Presented at 14th IMAC Conference, Dearborn, MI February, 1996

Using SDM to Train Neural Networks

for Solving Modal Sensitivity Problems
Brian J. Schwarz, Patrick L. McHargue, & Mark H. Richardson
Vibrant Technology, Inc.
18141 Main Street
Jamestown, California 95327

ABSTRACT [C] = (n by n) damping matrix (force/unit of velocity).

The Structural Dynamics Modification (SDM) algorithm is [K] = (n by n) stiffness matrix(force/unit of displacement).
very useful for solving the so-called forward variational [M] = [M] + [∆M] = (n by n) mass matrix of modified
problem for structures. That is, given changes in a structure.
structure's mass, stiffness, or damping properties, SDM [C] = [C] + [∆C] = (n by n) damping matrix of
efficiently yields the corresponding changes in its modal
modified structure.
properties.
[K] = [K] + [∆K] = (n by n) stiffness matrix of
There are several important classes of problems, however, modified structure.
that require solutions to the inverse variational problem, or [∆M] = (n by n) matrix of mass changes.
modal sensitivity problem. That is, given changes in a
structure's modal properties, what corresponding changes in [∆C] = (n by n) matrix of damping changes.
its mass, stiffness, and damping properties have taken place. [∆K] = (n by n) matrix of stiffness changes.
Applications such as structural damage detection, finite x t )} = acceleration response n-vector.
{(
element model updating using test data, and vibration {x( t )} = velocity response n-vector.
suppression or control through structural modification all
require solutions to the modal sensitivity problem. {x( t )} = displacement response n-vector.
Unlike the forward variational problem, the modal
{f ( t )} = excitation force n-vector.
sensitivity problem cannot be solved in a straightforward {X(jω)} = discrete Fourier transform of the displacement
manner. For most practical cases, its solution requires the response n-vector.
inversion of a rank deficient matrix, which creates
numerical difficulties. INTRODUCTION
Neural networks offer promise for solving the modal The underlying assumption of the modal sensitivity problem
sensitivity problem, because of their pattern recognition and is that changes in the vibration characteristics (modal
interpolation capabilities. In order to solve an inverse properties) of a structure are strongly coupled to changes in
variational problem, however, a neural network must be its physical properties. Modal testing itself assumes that the
"trained" using a set of solutions to its corresponding structure remains in a stationary condition throughout the
forward variational problem. Training a neural network test. That is, its modal properties must not change. Many
typically requires hundreds, even thousands of solution sets. times, this condition is difficult to maintain.
In this paper we show how SDM can be used to train a
neural network for solving the modal sensitivity problem. Most experimentalists have encountered mass loading
Because SDM only requires the modal parameters of the effects during a modal test. The apparently insignificant
structure, which can be obtained from a modal test or a mass of the measurement transducers (e.g. accelerometers)
finite element model, this method can be applied in a wide causes the modal frequencies to shift as transducers are
variety of experimental and analytical cases. moved from one point to another on the structure.
Temperature changes during the course of an all day modal
test can also cause modal frequencies to shift.
NOMENCLATURE
t = time variable (seconds). For all but the simplest of cases, matching boundary
jω = frequency variable (radians/second). conditions between finite element analysis and modal test
presents major difficulties. When the boundary conditions
n = number of measured DOFs. prescribed in the finite element analysis cannot be
[M] = (n by n) mass matrix (force/unit of acceleration). duplicated in the laboratory, the experimental modes don't

