(1) Material and Structural Analysis, Framatome ANP, 155 Mill Ridge Road, Lynchburg, VA, USA 24502
(2) Department of Civil Engineering, Texas A&M University, College Station, TX, USA 77843
Abstract
Accurately identifying damage is difficult after an extreme event since damage often cannot be directly measured. Sensors, however,
can collect data that reflects damage by measuring changes in structural response. Therefore, damage identification becomes an
inverse problem in which the goal is to find system properties using limited response information. The inverse problem is formulated
as an optimization problem and is solved using an implicit redundant representation genetic algorithm (IRR GA). Element damage is
introduced into the finite element model using a damage indicator to represent a loss of stiffness. The set of damage indicators are
optimized by minimizing the difference between frequency response functions obtained from the damaged structure and the updated
structural model. Although an accurate structural model may require thousands of finite elements, the number of damaged elements
is typically much smaller. The problem, however, is that the number of damaged elements and their locations are not known
beforehand, which leads to an unstructured optimization problem. To solve this problem the IRR GA uses redundant segments to
allow the number of design variables (damage indicators) to change during optimization. A damage scenario can then be represented
by only a small subset of damage indicators, instead of all possible damaged elements. The performance of the method on two-span
beam and multi-story frame structures was evaluated using several imposed damage scenarios. High damage identification accuracy
was obtained with limited sensor data even when measurement noise was present in the sensor data. The method also was able to
identify damage in larger systems that are difficult to solve using existing methods due to the large number of design variables.
1. Introduction
Overloading, corrosion, material aging, or other unexpected events are inevitable over the lifetime of many structures. Unfortunately
these acts result in structural damage. The goal of structural damage identification methods (SDIM) performing in this environment
is to accurately assess the presence of damage in order define rehabilitation and maintenance needs. Over the past twenty years,
significant research advances have been made in being able to detect, locate, and quantify damage in structural systems. Although
many of these studies found success and even application in the field, there remain two problems that restrict the use of many
proposed SDIMs: sensitivity to noise and difficulty in scaling well with problem size. Therefore, this research is focused on
developing a robust SDIM that uses minimal measurement information to efficiently identify damage in larger structural systems.
The damage identification problem used in this research is posed as an inverse problem in which the goal is to find system properties
using limited response information obtained from sensors. An IRR GA optimization method is applied to solve the inverse problem.
The SDIM presented provides a Level 3 Method [1] since it is capable of assessment by precisely locating and quantifying the
severity of damage. The same methodology is beneficial in other unstructured problem domains such as sensor optimization [2],
structural control, system identification, and the design of frame structures [3].
2. Background
The underlying principle of many SDIMs is that vibration signatures, e.g. frequency response functions (FRF) or modal data, are
sensitive indicators of structural integrity. This reasoning is based on early research efforts that found that damage changes the
dynamic stiffness of structures, and therefore, natural frequency measurements can be used to detect damage using a finite element
model [4]. In the last decade, the majority of research has focused on SDIMs that use modal information [5,6,7]. Several
researchers, however, have investigated using FRF in system and damage identification. Wang et al. [8] used FRF data obtained
before and after damage to develop an algorithm based on nonlinear perturbation equations, which were weighted at selected
locations and frequencies to minimize the influence of errors. An iterative procedure was proposed to overcome difficulties
associated with incomplete measurement data. An FRF-based SDIM was introduced by Lee and Shin [9] using the dynamic equation
motion for the damaged and undamaged beams. An iterative reduced domain strategy was developed to eliminate regions of
undamaged areas from the solution domain. Thyagarajan et al. [10] used gradient-based optimization with FRF measurements to
localize damage in a structural finite element model. Their research stated that a more robust optimization algorithm was desired for
noisy measurements and methods that reduce the optimization domain by eliminating undamaged elements would be beneficial for
larger problems.
More recently, genetic algorithms (GA) have been used in SDIMs. A common feature of all GA-based methods is that they
employ an error function between the measured data and the discrete analytical model. Mares and Surace [11] used GAs to
maximize an objective function based on the residual force method, which is computed from measured natural frequencies and mode
shapes. Dunn [12] studied the performance of GAs and stochastic hillclimbing on a finite element model identification problem in
which the error function between the measures and analytical FRFs was minimized. Moslem and Nafaspour [13] considered the
unstructured nature inherent in damage identification problems by applying a two-stage identification procedure in which areas with
possible damage are first identified and then only elements of that area are included in the optimization domain.
