2010 International Conference on Artificial Intelligence and Computational Intelligence

Damage Detection in Structures using Artificial Neural Networks

Shilei Zhang, Huanding Wang, Wei Wang Shaofeng Chen

School of Civil Engineering School of Transportation Science and Engineering
Harbin Institute of Technology Harbin Institute of Technology
Harbin, China Harbin, China
e-mail: [email protected], [email protected] e-mail: [email protected]
and [email protected]

Abstract—In order to select a sensitive input parameter for And the mapping between input parameters and damage
artificial neural networks in damage detection and construct states is established. At last, the artificial neural network is
an efficient and robust back propagation algorithm for damage inputed by the parameter of the structural dynamic
assessment, the application of neural networks to damage characteristics, which is obtained by measuring the damage
detection in structures is summarized and analyzed in this structure. And damage is detected by the output of the
paper. By discussing the use of natural frequency as a artificial neural network.
diagnostic parameter, natural frequency can rationally reflect Many scholars devote to study on the aspect. Kirkegaard
damage location but not provide enough information about and Rytter[2] use the back propagation algorithm to
damage degree. Mode shape and transfer function include
determine the damage location and degree of a steel beam by
abundance information about damage degree compared with
natural frequency but have a large measurement error. And
frequency variations before and after damage.Wu, Ghaboussi
three improved back propagation algorithms that are adaptive and Garrett[3] use the back propagation algorithm to identify
variable step-size algorithm, Levenberg-Marquart algorithm the damage of a three layer frame by a response spectra.
and homogeneous algorithm are introduced. The result shows Anantha and Johnson[4] carry out the damage assessment of
that Levenberg-Marquart algorithm harmonizes Gauss- composite materials by a fuzzy neural network. Kaminski[5]
Newton method with steepest descent method and tunes compares the effectiveness among natural frequency,
gradually to Gauss-Newton method when the result can not frequency variance and the ratio of regular frequency
converge to the minimum. Thus choosing complete vibration variances. Elkordy, Chang and Lee[6] use the variance of
modal parameters and using Levenberg-Marquart algorithm, mode shapes to identify the damage of a five layer frame.
structural damage can be effectively detected. Rhim and Lee[7] input the transfer function to detect the
damage of a cantilever beam by a multilayer perceptron
Keywords-artificial neural networks; damage detection; input model. Pandy and Barai[8] use a multilayer perceptron
parameter; Levenberg-Marquart algorithm model and the displacement under static loads to identify the
damage of a steel truss bridge. Mitsuru Nakamura[9] chooses
I. INTRODUCTION the relative displacement of interlayers as the input and the
restoring force of interlayers as the output. And the valicity
Based on the characteristics of human brains, the
artificial neural network is a parallel distributed processing of the method is studied by the result of a seven layer steel
system by cell bodies. The artificial neural network is also structure before and after the Hanshin-Awaji-daishinsai
called the connectionism mode, which can achieve the earthquake in 1995. Using the acceleration of a steel truss,
function of a nonlinear dynamical system. Now, many Chen and Kim train the neural network to identify the
researches based on model updating and singnal processing damage degree[10]. And the fuzzy neural network is
are carried out in the field of the structural dynamical combined to judge the damage of a beam[11]. Above results
damage. But there are some problems for above researches show that artificial neural networks have received increasing
such as easily influnce by the environment, the strong model attention in damage detection. However, above researchers
dependence and the poor tolerance to the fault of the system do not point out what kinds of input parameters is sensitive
and so on. However, in the field of structural damage and which algorithm may lead to false or unreliable output
results from such networks. In this paper, the application of
detection, the artificial neural network is used widely due to
artificial neural networks to damage detection will be
its good nonlinear mapping ability, the ability of solving
summarized and analyzed. How to select input parameters
inverse problems and the strong robustness.
for neural networks and some efficient and robust algorithm
The realizing way to identify damage by artificial neural
will be studied.
networks[1] is showed as follow. Firstly, based on the
analysis of the finite element method, the learning sample of II. PARAMETRIC STUDY
artificial neural networks is constructed by the structural
dynamic characteristics on different damage states. Then, the The selection of input parameters is very important to the
artificial neural network is trained by the learning sample. application of neural networks on damage detection. The
damage should be reflected sensitively by input parameters.

