990325master PDF

1
UNIVERSITY OF BRISTOL
DEPARTMENTS OF AEROSPACE AND CIVIL ENGINEERING
DYNAMIC FINITE ELEMENT MODELLING AND

UPDATING OF LOADED STRUCTURES
Paul David Greening
A thesis submitted to the University of Bristol for the degree of

Doctor of Philosophy in the Faculty of Engineering
March 1999
viii
CHAPTER 3 Modelling Of The Dynamic Behaviour of Loaded Structures...................................... 58
3.1 Introduction.................................................................................................................................. 58
3.2 Stress Stiffening........................................................................................................................... 60
3.2.1 Closed Form Solution.......................................................................................................... 62
3.2.2 Finite Element Formulation ................................................................................................. 66
3.2.3 Implementation of Stress Stiffening .................................................................................... 68
3.2.4 Updating of Stress Stiffening Effects Using Conventional Parameters............................... 69
3.3 Large Deformations ..................................................................................................................... 71
3.3.1 Finite Element Formulation ................................................................................................. 72
3.3.2 Effect of Boundary Conditions............................................................................................ 75
3.4 Effect Upon Validation / Updating .............................................................................................. 76
3.5 Concluding Remarks.................................................................................................................... 77
CHAPTER 4 Experimental Study Of The Dynamic Behaviour Of Loaded Structures.................... 83
4.1 Introduction.................................................................................................................................. 83
4.2 Test on Narrow Steel Plate .......................................................................................................... 84
4.2.1 Application and Identification of Loading .......................................................................... 84
4.2.2 Modal Analysis of Plate Specimen...................................................................................... 90
4.2.3 Comparison of Measured Dynamic Behaviour with FE Modelling .................................... 94
4.2.4 Identification of Zero-Load Resonant Frequencies ............................................................. 94
4.2.5 Investigation of FE Experimental Mismatch .................................................................... 97
4.3 Small Framework......................................................................................................................... 98
4.3.1 Arrangement of Framework ................................................................................................ 98
4.3.2 Experimental Results ........................................................................................................... 99
4.3.3 Finite Element Modelling .................................................................................................. 102
4.4 Concluding Remarks.................................................................................................................. 106
CHAPTER 5 Model Updating OF Load-Dependent Structural Properties..................................... 124
5.1 Introduction................................................................................................................................ 124
5.2 Updating of Stress Stiffening Effects ........................................................................................ 125
5.2.1 Theory................................................................................................................................ 126
5.2.2 Practical Application of Stress Stiffening Updating.......................................................... 130
5.2.2.1 Identification of Force Magnitude in Axially Loaded Beam....................................... 130
5.2.2.2 Identification of Loads in Small Framework............................................................... 134
5.2.3 Implementation Using Commercial Software ................................................................... 139
5.3 Updating of Deformation Properties.......................................................................................... 139
5.3.1 Eigenvalue Sensitivity to Rigid Body Rotation................................................................. 140
5.3.2 Implementation of Elemental Rotation Updating .............................................................. 144
5.3.3 Updating of Deflection Profile Magnitudes ...................................................................... 147
5.3.4 Implementation of Profile Updating Method..................................................................... 150
5.4 Comparison With Established Updating Parameters ................................................................. 153
CHAPTER 6 Experimental Updating of Loaded Structure .............................................................. 170
6.1 Introduction................................................................................................................................ 170
6.2 Experimental Details.................................................................................................................. 171
6.2.1 Description of Test Specimen............................................................................................ 171
6.2.2 Testing Procedure .............................................................................................................. 171
6.3 Measured Loading on Three Frames ......................................................................................... 172
6.4 Modal Analysis .......................................................................................................................... 174
6.4.1 Mode Identification ........................................................................................................... 174
6.4.2 Perturbation to Resonant Frequencies Under Loading ...................................................... 176
6.4.3 Variability of Results......................................................................................................... 177
6.5 Construction and Validation of MATLAB FE Model ............................................................... 178
6.6 Initial Correlation....................................................................................................................... 180
6.7 Static Updating of Load Dependent Properties ......................................................................... 182
6.7.1 Verification of Measured Displacement ............................................................................ 184
6.7.2 Frequency Perturbations From Different Analysis Approaches........................................ 185
6.7.3 Comparison of Statically Updated Model With Dynamic Data ........................................ 187
ix
6.8 Dynamic Updating Of Framework Structure............................................................................. 189
6.8.1 Updating of Stress Stiffening ............................................................................................ 189
6.8.1.1 Initial Investigation...................................................................................................... 190
6.8.1.2 Updating With Offset .................................................................................................. 194
6.8.1.3 Effect of Eigenvalue Selection .................................................................................... 195
6.8.2 Updating of Permanent and Transient System Parameters................................................ 197
6.8.3 Updating Using All Experimental Results......................................................................... 201
CHAPTER 7 Conclusions ..................................................................................................................... 228
7.1 Introduction................................................................................................................................ 228
7.2 Literature Review ...................................................................................................................... 228
7.3 Effect of Load on Structures...................................................................................................... 229
7.4 Static Updating of Loaded Structures ........................................................................................ 230
7.5 Dynamic Updating of Loaded Structures .................................................................................. 231
7.6 Future Work............................................................................................................................... 232
7.7 Epilogue..................................................................................................................................... 233
References ............................................................................................................................................... 234
APPENDIX A Finite Element Formulation......................................................................................... 240
APPENDIX B Experimentally Identified Modes ................................................................................ 244
LIST OF FIGURES
Figure 2.1 Beam Under Axial Load........................................................................................................ 55
Figure 2.2 Beam Under Shear and Rotational Loading .......................................................................... 55
Figure 2.3 Transformation of Axes......................................................................................................... 55
Figure 2.4 Schematic of Experimental Arrangement; Testing of Narrow Plate ..................................... 56
Figure 2.5 Residual Values ..................................................................................................................... 56
Figure 2.6 Updated Parameter Sets ......................................................................................................... 57
Figure 3.1 Tangent Stiffness ................................................................................................................... 78
Figure 3.2 Forces Acting on A Small Portion of Beam .......................................................................... 78
Figure 3.3 Lateral Movement of Small Beam Element Under Constant Axial Load............................... 78
Figure 3.4 Fixed-Fixed 13 Element Beam .............................................................................................. 79
Figure 3.5 Updating Overall Stiffness of Loaded Beam Using First Eigenvalue ................................... 79
Figure 3.6 Updating Overall Stiffness of Loaded Beam Using Four Eigenvalues ................................. 80
Figure 3.7 Rigid Body and Straining Displacements of 2D beam .......................................................... 80
Figure 3.8 Iterative vs. Incremental Tangent Stiffness Calculation ........................................................ 81
Figure 3.9 Measured Deflection of Plate ................................................................................................ 81
Figure 3.10 Change in Natural Frequency Magnitude of Deformation .................................................. 82
Figure 4.1 - Experimental Arrangement .................................................................................................. 107
Figure 4.2 - Co-ordinate System and Strain Gauge Locations on Plate................................................... 107
Figure 4.3 - Observed Stresses Distributions on Plate Cases A to D.................................................... 108
Figure 4.4 Application of Load to Finite Element Model..................................................................... 109
Figure 4.5 Stress Distributions from Different Stress Models .............................................................. 109
Figure 4.6 - Point Receptances from Narrow Plate Under Varying Loading .......................................... 110
Figure 4.7 FRAC Between Load Cases 1 and 2 Varying .................................................................. 110
Figure 4.8 Linear Receptance Case A and Case B................................................................................ 111
Figure 4.9 Frequency Shifted Receptances 0-800Hz ............................................................................ 111
Figure 4.10 Frequency Shifted Receptances, First Mode ..................................................................... 112
Figure 4.11 Effect of on FRAC of Case A and Case B Using Log Receptance ................................ 112
Figure 4.12 - Percentage Increase From Nominal Zero Load.................................................................. 113
Figure 4.13 - Exploded View of Frame ................................................................................................... 114
Figure 4.14 Front View of Redundant Frame Showing Measurement Points ...................................... 114
Figure 4.15 Finite Element Model of Framework................................................................................. 115
Figure 4.16 Point Receptance Load Cases 0 & 3.................................................................................. 115
Figure 4.17 Perturbation of Resonant Frequencies Experimental Observations................................... 116
Figure 4.18 Experimentally Identified Mode Shapes............................................................................ 117
Figure 4.19 MAC Between Experimental Load Cases 0 and 3............................................................. 118
Figure 4.20 - Finite Element Mode Shapes.............................................................................................. 119
Figure 4.21 - MAC Between Experimental Case 0 Modes and FE Prediction ........................................ 120
Figure 4.22 Correspondence of Modes Between Case 0 and FE Models ............................................. 121
Figure 4.23 Auto-MAC of FE prediction of First Fifteen Modes of Vibration .................................... 121
Figure 4.24 Closely Correlating Modes (MAC > 0.8) in FE Model ..................................................... 122
Figure 4.25 - Construction of FE Model.................................................................................................. 122
Figure 4.26 Experimental And Analytical Relationship Between Load and Resonant Frequency....... 123
Figure 5.1 Static vs. Dynamic Approaches to Updating ....................................................................... 157
Figure 5.2 - Fixed-Fixed 13 Element Beam............................................................................................. 157
Figure 5.3 Receptance of Experiment from FE Model...................................................................... 158
Figure 5.4 Identification of Stiffness and Stress Stiffening Using Two Modes.................................... 158
Figure 5.5 Effect of Multiplicative Noise & Eigenvalue Choice on Updating ..................................... 159
Figure 5.6 Effect of Additive Noise and Eigenvalue Selection on Updating Success .......................... 159
Figure 5.7 Numbering System for Framework Spars ........................................................................... 160
Figure 5.8 Sensitivity Matrix, {p} =0 ................................................................................................... 160
Figure 5.9 Identified Loads From No Initial Knowledge of Loading ................................................... 161
Figure 5.10 Loads Identified From Noisy Experimental Data .............................................................. 161
Figure 5.11 Load Identification With Various Initial Condition .......................................................... 162
Figure 5.12 EISM to Estimate Spar For Static Load Determination..................................................... 162
Figure 5.13 - Rotation of Single Beam Element ...................................................................................... 163
xi
Figure 5.14 - Perturbations to Resonant Frequencies .............................................................................. 163
Figure 5.15 - Sensitivity of Four Resonant Frequencies to Rotation of Element Six .............................. 164
Figure 5.16 Convergence Of Element Rotation on Experimental Value........................................... 164
Figure 5.17 Four Possible Solutions to Two Parameter System ........................................................... 165
Figure 5.18 Relationship Between Element Rotations and Central Displacement ............................... 165
Figure 5.19 Effect of Sinusoidal Static Displacement on Resonant Frequency.................................... 166
Figure 5.20 Mode Swapping With Large Displacement....................................................................... 166
Figure 5.21 Updated Solution initial=0.02............................................................................................. 167
Figure 5.22 Convergence Upon Best Fit Solution............................................................................. 167
Figure 5.23 - Converged Solution Using Two Profile Updating Parameters........................................... 168
Figure 5.24 - Updated Solution Using Two Profile Parameters; Various Initial Deflections .................. 169
Figure 6.1 Spar Numbering Convention and Excitation Node ............................................................. 206
Figure 6.2 Strain Gauge Positions Frames B, C and D ......................................................................... 207
Figure 6.3 Axial Load Distribution in Frame B, Load Case 2 .............................................................. 208
Figure 6.4 Bending Moments at Centre of Six Spars, Frame B............................................................ 208
Figure 6.5 Bending Moment Distribution Spars 1 and 2 ...................................................................... 209
Figure 6.6 Exaggerated Static Deflected Shape .................................................................................... 209
Figure 6.7 All Measured Responses Frame B, Case 2 .......................................................................... 210
Figure 6.8 Fourteen Identified Resonant Frequencies; Frame B Load Case 2...................................... 211
Figure 6.9 MAC > 0.8 Between FE and Experimental Modes, Zero Load........................................... 212
Figure 6.10 Modal Perturbation Under Loading, Three Independent Cases First Five Modes ............ 213
Figure 6.11 Point Receptances From Three Nominally Identical And Unloaded Frames .................... 214
Figure 6.12 Local Axis Definition ........................................................................................................ 214
Figure 6.13 - Modal Perturbations Using Different Loading Models; 1kN Applied Load...................... 215
Figure 6.14 Stress Stiffening and Large Deformation Effect on Resonant Frequencies........................ 216
Figure 6.15 FE and Experimental Load vs., Frequency Relationship................................................... 217
Figure 6.16 MAC Between Original and Improved FE Models ........................................................... 218
Figure 6.17 MAC Between Original and Improved FE Models ........................................................... 219
Figure 6.18 MAC Between FE and Frame B; Zero Load ..................................................................... 219
Figure 6.19 MAC Between FE and Frame B; 514N Load.................................................................... 220
Figure 6.20 MAC Between FE and Frame B; 1135N Load.................................................................. 220
Figure 6.21 MAC Between FE and Frame B; 1618N Load.................................................................. 221
Figure 6.22 FE and Frame B; Load vs. Frequency Relationship .......................................................... 222
Figure 6.23 Updated / Identified Loads ................................................................................................ 223
Figure 6.24 Convergence Upon Identified Loading ............................................................................. 224
Figure 6.25 Updated Parameters After Removing Zero Load Offset ................................................... 225
Figure 6.26 Utility of Modes in Producing Converged Solution .......................................................... 226
Figure 6.27 Identified and Measured Loads; Frame B Load Cases 1-3................................................ 226
Figure 6.28 Identified Loads; Frames C and D, Load Cases 1-3 .......................................................... 227
Chapter 1 Introduction
CHAPTER 1
INTRODUCTION
1.1
Prologue
Computer models of structures provide the engineer with an extremely powerful tool for
understanding their behaviour.
A great deal of information about structural
performance can potentially be obtained from simulations without the expense of

undertaking structural testing of prototype structures. Unlike practical experimental
testing, the number of scenarios which can be examined are almost boundless; it
becomes possible to load a structure up to - and beyond - failure as often and in as many
ways as the engineer desires!
The particular ability to model dynamic structural behaviour allows structures to be
designed in the most cost effective and safe manner. This mitigates against the adverse
effects of excessive vibration. There is, of course, an economic case for producing ever
lighter, cheaper structures. However, this must not be achieved at the expense of safety,
reliability and durability, each of which is strongly influenced by structural dynamic
behaviour.
Computer-models can be built with some confidence from knowledge of the structure's
geometry and material properties. This confidence in the quality of the model can be
enhanced if the computer model can in some way be compared with experimental
measurements from a prototype structure.
A mismatch between experimental and
analytical measurements motivates a process whereby the analytical model is altered to

provide a closer agreement with experimental readings. The hope and expectation is
that the altered model provides an improved representation of the prototype structure.
2
analytical model of a structure from test data further
The ability to alter an
opens the possibility that the finite element model can be treated as a store of
knowledge. This encyclopaedia of the structures condition allows the computer
model to be re-adjusted to account for time varying changes to structural response. In
terms of structural monitoring, dynamic measurements can be thought of essentially as a
signature of a structure.
Thus, changes to this signature represent change to the
structure itself.
1.2
Structural Modelling
A closed form differential equation approach to modelling the behaviour of structures is

possible only in the simplest cases, such as beams and plates. In the late fifties an
approximate approach involving discretisation of structures into a potentially large
number of sub elements, whose behaviour is known, came to prominence. The so
called finite element (FE) method allows the stiffness and mass distribution of a
structure to be described in matrix terms with rows and columns representing the active
degrees of freedom.
A finite element model of a structure can be used for calculating response to both static
and dynamic loading.
Many of the choices have to be made by an experienced
engineer. Estimates of material properties will often be textbook values and geometric
properties may be based upon the initial design rather than the measured configuration.
The finite element model and associated predictions of static and dynamic behaviour of
any structure is not unique. Indeed the well known DYNAS survey [2] and a more
recent survey by Lloyd's Register [3] show large variations in prediction of dynamic
structural behaviour by a number of finite element models of the same structure by
different practitioners. Several important factors can be identified as being responsible
for poor prediction of dynamic response. These principally include:
mis-estimation of structural material properties;
inaccurate modelling of structural geometry;
poor choice of element type and quantity required; and

difficulty in modelling complex structural components, the most common
and widespread being the pitfalls which attend the modelling of structural
joints.
Further errors in dynamic prediction of finite element models can arise due to model
reduction. This is undertaken for FE models with very large numbers of degrees of
freedom to reduce the size of the eigenvalue problem.
1.3
Dynamic Measurements
Vibration is a characteristic common to all engineering structures, from the smallest

electrical components to the largest bridges and dams. The effect of vibration upon
structural performance can be instrumental in causing fatigue, discomfort or ultimately
structural failure.
The dynamic characteristics of a structure - even measured at few points spatially on a
structure - offer a great deal of information about the structural form. Consequently, the
use of dynamic data for characterising structural behaviour has long been popular. The
well known example of striking solid train wheels and assessing their integrity based on
the audible response of the wheel is a good example.
The field of modal analysis is concerned with characterising the dynamic behaviour of
structures from experimental data. The methodology for dynamically testing structures
and identifying modal behaviour is set out clearly in [4]. Dynamic test data are most
commonly described in terms of a modal model consisting of resonant frequencies and
mode-shapes. A process of modal identification is required to process raw measured
data into this format.
Corresponding predictions of dynamic behaviour can be
determined from a finite element model using eigenvalue extractions techniques.