In the research presented, the damage identification problem is formulated as an optimization problem, which is solved using
both a SGA [14] and an implicit redundant representation (IRR) GA [3], which takes advantage of the unstructured nature of damage
1
identification. The IRR representation, which allows the dynamic change of optimization variables, was applied in [15]. In their
approach, however, the authors utilized static equilibrium equations instead of vibration data to set up the optimization problem.
Kφ j = ω 2j Mφ j (1)
where M and K are the mass and stiffness matrices of the undamaged structure, ωj is the jth circular natural frequency of the structure,
and ϕj is the corresponding n × 1 eigenvector. The expression for the receptance matrix, R, defined as a function of ω can be
determined as.
⎛ 1 ⎞ T (2)
R = Φ diag ⎜ 2
⎜ ω − 2iωζ ω − ω 2 ⎟⎟
Φ
⎝ j j j ⎠
where ζj is the jth modal damping ratio, diag() represents a diagonal matrix, and Φ indicates the eigenvector matrix representing the
eigenvectors. The analytical computation of this signature uses modal decomposition to reduce the computational expenses incurred
by using the elementary definition of frequency response functions. Using the receptance matrix, the corresponding mobility and
accelerance matrices can be obtained.
( )
k j = 1 − xj k j , 0 ≤ xj ≤ 1
dam s s (3)
where kj
dam is the stiffness matrix of the jth damaged finite element, kj is the stiffness matrix of the jth finite element without damage,
and x sj is the damage indicator of the jth finite element. The positive damage indicators may be interpreted as percentile damage
values from zero to 100 percent. The equation could be modified to accommodate stiffness increases due to material hardening.
2
where j is the excitation DOF, k is the DOF where the response is measured, Hjk is the jkth FRF function in the finite element model,
th
H jk is jk measured FRF function on the damaged structure, k1, k2,…kn are the DOF’s where measurements are taken, ϖ0 and ϖ1 are
the lower and upper frequencies of the measured frequency range, and the symbol | | indicates complex magnitude. In this research it
is assumed that there is only one excitation location, but measurements may be taken simultaneously at multiple locations. If there
are multiple excitation locations, then another summation on j can be taken to obtain a valid objective function. The Hjk FRF is a
function of the element stiffness matrices, and therefore, it also is a function of the damage indicators. Since the goal of damage
identification is to find the damage indicators for each finite element; consequently the damage indicators become the unknown
variables to be optimized.
Encoded damage
indicator for
Element 2
1 0 0 1 0 1 1 0 0 0 1 0 1 1 0 1 1 0 1 1 ... 1 0 0 1 0 1 1 0 0 0
3
Figure 1. Fixed representation for damage identification
A typical IRR GA representation is depicted in Figure 2. In this representation, three parts of the string can be identified: a
predefined gene locator (GL) pattern; a gene instance; and redundant segments. Each gene instance consists of two parts, a segment
encoding the finite element number (location of damage) and a segment encoding the damage indicator for that finite element
(magnitude of damage). Redundant segments occur between any two consecutive gene instances. The redundant segment may
become a useful gene instance in later generations through the actions of genetic operations. Gene instances encode the parameter
values similar to other GAs. To decode a damage indicator, the IRR string is parsed until a GL pattern is found that identifies a gene
instance. The selected GL pattern used in this research is a [1 1 1]. Then the finite element number identifying the damage location is
decoded along with the corresponding damage indicator value. If more than one gene instance identifies the same finite element then
the average of the damage indicators is taken. For the IRR GA, each individual in the population still represents one complete
solution, but the solution is defined only by the damage indicators for a small subset of the finite elements in the model. For the IRR
GA, the string length is obtained from the number of significant digits for the damage indicator and the user defined expected number
of gene instances in the string. Further details concerning the IRR GA are provided in [3].
Gene instances
...