Page 1 of 6
Presented at 14th IMAC Conference, Dearborn, MI February, 1996

match the analytical modes. More complex physical space. The SDM algorithm solves it in modal space, using
changes, such as the gravitational effects on joint stiffnesses the modes of the unmodified structure instead of the
in spacecraft, or the complex aerodynamic interactions of unmodified mass, stiffness, and damping matrices [2].
flight flutter in aircraft, also cause changes in the structural
The primary advantage, and assumption, of the SDM
modes of vibration.
approach is that the dynamics of the unmodified structure
In recent years, numerous researchers have proposed can be adequately represented by relatively few of its
solutions to the modal sensitivity problem. Entire fundamental (lowest frequency) modes. The solution of the
conferences have been held of the subject of structure forward variational problem can also be considered as a
damage detection using vibrational changes [11]. As modal transformation from physical space to modal space.
testing has become more widespread, finite element model
Representation of structural dynamics in terms of mass,
updating using modal test data has also become popular.
stiffness, and damping matrices typically requires much
Vibration problems are often caused when one or more more data than a modal representation. For example, to
modes are located too near to an operating frequency or represent the dynamics of a structure with a 1000 DOF
excitation frequency of a machine or structure. The model using its first ten (normal) modes would require,
difficulty with problems of this type is that moving one
(10 modes) X ( 1000 DOFs per shape + frequency +
mode, (by adding a stiffener, for instance), also causes other
damping) = 10,020 numbers.
modes to shift in frequency, thus causing a different,
perhaps greater vibration problem. One of our major
expectations for this trained neural network is that it will (The use of complex modes would approximately double
find stiffness solutions that change the frequency of a this number). To represent the same structure dynamically
using real symmetric mass, stiffness, and damping matrices
problem mode without also causing the frequencies of other
would require,
modes to shift to other problem areas.
3 matrices X ( 1000 X (1000 + 1) / 2) = 1,501,500 numbers.
The authors have tried to solve modal sensitivity problem in
the past, using pseudo-inverse [3]-[5], and other
approximation methods [6], but with limited success. In other words, the physical model requires 150 times as
Neural networks have also been applied to this problem [7] - much data as the modal model! This comparison also
[9], but in a different manner than is described here. illustrates the difficulty of solving the inverse variational
problem, or of transforming from modal space back to
physical space.
THEORETICAL BACKGROUND
Real world structures have an infinite number of degrees of
To re-state the two problems under consideration:
freedom. Therefore, to perfectly match the dynamics of a
Forward Variational Problem Given changes in a real structure, equation (2) would require infinite
structure's mass, stiffness, and damping properties, find the dimensional matrices, and would yield an infinite number of
corresponding changes it its modal properties. modes. In practice, of course, we approximate the dynamics
of infinite structures using finite dimensional matrices and
Inverse Variational Problem Give changes a structure's
finite numbers of modes.
modal properties, find the corresponding changes in its
mass, stiffness, and damping properties. Nevertheless, solving the inverse variational problem,
requires transforming from modal coordinates (with
If the equations of motion of the structure are stated as the
relatively little data), to physical coordinates (with large
usual statement of Newton's second law,
matrix representations), even for simple cases.
x t )} + [ C]{x( t )} + [ K ]{x( t )} = {f ( t )}
[ M ]{( (1)
Any direct solution of the inverse variational problem
requires the inversion of the flexibility matrix (or an
then the solution to the forward variational problem can be equivalent operation), to obtain the stiffness matrix [1].
stated as the unique solution to a matrix eigenvalue Since only a relatively small number of the modes of a real
problem, structure are usually known (are measured), the flexibility

[[M]( jω )2 + [C}( jω ) + [K ]]{X( jω )} = {0}

matrix will be rank deficient, and matrix inversion is
(2) impossible. So, we seek a solution to the inverse variational
problem which preserves the unique relationship between
changes in modal and physical properties, and provides
Notice that the above equation contains the matrices of the
discrete mass, stiffness, and damping changes, instead of the
modified structure, and that the equations are homogeneous
matrices themselves.
(external forces are zero). The above equation can be
solved, either in physical space or modal space, for the In the past few years, "standardized" neural network
modes of the modified structure. Most finite element software has become available that can be used "right out of
analysis software packages solve this problem in physical the box". For this work, the NeuralWindows software by