Comparison of experimental and FE data in the modal domain is used extensively in
this thesis.
1.4
Correction of FE Model Using Dynamic Data
Every attempt should be made to use realistic, or better still, measured parameters in the
process of building the finite element model as section 1.2 has indicated. However,
there is still likely to be some error in the finite element model.
The comparison of predicted structural dynamic behaviour with experimentally
measured data allows a great deal of insight into the likely sources of error in the finite
element model. This has been motivated by the requirement to improve the finite
element model of a structure. A large number of techniques have been developed
whereby analytical FE models of structures are altered such that their dynamic
characteristics become a closer match of experimentally determined behaviour.
Today, the correction of finite element models in this way is extremely widespread
especially when the main purpose of the finite element model is to understand dynamic
rather than static behaviour. A range of techniques exist for altering finite element
models. At the most simple, it is very common to make a small number of changes to
the overall properties of a finite element model in a number of iterations. This type of
process involves a large amount of intervention from an engineer to assess the level of
improvement in the dynamic predictions of the FE model. The engineer also has to
ensure that changes made to the finite element model are realistic. While a perfectly
acceptable approach, this manual technique is labour intensive and relies on the
intuition of the engineer to identify possible sources of error in the finite element model.
Finite element model updating is an alternative to these manual techniques.
methods allow changes to be made in an
The
sense. Several different approaches
to model updating have been developed. Methods regarded as being of most use
involve an optimisation scheme whereby a number of structural parameters (updating
parameters) are altered to minimise an objective function. Limitations upon the amount
of information available from experimentation curtail the number of updatable
parameters.
Factors upon elemental mass and stiffness are undoubtedly the most
popular updating parameters. The procedure is iterative with changes made to the finite
element model at each step. The objective function is a formulation of the differences
in dynamic behaviour between the experimental data and the finite element model. This
is recalculated at each stage of the iteration. While continuing to be an active area of
research, efforts to achieve robust model updating have achieved little success.
One important aspect of finite element model updating is that there exists much
confidence in the experimental dynamic data and less in the finite element model itself.
The concentration upon the finite element model as being in error stems partly from the
fact that finite element modelling and modal analysis are often treated as separate
disciplines. It is common for the two areas to be conducted by different departments in
research and development organisations. An awakened interest in the possibility that
any single set of dynamic data is not uniquely representative of the structure offers new
challenges to model updating. This thesis takes up the mantle by considering how finite
element model updating can be adapted to account for time varying (transient) changes
in dynamic behaviour.
1.5
Variations in Dynamic Readings
Given the inevitable uncertainty in the veracity of the initial finite element model, the
motivation for using experimental data to improve upon initial assumptions is clear.
The experimental dynamic data are thus treated as being
representative of the
structure. Additionally, the collection of large numbers of sets of experimental data or

the construction of a number of prototypes is an expensive undertaking. It is not
surprising then that the potential variance in dynamic measurements either from:
(a)
a single structure under a variety of conditions; or
(b)
nominally identical structures;
have been given comparatively little attention. Fregolent
[5] specifically consider
case (b). They provide a case study of using an FE model to characterise the behaviour
of a family of structures with nominally identical properties.
Several papers by
Imregun, for instance [6] and [7] consider a similar problem.

The experimental work presented by Fregolent
, as is common practice, seeks to
test the experimental structure in very carefully controlled laboratory conditions. The
purpose is to produce noise-free and consistent results. Generally, structures are tested
in isolation from the surroundings. The quest for less obtrusive supporting conditions is
an active research area; Carne and Dohrmann [8] for instance investigate the effect of
supporting conditions in some detail.
The result of testing structures in free-free conditions is twofold. A good consistency of
results can certainly be achieved. However, the results only represent the test structure
under the specific conditions which exist at the time of testing.
In both the laboratory regime, and especially in more practical situations, the variation
of dynamic data from the same structure - case (a) above - is increasingly being
recognised as an important factor which must be considered. For instance, recent work
by Woon and Mitchell [9] found that ambient laboratory temperature was responsible
for subtly altering material properties and thus changing the dynamic behaviour of a
simple plate specimen. At the other extreme engineers attempting to characterise the
dynamic behaviour of large civil engineering structures such as bridges, buildings and
dams frequently report differing dynamic behaviour being identified from the same
structure at different times.
Alampalli [10] for instance, presents widely variable
baseline results from tests on a highway bridge under different conditions.
The
variation in measurements is attributed to the in-service environment, signal processing

factors and variation amongst operators. The dynamic measurement of the bridge was
motivated by the hope that changes to the vibration signature would enable areas of
damage or deterioration to be located.
Most disturbingly, it is reported that the
variations in baseline modal data can be greater than the changes resulting from
damage.
If the application of model updating strategies outside of the laboratory is to become

practicable, a clear requirement exists for a unique baseline measurement of the
dynamic behaviour to be available or for the factors which alter the dynamic behaviour
to be accounted for in finite element models.
1.6
Load Dependence of Dynamic Behaviour
Amongst the plethora of factors which effect dynamic behaviour, the effect of load
upon structures stands out as being an important factor requiring closer consideration.
Structures by their very nature are called upon to carry loading. Table 1.1 for instance,
sets out a selection of possible structural loading scenarios. The effect of loading is
well known to cause changes to dynamic behaviour. The influence of axial load on a
columns transverse vibration as well as the change in pitch of stringed instruments with
wire tension are both good examples.
Table 1.1 - Sources of Static Stress in Structures

The specific effect of loading upon dynamic behaviour when comparing with finite
element models appears to have been overlooked for several reasons. In laboratory
conditions, every attempt is made to isolate a test structure from its surroundings, which
results in no load being transmitted.
For in-situ testing, there is a generally low
expectation of accurate results leading to a culture where the reasons for discrepancies
between expected and measured dynamic data are not specifically investigated.
1.7
Combined Static and Dynamic Load Identification
The consideration of the interaction between static and dynamic loading is particularly
opportune since developments in sensor technology are fuelling interest in monitoring
of both static and dynamic behaviour of
structures. The ability to monitor
structures in this way is principally bought about by improvements in strain gauge

technology.
The use of piezoelectric transducers to measure strain to very high resolution is
described in [11]. Such gauges have recently become available commercially, the
performance of one such product is described in [12] where resolution of signals of an
astonishing three orders of magnitude better than the common resistive type of gauges
are quoted. While this type of gauge is currently very expensive it is clear that the
future offers the possibility that strain will not continue to be the poor relation of
translational quantities in terms of measurement fidelity.
The use of optical fibres for structural monitoring offers the opportunity to monitor
structural behaviour at large numbers of locations about the structure, over a long period
of time and at relatively modest cost. The specific use of this equipment to measure
strain has received attention recently. A so-called Bragg grating system, for instance,
involves treatment of short sections of a fibre optic cable which when bonded to a
structure allows strain at that point to be determined remotely from processing of
emitted and reflected light signals. Such a system comprising 64 strain gauge channels
is described in [13]. The optical fibre in this case is mounted on the soffit of a bridge
and allows the response of the bridge to be sampled at frequencies of up to 360Hz. This
implies a useful range of up to 180Hz1 although in this case resolution at a range of up

to 45Hz was used.
A clear research opportunity exists to consider the interaction of the static and dynamic
components of strain on structures so that the static effects can be accounted for during
dynamic updating of finite element models. This area is investigated in this thesis.
1.8
Scope of Thesis
This thesis seeks to extend the possibility of using model updating techniques to
structures whose dynamic behaviour is affected by the transient effects described in the
preceding sections. Specifically the influence of structural loading is identified as a
source of variability of dynamic behaviour. As well as demonstrating the effect of
loading upon dynamic response, several methods of including this effect in finite
element models are explored to help to improve the chances of successfully identifying
permanent errors in the model.
The background to finite element model updating and in particular the evolution in
choice of updating parameters is presented in chapter 2. A simple case study using
measured experimental data is presented which allows some of the shortcomings of
model updating to be demonstrated.
Chapter 3 introduces the changes to finite element modes which arise from static
loading using a nonlinear geometric approach.
The potential effects on dynamic
readings due to static effects are also outlined.

The methods introduced in chapter 3 are implemented in chapter 4 with respect to
measured static experimental data. The results are compared with experimental data.
Loading is measured accurately but indirectly using strain gauges allowing the issue of
1 This is the Nyquist frequency. For this particular structure, the 0-180Hz range allows a great deal of the
dynamic behaviour to be spanned. The sampling rate of strain data using this particular technique is a
function of the number of gauging positions along each optical fibre.
10
load identification to be addressed practically. Additionally, the difficulty in identifying

a set of unique modal data corresponding to zero loading is addressed.
Chapter 5 introduces a novel approach which allows load dependent structural
properties to be updated from dynamic measurements. The methods are validated using
a number of simulated experimental case studies.
A comprehensive experimental case study is presented in chapter 6 in which
experimentally measured dynamic data is used to update a finite element model. A
methodology is proposed for dynamic model updating in the presence of static
structural loads. In addition, the author proposes a method for identifying static loads
from dynamic data. The results are discussed in detail and directions for future research
indicated.
1.9
Notes on Thesis
Nearly all of the finite element work described in this thesis was performed in
MATLAB [14] using the finite element toolbox CALFEM [15] developed as a teaching
tool at Lund University, Sweden.
The basic package - developed for educational
purposes - was augmented substantially by the author to include 3D stress stiffening

effects and nonlinear geometry capabilities.
Additional FE modelling work and
validation of the MATLAB code was performed using ANSYS version 5.4 [16]
Throughout this thesis two example structures are introduced and revisited several
times.
While any particular experimental arrangement cannot be expected to be
representative of all structures, the two experiments were designed to allow many
aspects of the effect of loading on structures to be examined. Both arrangements are the
subjects of extensive analysis which in turn is compared with a large volume of
experimental work.
For reasons of brevity and simplicity, the numerical examples and experiments
presented in this thesis relate to the study of frameworks made up of beam type
elements.
11
They are undoubtedly the most popular element type and particularly
important in modelling of civil engineering structures.
Where appropriate the
application of the methods to other element types is described.

The first stage in the validation of new updating methods is generally carried out by
simulating experimental data, with the use of experimentally measured data something
of a rarity. Both methods are used in this thesis with the distinction being made as clear
as possible.
Simulated experimental data are referred to by enclosing the word
experimental in quotation marks. Whenever measured data is referred to, in all cases it
derives from experimental modal analysis.
The standard notation for modal testing and analysis set out in [17] is used wherever
possible throughout this thesis.
Chapter 2 Finite Element Model Updating Using Experimental Data
12
CHAPTER 2
FINITE ELEMENT MODEL UPDATING USING EXPERIMENTAL DATA
2.1
Introduction
As the previous chapter has outlined, the goal of finite element model updating is to use
experimental data to improve the accuracy and hence quality of an analytical finite
element model of a structure. Much work has been expended in the last twenty years in
developing and improving methods to help to meet this target. While a vast amount of
literature related to the mathematical treatment of the model updating problem exists,
evidence of the use of updating techniques being applied to measured data to solve
practical problems is - by contrast - scant.
The purpose of this chapter is to introduce some of the concepts which will be used later
in the thesis as well as reviewing the state of the art in related fields. The following
sections present a brief introduction to the finite element method and to modal testing.
Developments in finite element model updating are then considered with particular
reference to the evolution of parameters amenable to updating. A review of literature
related to the use of measured experimental data to update finite element models is also
presented.
The chapter is concluded by means of a simple case study which allows some of the
topics related to conditioning of updating problems, choice of parameters and validity
of the solution to be examined with reference to experimentally determined data.
2.2
The Finite Element Method
The development of the finite element method has its roots in the aircraft industry in the
1950s where the problem of analysing complex aeroplane structures stretched the

available analytical methods of the day.
13
Courant [18] is generally credited with
proposing a method which was destined to be highly suited for implementation on

digital computers. Papers by Turner
[19] and Argyris and Kelsey [20] laid the
foundations for the finite element method as it is used today. The term finite element
was first used by Clough
in 1960 [21]. The use of finite element methods has
burgeoned in the intervening period. Huge rises in the size and complexity of structural
problems amenable to FE modelling as well as an increase in the variety of elements
available to the engineer have been observed. Today a vast amount of literature is
devoted to the subject; Zienkiewicz [22] and Bathe [23] for example provide a
comprehensive coverage of the field.
The finite element method has been adopted by a number of areas of engineering such
as heat transfer and magnetic field analysis. In this thesis, the original application of
structural engineering is considered. In this domain, the technique broadly consists of
discretising a structure into a number of small substructures.
This allows the
displacement or stress in these elements to be approximated, the latter being the most
common approach. These elements must then be assembled in such a way that stresses
are continuous across element interfaces and the internal stresses are in equilibrium both
with each other and with the applied loads. The finite element method can thus be
thought of as a two stage process, the first being the construction of finite elements and
the second their assembly into structural matrices.
Many aspects of the finite element method can be best described by means of example.
To this end the construction of 2D beam stiffness and mass elements are described in
sections 2.2.2 and 2.2.4. The 2D beam element and its 3D counterpart are employed
extensively in this thesis; however the method by which the elements are generated can
be applied to a wide range of different element types. The following sections briefly
outline other relevant aspects of the finite element method in particular the issue of
transformation of finite element co-ordinates which forms the basis of a deformation
updating technique presented in chapter 5.

2.2.1
14
Matrix Structural Analysis
The most common problem arising in structural analysis is to determine the deflection
arising from a set of static loads. If the loads at a number of points about a structure - or
- are defined by a vector
corresponding points are similarly defined
and the displacements at the
, a matrix stiffness
is required to
relate the load and displacement.

(2.1)
Of most interest to dynamicists is a similar formulation which includes inertia and
damping terms
,
(2.2)
where [ ] is the mass matrix describing the distribution of mass about the structural
degrees of freedom and
and
are the first and second derivatives of the
displacement with respect to time. Note that the force applied to the system is now a
function of time. While mass and stiffness of a structure are measured and relatively
easily derived, the mechanism whereby energy is lost through damping is less easily
modelled. The viscous damping model represented by matrix
in equation (2.2) is
commonly but by no means exclusively used2, being proportional to velocity.

Structural damping is an important area of structural dynamics which has deservedly
received much attention. Of most importance is that damping dominates the
of vibration around resonance [4]. Since this thesis is more concerned with modal
than amplitude dependency the consideration of identification of damping
parameters will not be discussed in any great detail herein.
Hysteretic (or structural) damping is also commonly used, being represented by the matrix
equation
in the
15
equation of motion from (2.2) is
The
(2.3)
For free (unforced) vibrations the following relationship is obeyed

0,
(2.4)
the solution to which can be written in the form

{ }
(2.5)
are the resonant frequencies. Substituting back into (2.4) leads to the
where the
well known eigenvalue problem

,
(2.6)
where
2
and
(2.7)
can be thought of the
frequencies
corresponding to the system natural
While the eigenvalues have an exact relationship with the resonant frequencies, the
eigenvectors are arbitrarily scaled; it is common practice to define a uniquely scaled set
of eigenvectors such that
.
(2.8)
This results in
( ),
where
is the matrix of
(2.9)
eigenvectors.