: Redundant segments
1 1 1 0 0 0 1 0 0 1 0 1 0 0 0 1 0 0 1 0 1 0 0 0 0 1 0 1
Gene locator
(GL) pattern
4
∂H (ω ) ∂Z (ω )
= − H (ω ) H (ω ) (5)
∂x sj ∂x sj
where H(ω) is an FRF matrix and Z(ω) is the inverse of the FRF matrix under investigation. Equation 5, however, is computationally
prohibitive for large structures. In these cases, a numerical differentiation approach for the calculation of the FRF function
sensitivities can be used.
∂H km (ω ) H km (ω ) x sj = 0 − H km (ω ) x sj =∆x sj
≈ (6)
∂x sj ∆x sj
where Hkm(ω) is an FRF function (kmth element of H(ω)) for which the excitation is located at the kth degree of freedom (DOF) and
the measurement is taken at the mth DOF, and ∆x j is a small perturbation at element j.
s
The diagnostic information measure was defined for a pair of excitation, kth, and measurement, mth, DOFs using the FRF function
sensitivities with respect to the damage indicator vector xs.
2
ne ⎛ ϖ1 ∂H km (ω ) ⎞
ιkm = ∑⎜ ∫ dω ⎟ (7)
⎜ ∂x js ⎟
⎝
j =1 ϖ 0
⎠
where lkm is the information contained in the Hkm FRF function, ϖ0 and ϖ1 are the lowest and highest measured frequencies,
respectively; and ne is the number of finite elements in the model. The information measure is defined with respect to all finite
elements in the model; therefore this function gives an overall quality measure of the information contained in the measurement data.
The total diagnostic information, I, contained in nmeas sensors, assuming only one excitation DOF, is defined as the sum of the
squared lkm terms obtained considering each measurement location using Equation 7.
0 1 1 0 0 0 0 0 0 1 0 1 0 0 1 1 0 0 1 0 0 1 0 ... 0 0 1
ON/OFF switches
corresponding to sensor
locations at elements
5
the fixed GA representation is simply a string in which all damage indicators are encoded with zero values. The benefit of using a
zero damage individual in the initial population is that this solution may contain beneficial genetic information that enables the fixed
GA to find a better solution. In each trial, the benefit of the optimal excitation and sensor layouts that were evolved using multi-
objective genetic algorithms is also investigated [2]. Solving an inverse problem using limited information leads to uniqueness
problems of that solution, i.e. several solutions can be found satisfying the inverse problem formulation. Since noise in sensor data
further intensifies the difficulty of inverse problems, the effect of noise on SDIM accuracy was studied for 0%, 5%, and 10% noise.
Normally distributed random noise was added to the simulated FRF data with zero mean and a variance of unity. The number of
significant digits for all case study trials was 6, which required 20 bits to represent a damage indicator with a precision of about
9.44⋅10-7 or 9.44⋅10-5%. Finally, for all case studies the Young’s modulus of steel, E, is 207 GPa (30,000 ksi), Poisson’s ratio is 0.3,
and the density is 7780 kg/m3 (0.000728 lb-s2/in4).
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
69 71 72
46
3.66 m
(12 ft)
36
67 68 65 66
73
3.66 m
(12 ft)
63 64 59 60 61 62
21 6 7
19 5
4.57 m
(15 ft)
38 28 16 11 2 4 10 18
49 27 3 34
6
Table 2. GA Parameter Settings for SDIM Case Study Trials
TWSP (Fixed and IRR) FRM (IRR only)
Population Size 20 200
Tournament Size 4 (Fixed) 6, 8 (IRR) 8
Maximum Generations 200 300
Stop if no Improvement 20 20
Crossover Type Adaptive Adaptive
Crossover Sites 6, 8 6
Crossover Rates Primary: 0.9, Secondary: 1.0 Primary: 0.9, Secondary: 1.0
Mutation Type Non-uniform (Fixed) Uniform
Mutation Rate 0.005, 0.0075 0.005
Elitism Yes Yes
IRR - Number of gene instances was 5 and the gene locator (GL) length is 3
For 0% noise, the fixed representation was not able to found the global optimum after 200 generations for the TWSP case study.
Instead the fixed representation obtained a solution in which the correct damaged elements were identified with higher damage
indicator values (4.19 and 8.82%) than any other elements. The largest false prediction of 1.48% occurred at element 11. Since the
fixed representation encodes all of the damage indicators, there are many falsely identified damaged elements of low severity.