Page 2 of 6
Presented at 14th IMAC Conference, Dearborn, MI February, 1996

Ward Systems Group [10] was used in a Visual Basic realistic stiffness changes can only be made between a
program that performed the rest of the processing necessary subset of DOFs.
to train the network.
ELIMINATING CONFLICTING TRAINING SETS
To quote the NeuralWindows operating manual, "Neural
networks excel at problem diagnosis, decision making, Since we are only training the neural network with modal
prediction, and other classifying problems where pattern frequencies and stiffness changes, it is possible that two or
recognition is important and precise computational answers more sets of different stiffness changes will yield the same
are not required." set of modal frequencies. This will always be the case when
modification DOFs are chosen that correspond to nodal
Neural networks are made up of "neurons" arranged in
points (zero magnitudes) of the mode shapes involved. If
layers. A neuron is a simple input-output device with a
two training sets have different stiffness changes but the
built-in transfer function. In a simple feed forward network,
same modal frequencies, one of them must be eliminated.
the output of each neuron in one layer in linked to the input
on each neuron in the succeeding layer. Inputs to the entire We used the Modal Assurance Criterion (MAC) to
network are fed into the neurons in the first (input) layer. determine whether or not two sets of modal frequencies are
Outputs (answers) from the neural network are the outputs the same. MAC was developed for comparing two mode
of the last (output) layer. shapes., and is essentially a projection of one vector onto the
other. In this case, we simply assembled the modal
For this application, we used a three layer network, with an
frequencies of each training set into a vector, and applied
input, an output, and a hidden layer. Network training was
MAC to all pairs of training set vectors. Any pair of
done using the Backpropagation of errors method.
training sets with a MAC value above 99% was considered
to have the same frequencies.
MODAL SENSITIVITY PROBLEM
Once a pair of training sets with the same frequencies is
In structures, stiffness modifications cause modal frequency
found, one set has to be eliminated in order to avoid
changes, mass modifications cause modal frequency and
confusing the neural network. Another criterion must be
damping changes, and damping modifications cause modal
used to determine which of the two sets is more desirable.
frequency and damping changes. In all cases, mode shapes
We chose the set with the minimum stiffness metric, defined
may also change.
by:
To simplify this analysis, we will focus only on stiffness
changes, although the method is equally valid for mass and Stiffness Metric = SUM (Stiffness Change Magnitudes)
damping changes. When a local stiffness modification is
This metric will eliminate all training sets that cause no
applied to a structure, the mode shapes dictate which modes
changes in the modal frequencies, When one of them is
will "absorb" the modification, and hence be most sensitive
compared to the baseline training set, which has a Stiffness
to the modification. Therefore, realizing that mode shapes
Metric = 0, it will always be less desirable than the baseline
also play a significant role in stiffness modifications, we can
set.
define a modal sensitivity problem as follows:
AN EXAMPLE
Stiffness Sensitivity Problem
To demonstrate this method, the 2 DOF lumped parameter
Given a set of modal frequencies different from those of a
model shown in Figure 1 was used. This model has a closed
baseline (unmodified) structure, find the stiffness changes
form solution for modal frequencies [12].
required to yield the new modal frequencies.
This definition indicates how the neural network should be
setup in order to solve the problem. Each input neuron
represents a modal frequency of the modified structure.
Each output neuron represents a DOF pair and will yield an
amount of stiffness change.
The number of input neurons should always be within a
reasonable range, (1 to 50). On the other hand, for large
structure models (with hundreds of modal test points or
finite element nodes), output neurons cannot represent all of Figure 1.
the possible DOF pairs, which could range into the
thousands. A practical neural network size should contain Using this model allowed us to create training sets
perhaps hundreds of outputs, of DOF pairs. This means that independently of SDM, and compare the training sets with
a subset of DOFs must be chosen for the neural network. those given by SDM.
However, this requirement is often a benefit, since most This model has modal frequencies:

Page 3 of 6
Presented at 14th IMAC Conference, Dearborn, MI February, 1996

F1 = 12.355 Hz CONCLUSIONS

F2 = 52.171 Hz This approach offers much promise for solving realistic

modal sensitivity problems in practical situations. The main
difficulty with solving this problem is having sufficient
Training Data To train the neural network, the spring
modal data to train a network. In a testing situation, a few
stiffnesses were varied over the ranges:
sufficiently accurate modal frequencies can usually be
∆K1 ∈[-12348, 12154] obtained by curve fitting measurement data. Sufficiently
accurate mode shape data can usually be obtained as well.
∆K2 ∈[-30240, 29760] Modal damping is the most difficult parameter to accurately
estimate from test data.
where each stiffness change is relative to the baseline
We have shown with the use of a simple example that a
stiffnesses. The neural network was trained using a total of
neural network can be trained to give back usable stiffness
10000 training sets. Each stiffness change was varied in results by only using modal frequencies and stiffness
100 increments between its minimum and maximum values. changes for training. We have also pointed out that
The stiffness changes yielded frequencies in the ranges:
conflicting training sets can result, where two different
F1 ∈[ 2.51, 17.08 ] Hz stiffness changes yield the same set of modal frequencies.
When this occurs, one of the two conflicting sets must be
F2 ∈[ 31.93, 66.28 ] Hz eliminated to avoid confusing the neural network.
Another way to resolve this ambiguity in training sets is to
Neural Network Predictions After the neural network was use mode shape data also as inputs. Since the shape values
trained, its was fed 10000 pairs of modal frequencies, and it will often be different even when the frequencies are not,
output stiffness changes. The percentage error between the this will add uniqueness to each training set.
actual stiffness changes and the neural network predictions
are plotted in Figure 2., where the error was computed as the The difficulty with using mode shapes as inputs is that the
ratio of stiffnesses: size of the network will quickly grow larger, lengthening the
training time. To minimize this impact, only those shape
Error = Predicted Change / Actual Change values corresponding to the stiffness DOF pairs (output
neurons) should be considered first. In the case of the 2
The worst error for the K1 stiffnesses is about 6.5 %, for the DOF example used here, adding the mode shapes as inputs
K2 stiffnesses about 9%. would only increase the number of input neurons from 2 to
6.
In summary, neural networks appear to do a good job of
solving the modal sensitivity problem, if properly trained.
The SDM algorithm provides a straightforward way of
providing training data sets for a neural network. The
combination of these two methods should provide a
practical tool for solving a variety of vibration trouble
shooting, finite element model updating, and structural fault
detection problems.

Page 4 of 6
Presented at 14th IMAC Conference, Dearborn, MI February, 1996

Figure 2.A K1 Ratio (Predicted/Exact)

Figure 2.B K2 Ratio (Predicted/Exact)

Page 5 of 6
Presented at 14th IMAC Conference, Dearborn, MI February, 1996

REFERENCES
[1] Potter, R. and Richardson, M.H. "Mass, Stiffness and
Damping Matrices from Measured Modal
Parameters", I.S.A. International Instrumentation -
Automation Conference, New York, New York,
October 1974.
[2] Wallack, P., Skoog, P., and Richardson, M. H.
"Simultaneous Structural Dynamics Modification
(S2DM)" 6th IMAC Proceedings, Kissimmee,
Florida, February, 1988.
[3] Mannan, M.A. and Richardson, M.H. "Detection and
Location of Structural Cracks using FRF
Measurements" 8th IMAC Proceedings, Kissimmee
Florida, February, 1990.
[4] Mannan, M.A. and Richardson, M.H.
"Determination of Modal Sensitivity Functions for
Location of Structural Faults" 9th IMAC
Proceedings, Florence, Italy, April, 1991.
[5] Richardson, M. and Mannan, M.A. "Remote
Detection and Location of Structural Faults using
Modal Parameters" 10th IMAC Proceedings, San
Diego, California, February, 1992.
[6] Richardson, M. and Mannan, M.A. "Correlating
Minute Structural Faults with Changes in Modal
Parameters" 11th IMAC Proceedings, Kissimmee,
Florida, February, 1993.
[7] Kudva, J., Munir, N. and Tan, P.W. "Damage
Detection in Smart Structures using Neural
Networks" Smart Materials and Structures, Vol.1,
p100-112, 1992.
[8] Worden, K. Ball, A.D. and Tomlinson, G.R. "Fault
Location in a Framework Structure using Neural
Networks" Smart Materials and Structures Vol.2,
p189-200, 1993.
[9] Worden, K. and Tomlinson, G.R. "Damage Location
and Quantification using Neural Networks"
Engineering Integrity Assessment ed. [Link],
[Link] and [Link] p11-31, 1994.
[10] "NeuroWindows Neural Network Dynamic Link
Library" Ward Systems Group, Inc., Frederick,
Maryland, 21702 (301) 662-7950
[11] "Structural Damage ID Workshop" Charles Farrar,
coordinator., Los Alamos National Laboratory, Los
Alamos, New Mexico, September, 1995
[12] Blevins, Robert D. Robert E. "Formulas for Natural
Frequency and Mode Shape" Kreiger Publishing
Co., Malabar, FL. 1984

Page 6 of 6