29 23.93 24.17 23.55 21.99

A3 15.34 14.26 14.59 14.73 14.91 15.60 15.43 14.73 14.81 15.01 15.70

A4

Table 10.10 Optimal design comparison for the 10-bar planar truss (Case study 8) (continuous)

Method Kaveh and Rahami [5] Shih [9] Rajeev [10] Kaveh and Hassani [8] Kaveh and Ahmadi [25]

Table 10.9 Optimal design comparison for the 10-bar planar truss (Case Study 7) (discrete)

0.1 0.1 0.1 0.1 0.10 0.14 0.1 0.1 0.1 0.1 0.1

A5

A5 1.62 1.62 1.62 1.62 1.62

0.46 0.1 0.1 0.36 0.10 0.24 0.21 0.1 0.41 0.22 0.5

A6

A6 1.62 1.62 1.62 1.62 1.62

7.48 8.39 8.59 8.55 8.70 8.350 7.65 8.54 7.547 7.63 7.55

A7

A7 7.97 7.97 14.20 11.50 7.97

20.96 20.74 21.07 21.11 21.08 22.21 20.98 20.95 21.05 21.65 21

A8

A8 22.9 22.00 19.90 22.00 22.90

21.70 19.69 20.96 20.77 21.08 22.06 21.82 21.84 20.94 21.32 22

A9

A9 22.00 22.00 19.90 19.90 22.90

0.1 0.1 0.1 0.32 0.19 0.1 0.1 0.1 0.1 0.1 0.1

A10

A10 1.62 1.62 2.62 1.62 1.62

424 10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

10.5

Minimum Weight

425

Fig. 10.11 Geometry of a 25-bar space truss (Case Study 8)

Table 10.11 Member grouping

Group number 1 2 3 4 5 6 7 8

Members 1–2 1–4,2–3,1–5,2–6 2–5,2–4,1–3,2–6 3–6,4–5 3–4,5–6 3–10,6–7,4–9,5–8 3–8,4–7,6–9,5–10 3–7,4–8,5–9,6–10

Table 10.12 Design data for a 25-bar space truss (Case Study 8) Design variables Size variables A1;A2;A3;A4;A5;A6;A7;A8; q1; q2 q3; q4; q5; q6; q7 Material property and constraint data Young’s modulus: E ¼ 1e7 psi Density of the material: ρ ¼ 0.1 lb/in3 ¼ 0.00277 kg/cm3 Stress constraints |σ i| * 40 ksi (275.8 MPa); i ¼ 1, . . ., 25 Displacement constraint in the directions of X and Y in the co-ordinate system |Δi| * 0.35 in (0.8890 cm); i ¼ 1, 2 List of the available profiles Case 1: (Discrete sections) Ai ¼ {0.1, 0.5 ) I (I ¼ 1,2,. . .,76), 39.81, 40} in2 Ai ¼ {0.6452, 3.2258 ) I (I ¼ 1, 2, . . ., 76), 256.8382, 258.0640} cm2 Case 2: (Continuous sections) Ai +0.1 in2 (0.6452) Loading data Node Px: kips (kN) Py: kips (kN) Pz: kips (kN) 1 "10 (44.48) "10 (44.48) "10 (44.48) 2 0 "10 (44.48) "10 (44.48) 3 0.5 (2.224) 0 0 6 0.5 (2.224) 0 0

Method Rajeev [10] Erbatur [18] Kaveh and Kalatjari [19] Kaveh and Rahami (C.1) [5] Kaveh and Rahami (C.2) [5] Kaveh and Ahmadi [25] (C.1) Kaveh and Ahmadi [25] (C.2)

Weight lb (kN) 546.01 (2.4287) 493.80 (2.1964) 480.23 (2.1361) 479.75 (2.1340) 467.6293 (2.0800) 479.75 (2.1340) 467.0660 (2.0774)

A1 0.1 0.1 0.1 0.1 0.1 0.1 0.1 A2 1.8 1.2 0.1 0.5 0.1 0.5 0.1001

Table 10.13 Optimal design comparision for the 25-bar space truss (Case Study 8) A3 2.3 3.2 3.5 3.0 3.7598 3.0 3.6979