Hillclimbing found the global optimum. In comparison for 0% noise, the IRR GA found a near-global optimum with only two false
predictions of low severity. Only a few hillclimbing iterations were required to find the global optimum.
The damage identification results obtained considering 5% noise are presented in Figure 7. The predicted damages in element
6 and 10 are slightly underestimated with gradually increasing magnitudes as the noise level increases. The IRR GA without initial
seed performed the best; finding the global optimum solution 3 times out of 4 (in the sense that hillclimbing could not improve on it)
with the optimal sensor layout.
With 10% noise imposed, the IRR GA representation using the non-optimal sensor layout predicted the following damages in
the true damaged elements: 5.9% in Element 6 and 9.5% in Element 10 as shown in Figure 7. Typically, two elements were falsely
identified with damage severities less than 1%. Therefore the SDIM accuracy degraded with measurement noise. Damage
predictions using the non-optimal layout tended to overestimate the damage in element 6 (5%), while underestimating the damage in
element 10 (10%). These results suggest that the presence of noise has the effect of equalizing damage levels in the SDIM
predictions using the non-optimal layout. In comparison, the IRR GA representation using the optimal sensor location predicted a
4.9% damage in Element 6 and 10.0% damage in Element 10 in the 10% noise trials. In addition, using the optimal layout reduced
the number and severity of falsely identified elements, since no false damage predictions were obtained typically. The average
fitness of the initial population increased by a factor of 3.9 using the optimal layout, which indicates that the measurements used to
detect damage contain more diagnostic information.
0.00%
6 10 11 14 15 16 17
Element number
Figure 6. TWSP SDIM results for 5% comparing type of representation and sensor locations
6.00%
4.00%
2.00%
0.00%
6 10 11 14 15 16 17
Element number
Figure 7. TWSP SDIM results for 10% noise using IRR GA representation comparing sensor locations
7
8. Discussion of Unbraced Frame Case Study Results
Multiple damages with different severities result in increased uniqueness problems that make it increasingly difficult to find the true
damaged elements and severities. The case study investigated for the unbraced frame problem is listed in Table 1. There are six
damaged elements all located in the first floor beams at the joints, which simulates a damage scenario that may occur in frames sized
according to the strong column, weak beam design concept. The damage severity at the exterior beam joints is 20% and the interior
beam joints have 10% damage on each side of the interior joints. The GA input parameters used are listed in Table 2.
Case study trials show that the IRR GA continues to perform well, while no valid results could be obtained using the fixed GA
representation. Previously in the TWSP case study, the initial population of fixed representation SGA had to be seeded with the zero
damage individual in order to perform well. In comparison, the good performance of the IRR GA can be traced to its adaptive
characteristic that make it well suited for damage identification problems.
For 0% noise trials, the number of correctly identified elements out of the six true damaged elements was 4, 5, and 6 using 3, 4,
and 5 optimum sensors, respectively. Three interior beam elements (4, 10 and 11) and one exterior beam element (18) were correctly
identified as damaged in all three trials and the accuracy of the severity estimates increased when more sensors were used. By using
four sensors instead of three, the additional exterior beam element (28) was identified with a severity somewhat higher than the
imposed 20%. The addition of the fifth sensor provided enough information for the SDIM to uniquely identify all six damaged
elements without any false identification as shown in Figure 8. After hillclimbing, the global optimum was found using the five-
sensor optimal layout. By increasing the number of sensors, the SDIM results become more unique. The benefit of the optimal
layout is that only a fraction (5) of the total number of possible sensor locations (144) was required to accurately locate and quantify
multiple damages with different severities.
Additional trials were performed assuming 5% measurement noise. Using a 9-sensor optimal layout, all six true damage
elements were identified with damage indicator values reasonably close to the inflicted damages as depicted in Figure 9. After
hillclimbing, however, element 28 was assigned a zero damage indicator reducing the correctly located elements to five. The
information contained in the measurements using nine sensors was not sufficient to uniquely locate all damage elements with high
noise levels. Based on the trials performed, it is concluded that that increasing the number of sensors may not always facilitate
damage identification, although typically both localization and quantification of damages becomes more accurate with increasing
number of sensors. The results of trials investigated for the noisy measurements suggest that there is an optimal sensor configuration
that is dependent on the damage scenario. In other words, different damage locations may impose a different set of optimum sensor
locations. For certain civil engineering structures, such as frames, the designer may have a good understanding about possible
damage locations before measurements are taken. For instance, joints on lower floors are more likely to be damaged than joints at
higher elevations. Although the information measure defined in this research was formulated with respect to all finite elements in the
model, this measure could be prioritized based on heuristics.