2.2.2
16
The Two Dimensional Beam
Elemental stiffness matrices are most commonly created using assumed displacement
fields, the method for which is given in appendix A. The derivation presented therein is
general and can be applied to any element type and proposed displacement field. The
construction of the elemental stiffness matrices of a 2D beam is described in this section
using the techniques and notation set out in appendix A.
Consider first the axial loading of such a beam, shown in figure 2.1. From equation
(A.13) in appendix A, the displacement of a general point in the bar is related to the end
displacements of the bar by the
(2.10)
For the axial displacement of the beam, the shape function is trivial and can be inserted
directly into (2.10) leading to
1
(2.11)
From equation (A.15) the
matrix -
- is defined as
(2.12)
thus for the 2D beam

1
(2.13)
The elemental stiffness matrix from (A.17) is

1
0
1
1
1
1
(2.14)
which is intuitively correct and could have been ascertained directly. However, it is the
exception rather than the rule that elemental matrices can be constructed by
observation.
17
The lateral displacement of a beam undergoing shearing and rotational loading (shown
in figure 2.2) can be shown from elementary theory to be of cubic order and hence the
generalised expression for lateral displacement can be expressed as
1
1
(2.15)
2
2
A set of cubic functions [24] is required to express the shape functions
(2.17)
(2.16)
(2.18)
and
2
4
3
2
(2.19)
It is important to note that these expressions are assumed and are not necessarily exact
or unique. Thus while leading to a viable finite element, the result will be one of an
infinite number of similarly viable but subtly different elements, this non-uniqueness is
exploited by so-called generic element updating parameters described in section 2.3.3
and subsequently in section 5.4.
The strain displacement matrix is therefore
6
12
6
2
12
6
2
(2.20)
and the form of (A.17) which is applicable to beam elements [24] is

.
(2.21)
18
Combining (2.20) and (2.21) and multiplying out gives

12
6
12
6
4
2
12
6
12
6
2
(2.22)
Combining (2.14) and (2.22) to find the elemental stiffness for a full 2D beam element
which can resist shear and moment as well as axial load gives
0
12
6
3
12
12
6
3
2
2
6
3
2
2
0
0
(2.23)
12
6
3
4
2
This is the well known and very widely used Euler-Bernoulli formulation for a two
dimensional cubic beam element.
2.2.3
Stress and Strain
Stress and strain are calculated by most finite element software as a function of nodal
displacements. From equation (A.20) in appendix A, we recall that the general strain
is related to the nodal displacements
for a particular element by its strain
displacement matrix
.
(2.24)
The relationship between stress and strain for zero initial stress is given by (A.10) as
(2.25)
19
and so from (2.24) and (2.25)

.
(2.26)
Required stresses are calculated element by element. Each point within the element will
have a number of components of stress3 associated with it and thus some user input is
required to specify which stress is required. Stress can be defined in either local or
global co-ordinates depending on the element type. For instance, the stress caused by
beam flexure is given in the elements local co-ordinates since this type of stress is, by
definition, in the direction of the beams axis.
Returning to the example of the 2D beam, the standard equation for flexural stress is
given by
,
where
(2.27)
is the distance from the beams neutral axis and
is the second moment of
area of the section. Recalling that the product of the strain displacement matrix and the
elemental nodal displacements gives the curvature, then the following relationship can
be written
2
(2.28)
From (2.27) and (2.28) it can be seen that

.
(2.29)
This formulation for stress is used to estimate the stress on the outer surface of a beam
element and will be used as a comparative measure with the experimental observations
in chapter 4.
3 Three
for a 2D element and 6 for a 3D element.

2.2.4
20
Construction of Mass Matrices
Using finite elements for dynamic analysis requires consideration of inertial forces
which arise from mass undergoing acceleration. Traditionally mass was considered as
discrete particles distributed about a structure, resulting in lumped mass matrices an
approach which results in a non-continuous inertia distribution.
A consistent
formulation of the mass matrix [25] can be shown to be

(2.30)
where
is the material density and other symbols are defined in appendix A.
Substituting the shape functions for the 2D beam (2.11) and (2.16)-(2.19) into (2.30)
and integrating over the volume leads to the elemental mass matrix
140
0
0
420 70
0
0
2.2.5
0
156
22
0
54
13
0
70
0
22
0
54
4 2
0 13
0
140 0
13
0 156
2
3
0 22
0
13
3 2
0
22
4 2
(2.31)
Co-ordinate Transformation
The formulation of the elemental matrices described above has - for convenience and
generality - been relative to a set of axes local to the element itself. To convert
elemental stiffness matrices from the local set of co-ordinates ( , , ) within which
they were formulated into the global co-ordinates ( , , ) of the structure of which
they must form a part a transformation is required. This is given by
(2.32)
21
Figure 2.3 shows the local and global axes of a three dimensional bar or beam type of
element.
Forces in elemental co-ordinates can similarly be transformed:
(2.33)
Equating work done in the local co-ordinates with that done in global co-ordinates, it
can be shown that
.
(2.34)
The transformation matrix, T, can then be seen to be orthogonal, i.e.

(2.35)
Considering work done once more and now including the local elemental stiffness
1
and global elemental stiffness

,
(2.36)
therefore
.
(2.37)
It can be shown [26] that the elements of
are the trigonometric identities for
direction cosines. Thus for the two dimensional beam example, the transformation
is given by
0
2
(2.38)
0 ,
(2.39)
where
0
0
22
the terms of which are the elemental direction cosines such that
2
(2.40)
and
2
(2.41)
If the angle between the local x-axis and global x-axis is
then the transformation
matrix becomes
cos( ) sin( )
sin( ) cos( )
0
0
0
0
0
0
0
0
0
0
1
0
0
0
0
0
0
0
0
0
cos( ) sin( )
sin( ) cos( )
0
0
0
0
0
0
0
1
(2.42)
The transformation matrix for a three dimensional beam is found in a similar manner to
be given by
0
T
0
0
0
0
0
0
0
0
0
,
0
(2.43)
where
(2.44)
with the direction cosines having the same definition as in the two dimensional beam
case.

2.2.6
23
Construction of Global Matrices
Once transformed into global axes, elemental structural matrices can be constructed into
global structural matrices where each row and column in the global matrices represents
a degree of freedom of the modelled structure. For connectivity of elements some
overlap of elemental matrices will occur in the global matrices.
For more
comprehensive details the reader is referred to an elementary text on finite element

implementation such as [22] or [23].
The most recent model updating technologies display a trend towards altering
parameters at an elemental level. Therefore, it is important to consider the model in
terms of elemental matrices and their connectivities since the latter are lost once the
transformation into global matrices has been performed.
The storage space and
processing time involved in storing the structural matrices is insignificant compared

with the matrix inversions and eigen-solutions which are commonly performed a
number of times during updating.
2.2.7
Finite Element Modelling for Model Updating
The philosophy of model updating is to improve finite element models of structures.

However, the functionality of the finite element method itself is largely overlooked in
the literature. Therefore, the procedural aspects of the finite element method have been
dwelt upon in the previous sections.
With ever increasing computer power and the cost of storing data dropping the size of
finite element models are increasing. It is not uncommon for a finite element model to
contain several hundred thousand degrees of freedom. Most of the credible examples of
model updating however (section 2.6), consist of making changes to relatively simple
FE models whose construction is trivial and does not require the power of a commercial
FE package.
24
Updating of real-life finite element models which might well consist of tens of
thousands of degrees of freedom requires that only a limited amount of structural
information can be stored and processed at once.
As later sections will demonstrate, the only involvement of the commercial FE package
in the model updating process is to generate the initial structural matrices. It is the
opinion of the author that future model updating techniques should enjoy a more
intimate relationship with the finite element method. To this end, the use of finite
element methods to identify loading on structures and to determine accurate deflected
shapes of real structures as part of a model updating strategy are investigated in the
forthcoming chapters.
2.3
Model Updating Technologies
The following sections outline and explore some of the main issues relating to the
current state of the art. They are concerned with the application of updating procedures
to practical situations with particular reference to choice of updating parameters. It is
not the intention to give a comprehensive review of the field. For this the reader is
directed to review papers such as those by Natke [27], Visser and Imregun [28] and
Mottershead and Friswell [29] as well as a book by Friswell and Mottershead [30].
These give more comprehensive coverage of the very large amount of literature related
to model updating.
2.3.1
Direct Methods
The earliest methods which fall under the model updating heading are known as
The first advocates of these were Berman and Nagy [31] and Baruch [32].
These methods consist of updating a set of structural matrices by making change to
of the terms in a single step. The result is that while the updated finite element model
will exactly reproduce the experimental results, there is no guarantee that the model will
correctly predict the structural response in other loading or testing configurations. This
25
is inevitable since the changes have not been made with reference to the physical
relevance of the structural matrices.
[33] have recently examined in some detail the effect of making
Levin
changes to a set of structural matrices. They conclude that the inclusion of such terms
is likely to result in the representation of a structure containing grounded springs which
can have a profound and unrealistic effect on dynamic response.
Direct methods have been largely superseded by sensitivity and other forms of
optimisation methods. The resulting updated models are constrained to be physically
realisable and are outlined in the following sections.
2.3.2
Sensitivity Methods
Another generation of methods requires the determination of the sensitivity of a set of

updating parameters to differences in dynamic behaviour between analytical and
experimental dynamic data. The techniques yield an expression of the form
.
where
(2.45)
are a set of alterations to a set of updating parameters upon which the
structural matrices are dependent,
are a set of differences in dynamic behaviour
between an analytical model and experimental observations and
is the sensitivity
of observation
to change in updating parameter
The form of
depends on the choice of updating parameter (section 2.3.3). Apart
from certain exceptional cases, the changes in parameter value have nonlinear
relationships with the changes in observed dynamic behaviour. In such cases a first
order approximation of
is used and the best choice of
to match the
are
found iteratively.
These methods can be subdivided into techniques which use modal data and those
which use frequency response functions as the system descriptor. The methods both
continue to enjoy popularity and are examined practically in section 2.7. Furthermore
26
the ability of the eigenvalue sensitivity method of model updating to account for the
effect of load on structures is described in chapters 5 and 6.
2.3.2.1 Methods using Eigenvalues and Eigenvectors
The comparison of observed resonant frequencies and - to a lesser extent - mode shapes
with their analytical counterparts pre-dates model updating as the most obvious and
direct method of comparing analytical predictions of dynamic behaviour with
experimental observations. Indeed, manual methods of altering aspects of the analytical
model to bring predictions of resonant frequencies into agreement with observed values
was and is standard practice. It is logical then that early methods of updating should
seek to find the sensitivity of system eigenvalues to changes in certain structural
parameters.
The rate of change of eigenvalues with respect to structural parameters,
derived as follows [34]. Considering
and
, can be
which are the solutions to the
equation (2.6) derived in section 2.2.1.

0 ,
(2.46)
where the eigenvectors are mass-normalised such that

1.
(2.47)
If either or both of the structural matrices is a function of a parameter

differentiating (2.46) with respect to
gives
[
Pre-multiplying by
(2.48)
makes the right hand side of (2.48) vanish due by an
orthogonality argument. Re-arranging leads to

,
(2.49)

which is the rate of change of the
system eigenvector to the
27
updating parameter.
This relationship is used extensively in chapters 5 and 6 of this thesis.

Methods have also been developed which also employ discrepancies in eigenvectors.
This is partly to address the problem that only a relatively small number of resonant
frequencies are available for updating. The use of eigenvectors to update structures
suffers from their being relatively insensitive to structural modifications as well as the
level of uncertainty in their value when determined experimentally.
2.3.2.2 Methods Using Frequency Response Data
To address the issue of lack of information provided by the potentially small number of
resonant frequencies in a measured frequency range, researchers have turned to using
the response data directly. This has the additional benefit that there is no requirement
for time consuming and potentially inaccurate modal identification to be performed. A
further feature of the method is that some residual information from frequencies outside
the measured range is implicit within the experimental data.
The development of the so-called response function method of model updating is
generally attributed to Lin and Ewins [35] with a significant later contribution from
Fritzen [36]. The method employs the sensitivity of structural parameters to differences
in experimental measurements and analytical predictions of frequency response
functions.
The sensitivity matrix relates the structural updating parameters to the
changes in receptance by
,
[ ( )]
(2.50)
where the change in receptance term on the right hand side is defined as
( )
( )
( )
The sensitivity of the receptance at the

parameter
(2.51)
degree of freedom to the
at a given frequency point can be shown [37] to be
updating

2
( )
28
(2.52)
An appropriate choice of updating parameter leads to an iterative solution to find the set
of values which minimise the difference between experimental and FE predictions of
receptance at given frequency points.
A number of sensitivity matrices at as many frequency points as one wishes to compare
between experimental and analytical models can be constructed and stacked as
(
(
(
)
2)
1
2
(2.53)
to produce
equations in
unknowns where
and
are
respectively the number of degrees of freedom at which responses are available, the
number of frequency points and the number of (unknown) parameter values.
At first glance, the generally greatly overdetermined system of equations given by
(2.53) allows a large number of parameters to be updated with some confidence. In
practice there is a large amount of inter-dependence of the rows of the sensitivity
matrix. A number of techniques exist for selecting the most useful frequency points for
instance [38]. It is the opinion of the author however, that the extra information yielded
from considering differences in response is not significantly greater than that from
changes in eigenvalues.
2.3.3
Selection of Updating Parameters
The issue of choice of updating parameter is crucial in the development of finite

element model updating.
The brief history of the subject is punctuated by the
introduction of different updating parameter schemes. It is a general topic which

applies to whatever choice of comparison between experimental and analytical data is
made. Chapter 5 of this thesis investigates new types of updating parameter. It is
29
important to understand the motivation and requirements which determine the choice of
these parameters.
Considering a Taylor expansion about the original structural matrices in terms of a set
of updating parameters, the
of which is defined as
A first order approximation
for the updated systems matrices is found to be

(2.54)
1
and
.
(2.55)
where the summation sign represents matrix building4. If the first order derivative of
the structural matrices with respect to the chosen updating parameters can be found then
an approximation to the changes required to the structural matrices can be derived. A
best solution for
can be established by a process of optimisation of a generally over-
determined yet ill-conditioned problem (commonly by using singular value

decomposition described in section 2.4).
A particular case which has commonly been applied in model updating is to define the
updating parameters as factors of individual or groups of elemental stiffness or mass
matrices. The updated structural matrices are given by
(2.56)
1
and
,
(2.57)
where
and
are, respectively, the
elemental stiffness and mass matrices.
This equation is strictly true for elemental matrices having the same dimensions as the global value.
30
It is common for the updating parameters to be chosen to factor a group of elements or

super-element. This is because the number of individual elements in an FE model may
be very large and the amount of data available to inform the choice of updating
parameters relatively small.
Mottershead
[39] have considered a different approach whereby
of elements are updated principally to enable identification of joint

behaviour. Joints are known to be a likely source of error in the FE model. Specialised
finite elements have been developed in which the elemental formulation is recast in
terms of a structural parameter change.
This is likely to account for differences
between experimental and analytical models.
For instance the flexural and shear
stiffness of a 2D beam (equation (2.22)) is modified to include a flexible length

rigid part
and a
and is given by [40] as

12
6
12
12
12
12
4
12 6
12
2
6
4
(2.58)
with the corresponding mass matrix given by

156
[
156
156
22
44
54
4
54
420
The offset parameter -
13
156
13
13
22
42
(2.59)
- therefore becomes the updating parameter. It is noted that
while varying the value of
apparently has little effect on the stiffness matrix, the
flexibility (to which the lower modes of vibration are most sensitive) is changed
significantly by perturbations of the offset parameter.
A similar method involving the development of specialised finite elements to update a
particular model of a rubber seal is described by Ahmadian et al [41]. A finite element
model of the rubber component is built from first principles. Its dependence upon a
number of user-defined parameters (in this case two) can then be updated.
31
A more general approach for parameter selection, also motivated by the perceived
requirement to update structural joints, has been suggested by Gladwell and Ahmadian
[42].
They consider families of elemental stiffness and mass matrices that are
physically realisable and which will alter the overall dynamic characteristics of a
structure of which they are a part. Thus the
is any element of a family
which has certain properties in common with an initial elemental matrix. This approach
allows a great deal of scope in the changes one can make to elements while ensuring
that the element remains physically realisable.
It is therefore very attractive.
comparison of generic updating parameters with load-dependent updating parameters is

undertaken in chapter 5 (section 5.4).
While the generic element method appears to be deservedly gaining in popularity, many
practitioners continue to factor elemental stiffness and mass matrices as described by
equations (2.56) and (2.57). Given that information available to inform the updating
process is scarce, the choice of suitable and sensible parameters for model updating is
of paramount importance.
2.4
The use of Singular Value Decomposition in Model Updating
Singular value decomposition (SVD) is a very powerful tool in the consideration of the
equations which arise in the formulation of model updating problems. In the following
sections the use of SVD for optimising over and under determined multiple parameter
problems is considered. In addition the method is applied to estimate the conditioning
of sensitivity equations.
The method is described in [43] and its use in optimisation in dynamical systems is
comprehensively set out by Maia [44]. The singular value decomposition comprises
recasting a matrix into the product of two orthogonal matrices [U] and [V] as well as a
set of singular values
thus:
1
...
...
(2.60)

The columns of
vectors of
and
32
are respectively said to be the left and right singular
.
consists of a set of min
The matrix
singular values
, ... arranged in
descending order along the leading diagonal,

0
1
2
(2.61)
It is very common for updating procedures - and certainly those employed in the later
stages of this thesis - to require the solution of an equation of the form
(2.62)
where
is a set of measured values and
is a set of sensitivities of the measured
values
to changes in a set of parameters
. If the rank of
number of elements in
problems, both
solution to
and
is greater than the
the problem is overdetermined. Since in the case of updating

contain experimental data or derivatives of it, an exact
is unlikely to be obtainable. In such cases other criteria must be used to
select a set of values of
. While a set of equations for determination of updating
parameters can appear over-determined, the sets of equations are likely to be nearly
linearly dependent on one another. In other words the whole problem is
The SVD allows a quick insight into the rank of an updating problem. It can be shown
[43] to be
where
1
...
... 0
(2.63)
and
1
where
(2.64)
is a small threshold value below which sets of equations can be thought of as
not contributing usefully to the overall solution. A comparison of the ratio of the
33
singular value to the first in practical terms gives an indication of the practical rank of
the problem.
Turning to the solution of (2.62) the clearest choice is to minimise the value of the
,
, defined as
(2.65)
where the parameter
indicates that several different residuals can be defined based
upon type of norm. Most commonly the two norm or the so called Frobenius norm are
employed.
Golub and Van Loan [43] show that for any problem of the type, including the
underdetermined case, there exists exactly one
which minimises the sum of the
squares of the residuals,
min(
(2.66)
They show that for