A4 0.2 0.1 0.1 0.1 0.1 0.1 0.1005

A5 0.1 1.1 2.0 2.0 1.8552 2.0 1.8687

A6 0.8 0.9 1.0 1.0 0.7755 1.0 0.7888

A7 1.8 0.4 0.1 0.1 0.1408 0.1 0.1420

A8 3.0 3.4 4.0 4.0 3.8460 4.0 3.8612

426 10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

10.5

Minimum Weight

427

Table 10.14 Design data for a 120-bar space dome (Case Study 9) Design variables Vaariables: A1;A2;A3;A4;A5;A6;A7; q1; q2; q3; q4; q5; q6; q7; q8; q9 Material property and constraint data Young’s modulus: E ¼ 30,450 ksi ¼ 210,000 MPa Density of the material: ρ ¼ 0.288 lb/in3 ¼7971.810 kg/cm3 For all members: 0.775 * Ai * 20 in2; i ¼1, . . ., 120 Constraints pﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃﬃ r ¼ 0:4 ) A λi ¼ lri Stress constraints For tensile members Fa * 0.6 Fy and λi * 300 For compressive members λi * 200 h* + i * + λ3i λi 3λi 5 for λi * Cc Fa ¼ 1 " 2C 2 Fy = 3 þ 8C " 8C3 c c

c

2

E Fa ¼ 12π for λi * Cc 23λ2 i

σ i < 58.0 ksi (400 MPa); i ¼1, . . ., 120 Displacement constraint in the directions of X, Y anc Z in all unsupported nodes |Δi| * 0.1969in

Case Study 9: A 120-bar Dome. A 120-bar dome structure is considered in this example. This structure has 9( of statical indeterminacy. The necessary data for design, and constraints are shown in Table 10.14. Optimal design comparison for the 26-story tower is obtained in Table 10.15. Geometry, ordering and member grouping structure are shown in Fig. 10.12. Loading condition is considered as: 1. A vertical load at node 1 equal to "13.49 kips ("60 kN) 2. Vertical loads at node 2 through 14 equal to "6.744 kips ("30 kN) 3. Vertical loads at the rest of the nodes equal to "2.248 kips ("10 kN) Redundant forces are considered as the reactions at nodes 39, 43 and 47. For the present approach the maximum stress ratio is equal to 0.9552 and the maximum displacement using the exact displacement method is equal 0.17335 in, and the maximum displacement using the present method is calculated as 0.17339 in. In this example, when the displacement method is utilized as an analysis procedure, the unknowns change from redundant forces to nodal displacements. Then number of unknowns drastically increase from 9 redundant forces to 111 nodal displacements. This imposes a highly computational cost on the optimization procedure. Equation 10.24 will be used to analysis using displacement method.

Element group Kaveh et al. (IACS) [20] Kaveh and Talatahari (PSOPC) [21] Kaveh and Talatahari (PSACO) [21] Kaveh and Talatahari (HPSACO) [21] Kaveh and Talatahari (HBB-BC) [22] Kaveh and Talatahari (CSS) [6] Kaveh and Ahmadi [25]

A1 3.026 3.040 3.026 3.095 3.037 3.027 4.795

A2 15.060 13.149 15.222 14.405 14.431 14.606 5.153

A3 4.707 5.646 4.904 5.020 5.130 5.044 4.527

Optimal cross-sectional areas (in2)

Table 10.15 Optimal design comparison for the 120-bar dome (Case Study 10) A4 3.100 3.143 3.123 3.352 3.134 3.139 4.353

A5 8.513 8.759 8.341 8.631 8.591 8.543 1.650

A6 3.694 3.758 3.418 3.432 3.377 3.367 3.533

A7 2.503 2.502 2.498 2.499 2.500 2.497 3.888

Best weight(lb) 33,320.52 33,481.20 33,263.90 33,248.90 33,287.90 33,251.90 28,712.69

428 10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

10.5

Minimum Weight

429

Fig. 10.12 A 120-bar dome (Case Study 9)

normð½K'fXg " fFgÞ ¼ 0

ð10:24Þ

where K is considered as the stiffness matrices of the structure. X is considered as the nodal displacement vector and F is the nodal forces vector. Case Study 10: A 26-story Tower. The main aim of the present method is to avoid the computation of the inverse of the large-scale structures matrices. This method must be applied to the large-scale structures to show the superiority of the present

430

10 Simultaneous Analysis, Design and Optimization of Structures Using Force. . .

Table 10.16 Design data and constraints of 26-story truss (Case Study 10) Design variables Variables A1;A2;A3;. . .;A59; q1; q2;. . .; q246 Material property and constraint data Young’s modulus: E ¼ 1e7 psi Density of the material: ρ ¼ 0.1 lb/in3 ¼ 0.00277 kg/cm3 Stress constraints |σ i| * 25 ksi(172.375 MPa); i ¼1, . . ., 942 Displacement constraint in the directions of X and Y in the co-ordinate system |Δi| * 15 in (about 1/250 of the total height of the tower) for the four nodes of the top level in the x, y and z directions List of the available profiles Ai