Simulated 10% Damages in Elements 4, 10, 11 and 16, and 20% Damages in
Elements 18 and 28
25.00%
Trial 5, GA
20.00%
Percent damage
Trial 5, Hillclimb
15.00%
10.00%
5.00%
0.00%
4 10 11 16 18 19 20 21 28 48 51 75
Element number
Figure 8. FRM SDIM results using IRR GA and a 5-sensor optimal layout (0% noise)
Simulated 10% Damages in Elements 4, 10, 11 and 16, and 20% Damages in Elements
18 and 28
50.00%
Trial 9, GA
40.00%
Trial 9, Hillclimb
Percent damage
30.00%
20.00%
10.00%
0.00%
2 4 10 11 12 13 16 18 19 21 24 28 38 43 56 67
Element number
Figure 9. FRM SDIM results using IRR GA and a 9-sensor optimal layout (5% noise)
There are at least three distinguishable benefits of using optimum measurement layouts in damage identification. First, the
damage-balancing phenomenon (i.e. predicted damage indicators of unequal damages tend to approach to a value in between the
inflicted damage severities) that is observed for the non-optimal sensor location cases is not apparent in the optimum sensor location
cases. The number and severity of falsely identified elements is also very much reduced. In addition, the consistency of the results is
8
maintained at all damage levels and the accuracy of damage identification is less sensitive for noise in the measurements.
The performance of fixed representation GA degrades as the number of variables increases due to the larger search space that
needs to be explored. For example, the size of the search space considering a 20-bit encoding of the damage indicators is
S = ( 220 ) = 6.668 ⋅10240 , which is extremely large. In comparison, the IRR GA is able to effectively reduce the search space by
40
adjusting the number of gene instances present in the string. For example in a trial with a single damaged element, the best individual
in the initial population may have 8 gene instances (damaged elements), which is equivalent to a ( 220 ) = 4.562 ⋅10192 times reduction
32
in the search space when compared to the fixed GA. In the final population, there may be only a single gene instance in the best
individual, which means that the search space has a size of 2 20 = 1, 048, 576 , which is much smaller then the initial search space or
the search space for the fixed representation.
9. Conclusions
This research addressed the problem of structural damage identification using the inherent information contained in FRF vibration
signatures. A robust and efficient SDIM that only requires minimal measurement information to precisely locate and quantify the
severity of damage was developed. The SDIM adjusts the stiffness properties of the undamaged linear discrete analytical model by
minimizing the difference between the measured and analytically computed FRFs. The problem of damage identification is
formulated as an unconstrained optimization problem using continuous variables representing the stiffness reductions of elements.
Two GA representations were evaluated on simulated damage case studies involving a two-span continuous beam and an unbraced
frame structure. The performance of the fixed GA representation was severely impacted by the size of the search domain and seeding
the initial population was required to obtain a solution. The accuracy of damage identification using the fixed GA also degraded with
increased measurement noise.
The IRR GA in comparison was able to evolve global or near-global optimum solutions in noise-free trials. In the presence of
noise, the IRR GA results were considerably less sensitive than the fixed GA results. As a consequence, on average the IRR GA
converged faster than the fixed representation GA and its ability to find the true damaged elements was much more stable and
accurate. Even for large frame problems, the IRR GA was capable of identifying the true damaged elements due to encoding only a
small subset of all possible finite elements. In addition, seeding the initial population with the zero damage individual was not
necessary to find the solution for IRR GA, but was advantageous in certain situations.
Trials using non-optimal and optimal sensor layouts were performed to identify the benefits of using the optimal sensor layouts
with the FRF-based SDIM. The solution to the multi objective optimization problem was a set of optimal sensor layout designs,
which defined tradeoffs between the two objectives. Overall, trial results show that there was an increase in the measurement
information collected using the optimal layouts. However, the diagnostic information contained in the measurements did not
increase proportionally with the increase in the number of sensors. For all trials, the ability of the SDIM to uniquely identify
damaged elements was enhanced using the optimal sensor layout designs, even in noisy measurement environments. Significant
improvement in the robustness of the damage identification method was found when the optimal sensor layout designs were used for
detecting multiple damages in larger frame structures.