( )
(2.67)
the least squares estimate of
is given by
( )
(2.68)
resulting in
2
(2.69)
( ) 1
The vector
projection of
can be thought of as a predictor of

onto the range of
The pseudo inverse matrix
and is the orthogonal
where
(2.70)
34
can be found to be
(2.71)
where
0
1
1
(2.72)
is the unique Frobenius norm of the residual
The pseudo inverse of

min
[ ]
Furthermore, if the rank of

1
(2.73)
then
(2.74)
clearly for the particular case

1
2.5
(2.75)
Issues Related to Using Experimental Data
The details of experimental dynamic testing of structures and the subsequent

identification of a dynamic model of a test structure will not be covered in detail here.
A comprehensive coverage of the area can be found in [4]. The extent to which a single
set of identified data can be thought of as uniquely defining a structure is currently an
area of active research. The variance of predictions of dynamic data from nominally
similar5 structures falls within this definition.
The effects of essentially transient
changes to a structures dynamic behaviour resulting from loading - which is the

principal topic of this thesis - being an important example.
5 The expression
is used a number of times throughout this thesis and refers to the fact
that a single FE model will be used to represent a single structure under any conditions as well as any
identically manufactured structure.
35
The effect of noise on experimental data along with the other well known difficulties
such as lack of frequency bandwidth and limitations on the number of measurement
locations is blamed extensively for any lack of success in using measured data to update
finite element models. The term can be used as something of a catch-all into which any
number of unidentified factors whose effect is to hinder successful model updating
using experimental data can be thrown. Different predictions of dynamic behaviour can
occur from the same structure at different times or from nominally identical structures.
However, the factors which cause this variation have only recently received some
attention and are discussed below.
Cafeo
[45] set out to consider the variability of test measurements in a methodical
manner, their motivation being the reduction of tests required while maintaining
confidence in the results. They observe that the variability in modal test data results
from either inconsistencies in the modal/test analysis procedure or in the structure that
is being tested. The former is investigated with reference to a set of 9 tests on 7
vehicles. Specifically the test/analysis procedure is subdivided into the following:
(a)
Linearity Considerations;
(b)
Shaker Attachment Method;
(c)
Measurement System;
(d)
Accelerometer Placement; and
(e)
Signal Processing Considerations.
The study of these individual components of a dynamic test provides the basis for a
thorough investigation in to the causes for test-to-test error. While they are not studied
in any great detail, the dynamic behaviours of the seven nominally identical vehicle
specimens are seen to vary quite considerably. Further investigation of the test-to-test
variability problem [46] leads to some very useful guidelines on methods to avoid
variability of measurements resulting from inconsistencies in the measuring and
analysis of structures.
36
Balms [47] considers the problem of a single FRF model being required to represent a
real structure. He notes that the behaviour of a single structure can be altered by time
varying factors such as ageing, temperature effects, loading conditions. Also, he states
that variation amongst a population of nominally identical structures can occur due to
manufacturing tolerances, residual stresses and welding point positions. The results of
the GARTEUR round-robin exercise whereby a number of laboratories tested a single
structure each with their own testing equipment are examined. The relative simplicity
of the model and the availability of free-free testing in laboratory conditions led to the
observation that differences in dynamic readings arose from slight differences in test
equipment and procedures.
De Clerck [48] notes the difficulty of comparing multiple sets of nominally similar data
and describes a method which statistically compares a number of sample modal
analyses. A relatively simple formulation involving the term by term comparison of the
sample eigenvector
with the average of the population is given by
(2.76)
A more sophisticated method involves the singular value decomposition (SVD) which is
covered more fully in section (2.4) below. The SVD of the set of eigenvalues is taken
1
.. .
(2.77)
and the average of the n eigenvectors in this way is given by

1
(2.78)
with the maximum normalised singular vector,

11
2
(2.79)
shown to give a measure of the consistency between a set of measured modal data. The
remaining singular values are found to represent the spatial variance of the set of
eigenvectors.
37
This approach allows the variation of mode shapes with respect to variability of
components of the model to be considered. The usefulness of these methods was
proved with a simple FE model whose characteristics could be changed and dynamic
response re-calculated many times and there appears to be a clear requirement to test the
effectiveness of the method in a real-life situation.
Ziaei-Rad and Imregun [6] appear to be amongst the first to consider the specific area of
understanding what level of noise in the experimental data can still allow a finite
element model to be successfully updated. They describe the problem in terms of the
range of error arising from the measurement of a single specimen. They acknowledge
that the method is equally valid in considering the variability in dynamic behaviour
between nominally identical structures. The method considers the response function
method of section 2.3.2.2 with a noise model of the form
1
where the dynamic stiffness matrix
(2.80)
is given by
and the subscripts
(2.81)
and
denote analytical (FE) prediction and experimental
measurement respectively. An upper bound upon the error matrix is found to be given
by
2
where
(2.82)
in this case denotes the Frobenius norm of the error matrix. This expression
indicates that the level of noise in the experimental data (or discrepancy between test to
test measurements) which can be tolerated by the response function updating technique
is a function of the type of mathematical model employed.
The common choice of updating factors of elemental mass and stiffness matrices
(equations (2.56) and (2.57)) lead to a formulation of the sensitivity equation (2.45)
which includes terms from the error matrix

,
where
38
(2.83)
are noise affected p-values. The solution is found by finding the pseudo
inverse (described in section 2.4 and denoted +) of the sensitivity matrix

.
(2.84)
Ziaei-Rad and Imregun show that for the case of the norm of the sensitivity error matrix
being much smaller than that of the sensitivity matrix itself then a solution is possible.
For the converse case they show that the extent of the noise leads to equation (2.84)
becoming increasingly ill-conditioned. This method allows a statistical consideration of
the variability in experimental data to be used to assess the likelihood of success in
performing model updating.
2.6
Updating Applied to Real Structures
Despite the large amount of literature published on the subject of finite element model
updating, there is a relatively small amount of information describing the experiences of
applying such techniques to measured data. Instead many of the methods devised have
been demonstrated and validated using only simulated experimental data - that is to say
that the data originate from an analytical model rather than from a real structure. The
use of simulated data offers the analyst the chance to assess in a systematic and rigorous
way the effects of various perceived aspects of experimental data.
Many of the
limitations which apply to experimental data can be accounted for and investigated in
simulations. These include:
truncation of experimental data in the frequency range; and
limitation on the number of measurement points upon a structure.
Other factors however can be less well understood by means of the use of simulated
case studies - for instance:
39
interaction between the test structure and excitation and measurement

equipment; and
noise resulting from the experimental process.
The importance of these factors is well known as is their detrimental effect upon the
success of any model updating strategy.
A number of authors have described experiences in applying
model updating
techniques in practical situations; a review of the literature follows. Much literature

uses the word updating to describe processes which amount to little more than manual
alterations to structural matrices. The following examples differ in that the updating is
in that structural parameters are altered in a number of steps according to
set of rules.
Hoff and Natke [49] appear to be amongst the first to apply model updating techniques
to practically measured experimental data in 1988. A radar tower was modelled using
3D beam elements and a lumped parameter model of the foundation behaviour resulting
in some 52 degrees of freedom. Unsurprisingly, calculations showed that eigenvalues
and eigenvectors were most sensitive to the stiffness in the region of the towers
foundation. These were used as updating parameters and converged updated values for
these parameters were obtained from eight identified resonant frequencies.
Imregun
[50] considered a box like structure in 1993. The response function
method (RFM) was used to update the finite element model of a sheet metal box
comprising 716 active degrees of freedom. Some of the material properties of the finite
element model were altered manually at the correlation stage. A limited number of
responses (one direction only) were used to update the properties of two of the sides of
the box but with limited success despite the choice of multiple sets of frequency points.
The authors concluded that the scale of the discrepancy between measured and
predicted FRFs was beyond that which could reasonably be reconciled.

At around the same time, Lammens
40
[51] considered the use of measured
experimental data for model updating. They took the cautious approach of using very
simple models - namely a PVC plate and a small frame structure. A method employing
the difference between pairs of frequency response functions was used to update a small
number of elementary updating parameters. The influence of frequency point selection
was considered practically as well as the use of including damping in the analytical
model. While finding that updating using real experimental data is possible - at least at
a very simple level - they arrive at a number of interesting conclusions. Firstly that the
choice of frequency points is important in influencing the likelihood of a successful
update. Secondly it was observed that the amount of information available from the
structure limits the number of updating parameters which can be updated. They also
found that zero damping in the analytical model produces more robust convergence than
attempting to include damping parameters in the updating process.
Imregun
[7] investigated both the response function method and the eigen-
sensitivity methods of model updating using experimental data from five nominally
identical L-shaped plates with spot welded connections. A significant observation from
this work is that the five specimens exhibited remarkably different dynamic behaviour.
The assumption that the global correction matrices can be found by multiplying
individual element matrices by scaling factors (equation (2.56) and (2.57)) is used.
Updating is performed using both the response function method (section 2.3.2.2) and an
eigen-sensitivity method (2.3.2.1). The eigen-sensitivity method was found to fare
better when augmented in a two stage approach with certain constraints. While the
method based on response function also produced satisfactory results, the choice of
frequency points was found to have a strong influence on the likelihood of success. The
choice of FRF estimator led to the response being ill-defined around resonance and
therefore not contributing to a well conditioned problem.
A further set of experiments was presented by Imregun
[52] with two nominally
identical structures built to compare with a simple FE model. The correlation between
41
one of the sets of data and the finite element model was found to be very poor and
hence was not used in the subsequent model updating procedures a worrying
observation.
Updating was performed using the response function method with
converged results only occurring after rejecting noisy experimental FRFs and rejecting
the use of damping parameters. It was noted that a number of sets of p-values led to
converged solutions clearly demonstrating that the solution is non-unique.
The
computational effort is investigated and shown to require the least square inversion of a
matrix of size
where
and the inversion of a matrix of size
equals the number of degrees of freedom,
points considered and
times
the number of frequency
is the number of p-values. It is noted that this amount of
heavy computation is practical only for very small models.

The application of model updating techniques has not been confined to small structures
tested under laboratory conditions. Imregun and Agardh [53] considered responses
measured from a three span road bridge.
Sixty five measurement positions were
employed and 12 modes were extracted in the 0-40 Hz frequency range and compared
with results on a 1050 degree of freedom finite element model with reasonably good
correlation found between pairs of modes. The noisy nature of the data led to the
decision to use the eigen-sensitivity method to update elemental stiffness factors
(equation (2.56)) with successfully converging solutions. The conclusions are upbeat
suggesting that the method could be used for damage detection. The authors note that
the use of dynamic data has become popular as a method for monitoring bridges and
other large structures for signs of damage. The finite element model of a structure can
then be thought of as a repository of information regarding current levels of damage and
deterioration which can be periodically updated using dynamic test data.
Waters [54] concluded an in depth study of the FRF sensitivity method by using high
quality experimental data from a plate wing model. High order frequency responses
were used and frequency points were selected to be away from resonance to avoid illconditioning. The plate was initially cantilevered but the clamping arrangement and the
42
bench upon which the specimen was mounted were found to influence the dynamic
response so free-free testing was used instead. The effect of interaction of the structure
and the shaker and stinger was also found to significantly pollute the experimental data.
The stiffnesses and masses of groups of elements as well as an element representing the
mass of the force transducer were updated and convergence was achieved.
Mottershead [55] summarised several case studies of application of updating of the
geometric joint properties described in section 2.3.3 to practical situations. The finite
element of a welded joint was successfully updated by increasing the initial offset
parameter by 3%. In a variation upon this theme a specialised element for modelling
the rubber interface between a car window and adjacent bodywork was developed and
two of the elemental parameters adjusted to give the closest fit to a set of experimental
data. The veracity of the updated rubber seal model was tested by re-testing the seal
with a metal sheet replacing the car window. The closeness of the agreement gives
much confidence in the accuracy of the updated model in a way that is not generally
available.
Mottershead and James [40] considered both joint geometry and generic elements to
update a three storey aluminium space frame structure. The sensitivities of resonant
frequencies to three elemental eigenvalues were considered. These contributions from
these updating parameters, however, added exclusively positive values to the overall
sensitivity matrix. Since the experimental results did not lie exclusively above or below
the finite element model prediction, no combination of the elemental eigenvalues could
adjust the resonant frequency prediction of the finite element model. To address this
problem the mass of the structure adjacent to its joints was also considered although it is
acknowledged that there was no reason to suppose that the mass was in any way mismodelled in the original model.
More recently Imamovic [56] considered the practicality of updating FE models using
measured experimental data in some detail. He concluded that methods employing
sensitivity of parameters to eigenvalues offer the best prospect of updating finite
43
element models of structures using measured experimental data. Several practical case
studies are considered. Experimental data from a cantilever plate6 mounted in a large
rubber block was obtained from several sources and the average experimental data were
used and compared with a finite element model of the structure where the principal
plate component was modelled using 60 plate elements. For the first seven correlated
modes, the initial finite element model produced resonant frequencies which were of
greater magnitude than their experimentally observed counterparts by factors of
between 7 and 27%. Due to the limited availability of resonant frequencies observed,
the stiffnesses of five plate elements closest to the grounded end of the cantilever were
used as updating parameters with a converged solution arising. Alterations to the
stiffnesses of these elements of between 40% and +20% are required to bring the FE
prediction into agreement with the observed data. The author inferred that much of the
discrepancy lay in the modelling of the joint between plate and foundation. While this
may be true it is also clear that resonances of the cantilever (as opposed to the low
magnitude resonances resulting from the straining the rubber foundation) are most
sensitive to the stiffnesses of this part of the structure. They would thus be changed
most readily by the updating method whether or not they actually contribute to the
discrepancies between FE prediction and measured dynamic data.
A further study involved a very large (92 000 DoF) model of an aeroplane engine
casing. Firstly a set of super-elements was chosen where elements were considered to
share properties. Multiple combinations of stiffness factors were considered with some
300 of these producing converged solutions. The best solution was chosen based on
issues such as good modal assurance criterion (MAC) [57] values between correlated
mode pairs (CMPs) and improvement in the agreement of modes not included in the
updating process.
The test data derived from a simple structure designed and built by Lloyds Register and tested by three
independent companies
44
These applications of model updating procedures to experimental data represent a

useful attempt to apply immature technologies in the hostile environment in which they
must eventually operate. The mismatch between the amount of information from a
modal test and the range of possible parameters in a finite element model still appears to
be the largest obstacle. Some converged solutions suggested in the literature appearing
at best to be rather arbitrary. It is the opinion of the author that the limited scope for
change of finite element models common to all of the examples described above means
that the wrong problem is being solved. Specifically, the choice of updating parameter
is crucial. Specialised updating parameters which allow the scope to alter a finite
element model to represent any possible structural configuration appear to present a
route to successfully implementing model updating in practice. This area of research is
pursued in chapters 5 and 6 of this thesis. In the mean time, a case study is presented
which allows some of the issues raised in this section to be demonstrated.
2.7
Case Study
As an introduction to the consideration of using experimental data to update finite

element models a simple case study is presented here. This is included to highlight the
more general features and drawbacks of model updating. The test data derive from a
real structure tested on several occasions. The use of real experimental data allows
some conclusions to be drawn about the requirement for repeatability of results.
It is prudent that in considering the possibility of updating real structures one should
start with a very simple structure allowing the areas of uncertainty to be isolated and
studied in detail.
The experimental specimen is a narrow mild steel plate of dimension 850
100
5mm.
Further work on this structure is described in chapter 4; a full description of the

experimental data collection techniques is deferred until that point. The tests upon the
narrow plate described in this section essentially validate the FE model.
45
Dynamic measurements were taken with the plate hanging in free-free conditions, the
experimental arrangement is shown schematically in figure 2.4. The response of the
plate was measured along the centre line of the plate using a scanning Doppler Laser
Velocimeter. Excitation was applied at a point 250mm from one end. Frequency
response functions between the load point and up to 18 response points along the length
of the plate were calculated.
2.7.1
Identified Resonant Frequencies
Dynamic measurements were conducted on three independent occasions. The same