978-0-7695-4225-6/10 $26.00 © 2010 IEEE 207

DOI 10.1109/AICI.2010.50
In the past, mode data and static responses are choosed as ⎡ J ⎤
input parameters. And they can be divided into several kinds Δωi2 = ⎢ ∑υ Tj ( Φ i ) Δk jυ j ( Φ i )⎥ / ( ΦT M Φ ) (10)
as follow. ⎣ j =1 ⎦
can be deduced, where J represents the amount of damage
A. Two order frequency variation ratio elements.
Frequency is a sensitive parameter on damage detection. For the Nth damage element,
And frequency has a good measuring accuracy. But Δωi2 = ⎡⎣υ NT ( Φ i ) Δk Nυ N ( Φi ) ⎤⎦ / ( ΦTi M Φ i ) (11)
frequency can not be used to identify the damage of
symmetric positions on a symmetrical structure. can be obtained by (10).
Cawley’s research shows that two order frequency Assuming that Δk N = α N k N ,
variation ratio is only related to the damage location[12].
Based on the characteristics of structures, Δωi2 = ⎡⎣α Nυ NT ( Φi ) k Nυ N ( Φ i )⎤⎦ / ( ΦTi M Φ i ) (12)
δω = f (δ K , r ) , (1) can be deduced.
The square of the frequency variation is related to the
where r represents the damage location. damage degree and location of the element. If the ith
To expand (1) and ignore the higher order term, frequency and the jth frequency are used,
∂f ( 0, r )
δω = f ( 0, r ) + δ K Δωi2 ⎡υ N ( Φi ) k Nυ N ( Φ i ) ⎤ ⎡υ N ( Φ j ) k Nυ N ( Φ j ) ⎤
T T
(2)
∂ (δ K ) =⎢ ⎥/⎢ ⎥ (13)
Δω 2j ⎢⎣ ΦTi M Φ i ⎥⎦ ⎢⎣ ΦTj M Φ j ⎥⎦
can be obtained.
Because f ( 0, r ) = 0 ， can be gotten.
It shows as in (13) that the ratio of the frequency
∂f ( 0, r ) variation square is the function of damage locations, when
δω = δ K = δ Kgi ( r ) (3)
∂ (δ K ) damage degrees are near.
can be deduced. C. Frequency variation ratio
Assuming that the variation of the stiffness is
Kaminski compares the effectiveness among the natural
independent to frequency,
frequency, the frequency variation and the regular frequency
δωi / δω j = gi ( r ) / g j ( r ) = h ( r ) (4) variation ratio to damage detection[5]. The result shows that
can be obtained. the ratio of the frequency variation
So, according to (4), two order frequency variation ratio FFCi = ( fui − f di ) / fui (14)
is only the function of the damage location.
is related to the damage degree and the damage location, but
B. The square of the frequency variation the regular frequency variation ratio
The square of the frequency variation is related to the ⎛ p ⎞
damage location and the damage degree. However, the ratio N FCRi = FFCi / ⎜ ∑ FFC j ⎟ (15)
of the frequency variation square is only related to the ⎝ j =1 ⎠
damage location [13]. is only related to the damage location.
The characteristic equation of the structural motion is D. Mode shape, mode curvature and transfer function
(
K − ω2M Φ = 0 . ) (5) According to the relationship between displacements and
The perturbation equation of (5) is strains, each displacement mode shape is corresponding to a
(( K + ΔK ) − (ω 2
)
+ Δω 2 ) ( M + ΔM ) ( Φ + ΔΦ ) = 0 . (6) strain mode shape. When the damage happens, there is a
prominent stress redistribution nearby the damage location.
Let ΔM equal to zero.To expand (6) and ignore the That will cause a significant change of the strain mode
quadratic term, shape. The damage can be located by comparison with every
(
Δω 2 = ΦT K Φ / ΦT M Φ )( (7) ) order strain mode shape before and after the damage.
According to the theoretical analysis as in [14], high order
can be gotten. mode shapes are easier to identify the structural damage than
The overall stiffness matrix can be decomposed to the low order mode shapes. But exact high order mode shapes
element stiffness matrix. And the element deformation is actually are not easy to be obtained. The error of mode
related to the structural mode shape. And shapes has an enormous implication to the result. Some
ν m (Φ) = f (Φ) . (8) researchers point out that the measurement error of mode
For the ith mode, there are shapes matches with the mode shape variation caused by the
M structural local change[15] .
ΦTi K Φi = ∑υmT ( Φi ) kmυm ( Φi ) , (9) The relationship between curvature and bending moment
m =1 is v′′ = M / EI . When the damage happens, EI decreases
where M is the amount of elements.
and v′′ increases. Therefore, the damage can be located by
To take (9) into (7),
the variation of v′′ . Chance, Worden and Tomlinson[16]