The FRF-based SDIM developed in this research is distinct from other SDIMs in several aspects. Damage can be localized and
quantified using only a small subset of the possible measurement locations using the IRR GA representation. The SDIM developed is
robust and is able to accurately locate and quantify multiple damages in relatively large-scale structures. The SDIM provides
reduced sensitivity to noise, which allows greater accuracy in predicting damage under realistic noise levels. This research shows
that IRR GAs support the development of highly-sophisticated systems capable of identifying crucial changes in the environment. In
the future, the use of advanced GAs will help identify new solution techniques, alter the way in which we define the problems, and
will broaden the boundaries of smart design technologies.
10. References
1. Rytter, A., (1993). “Vibration based inspection of civil engineering structures.” Doctoral dissertation, Department of Building
Technology and Structural Engineering, University of Aalborg, Denmark.
2. Liszkai, T.R. (2003). “Modern heuristics in structural damage detection using frequency response functions,” PhD Thesis, Dept.
of Civil Engineering, Texas A&M University, College Station, TX.
3. Raich, A.M. and Ghaboussi, J. (1997). “Implicit redundant representation in genetic algorithms.” Evolutionary Computation 5
(3), 277-302.
4. Cawley, P., and Adams, R.D., (1979). “The location of defects in structures from measurements of natural frequencies.” Journal
of Strain Analysis, 14(2), 49-57.
5. Kim, H.M., and Bartkowicz, T.J., (2001). “An experimental study for damage detection using a hexagonal truss.” Computers
and Structures, 79, 173-182.
6. Kim, J.T., and Stubbs, N., (2002). “Improved damage detection identification method based on modal information.” Journal of
Sound and Vibration, 252(2), 223-238.
7. Law, S.S., Chan, T.H.T., and Wu, D., (2001). “Efficient numerical model for the damage detection of large scale structure.”
Engineering Structures, 23, 436-451.
8. Wang, Z. Lin, R.M. & Lim, M.K. 1997. Structural damage detection using measured FRF data. Computer Methods in Applied
Mechanics and Engineering 147: 187-197.
9. Lee, U. & Shin, J. 2002. A frequency response function-based structural damage identification method. Computers & Structures
80: 117-132.
10. Thyagarajan, S.K. Schulz, M.J. & Pai, P.F. 1998. Detecting structural damage using frequency response functions. Journal of
Sound and Vibration 210 (1): 162-170.
9
11. Mares, C., and Surace, C., (1996). “An application of genetic algorithms to identify damage in elastic structures.” Journal of
Sound and Vibration, 195(2), 195-215.
12. Dunn, S.A. 1998. The use of genetic algorithms and stochastic hill-climbing in dynamic finite element model identification.
Computers & Structures 66 (4): 489-497.
13. Moslem, K., and Nafaspour, R., (2002). “Structural damage detection by genetic algorithms.” AIAA Journal, 40(7), 1395-1401.
14. Goldberg, D.E. 1989. Genetic algorithm in search, optimization, and machine learning. Reading, Mass.: Addison-Wesley Pub.
Co.
15. Chou, J-H., and Ghaboussi, J., (2001). “Genetic algorithm in structural damage detection.” Computers and Structures, 79(14),
1335-1353.
16. Holland, J.H. 1975. Adaptation in natural and artificial systems. Ann Arbor, MI: The University of Michigan Press.
17. Michalewicz, Z. 1996. Genetic algorithms + data structures = evolution programs. Berlin – New York: Springer-Verlag.
18. Coello, C.A.C. (2001). “A short tutorial on evolutionary multiobjective optimization.” First International Conference on
Evolutionary Multi-Criterion Optimization, EMO 2001, LNCS 1993, Springer-Verlag, Berlin Heidelberg, 21-40.
19. Srinivas, N. and Deb, K. (1994). “Multiobjective optimization using nondominated sorting in genetic algorithms,” Evolutionary
Computation, 2(3), 221-248.
10