equipment was used during each test but was disassembled in between sessions. The
tests will be referred to as A, B and C. The three test programmes differed in several
ways, a combination of point measurements7 and complete measurements were taken
during the three sets of tests and during test B point FRFs were measured at different
levels of forcing. Tests A and B involved measurements in the 0-800 Hz frequency
range while test C considered the range 0-400 Hz.
Two different modal identification methods were used. The global method outlined by
Fillod et al. [58] was employed to analyse the full sets of 18 FRFs while a line fitting
method [59] was used to identify frequencies from single point measurements. In
addition to using a large set of response measurements the global method also allows
identification of multiple modes. The line fit method by contrast might be used where a
single response is available and modes are sufficiently spaced that interaction is
minimal. The global method would be expected to provide the better quality estimate of
resonance frequencies since most information is used.
While these two modal
identification methods are common and representative, they are both members of an
increasingly large family of such methods. For a comprehensive and recent review of
the field one should turn to chapter 4 of [60].
7 Measurement
point and excitation point are coincident
46
Table 2.1 summarises the tests conducted as well as showing the results of modal
analyses performed on data from the structure. The two pairs of resonant frequency
data gained using the global method were extracted from the same data but at different
times.
Table 2.1 Identified Resonant Frequencies
The mean and standard deviation8 of the value estimated for each resonant frequency
are shown in table 2.2. The scarcity of data and the relative inter-dependence of
readings9 relating to the circumstances of their collection makes it difficult to draw very
firm conclusions regarding the variability of data. It is however interesting to note that
any one of the modal analyses would have allowed the modal analyst to be very
confident that the correct readings for the plate had been attained.
2
1
8 mean
=
and standard deviation =
number of observations.
, where
is the
observation and
is the
9 Each column for instance should be treated as independent of the other columns since the readings were
taken and to a lesser extent the modal analyses were performed at the same time.
47
Table 2.2 Comparison of Variation of Experimental Modal Analyses
The data suggest that certain modes can be extracted with some more consistency than
others, the standard deviation of the prediction of modes 3 and 6 is 20% of the standard
deviation of the first mode.
This difference in consistency could result from a
combination of a number of factors such as noise in certain parts of the frequency range,
close modes, lower levels of response etc. It is not intended to pursue the consideration
of the reasons for inconsistency of extraction of modal parameters.
2.7.2
Finite Element Modelling of Beam
The behaviour of the narrow plate which has been discussed in the previous section is
modelled as a set of two dimensional beam elements of the type described in sections
2.2.2 and 2.2.4 to keep the problem at a relatively elementary level. The principal
parameters used in the finite element model are shown in table 2.3.
Table 2.3 - Material and Geometric Properties Chosen in Updating
48
The predictions of resonant frequencies yielded by this simple method are shown in
table 2.4 along with the average experimental observations for comparison. The plate
was modelled with 17 two dimensional Euler-Bernoulli beam elements such that the
node points matched the measurement points on the structure described above.
Additionally the plate was modelled with twice as many beam elements and further with
34 elements the results are shown in table 2.4 along with the closed form prediction
[61]. The analytical readings, like the identified experimental readings, show some
variety.
Table 2.4 FE Predictions of Resonant Bending Frequencies
The coarsest discretisation using 17 elements each of length 50mm produces excellent
agreement with the experimental data. This would be regarded as a good starting point
for model updating. It is shown in section 2.2.2, however, that the finite element
method using elements derived from the displacement method described in appendix A
will
produce an overestimate of resonant frequencies. This is acknowledged and
in at least one case it has been studied [62]. However, the cause of mismatch between
finite element prediction and experimental observation is rarely taken into account.
This is because improving the finite element prediction dramatically increases the
processing time required to extract the eigen-solution. Some practitioners [63] employ
a rule of thumb that the first third of the resonant frequencies produced by a finite
49
element model can be regarded as mesh convergant. The similar FE model with double
the number of elements gives a very slightly lower prediction of the resonant
frequencies as one would expect.
However, one might argue that the elemental
formulation is less valid when elemental length is less than width.
2.7.3
Updating of Finite Element Model
As the proceeding sections have described, there is some uncertainty about which
experimental data should be used for updating. Furthermore the choice of parameters to
be updated is by no means simple.
In this case the test structure is sufficiently
uncomplicated that any error in the finite element model is expected to lie in global
properties such as mass and stiffness or in the added mass of the shaker and force
transducer arrangement.
To assess the effectiveness of updating using experimental data, the stiffness of the
beam between 200 and 250mm from one end (element 5 of the 13 beam element model)
was reduced by 23.1%. This value of artificial error implies that a factor of exactly 1.3
must be applied to the stiffness of this area to bring it to the same value as the rest of the
beam. The artificial error is analogous to an initial mis-modelling but allows the
usefulness of updating strategies to be quantitatively examined.
Eight updating runs were performed with the effect of some of the principal factors
varied. Three sets of experimental data were used, the description of each updating run
is given in table 2.5. It is important to note that each of the different problems are
realistic and that
could be chosen in practice.
50
Table 2.5 Initial Parameters For Experimental Data Updating
The three sets of experimental data referred to are the mean and extreme identified
resonant frequencies shown in table 2.4.
Only stiffness parameters are updated with the updated stiffness matrix being given by
1
(2.85)
where the second summation represents the elements which are not factored by the
updating parameters
Since only six resonant frequencies were available, updating of any more parameters
than this number of (independent) parameters leads to an underdetermined problem.
Two sets of updating parameters were chosen, the three stiffnesses of respectively
elements 3, 5 and 7 leading to an overdetermined problem and stiffness of elements 3
through 10 leading to an underdetermined problem.
Updating was carried out using the eigenvalue sensitivity method described in section
2.3.2.1
Convergence was assumed to have occurred when the largest change in any of the
updating parameter dropped below 1% of their original value and was achieved in each
case within five or six iteration representing quite robust convergence.
51
Figure 2.5 shows the value of the residual with each iteration. The residual is defined in
general terms in section 2.4, in this context it is defined as
2
where
are the six values of system eigenvalue at the
(2.86)
updating iteration. The
residual is therefore an indication of the extent to which the updated model matches the
experimental data. The least successful updating run is seen to be number 7, the large
error arises since it is does not attempt to minimise the difference between experimental
and finite element predictions of modes 4, 5 and 6. Updating run 8 on the other hand
which should suffer from the same problem matches all of the frequencies. While a
very simple example this is phenomena is likely to occur in practice when only a small
number of modes are available. Of the sample of measured resonances those occuring
at higher frequencies are seen to be much more useful in making changes to the finite
element model than the lower order modes.
The choice of updating parameters is also seen to be important in reducing the residual.
The under-determined runs 4 to 7 which alter 8 elemental stiffnesses using the six
resonant frequencies are successful as is run 8 which only alters modes 3, 5 and 7 using
modes 4, 5 and 6. The latter result suggests that attempting to minimise the difference
between the FE and experimental predictions of higher order modes is more useful than
using all of the modes.
A range of updating parameters arise from the eight runs and are displayed in figure 2.6
with the target value of stiffness shown as a dashed line. The updated parameter values
are seen to vary widely.
The variation of experimental data is seen not to affect the converged solution for the
overdetermined cases (1 to 3) although the final parameter values in the other cases are
seen to alter dramatically as a result of data variation.
One might argue that the variation in updated parameter values is to be expected from
altering a parameter which was not thought to be in error. This problem with the initial
52
FE model being artificially damaged is completely analogous to the problems

encountered in reality. If the stiffness of one element of the beam had indeed been misestimated the insight into the level of correctness of the initial FE model would not be
available.
On a more positive note, most of the updating runs successfully identified the location
of error in the initial finite element model with the overdetermined data producing the
consistently better results
2.8
Concluding Remarks
The background to the generation of finite elements using the displacement method
based on the theory of minimum displacement has been set out. It has been argued that
an understanding of the construction of elements from first principles is vital. The
method of estimating stress at locations about a structure from elemental displacements
has been demonstrated and will be employed to useful effect in the following chapter.
The transformation of elements from local to global co-ordinates has been set out. The
application of this simple congruent transformation can account for structural
deformations and this transformation is considered as an updating parameter in chapter
5.
The reasons for variability in observed experimental data have been discussed.
Literature on the subject has been reviewed and a number of approaches to
quantification of test-to-test error have been described. The review of the field of
variability of experimental data reveals that a methodical approach to the consideration
of the factors which cause sets of data from nominally similar structures to differ will
continue to provide a very fruitful line of research.
The development of model updating techniques has been briefly reviewed. Emphasis
has been placed on sensitivity methods with the eigenvalue sensitivity method and the
response function method outlined in detail. The use of Singular Value Decomposition
53
to solve the equation relating small changes in observed dynamic behaviour to changes
in updating parameter has been advocated and described.
The selection of updating parameters is considered in some detail. There has been some
conservatism in the choice of updating parameters. Recent methods which update
geometric parameters appear to offer a more versatile method of changing finite
element models, particularly in specific regions such as joints. This versatility comes at
the expense of considerable engineering judgement being required to assess the number,
location and type of updating parameters to use on a case-by-case basis.
A review of the application of
model updating methods to practical real-life
situations has been presented. It has been observed that practitioners have generally
been able to alter FE models such that they bring FE predictions of dynamic behaviour
into close agreement with
behaviour. However, the updated solutions have
been seen to be far from unique. Disparity between sets of experimentally collected
data from nominally identical structures has been observed. In addition a discussion of
the difficulty in discerning which amongst a large population of sets of parameter
values make the best changes to the finite element model has been included. This leads
to the conclusion that the updated models are rather arbitrary and may only fare better
than the original model in matching the observed dynamic data.
Some studies which test the usefulness of the updated model by using tests which are
independent of the updating process have been reported. These give rise to optimism
that the wealth of experience in developing finite element updating techniques can
eventually be successfully be translated into the practical arena.
A practical case study has allowed some of the issues related to applying updating
procedures to measured experimental data to be examined. By creating a known error
in the initial finite element model, it has been possible to show that experimental data
can be used to correctly make alterations to the finite element model. It was shown
however that the nature of the updated model is very sensitive to the updating
parameters chosen.
54
Variation in experimental data is seen not to decrease the chance of obtaining a

converged and plausible updated model. The variability, choice of updating parameter
and number of modes all appear to have a similar effect in leading to non-uniqueness of
the solution.
Higher order modes have been seen in the experimental case to provide the most useful
information in updating elemental stiffness parameters.
Figure 2.1 Beam Under Axial Load
Figure 2.2 Beam Under Shear and Rotational Loading
Figure 2.3 Transformation of Axes
55
56
Figure 2.4 Schematic of Experimental Arrangement; Testing of Narrow Plate
Figure 2.5 Residual Values
Figure 2.6 Updated Parameter Sets
57
Chapter 3 Modelling of the Dynamic Behaviour of Loaded Structures
58
CHAPTER 3
MODELLING OF THE DYNAMIC BEHAVIOUR OF LOADED STRUCTURES
3.1
Introduction
Structures by their very nature are likely to experience a variety of loading conditions
during their service lifetimes. The effect of static or quasi-static10 loading upon a
structure will at the very least alter the magnitude and direction of internal stresses as
well as causing deflection. Since a structures dynamic behaviour is a function of these
characteristics it will be dependant on the loading conditions. It is therefore commonly
understood that structures undergoing loading will experience changes to their dynamic
behaviour. This very phenomenon is exploited when tuning a stringed instrument for
example, as the pitch of a tensioned wire is observed to change with its axial loading.
The common practice of modal testing of structures in unconfined free-free conditions
minimises the interaction of structural components with the outside world.
The
consequence is that the dynamic behaviour of the test structure itself can be considered
in isolation.
The result of this approach is that the effect of load-dependence of
dynamic behaviour upon validation and updating of finite element models has been
largely overlooked. Given that structures are likely to experience a range of loading, we
take as premises that:
(a)
loading can effect dynamic readings; and
(b)
the structural changes which result from loading can be accounted for
in a finite element model.
10 This implies that while loading might change during time, the rate of change is sufficiently small that
inertial forces need not be taken into account when considering equilibrium. This definition applies
whenever static loading is referred to later.
59
Thus, it is incumbent upon the engineer to investigate the possibility of altering finite
element models to reflect the loading conditions.
Two principal factors altering the dynamic response of loaded structures will be
considered in this chapter, they both fall under the general heading of
. The first is stress (or geometric) stiffening in which axial load in slender
beams and membrane forces in thin plates or shells interact with small out of plane
deflections. The second concerns11 a structures form while under static loading which
will deflect and hence alter the response. While it is not inconceivable that static
loading will affect the dynamic response of a structure in other ways - for instance the
damping at joints - these are less easily defined and modelled and will not be considered
in this thesis.
It is important to note that although the consideration of large displacements falls into
the category of nonlinear geometry, this term refers to the
force and displacement.
relationship between
The solution method itself involves approximating the
nonlinear relationship by a number of linear steps. The validity of performing an

eigenvalue analysis upon a structure which has been altered in the nonlinear sense still
holds. Figure 3.1 shows schematically that for levels of dynamic loading which are
small compared with the static loading, the stiffness of the structure can accurately be
represented by the
In the experimental studies in the following
chapers the amplitude of the dynamic load is of the order of 0.1% of the static loading.
Both stress stiffening and large deformation can be accounted for in finite element
analyses.
They each require at least a single reformulation of the finite element
structural matrices.
For large load levels, flexible structures, or if high solution
accuracy is required, the structural matrices might require reformulation a large number
11 A third factor concerning the bowing of structural members (or finite elements) completes a trio of
factors quoted by most authors, and is investigated in detail in [64]. This deformity necessarily involves
large elemental strains. In the context of finite element modelling this effect is an extension to
discretisation inaccuracies and thus while its effect is noted it is regarded as being insignificant compared
to stress stiffening and geometry changes.
60
of times. Each of these actions potentially increases the cost of performing the analysis.
It is interesting to note that in the field of validating and updating finite element models
using dynamic data, facilities for including the effects of nonlinear geometry in modal
analyses offered by most commercial finite element packages are rarely12, if ever, used.
As the previous chapters have discussed, finite element model updating seeks to make
changes to a finite element model. These result in the finite element model not only
producing a closer dynamic response to the measured behaviour but being a closer
facsimile of the structure itself. More specifically, the types of updating parameters
which can be changed by finite element updating schemes have evolved to address the
supposed sources of error in the FE model (see chapter 2). This chapter investigates the
changes in FE models required to account for loading so that they can be included in the
arsenal of updating parameters. The implementation of alterations to finite element
models to account for loading upon a practical experimental example is described in
chapter 4.
The phenomena of stress stiffening and large deflections are described in detail in the
following sections. The derivation of the finite element formulation of each effect is
demonstrated and their interdependence is described. The remainder of the chapter
discusses the influence of these effects on dynamic modelling of structures and the
practicalities of their inclusion in validation and updating procedures.
3.2
Stress Stiffening
Stress stiffening arises due to the interaction of an axial load on a thin structural
member13 with transverse motion. Lateral deflections are assumed to be sufficiently
small that the axial load can be regarded as constant throughout the transverse motion.
The influence of stress stiffening in the area of stability analysis of columns and shells
12
The author has been unable to find references for instances where the finite element model of a
structure is altered prior to dynamic validation.
13 Generally applicable to beams and plates. In the latter case it is a membrane load which interacts with
transverse movement.
61
as well as the vibration of slender structures is well established. Euler [65] took an
interest in the subject in the mid eighteenth century; the ultimate load sustainable by a
perfect slender column which bears his name results from a consideration of stress
stiffening. The formulation is a function of the geometry of the slender structure and of
the axial/membrane force itself but
of the elastic properties of the structure.
The importance of the phenomenon in a structural dynamics context is also well known.
Lurie [66], for instance, explored the relationship between lateral vibration and the
elastic stability of beams and plates in 1952. The relationship between frequency and
end load were investigated experimentally on both struts and plates. In the former case
the behaviour of two simple frameworks were examined.
The use of vibration data to assess the loading conditions and fixity of columns has
been of perennial interest to engineers. Stephens [67] appears to be the first to consider
the problem presenting a paper which claims that both parameters can be identified
from the first resonant frequency although some corrections to Stephens work were
made some years later by Lurie [68]. Closer to the present day Plaut and Virgin [69]
consider the steady state linear response to transverse harmonically varying load. The
influence of an axial load specifically on the response is investigated for several
boundary conditions.
The work of Lurie and others pre-dated the widespread use of the matrix methods in
structural analysis and in particular finite element methods. Therefore, the complexity
and range of structures which were amenable to analysis using simplified governing
differential equations of motion were rather limited. The widespread implementation of
the finite element method has revolutionised the size of problems upon which the
effects of stress stiffening can be determined. As the following sections (principally
section 3.2.2) will show, stress stiffening can be incorporated in the finite element
method by the addition of extra terms to elemental stiffness matrices.
The following sections firstly derive the influence of axial loading on the closed form
solution. This is followed by a description of the technique that implements the effect
62
in the finite element method. In common with previous and subsequent chapters both
methods will be illustrated by means of the Euler-Bernoulli beam.
3.2.1
Closed Form Solution
Before investigating how stress stiffening effects are accounted for in the finite element
method it is instructive to demonstrate how the effect can be included in closed form
solutions. Note that the range of structural forms whose behaviour can be modelled
using closed form differential equations is limited.
The following relates to the
development of the Euler-Bernoulli beam.