208
prove that mode curvatures are better than mode shapes to A. Adaptive variable step-size algorithm
damage detection. The mode curvature is a very sensitive The difference between adaptive variable step-size
parameter for the damage. But the measurement mode shape algorithm and the standard back propagation algorithm lies
usually can not reach a high precision so that limits the in that the learning step is rectified by the change of the error
application of mode curvatures. curve. Due to the nonuniform of the gradient for an
The adventage of using dynamic responses and modes to approximate error curve on the standard back propagation
identify the damage is the input parameter can be obtained algorithm, the convergence becomes slow, when η is a small
easily. But there are also several shortcomings. Firstly, only value. When η is a large value, an oscillation takes place at
using frequency information is difficult to identify the the valley region. The adaptive variable step-size algorithm
damage of symmetric positions on a symmetrical structure. can correct the problem effectively. The theoretical basis of
Secondly, high order mode shapes can not be measured the adaptive variable step-size algorithm is the advance and
easily. In order to detect damage degrees, a large of retreat method on evolutionism, which computes two
comparisons between the new and the old model have to be
carried out. At last, there may produce a blurring result successive training errors[25]. If E ( k ) > Emin ∗ er ,
unless using modes with a high strain energy. However, the ⎧⎪( k + 1) = η ( k ) ∗ de α =0
transfer function has complete information and reduces the then ⎨ (18)
number of sensors. During the phase of training, the ⎪⎩( k + 1) = η ( k ) ∗ in α =α
structural model is classified by the damage location and the where Emin is the smallest error on the kth iteration; er is
damage degree. On the identification stage, the unknown
the learning rate based error rebound; de and in are
damage structure is classified into the most similar
respectively the decrease rate and increase rate of learning
category[17]. But for an actual structure, it is difficult to
steps; α is the coeffcient of momentum terms.
apply enough control forces or excitations as inputs. And the
environmental random vibration is difficult to be measured. B. Levenberg-Marquart algorithm
The input is difficult to be determined so that is not easy to During the process of solving a nonlinear equation, the
get the transfer function. Gauss-Newton method has two order convergence speed.
E. Other input parameters But the Hessian matrix may become an odd matrix so that
the iteration can not be continued. Therefore, the stability of
Pandy and Bara[8] identify the damage of a truss bridge
the Gauss-Newton method is not good. The Levenberg-
by the structural node static response. Their research shows
Marquart algorithm is the harmonization between the Gauss-
that the static measured data usually has a high precision.
Newton method and the steepest descent method. When the
But some parts of a large structure can not be measured.
result keep away from the minimum, the Levenberg-
Based on the time domain theory, Mitsuru Nakamura[9] puts
Marquart algorithm tunes gradually to the Gauss-Newton
forward to use the structural story drift and story speed as
method[26].
input parameters and resiliences as output parameters. The
The equation of the Levenberg-Marquart algorithm is
sensibility of six kinds of input parameters to damage
detection is compared. And the results from low to high are W ( k + 1) = W ( k ) + Pk (19)
respectively displacement mode, natural frequency,
Pk = −2 ( H + μ DH ) ∇E (W ( k ) ) ,
−1
(20)
displacement, transfer function, mode curvature and strain
function[18]. where H is the Hessian matrix of the energy function E; DH
is a diagonal matrix which is composed by the diagonal
III. IMPROVED BACK PROPAGATION ALGORITHMS element of H; ▽E is the derivative matrix of E. If
Above studies mainly focus on the sensitivity of input E (W ( k + 1) ) ≥ E (W ( k ) ) , then μ = 10μ . Else, μ = μ /10 .
parameters of neural networks. Beside the influence of input
parameters, the algorithm also has an impact on the result of C. Homogeneous algorithm
damage detection. Now, the back propagation algorithm is
The homogeneous algorithm is an efficiency method on
used widely in damage detection. The standard back
the theory of the differential topology. The homogeneous
propagation algorithm uses the law of chain derivative to
back propagation algorithm brings the idea of advancements
calculate the gradient of the error square sum[19-23]. The
in a regular order into the back propagation algorithm ,
steepest descent method with a regular step is used to modify
which is helpful for solving the minimum of the energy
the weight of artificial neural networks to minimize
function. The global minimum can be obtained by tracking
errors[24]. The algorithm is
the change of the energy curvature and the curvature
ωij (n + 1) = ωij ( n ) + Δωij ( n ) (16) trajectoty of the least value[27]. The first step of the linear
Δωij ( n ) = ηδ i ( n ) yi ( n ) , (17) homogeneous algorithm with a teacher is to make sure the
teacher homogeneous function T,which is
where ωij ( n ) is the weight from the ith node to the jth node T ( t ) = (1 − t ) Tb + t ∗ Te , (21)
during the nth iteration; Δωij ( n ) is the increment of the
where t is the variation to form the transition function and
weight; η is the learning step; δ i ( t ) and yi ( t ) are
transits from 0 to 1, t∈[0,1]; Tb and Te are respectively the
respectively the local gradient and the output of the ith node.
initial comformed teacher and the given teacher.