The basic governing equation for a beam can be found for instance in [70]. The loads
acting on a small section of beam of length
are shown in figure 3.2. Taking
moments about the point marked P gives
0,
where
is the shear force and
(3.1)
the bending moment. This leads to
(3.2)
Continuing by resolving vertically yields

2
2
(3.3)
Also recall that a simplified relationship14 between bending moment and curvature can
be expressed as
2
2
(3.4)
14 This formulation neglects the influence of shear strain, the result being the so called Euler-Bernoulli
beam model. If more accuracy is required an approach which considers the shear strain results in the
Timoshenko beam formulation.
63
Combining (3.2), (3.3) and (3.4) leads to the familiar partial differential equation
describing the dynamic behaviour of a straight two dimensional beam without external
loading.
2
2
2
(3.5)
For a uniform beam this can be further simplified to

4
2
4
(3.6)
This can be solved for various boundary conditions. Substituting ( , ) as the product
of shape
( ) and harmonic component sin(

( , )
( ) sin(
) gives
(3.7)
substituting (3.7) in (3.6) leads to

4
0.
(3.8)
Substituting
(3.9)
provides a suitable solution for (3.8) if

2
4
0
(3.10)
which has the four roots

2
and
where
1
4
(3.12)
The general equation of motion is therefore

1
sin
cos
sinh
cosh
(3.13)

The unknowns,
, ... ,
64
, can be determined by applying the boundary condition
constraints.
Equation (3.2) can be modified relatively simply to take account of the effect of a
axial load, N shown in figure 3.3. This assumes that the transverse deflections
during vibration are small enough that no time varying axial load needs to be
considered:
(3.14)
Using this modified version of shear force we obtain a new partial differential equation
describing the motion of a bar experiencing an axial load:
4
2
4
2
2
(3.15)
Exact solutions are only available for certain end conditions. If, for instance, the ends
of the beam are considered as being pinned then
( , )
sin
sin
(3.16)
is a solution to (3.15) provided

4
2
0.
(3.17)
The ability of the time dependent term in equation (3.16) to satisfy the modified
differential equation of motion shows that the axial load does not affect the simple
harmonic nature of the solution. However, the natural frequencies are altered with the
resonant frequency being given by
2
1
2
(3.18)
The resonant frequencies are seen to be a function of the axial load N. Substituting the
well-known Euler buckling load

2
65
2
2
(3.19)
leads to
1
2
(3.20)
The relationship between transverse vibration and elastic stability becomes very clear.
It is seen that
2
(3.21)
The mode shapes can also be shown to be affected by stress stiffening. Experimental
results by Lurie [66] show close results for tests on beams. For a plate15 he finds
significant nonlinearity which he surmises is a result of initial curvature of the test
specimen16. Massonnet [71] experienced similar results when considering the effect of
circumferential loading of circular plates.
He concludes that the perturbation to
resonant frequencies are influenced by initial deformation.

Many authors consider only the change in the first resonant frequency with a view to
studying the stability of the structure. Indeed the literature related to estimation of
buckling load from frequency characteristics appears to concentrate exclusively upon
the first mode of vibration. To the modern structural dynamicist, perturbations to any
frequencies are of importance. The observation that
of the flexural modes will be
affected by axial load as shown in (3.21) bodes well for including loading effects in
updating procedures.
15
This need not apply to exclusively to plates, a 3D beam with low second moment of area in one axis
would demonstrate the same behaviour.
16
The initial out-of-shape of structures is considered in a practical sense in chapter 4.

3.2.2
66
Finite Element Formulation
The closed form solution described in the previous section provides great insight into
the influence of axial load upon simple structures. The inclusion of these effects into
matrix methods of structural analysis allows stress stiffening effects to be considered for
much more complex structures.
The application of the phenomenon described in the previous section was applied to the
matrix method of structural analysis of both plates and beams by Gallagher and Padlog
[72].
Continuing to consider the example of beam-type elements, a method for
determining the stress stiffening effects on finite beam elements is described in [25].
The strain of a beam in bending, with only the normal stresses taken as contributing to
the energy stored is given by
2
2
where
1
2
(3.22)
is the axial change in length, ( ) is the transverse displacement and
is the
distance from the neutral axis. The first term of (3.22) is the axial strain term, the
second term is familiar from beam theory relating to the energy stored in the beam in
curvature while the third term expresses the lengthening of the beam element due to its
rigid body rotation.
The strain energy is given by
2
(3.23)
Substituting (3.22) into (3.23) leads to

2
2
1
2
For constant cross sectional area,
(3.24)
, expanding and omitting higher order terms gives
2
0
(3.25)
67
Recalling the relationship between general lateral displacement,
and end
displacements from equations (2.16) - (2.19)

1
(3.26)
2
2
differentiating with respect to gives

1
6
2
6
2
(3.27)
2
2
and further
1
2
2
12
6
2
12
1
2
(3.28)
2
1
Substituting (3.27) and (3.28) into (3.25) and minimising the total potential energy with
respect to the nodal displacements leads to the elemental beam stiffness previously
derived in chapter 2 equation (2.23) - plus the stress stiffness matrix
0
0
6
5
0
10
2 2
15
0
0
0
0
0
6
5
10
2
10
0
6
5
where
30
(3.29)
10
2 2
15
is the axial load. Like the formulation for the 2D beam element derived in the
previous chapter, the formulation is not a unique representation of the effect of stress
stiffening but is commonly used in commercial FE software.
If torsional stress stiffening is ignored, a similar expression can be gained for the
modification to a three dimensional beam element using a symmetry argument.

0 0
6
5
6
5
0
0
10
0
2 2
15
10
2 2
15
0
0
0
6
5
10
6
5
10
10
30
6
5
6
5
10
0
10
0
2 2
15
68
30
(3.30)
10
0
0
0
2 2
15
3.2.3
Implementation of Stress Stiffening
Most commercial packages allow the implementation of stress stiffening although it is

not generally applied by default since modification to the stiffness matrix after at least
one initial load step is required to determine axial loading. The perturbation of natural
frequencies can easily be determined by modifying elemental stiffnesses by the
expressions such as those given in equations (3.29) and (3.30), for a 2D and 3D beam
respectively.
There is no limit to the range of loading to which the technique accurately applies
although the derivation holds only for small rotations from the original element shape.
For situations where significant elemental rotation occurs the change in geometry of the
structure must be taken into account; this is investigated in section 3.3.
The question of whether the extra processing time and hence expense involved in
including stress stiffening effects has occupied a number of authors. Ryu
[73]
reviewed methods for determining whether stress stiffening effects should be included
in dynamic simulations of rotating multi-body structures. The method proposed was
intended for assessing whether stress stiffening reaches levels where static buckling is
likely and is therefore not sensitive enough for assessing slight frequency perturbations.
69
Finite element model updating is a computationally intensive procedure. Therefore the

extra cost in performing a loading step and subsequent eigenvalue analysis does not
represent a significant extra cost and is an effective and useful exercise. This is
particularly true in situations where dynamic data from a loaded structure are used to
validate or update an FE model.
The engineer must consider all possibilities of loading to the structure including thermal
effects.17 The application of the worst case of loading (as one would perform in the
design process) should include stress stiffening effects. The resulting system matrices
would lead to an eigen-solution to determine the dynamic behaviour. This will enable
the sensitivity of frequency response with respect to loading conditioning to be
reasonable easily assessed. This information allows the engineer to discern whether the
perturbations to resonant frequencies are of a scale which might affect a subsequent
validation exercise. If the resonant frequencies are found to be sensitive to the loading
level then the analyst must consider how best to monitor the load on a structure. This
enables the effect of load to be taken into account. Issues relating to the identification
of loading and the sensitivity of resonant frequencies to load are investigated in the
following chapter.
3.2.4
Updating of Stress Stiffening Effects Using Conventional Parameters
As the term
implies, the effect of an axial load is to change apparent
transverse stiffness. The changes to structural matrices required to model the stress
stiffening effects correctly cannot be accounted for by conventional finite element
updating parameters. Therefore, the updated model may exhibit the correct modal and
response properties, but on a physically unjustifiable basis [74].
17 If the structure being tested is likely to experience a range of temperatures then the thermal expansion
coefficient for all structural materials should be determined and included in the structures properties in
the finite element model.
70
A simple case study allows the effect of stress stiffening on conventional model
updating procedures to be investigated. Consider the beam structure shown in figure
3.4 consisting of 13 elements modelled with simple 2D beam elements. The section and
material properties are the same as the narrow plate structure described in the previous
chapter (section 2.7). The structure has fixed ends and the whole beam is subjected to
an axial load of 5kN18.
The first four bending resonant frequencies of the beam
predicted by finite element modelling are shown in table 3.1 along with the
corresponding frequencies from an unloaded beam.
Table 3.1 Change in Resonant Frequency Under Loading
The bending frequencies are seen to be shifted by the approximately the same absolute
value. It is well known that altering the overall structural stiffness on the other hand
increases (or decreases) all frequencies by a common
A rudimentary model update using the sensitivity of the first resonant frequency to the
overall stiffness leads to the observation that the stiffness of the initial finite element
model needs to be decreased by around 25% (or the mass increased by 25%). Figure
3.5 shows the point receptances (the excitation point being shown in figure 3.4) arising
from matching the first resonant frequency of the updated structure with the
experimental counterpart. It is very clear that the updated model - while meeting the
single criterion set out in the updating procedure - is a significantly worse dynamic
representation of the structure than the original finite element model. This is the worst
18
The Euler buckling load in this case is 20.5kN.
71
case scenario and the inconsistency of the updated dynamic behaviour with the higher
experimentally observed frequencies would be noted. If the first four frequencies are
used to update the overall stiffness a reduction of 5% is reported, the resulting
receptance is shown in figure 3.6.
In practice if more than one frequency is used to update the overall mass or stiffness of
the structure, a poor final residual value should indicate that the observed frequency
perturbations do not derive from mis-modelled overall stiffness.
3.3
Large Deformations
When structural loading is sufficiently large (or the structure particularly flexible) that
the geometry of the structure changes during the load step, the stress stiffening
approximation to the structures loaded characteristics may not be sufficient. To account
for the deformity, equilibrium equations must be written with respect to the deformed
geometry
. Although it is assumed that no components
of structures undergoing modal testing will stray into plastic ranges, the solution
procedure for geometric nonlinearities shares some similarities with the treatment of
material nonlinearity.
Commonly the structural deflection is determined in a single step from the structural
matrices and loading. The stiffness matrix - which was constructed from the initial
estimates of geometry and material properties - is assumed to describe adequately the
stiffness distribution representative of the structure for the entire load step. Cases exist
where combinations of large loading and structural flexibility make it difficult to
determine the deflected geometry particularly accurately. In instances such as these the
assumption that the stiffness remains constant throughout the load step might be
inadequate.
An important concept is introduced at this point. The displacements experienced by an
element can be split into those which cause straining of the element and those which
cause rigid body motion of the element. That is

.
72
(3.31)
of any
This is a simple yet important relationship. The rigid body displacement
finite element consists of both translation and rotation components although only the
latter has any effect on the elemental matrices. Therefore, the geometry for each
element is a relatively simple rotational transformation. Figure 3.7 shows the rigid
body and straining deformations for a two dimensional beam where the accurate
determination of the rigid body rotation
is the most important term since it defines
the deformed shape and thus the dynamic response.

Most commercial finite element packages allow the implementation of both material
and geometric nonlinear techniques. Indeed some packages, for instance ANSYS [16],
have developed code to address this problem.
3.3.1
Finite Element Formulation
The simplest way to model the nonlinearity resulting from static loading is to discretise
into a number of linear steps. This involves recalculation of the structural matrix at
each point as suggested by Turner
[75]. The displacement at the
load
step is given by
1
1
(3.32)
The effect of loading upon an element can be included after the first load step and thus
the effect of stress stiffening can be included in the nonlinear analysis. Improved
results can be obtained by increasing the number of load steps. An example of this
approach for a single degree of freedom system is shown in figure 3.8 where the upper
dotted line shows the result of recalculating the stiffness parameter at several discrete
points. It is clear that a purely iterative approach can lead to a relationship between
stiffness and deflection which diverges from the true relationship. As the size of the
load step decreases towards zero however the predicted line will converge upon the true
one.
73
Another approach which is set out by Jennings [76] is to combine the previously
described incremental approach with a series of equilibrium iterations.
A very large amount of literature is published on the subject. Much of this - [77] being
a good example - is related to determining accurately the force displacement
relationship for very large elemental rotations. This situation is most likely to occur
near collapse. As far as the current discussion is concerned, the interest is not in
extreme deformations but small deformations to which the dynamic behaviour of the
structure can be sensitive.
In this case elemental rotations while needing to be
determined accurately are not of a large magnitude.

For illustrative purposes, a method for establishing the nonlinear relationship between
load and displacement will be described. The method is an iterative solution and
involves moving co-ordinates. It is clearly set out by means of an example in [26].
Considering a structure exhibiting small deflections under the action of small loads
,
where
(3.33)
are sufficiently small changes in deflection that
is linear. The forces
comprise the specified loading plus loading applied at nodes by elemental distortion.
For equilibrium we must achieve
0.
(3.34)
At this point, the example of stiffness changes to a 2D beam will be considered in

detail; note however that the method is general. The straining displacement of such a
2D beam can be described by two end rotations and an axial displacement relative to
suitable local axes
0 0
where definitions of
and
(3.35)
are shown in figure 3.7. The loading upon a single
element resulting from the deformations are found to be

.
74
(3.36)
Summing over all elements and substituting into (3.32) to find the global displacements
gives
1
(3.37)
where
represents matrix building of
elements rather than straight summation
and
is the external static loading on the structure.
The overall structural
deformations are updated at each step using the relation

X
i 1
i 1
(3.38)
The most convenient convergence criterion is to consider the incremental

displacements. Bergan and Clough [78] recommend that convergence should be based
on the ratio:
max
where max
is the maximum displacement of the same type as
(3.39)
. The lower
this value is set, the tighter the convergence. The choice of convergence criteria
requires engineering judgement. If the deformed structural matrices are to be used to
perform a subsequent dynamic analysis, then the sensitivity of eigenvalues and possibly
mode shapes to small changes in structural geometry must be considered.
Nonlinear geometry methods that take account of large deflections are a superset of
those described in the previous section. These account for the stress stiffening effect.
That is, the stress stiffening elemental stiffness modifications can be included in the
iterative procedure. These account for the transverse lack of stiffening affecting both
the static solution and the subsequent resonant frequency.

3.3.2
75
Effect of Boundary Conditions
The influence of deformation of structural members upon vibration characteristics is

itself very strongly influenced by the boundary conditions. This can be demonstrated
by means of a simple simulation based on a real situation. A plate such as the one
described in section 2.6 was ordered from a standard supplier and found to contain a
noticeable camber arising from the manufacturing process. The plate deflection is
plotted with equally scaled axes in figure 3.9. It can be seen that while discernible, the
camber is far from spectacular. The maximum deflection is 4mm over the span of
850mm.
The plate was modelled using seventeen 2D beam elements as described in chapter 2,
with and without the deformation included and with clamped and free boundary
conditions. The results are shown in table 3.2 from which it can be seen that there is no
difference to the frequency perturbation for the free plate but a considerable (33%)
increase in the fundamental resonant frequency for the completely fixed situation. The
increase in the first natural frequency with amount of deformation is shown in figure
3.10.
Table 3.2 Interaction of Boundary Conditions and Deformity
The observation ties in with early analytical work on the vibration of curved bars
[79,80] showing that for a beam with symmetrical curvature the first mode of vibration
is most strongly altered.
This observation has interesting and significant repercussions on the process of
validation of finite element models of substructures. If the plate is tested independently
of a structure into which it will eventually form a part, the response of the deformed
76
plate will be indiscernible from that of the straight structure. However, if the deformity
of the plate exists when assembled into the superstructure, the stiffness of its constraints
will affect its dynamic response considerably.
Generalising from this simple case by considering a more complex structure containing
components which are deformed. The deformation could be the result of any one of a
number of factors such as manufacturing imperfections or simply in response to inservice static loading. The dynamic behaviour is likely to be altered but unlike the
simple beam example described above whose first mode is sensitive to the deformation,
a wider range of modes are likely to experience perturbations. This hypothesis is tested
by means of an experimental case study in the following chapter (section 4.3).
3.4
Effect Upon Validation / Updating
As the foregoing discussion has shown, the effect of stress stiffening upon structures is
to alter the apparent transverse stiffness of structures subject to axial loading. The
result is that all of the
resonant frequencies are shifted in the same direction
while extensional frequencies are not altered.

The changes to elemental matrices to account for the change in behaviour derived in
section 3.2.2 cannot be included in an updated finite element model using any
conventional updating parameter. Limiting the modifications to elemental stiffness,
density, or cross sectional area is inadequate.
The highly underdetermined nature of updating parameters is such that a huge number
of sensible changes to arbitrary sets of updating parameters will result in an apparently
successfully updated model.
However, if loading is present in a structure then
modifying its dynamic behaviour through a model update is invalid.
The stress
stiffening matrix alterations necessary for making justifiable changes to the original
finite element model are omitted.
The same observation can be made for the transformation resulting from the largedeflection transformation. In this case the transformation consists of a rigid body
77
rotation. To take account of the changes to an elemental matrix caused by loading in an

updating procedure the updated matrix would take the form
,
where
(3.40)
is a rotational transformation of the type described in chapter 2.
The
simplicity of this relationship belies the computational effort which is required to derive
and
. This is significant and the relationship is exploited in chapter 5 which
considers the use of these quantities as updating parameters.

3.5
Concluding Remarks
The factors which are responsible for changes to dynamic behaviour of structures have
been set out. Methods for incorporating these techniques into finite element models
have been outlined. The specific case of the 2D beam element is considered in some
detail.
The relative importance of two phenomena, namely stress stiffening and structural
deformity, in altering dynamic behaviour of structures has been investigated.
The apparent stiffness of boundary conditions are shown to be crucial in influencing the
extent to which the dynamic behaviour of a deformed structural component is altered. It
is shown that the a deformed component tested in free-free condition displays much the
same dynamic characteristics as its undeformed counterpart.
The phenomena which result in changes to dynamic behaviour under loading have been
shown to be encapsulated by relatively simple alterations to finite element structural
matrices. Conventional model updating techniques using elemental mass and stiffness
used to update loaded structures are invalid.
Figure 3.1 Tangent Stiffness
2
2
Figure 3.2 Forces Acting on A Small Portion of Beam
Figure 3.3 Lateral Movement of Small Beam Element Under Constant Axial
Load
78
79
Figure 3.4 Fixed-Fixed 13 Element Beam
Figure 3.5 Updating Overall Stiffness of Loaded Beam Using First Eigenvalue
80
Figure 3.6 Updating Overall Stiffness of Loaded Beam Using Four Eigenvalues
Figure 3.7 Rigid Body and Straining Displacements of 2D beam
Figure 3.8 Iterative vs. Incremental Tangent Stiffness Calculation
Figure 3.9 Measured Deflection of Plate
81
Figure 3.10 Change in Natural Frequency Magnitude of Deformation
82
Chapter 4 Experimental Study of the Dynamic Behaviour of Loaded Structures
83
CHAPTER 4
EXPERIMENTAL STUDY OF THE DYNAMIC BEHAVIOUR OF LOADED
STRUCTURES
4.1
Introduction
The previous chapters have discussed how changes to dynamic behaviour resulting from
loading to structures can occur. Finite element techniques for taking account of the
changes to structures due to loading have been developed. To assess the validity of
finite element models which have been altered in this way, some practical validation of
the statically updated finite element model is required. To this end, this chapter
presents two experimental case studies in which the dynamic behaviour of structures
subjected to controlled loading is investigated. The results are compared with the FE
modelling techniques described in the previous chapter.
The first case involves the consideration of a slender structure whose state of loading
can be altered due to its inclusion in a relatively substantial testing structure. This
arrangement is directly analogous to the situation where a structural component or
substructure19 - is tested and its FE model validated in free-free conditions before being
included in a larger assembly.
The direct identification of static loading to a structure is likely to be impractical in a
real-life situation. This is because the loading can arise as a result of such wide range of
factors. The experimental study describes a method of identifying load indirectly from
surface mounted strain gauges. The example shares clear similarities with methods
which measure bridge strain using optical fibres [81].
19 A sub-structure is a constituent part of a larger structure some of whose properties are determined
independently of its parent structure.
84
The second experimental example allows the influence of loading to be investigated in

isolation. The acquisition of spatial data enables some conclusions to be drawn about
the influence of the loading on mode shapes. Direct identification of loading is possible
in the second experimental set-up.
4.2
Test on Narrow Steel Plate
A test rig was designed and constructed [82] to apply loads to certain simple structures.
This section considers the behaviour of a narrow steel plate under loading. The freefree behaviour of the plate has previously been studied in chapter 2, being mild steel
and measuring 100mm wide by 5mm deep and 850mm long. The shape was chosen
such that the structure could reasonably be modelled using either plate or more
elementary beam finite elements.
The narrow plate underwent dynamic testing under several static loading regimes. The
test allowed some of the issues related to the determination of loading in real structures
to be investigated. The study demonstrates why the resulting changes to the structure
should be accounted for in a subsequent dynamic analysis.
4.2.1
Application and Identification of Loading
A schematic of the experimental arrangement is shown in figure 4.1. Loads are induced
in the test structure by changing the position of the end masses relative to the
adjustment bars by means of nuts on the threaded ends of these bars. The 100mm at
each end of the plate were clamped within the two large masses which comprise the test
rig producing an encastr connection. The test rig was designed to be sufficiently stiff
that the test specimen experienced boundary conditions approaching fixity. The two
end masses weigh 100kg each and the adjustment bars are solid and measure 40mm in
diameter.
As in most situations where direct experimental determination of loading onto
individual components is impractical or impossible, the loading on the structure within
85
the testing rig needs to be determined indirectly. To this end, small single-element
electrical resistance strain gauges were placed at a number of locations along either side
of the plate. These allowed the longitudinal strain - and hence stress - distribution on
each surface of the structure to be assessed. This in turn permitted identification of
loading applied to the plate relative to an initial set of conditions. The strain gauges and
associated wiring were sufficiently unobtrusive to have little effect on the plates
stiffness. As chapter 1 has outlined, this type of instrumentation strategy would offer a
versatile method of instrumenting pre-existing structures to help to characterise their
loading regimes through time. More specifically, substructures tested with free-free
boundary conditions could be instrumented to determine loading introduced though
boundary conditions. The measured loads could then be used for dynamic validation of
the in-situ finite element model.
Figure 4.2 shows the location of the strain gauges along the plate. It also indicates the
co-ordinate system and the single point at which excitation was applied and dynamic
response was measured. The strain gauges were placed in pairs on opposite sides of the
plate. Strain readings were logged continuously throughout the testing programme,
beginning with the plate resting in an entirely unconfined state and ending in the same
condition. Four different load regimes were introduced in to the specimen by adjusting
the length of the bars joining the masses. The four load cases will henceforth be
referred to as cases A, B, C and D. Note that case A was intended to represent the plate
with zero loading.
The discrete points in figure 4.3 show the measured stress
distribution along the top and bottom of the plate in each of the four load cases
assuming a linear relationship between stress and strain.
Identification of the loading applied to the plate from a knowledge of the stress
distribution is not a trivial task. The strong nonlinearity of the stress distributions raises
some doubt about whether the use of a simple linear model to estimate the stress - from
bending moment and axial load - would produce a sufficiently accurate prediction of
loading. The most rigorous method available using the finite element method is to
86
determine the deflected shape of the plate accurately using a nonlinear geometric
techniques described in the chapter 3.
The consideration of less computationally
intensive methods of determining loads from the measured stress distribution are
described at the end of this section.
The plate was modelled in MATLAB [14] using a finite element toolbox - CALFEM
[15] - with thirteen 2D Euler-Bernoulli beam elements (described in chapter 2).
Elemental stress is determined from the displacements of the nodes of each beam
element. The relationship is set out in equation (2.29)
,
(4.1)
where E is the Youngs modulus of the material,
is the strain displacement matrix,
are the nodal shearing and rotational displacements and
is the distance from the
neutral axis at which the stress is required.

Recall that one of the premises of the finite element method is that the stress is assumed
to vary linearly between elemental nodes. For the two dimensional beam element the
flexural stresses are given by
1
12
12
1
2
(4.2)
2
2
Rotation of the central axis of the beam allows the flexural stress to be cast entirely in
terms of two rotations
4
2
2
6
2
(4.3)
where the rotations are now relative to the local axis passing though both ends of the
deformed element. The local rotations
previous chapter.
and
can be seen in figure 3.7 from the
The longitudinal axial stress resulting from the x-direction
displacements of the ends of the element is known to be given by
87
(4.4)
where
2
(4.5)
Since the estimation of stress is not absolute20, any attempt to determine loading and
resulting deflected shape of the structure will also be non-absolute. To establish an
estimate of the absolute loading and hence the structures dynamic behaviour some
extra information is required. This would then enable identification of the deflected
shape Using only static observations two possibilities presented themselves; the first is
to regard case A - where the test rig was adjusted so that the end blocks were parallel
and thus the plate was close to being flat - as the datum. The alternative was to take
readings relative to the entirely unconfined plate. The unexpectedly large loading
observed in case A (top graph in figure 4.3) led to the adoption of the latter scheme.
The magnitude of the initial deflection was surveyed and found to be a smooth but nonsymmetrical curve with maximum eccentricity of 2.4mm. The measured shape of the
plate was included in an initial FE model which then corresponded closely to the state
of the plate at the point of assumed zero strain. A Nelder-Mead21 simplex direct search
method [83] was used to vary two bending moments M1 and M2 and the axial load Faxial
- shown in figure 4.4 - applied to the ends of the FE model. The minimisation was
applied to the difference between analytical and experimental stresses. The objective
function was taken as the 2-norm of the vectorised difference
FE
exp 2
(4.6)
20
The strain gauges can only give an indication of the change in strain relative to the condition of the
gauges point when monitoring begins.
21 Different optimisation routines are not discussed herein since it is generally accepted that the
conditioning of the problem - and clearly the number of parameters - are more important influences on
the speed of convergence than the choice of optimisation routine chosen.

where
and
exp
These were defined at
88
are the finite element and experimental stress distributions.

corresponding positions so that
1
2
(4.7)
The iterative approach to determining the effects of large deformations described in

chapter 3 was used with the displacement convergence (equation (3.39)) set at 1 10-6.
The result is a close match of the stress plots for each case. The longitudinal stress
distributions of along the length of the top and bottom of the plate arise from the
identified loading regimes. These are shown as lines in figure 4.3 with the experimental
readings appearing as discrete points. The closeness of the agreement gives much
confidence in the validity of the identified loads. This observation takes on particular
importance when considering the comparison of the FE prediction of loaded dynamic
behaviour with experimental observations. The end moments (M1 and M2) and axial
load (Faxial) giving rise to these distributions are shown in table 4.1, below.
Table 4.1 Identified Loading
The maximum overall deflections are determined by the nonlinear finite element
analysis to be 1, 3, 5 and 6mm respectively. The greatest of these is a mere 1% of the
overall length of the narrow plate specimen which is likely to be well within the
serviceability tolerance of most structures.
While producing coherent estimates of loading, the method of indirect determination is
far from efficient and probably not practical for application on real structures. Several
simpler methods of deriving stress distributions from loading provide different
89
estimates of the loading. The simplest route is via consideration of a linear stress
distribution.
Considering the plate specimen as a single beam finite element, the
bending load and the axial load become uncoupled. A more involved method is to use
the full finite element model. This is achieved by applying the bending moments and
subsequently the axial load, with the finite element model being reformulated after each
step but without considering equilibrium iterations. This will be referred to as the two
step approach. Note that both methods are special cases of the nonlinear geometrical
method described previously. The loads identified by application of these two schemes
to the data from case B are set out in table 4.2. The last two columns show the effect of
not including the knowledge of the initial deformity of the plate.
The number of individual arithmetic operations required for the identification of loading
by each scheme are also shown to give some comparison of the relative computational
efficiency. Whilst noting that the number of operations is a function of the starting
point of the minimisation process it is clear that a vast increase in computing power is
required to implement the more accurate schemes. The resonant frequencies which
would arise from each set of identified loads are calculated using the full nonlinear
analysis with tight convergence criteria and are also presented in table 4.2.
Table 4.2 Identified Loading and Resonant Frequency Predictions for Case B
90
There is certainly some variation in the prediction of loading which in turn results in
some inaccuracy in the prediction of resonant frequency.
However, from these
observations and from the resulting stress distributions which are shown in figure 4.5, it
is clear that - for this particular specimen - the expense of matching the stress to a large
amount of accuracy is not justifiable.
The results indicate that a rudimentary match of strain is required. Once loading is
identified the accuracy of the statically updated finite element model should be
calculated to as much accuracy as can be afforded. It is particularly important to note
that the spatial distribution of strain gauges is a crucial factor in allowing the stress
distribution to be characterised accurately.
4.2.2
Modal Analysis of Plate Specimen
Excitation was applied normal to the plane of the plate at the point shown in figure 4.2
by a small electro-magnetic shaker resting on a foam foundation. The response for each
of the four cases was measured at the same point in the same direction by means of a
laser velocimeter. It was not the intention to consider changes to mode shape resulting
from loading in this instance22. Figure 4.6 shows the FRFs in the form of point
receptances,
( )
( )
,
( )
(4.8)
taken from the plate under the four loading conditions.

The H1 FRF estimator was used to experimentally determine mobility. As a result some
definition is lost at resonance. Some evidence of interaction of the specimen with the
mechanical shaker is also evident. Considering only the four clear resonances, the first
corresponds to the first bending mode while following modes are the second bending,
22
Recent work by Link [84] lends weight to the view that mode shapes cannot be identified consistently
accurately enough to be considered as representative of a structure.
91
first torsion and third bending modes respectively. For comparison with the natural
frequencies predicted by the loadable two dimensional beam model, only the bending
modes are examined in the following discussions. A single input single output (SISO)
line fitting modal identification algorithm [59] was used to determine the bending
resonant frequencies. The results of the analysis are set out in table 4.3.
Table 4.3 Experimental Resonant Frequencies
There is clearly a large amount of variation in the measured responses resulting from
the loading process, particularly of the lowest natural frequency.
The choice of a method for comparing the sets of data is not a simple one, insufficient
spatial data exists to enable consideration of the variation in Modal Assurance Criterion
(MAC) [57] values between these sets of data. However, an alternative criterion, the
Frequency Response Assurance Criterion FRAC [85], has been suggested as a means of
assessing the similitude of the data by comparing frequency domain responses. The
FRAC formulation is analogous to the MAC and has similarities to the COMAC [86] in
that each value compares co-ordinate response. It is given by
*2
.
.
where the superscripts
and
(4.9)
refer to two response functions defined at frequencies
. The star symbol ( * ) represents the conjugate transpose. While the receptances
in (4.9) are single row (or column) complex quantities it is suggested in [85] that
it might be necessary to replace these quantities with their absolute values.
92
, can be included to account
It is suggested that a factor on either set of frequencies,
for a systematic mis-estimation of stiffness in an analytical prediction of
. The
modified FRAC becomes

* 2
max
.
Figure 4.7 shows the result of applying different values of
(4.10)
to the FRAC between
receptances from case 1 and case 2, formulating the FRAC using both the complex
receptances and absolute values. The observation that an overall change in stiffness is
required to maximise the level of FRAC is counter intuitive and arises since the
magnitude of the response of the first mode dominates the formulation. Figure 4.8
indicates this clearly on a plot of the absolute values of point receptance for cases A and
B against a linear scale. If the factor is included the responses can be clearly seen to be
very different. However, it is noticeable that the first mode of the stiffened response
converges on the mode of the response to which it is being compared (figure 4.9 and
figure 4.10).
This observation has some repercussions upon updating methods which use the
difference between response data (chapter 2 section 2.3.2.2). If the first mode of a real
structure a bridge for instance was to experience frequency shifts under loading then
a dilemma arises: if the differences in low order responses are included then there is a
risk of erroneously stiffening the entire structure. However, if they are ignored the
effect of loading will not be included in the updated finite element model.
Formulating the FRAC with a logarithm of the response leads to a value of
1 as one
would expect (figure 4.11). The result of using the familiar representation of the
receptance in decibel units is that the FRAC in all cases is very close to unity as table
4.4 clearly shows. The FRAC values are shown as percentages of unity for clarity.
93
Table 4.4 % FRAC Values Based on dB Receptance
Table 4.5 shows the corresponding FRAC values for the absolute values of the
receptances from the four load cases.
Table 4.5 % FRAC Values Based on Absolute Receptance
Given the experience with determination of the stiffness factor, it is understood that the
first mode contributes heavily to the low level of agreement between the pairs of
responses from different load cases. In such cases the use of FRAC values should be
treated with extreme caution.
The preceding paragraphs effectively demonstrate some of the problems which arise
when attempting to compare response data from nominally identical structures. A
comparison of resonant frequencies represents an alternative comparison. Accordingly
the implementation of methods which compare sets of identified resonant frequencies
are investigated in the following sections - principally in section 4.2.4.