209
 The learning procedure of the homogeneous algorithm is [8] P. C. Pandy and S. V. Barai, “Multilayer perception in damage
2 detection of bridge structures,” Computers and Structures, vol.54,
Ei (W , ti ) = T ( ti ) − O (W , I ) (22) pp.597-608, 1995.
[9] Mitsuru Nakamura, Masri Sami and Anatassios, “A method for
T ( ti ) = (1 − ti ) Tbi + ti ∗ Te (23) nonparametric damage detection through the use of neural networks,”
Earthquke Engineering and Structural Dynamic, vol.27, pp.997-1010,
Tbi = O (Wi −1 , I ) (24) 1998.
where O (W , I ) is the actual output of the neural networks. [10] S. S. Chen and S. Kim, “Neural Network based Signal Monitoring in
A Smart Structural System,” Smart Structures and Materials: Smart
Sensing, Processing and Instrumentation, J.S.Sirkis, SPIE, 1994, pp.
IV. CONCLUSIONS 176-186.
Damage detection based on artificial neural networks is [11] Xiaoyun Han and Ruiyan Liu, “Researches for beam’s fault diagnosis
a subject with a widespread application. How to select the based on neural network and fuzzy synthetic judgement,” Journal of
National University of Defense Technology, vol.18, pp.17-22, 1996.
input parameter of artificial neural networks is very
important to damage detection. The input parameter should [12] P. Cawley and R. D. Adams, “The location of defects in structures
from measurements of natural frequencies,” Journal of Strain
be sensitive and obtained easily. Natural frequency is easy Analysis, vol.14, pp.49-57, 1979.
to be acquired. It is reasonable to use natural frequency as [13] Fangqing Gao, Jianming Jin and Shuying Gao, “Damage detection in
input parameters. But information is not sufficiency ,if only structures by model analysis,” Journal of Southwest Jiaotong
using natural frequency. Mode shape has abundance University, vol.33, pp.108-113, Feb. 1998.
information but a relatively large measurement error. The [14] Saluwu, “Detection of structural damage through in frequency: a
transfer function includes complete information but not easy review,” Engineering Structures, vol.19, pp.718-723, 1997.
to be obtained. Meanwhile, the algorithm and the structure [15] U. Pabst and P. Hagedorn, “Identification of localized losses of
of the artificial neural network also have a large influence to stiffness in structures, structural dynamic of large scale and complex
system,” American Society of Mechanical Engineering, Design
damage detection. Results show that Levenberg-Marquart Engineering Division, vol.59, pp.99-104, 1993.
algorithm harmonizes the Gauss-Newton method with the
[16] J. E. Chance, K. Worden and G. R. Tomlinson, “Processing Signals
steepest descent method and tunes gradually to the Gauss- for Damage Detection using Frequency Response Functions,” Smart
Newton method when results keep away from the minimum. Structures and Materials: Smart Sensing, Processing and
Thus choosing complete vibration modal parameters as Instrumentation, J.S.Sirkis, SPIE, 1994, pp.187-198.
input and using Levenberg-Marquart algorithm, structural [17] Yingming Wang, “An adaptive neuro-fuzzy inference system for
damage can be effectively detected. bridge risk assessment,” Expert Systems with Application, vol.34, pp.
3099-3106, 2008.
ACKNOWLEDGMENT [18] M. Ruzzene, A. Fasana, L. Garibaldi and B. Piombo. “Natural
frequencies and damping identification using wavelet transform:
This work is supported by the health monitoring project application to real data,” Mechanical Systems and Signal processing,
of Yitong river bridge and the safety assessment project of vol.11, pp.207-218 , 1997.
Qiandaohu 1# bridge. [19] K. J. Janssen, K. G. Moons, C. J. Kalkman, “Updating methods
improved the performance of a clinical prediction model in new
REFERENCES patients,” Journal of Clinical Epidemiology, vol.61, pp. 76-86, 2008.
[20] Weizhuo Wang, John E. Mottershead, Cristinel Mares, “Mode-shape
recognition and finite element model updating using the zernike
[1] Wei Chen, Weilian Qu and Zhenggou Jiang. “The application of moment descriptor,” Mechanical Systems and Signal Processing,
artificial neural network to structural damage detection,” World vol.23, pp . 2088-2112, 2009.
Earthquake Engineering, vol.18, pp.146-149, Mar. 2002.
[21] J. A. Guajardo, R. Weber, J. Miranda, “A model updating strategy for
[2] P. H. Kirkegaard and A. Rytter, “The use of neural networks for predicting time series with seasonal patterns,” Applied Soft
damage detection and location in a steelmember,” Neural Networks Computing, vol.10, pp. 276-283, 2010.
and Combinatorial Optimization in Civil and Structural Engineering,
Edinburgh,UK, pp.1-9,1993. [22] M. Palmonella, J. E. Mottershead, “Finite element models of spot
welds in structural dynamics: review and updating,” Computers and
[3] X. Wu, Ghaboussi and J. H. Garrett, “Use of neural networks in Structures, vol.83, pp.648-661, 2005.
detection of structural damage,” Computers and Structures, vol. 42,
pp. 649-659, April 1992. [23] Alemdar Bayraktar, Baris Sevin, Ahmet Can Altunisik, “Effect of the
model updating on the earthquake behavior of steel storage tanks,”
[4] R. S. Anatha and V. T. Johnson, “Damage assessment of composite Journal of Constructional Steel Research, vol.66, pp. 462-469, 2010.
structures: a fuzzy logic integrated neural networks approach,”
Computers and Structures, vol.57, pp.491-502, 1995. [24] Zhong Yang, Ming Bao, Chuansheng Zhao. “Using back propagation
neural networks to correct the stiffness matrix in a finite element
[5] P. C. Kaminski, “The approximate location of damage through the model,” Journal of Vibration Engineering, vol.8, pp. 390-394, Dec.
analysis of natural frequencies with artificial neural networks,” 1995.
Journal of Process Mechanical Engineering, vol.209, pp.117-123,
1995. [25] Suwen Chen and Guoqiang Li, “Application of atificial neural
networks to damage identification of structures,” Journal of Vibration,
[6] M. F. Elkordy, K. C. Chang and G. C. Lee, “Application of neural Measurement and Diagnosis, vol.21, pp. 116-125, Jun. 2001.
networks in vibrational signature analysis,” Journal of Engineering
Mechanics, vol.120, pp.251-264, 1994. [26] Shengjun Suo, Guangwei Sun and Yan Wang, “The performance
comparison of several kinds of modified BP algorithm,” Journal of
[7] J. Rhim and S. W. Lee, “A neural network approach for damage Harbin University of C.E.&Architecture, vol.33, pp.80-83, Feb.2000.
detection and identification of structures,” Computational Mechanics,
vol.16, pp.437-443, 1995. [27] M. Mehrjoo, N. Khaji, H. Moharrami and A. Bahreininejad, “Damage
detection of truss bridge joints using artificial neural networks,”
Expert Systems with Application,vol.35 , pp.112-1131, 2008.

210