4.2.3
94
Comparison of Measured Dynamic Behaviour with FE Modelling
The loads identified in section 4.2.1 (see table 4.2) were applied to the 2D beam
representation of the narrow plate using the nonlinear geometric approach described in
chapter 3. An eigen-solution of the resulting set of transformed structural matrices
results in a set of resonant frequency predictions which are shown in table 4.6 along
with the experimental measurements.
Table 4.6 Analytical Predictions and Experimental Observations of Resonant

Bending Frequencies
It is noted that the finite element predictions consistently overestimate the experimental
data. Several factors are likely to result in the differences between the predicted and
observed resonances. The first is the effect of the discretised nature of the finite
element model which has been discussed previously. Additionally the lack of complete
fixity provided by the experimental supporting structure will have a significant effect on
the experimental resonances. These effects make it more difficult to assess to what
extent the changes to the analytical model reproduce the experimental behaviour.
It is clear however, that a similarly large amount of variation is manifest in the finite
element prediction as was measured experimentally. The following section develops a
method which facilitates an easier comparison between the sets of data.
4.2.4
Identification of Zero-Load Resonant Frequencies
An understandable comparison of experimental and analytical results can be ascertained

if the experimental dynamic behaviour of the unloaded structure is available. For this
95
experimental case - as with many industrial cases - it is almost impossible to recreate

the situation corresponding to the unloaded FE model. For the narrow plate currently
under consideration this would involve end constraints which straighten out the initial
deformity in the plate without introducing axial load.
This section sets out a technique for estimating the zero-load experimental load from
the FE model of the structure and several sets dynamic measurements under various
loading conditions.
If the
experimental resonant frequency from the
the corresponding analytical prediction
loading case is written
then the increase in the
and
identified
load case over the corresponding mode in the unloaded
resonant frequency from the

case is given by
E
(4.11)
with a similar expression for the analytical case

A
A ,
(4.12)
where the 0 subscript refers to the zero load case. It is common to consider percentage
increases between corresponding resonant frequencies.
Therefore, the normalised
experimental frequency increase can be considered. This is

E
0
0
(4.13)
For a given frequency from a single load case, a prediction of the corresponding zeroload resonant frequency can be found by minimising the differences between the
normalised increases with respect to
E
0
0
For a set of
minimising
0
0
load cases the
(4.14)
zero-load resonant frequency can be found by

E
1
0
0
0
0
with respect to 0 E where
0
0
(4.15)
A
1...
turn or the global sum over all

E
96
. This can be performed on each frequency in

measured identified frequencies
0
0
(4.16)
Using this method with all four sets of data yields predictions of bending resonant
frequencies shown in table 4.7 along with their FE counterparts.
The result of
minimising absolute frequency differences is also tabulated but can be seen not to be
significantly different.
Table 4.7 - Estimated Experimental Resonant Frequencies
The percentage increase of each natural frequency relative to the unloaded values are
displayed in figure 4.12. The minimisation procedure described above ascribes equal
importance to the perturbation of all of the modes. The results indicate that the zeroload resonant frequencies are estimated so that the perturbation to the higher order
modes most accurately matches the finite element model prediction. This phenomenon
is understandable in light of the observation that the first mode is most sensitive to
slight alterations in loading. Hence the prediction of this mode would be expected to be
less accurate. Other factors which lead to lack of agreement between dynamic and
experimental measurements and finite elements predictions are described in the
following section.
97
The technique presented here - while simple - has considerable practical appeal. Given
knowledge or possibly estimation of structural loading at the time of several runs of
modal testing, it is possible to estimate the set of resonant frequencies which
represent the loaded structure.
These are referred to as the
in forthcoming chapters.
4.2.5
Investigation of FE Experimental Mismatch
The previous sections give confidence that the loading upon the plate is the primary
factor in causing the observed perturbations to resonant frequencies. Several other
factors whose influence causes inconsistency between the experimental and FE models
must also be addressed.
Firstly, as the previous sections have acknowledged, the test rig not being infinitely
massive will interact with the plate to some degree. To investigate the influence of the
test rig on the dynamic behaviour of the plate, a simplified finite element model was
used with each the end blocks represented by four solid brick finite elements. The
plate specimen and each of the adjustments bars were modelled with 5 3D beam
elements.
As one would expect, the loss in apparent23 stiffness at the supports results in lower
predictions of resonant frequency. Table 4.8 shows that the effect upon the first three
bending modes.
Table 4.8 Experimental Modes Identified
23
This refers to the stiffness of supports perceived by the plate specimen.
98
While the result is a closer agreement between the analytical and experimentally
determined values of resonant frequency at zero load, some disagreement between the
measured zero-load resonant frequencies and the FE predictions is still evident. Several
other factors seem likely to be to blame.
The previous sections have assumed that the loading applied to the plate consists of
only axial and bending loads. While this was the intention during testing, the design of
the rig makes it difficult to prevent some component of torsional loading to be
experienced by the specimen.
A further source of error will arise from the stiffness of the physical connection between
the plate and the large restraining blocks.
4.3
Small Framework
To address some of the issues raised by the initial experimental investigation and to
circumvent some of its shortcomings, a second test was devised involving a cross
braced frame structure. The structure was supported in free-free conditions for dynamic
testing thus obviating the requirement of accounting for potentially complex interaction
with boundary conditions. An adjustable strut in the framework allowed loading to be
introduced into the structure in a controlled and directly measurable manner.
The following sections present several sets of dynamic readings taken from the
framework under different load levels and investigate some of the aspects of the
changes in dynamic characteristics. A more comprehensive study of three nominally
identical frameworks is undertaken in chapter 6.
At this stage the use of model
updating strategies to alter the finite element model of the structure will be studied.
4.3.1
Arrangement of Framework
An exploded view of the truss is shown in figure 4.13; the overall dimensions (between
joint centres) are 300mm by 500mm. The truss was constructed from 6
15mm mild
99
steel section spars which were bolted at their joints. Figure 4.14 shows the points at
which the structure was tested as well as the numbering and co-ordinate system. Three
strain gauging locations (also shown in figure 4.14) were used with the instruments
positioned in pairs on either side of the spars. These gauges were oriented to measure
the longitudinal strain. The framework was supported from two flexible strings to
minimise interaction of the test structure with the supporting structure.
The framework was deliberately designed to contain a single redundancy if considered
as a two dimensional pinned structure. However, it was anticipated that the noncoincidence of the centre lines of the constituent spars would result in bending moments
being transmitted at the joints. This leads to the requirement that the structure be
modelled in three dimensions. The practical observation was that tension induced in the
adjustable spar led readily to noticeable deflection of spars 1, 2, 3 and 4, allowing the
effects of deformity in addition to axial spar loads to be investigated.
4.3.2
Experimental Results
Four sets of dynamic readings were taken in the out-of-plane z-direction at different
levels of loading. Strain gauges on spar 5 adjacent to the turnbuckle indicated that axial
load in this strut were 440N, 820N and 1230N relative to the initial case. These load
cases will be referred to as cases 1, 2 and 3 with the zero-load initial conditions being
case 0. Experience gained in the previous experimental investigation (section 4.2.1) has
shown that difficulties can arise from judging an initial load case to represent zero
loading.
In this situation the turnbuckle slackened noticeably when no load was
induced in the axial member. The strain gauges were zeroed with the turnbuckle in this
state. The bolted nature of the joints did not preclude the possibility of some loading
and deformity being locked into the structures. However, this was thought not to be
significant compared to the known loading levels applied with the turnbuckle.
Broadband random excitation was applied to node 26 shown in figure 4.15 - in the zdirection with a small electro-dynamic shaker via a flexible rod or
and a force
100
transducer. Response (mobility) was measured in the same direction at the 24 points
shown in figure 4.14 in the 0-400Hz range.
Modal analysis was performed using the MODENT program within the ICATS [87]
suite of software. The global method [58] of analysing multiple frequency response
functions was used to extract natural frequency and mode shape estimates.
Figure 4.16 shows the point receptances measured from the extremal load cases 0 and 3.
The dynamic behaviour of the frame is seen to change dramatically under the loading.
The identified resonant frequencies from all of the load cases are shown in table 4.9.
Note however that the rows of the table do not necessarily imply correspondence of
modes.
The identified modes shown in the table are the result of an extensive
examination of the sets of responses. It was observed that the prominence of modes
altered noticeably with the level of loading. Two of the modes, shaded in table 4.9,
were particularly difficult to identify in the unloaded case.
Table 4.9 Experimental Modes Identified
A graphical representation of these results is given in figure 4.17. This shows each of
the rows of table 4.9 as a line against the values of identified load in the adjustment bar.
101
The graph gives some confidence in the tables rows representing correlated modes. Of
most significance is that all of the resonant frequencies of the structure identified in the
0-400Hz range are perturbed a great deal, by different amounts and
. The pairs of modes two and three as well as eight and nine are observed to
be converging upon one another. Also of interest is the observation that the scale of the
perturbation in absolute terms for each identified resonant frequency is of a similar
order. This observation is borne out in table 4.10 which shows the absolute and
percentage difference in the identified value of each mode between load cases 0 and 3.
Table 4.10 Differences Between Identified Frequencies - Case 0 and Case 3
The mode shapes corresponding to each of the resonant frequencies are shown in figure
4.18. As one would expect some modes - for instance two and three - are dominated by
resonance of a subset of the total number of members. Others represent more global
displacement patterns, for example modes one and six. Employing the modal assurance
criterion to compare the sets of experimental data yields interesting results. Figure 4.19
shows the MAC of the mode shapes identified in case 0 with those in case 3 and
indicates that there is good correlation between the pairs of modes numbered one to five
102
and ten to twelve. The correspondence between pairs six and seven and particularly
eight and nine show less clear correlation.
This observation suggests that the
deformities to the structure and the state of internal stress resulting from loading can
have a considerable effect on some of the mode shapes. The following sections allow
the phenomena to be investigated in more detail with reference to the finite element
model of the structure.
4.3.3
Finite Element Modelling
The framework was modelled using the ANSYS [16] commercial finite element
package. Three dimensional cubic beam elements were used exclusively to model the
structure. The model consists of a total of 34 elements and 192 degrees of freedom.
Soft grounded springs were attached to the top corners of the framework to prevent
numerical instability arising from calculation of eigenvalues corresponding to rigid
body modes of vibration. The node points at each of which there are six degrees of
freedom are shown in figure 4.15.
As the exploded view of the framework in figure 4.13 clearly shows, the struts which
constitute the truss lie in several layers. The co-ordinate system is shown in figure 4.15,
the z axis has its zero value at the mid-depth of back-most spar, 6.
summarises the relative location of all of the spars.
Table 4.11 Relative Depth of Spar Elements
Table 4.11
103
This non-coincidence of the lines of action of the spars under loading was accounted for
by including stiff beams with axes parallel to the global z axes.
An eigen-solution of the finite element model results in fifteen resonant frequencies in
the range 0-400Hz whose values are shown in table 4.12.
Table 4.12 Resonant Frequencies Predicted By FE Model (Zero Load)
The mode shapes corresponding to these modes are shown in figure 4.20.
The mode shapes predicted by the FE model were compared with the experimentally
determined values from case 0 using the MAC. The results of this standard procedure
are shown in figure 4.21. The results show that while for some of the modes at each
end of the frequency range of interest show good agreement, several modes around 180220Hz do not agree with one another with any certainty.
Figure 4.22 shows the
agreement of modes between the case 0 readings and the FE prediction in terms of
MAC values above 0.9, above 0.7 and above 0.5. Only seven of the twelve identified
104
modes can be readily matched with analytically derived mode shapes leaving
considerable uncertainty about the correlation of the remaining five modes.
Some insight into the poorly correlating modes is afforded by considering the autoMAC24 of the FE mode shapes. The results shown in figure 4.23, indicate that there is a
great deal of similarity between the very modes which had previously failed to correlate
well with experimental data.
Figure 4.24 illustrates the auto-MAC more clearly,
showing the locations of MAC values greater that 0.8 with the shaded points indicating
multiple mode correspondence. It is seen that the mode shapes of modes 6 and 9 are
both of a similar form to mode 7; additionally modes 8 and 12 are seen to be much
alike.
While the behaviour of the closely correlating modes is a matter of some interest, for
the purposes of using dynamic data to characterise a structure, the modes which
correspond to FE modes 6, 7, 8, 9 and 12 can be justifiably overlooked. For the
purposes of the current study, the closely correlating modes 1 through 5 will be
considered in isolation.
The loading upon the structure was applied to the ANSYS model with both the stress
stiffening and large-deformation effects described in chapter 3 taken into account. The
effect of loading the frame by shortening spar 5 was accounted for in ANSYS in two
stages as shown in figure 4.25. Spar 5 was loaded in tension separately from the rest of
the frame with the load resulting from the shortening of this spar being applied to the
frame independently of the spar itself. The elemental stiffness and mass matrices from
the two sub-models were output from ANSYS and combined in the MATLAB
environment.
Loads from 0 to 1200N were applied to the finite element model in this way. The
variation of the first five resonant frequencies for which good correlation between finite
24 The auto-MAC is the special instance of the MAC where two identical sets of mode shapes are
compared and is thus symmetrical.
105
element and experimental modes exists are shown in figure 4.26 along with the
experimental readings. The trends in both sets of data appear to be qualitatively similar
which inspires confidence that the changes to the structure caused by the loading are
accounted for by the changes to the finite element model resulting from the nonlinear
analysis.
To allow the accuracy of the finite element model to be determined, pairs of
corresponding modes from the zero load case experimental readings and the finite
element model are set out in table 4.13. These lower order modes conform to the
expectation that the finite element model discretisation will lead to over-estimates of
resonant frequency.
The larger error observed in the prediction of mode two is
explained by the fact that this mode consists predominantly of flexure of the adjustable
spar which contains the turnbuckle device whose characteristics were not included in
the simple model.
Table 4.13 Corresponding Modes Zero Loading
While giving encouraging agreement with the experimental data qualitatively, there is
clearly some error between the zero-load FE model and experimental measurements
whose causes must be investigated. The framework example is revisited in chapter 6
where the techniques described in chapters 3 and 4 are joined by a dynamic updating
technique to account for the effects of loading in the structure.

4.4
106
Concluding Remarks
The state of loading in two experimental case studies has been found to alter
significantly the dynamical behaviour of the structures. The techniques described in
chapter 3 account for both the deformation and stress stiffening resulting from loading.
These produced perturbations to predicted resonant frequencies consistent with
observed values.
An instance which demonstrates the indirect identification of loading using strain
gauges has been investigated. Several methods of employing the
finite element
model of a substructure to help to identify loading from discrete strain measurements

have been investigated. Deformations which are likely to give rise to changes in
dynamic readings have been seen to lead to nonlinear stress distributions. Therefore, an
appropriately large number of strain gauges must be used to characterise this type of
stress distribution.
The problem of determining the dynamic behaviour characteristics of the unloaded state
of a structure has been investigated. This scenario could quite plausibly be encountered
in a real-life situation. A new method for identifying the zero-load dynamic behaviour
of a structure has been introduced. The technique uses several sets loaded data and
allows the identification of loading using a finite element model of the structure.
The sensitivity of mode shapes to static loading has been studied in a specific
experimental case study. The
finite element model allows the vibration modes
with sensitive mode shapes to be excluded from validation exercises.
Figure 4.1 - Experimental Arrangement
Figure 4.2 - Co-ordinate System and Strain Gauge Locations on Plate
107
Figure 4.3 - Observed Stresses Distributions on Plate Cases A to D
108
Figure 4.4 Application of Load to Finite Element Model
Figure 4.5 Stress Distributions from Different Stress Models
109
Figure 4.6 - Point Receptances from Narrow Plate Under Varying Loading
Figure 4.7 FRAC Between Load Cases 1 and 2 Varying
110
Figure 4.8 Linear Receptance Case A and Case B
111

990325master PDF

Uploaded by

Copyright:

Available Formats

990325master PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

990325master PDF

Uploaded by

Copyright:

Available Formats

1

DYNAMIC FINITE ELEMENT MODELLING AND

Paul David Greening

A thesis submitted to the University of Bristol for the degree of

A great deal of information about structural

performance can potentially be obtained from simulations without the expense of

A mismatch between experimental and

analytical measurements motivates a process whereby the analytical model is altered to

The ability to alter an

Thus, changes to this signature represent change to the

A closed form differential equation approach to modelling the behaviour of structures is

Many of the choices have to be made by an experienced

poor choice of element type and quantity required; and

Vibration is a characteristic common to all engineering structures, from the smallest

Corresponding predictions of dynamic behaviour can be

determined from a finite element model using eigenvalue extractions techniques.

Correction of FE Model Using Dynamic Data

sense. Several different approaches

Variations in Dynamic Readings

structure. Additionally, the collection of large numbers of sets of experimental data or

a single structure under a variety of conditions; or

nominally identical structures;

have been given comparatively little attention. Fregolent

[5] specifically consider

of a family of structures with nominally identical properties.

Imregun, for instance [6] and [7] consider a similar problem.

, as is common practice, seeks to

Alampalli [10] for instance, presents widely variable

baseline results from tests on a highway bridge under different conditions.

variation in measurements is attributed to the in-service environment, signal processing

Most disturbingly, it is reported that the

If the application of model updating strategies outside of the laboratory is to become

Load Dependence of Dynamic Behaviour

Table 1.1 - Sources of Static Stress in Structures

For in-situ testing, there is a generally low

Combined Static and Dynamic Load Identification

structures. The ability to monitor

structures in this way is principally bought about by improvements in strain gauge

implies a useful range of up to 180Hz1 although in this case resolution at a range of up

The potential effects on dynamic

readings due to static effects are also outlined.

load identification to be addressed practically. Additionally, the difficulty in identifying

The basic package - developed for educational

purposes - was augmented substantially by the author to include 3D stress stiffening

Additional FE modelling work and

While any particular experimental arrangement cannot be expected to be

important in modelling of civil engineering structures.

Where appropriate the

application of the methods to other element types is described.

Simulated experimental data are referred to by enclosing the word

Chapter 2 Finite Element Model Updating Using Experimental Data

The Finite Element Method

Chapter 2 Finite Element Model Updating Using Experimental Data

Courant [18] is generally credited with

proposing a method which was destined to be highly suited for implementation on

[19] and Argyris and Kelsey [20] laid the

in 1960 [21]. The use of finite element methods has

This allows the

Chapter 2 Finite Element Model Updating Using Experimental Data