Beruflich Dokumente
Kultur Dokumente
www.elsevier.com/locate/engstruct
Received 27 October 2006; received in revised form 8 January 2007; accepted 8 January 2007
Available online 12 March 2007
Abstract
Artificial Neural Networks (ANN) have received increasing attention for use in detecting damage in structures based on vibration modal
parameters. However, uncertainties existing in the finite element model used and the measured vibration data may lead to false or unreliable
output result from such networks. In this study, a statistical approach is proposed to take into account the effect of uncertainties in developing an
ANN model. By applying Rosenblueth’s point estimate method verified by Monte Carlo simulation, the statistics of the stiffness parameters are
estimated. The probability of damage existence (PDE) is then calculated based on the probability density function of the existence of undamaged
and damaged states. The developed approach is applied to detect simulated damage in a numerical steel portal frame model and also in a laboratory
tested concrete slab. The effects of using different severity levels and noise levels on the damage detection results are discussed.
c 2007 Elsevier Ltd. All rights reserved.
Keywords: Damage detection; Neural networks; Uncertainties; Rosenblueth’s point estimate; Random noise; Modal data
1
µα = E(α) = (α̂++ + α̂−− + α̂+− + α̂−+ ) (2)
4
1
σα = [E(α 2 ) − [E(α)]2 ] 2 (3)
where E(α 2 ) is calculated using Eq. (2) with α 2 terms
substituted for the α terms. Using this approach substantially Fig. 2. Finite element model of the frame.
reduces the computational time in deriving the statistical mean
and standard deviation of each stiffness parameter for structural The input data consist of natural frequencies and mode shapes,
condition monitoring. and the output layers consists of Young’s modulus (E values)
The PDE can be estimated from statistical distributions to represent the stiffness parameter. The change of the stiffness
of the stiffness parameters of the undamaged and damaged parameter or the damage severity for each segment are denoted
models. For example, if the stiffness parameter (α j ) of the by a Stiffness Reduction Ratio (SRF), defined as
undamaged segment j is normally distributed with mean E(α j )
and standard deviation σ (α j ), the probability density function E0
SRF = 1 − (5)
can be obtained as illustrated in Fig. 1, where L α j is the lower E
bound of the healthy parameter. In this study, the confidence
where E is the Young’s modulus in the intact state and E 0 is
level is set to 95%, thus the lower bound is L α j = E(α j ) −
that at the damage level of interest.
1.645σ (α j ), which indicates that there is a probability of
In most ANN applications for damage detection, the training
95% that the healthy stiffness parameter falls in the range of
data are obtained from FE analysis, which involved generating
[E(α j )−1.645σ (α j ), ∞]. Similarly, for the stiffness parameter
large number of damage cases based on an initial baseline FE
of segment j in the damaged state (α 0j ), the distribution is again
model. Once the ANN model is well-trained, the testing data
assumed as normal with mean E(α 0j ) and standard deviation are then applied to the ANN model to obtain the locations and
σ (α 0j ), and the corresponding probability density function is severities of any damages. In most of the previous studies, both
also plotted in Fig. 1. The PDE is defined as the probability of training and testing data are assumed to be free from modelling
α 0j not being within the 95% confidence healthy interval. Thus and measurement error. In practice, however, modelling error
the PDE of segment j is and measurement noise are inevitable.
j
Pd = 1 − prob(L α j ≤ xα 0 ≤ ∞) According to Xia [15], the inaccuracy due to modelling and
measurement error can be overcome by taking into account
= prob(−∞ ≤ xα 0 ≤ L α j ). (4) the uncertainties through a statistical method. In this study,
PDE is a value between 0 and 1, and if the PDE of a segment modelling error and measurement noise are assumed to be
is close to 1, then it is most likely the element is damaged. If normally distributed with zero means and specific variance. The
the PDE is close to 0, damage existing in the element is very noise is applied in terms of coefficient of variations (COV). The
unlikely [13]. It should be noted that the stiffness parameters statistical properties of E value for each segment are obtained
of the undamaged and damaged state have normal distributions by using Rosenblueth’s point estimation method verified by
because the random variations in Eq. (1) are assumed as Monte Carlo simulation. A Kolmorogov–Smirnov goodness-of-
zero mean normally distributed random variables. This will be fit test (K–S test) is then applied to verify the distribution type
proved later. of E values. This is followed by calculation of the PDE of E
values for each segment.
3. Methodology
4. Numerical example
A multilayer perceptron with Levenberg–Marquardt back-
propagation algorithm is utilized to train the ANN model in To demonstrate the proposed method, the single span steel
this study. Sigmoid functions are employed as nonlinear activa- portal frame shown in Fig. 2 is used as an example. The
tion functions for all layers. To reduce the effect of overfitting, cross section of beam is 40.50 × 6.0 mm2 , and column is
the early-stopping method [14] is applied during the training. 50.50 × 6.0 mm2 . The span length and height of the frame are
N. Bakhary et al. / Engineering Structures 29 (2007) 2806–2815 2809
Table 2 the actual values. The results show that the ANN model is
E values for scenario 1 and scenario 2 able to provide good output, which indicates that the ANN
Segment 1 2 3 4 5 6 model is reliable in detecting damage from noise-free data.
Scenario 1 0.4× E 1.0× E 1.0× E 0.2× E 1.0× E 1.0 × E To investigate the reliability of the ANN model in predicting
Scenario 2 0.4× E 1.0× E 0.3× E 1.0× E 0.4× E 0.3 × E structural damage with noisy data, 2% and 15% random noise
are applied respectively to the frequencies and mode shapes of
the testing data. Applying these errors implies that the measured
Table 3 data is no longer exact. These data are then introduced to the
First three frequencies values for scenario 1 and scenario 2
trained ANN model. Fig. 5 shows the ANN output.
Undamaged Scenario 1 Scenario 2 The figure shows that for scenario 1, the false damage
Mode 1 4.628 3.9373 3.530 identification occurs at segments 2 and 3, while the stiffness
Mode 2 16.112 12.567 11.269 of segment 6 is overpredicted. The same situation occurred in
Mode 3 20.649 16.491 14.891 scenario 2, where the ANN model falsely identifies damage
at segment 2, overestimates damage at segments 1 and 3, but
underestimates at segment 6. This indicates that the common
ANN model trained with simulated vibration parameters from
finite element model cannot give reliable structural damage
prediction if the errors exist.
Fig. 4. ANN prediction for scenario 1 and scenario 2 compared to the actual value using noise-free input data.
Fig. 5. ANN prediction for scenario 1 and scenario 2 compared to the actual value using noisy input data.
Table 4
Probability of damage existence (%) for every segment of scenario 1 and
scenario 2
Segment
1 2 3 4 5 6
Scenario 1 96.0 38.5 3.0 99.8 0.00 0.00
Scenario 2 99.7 1.0 100.0 13.5 99.7 99.9
Fig. 8. Mean values and coefficients of variation (COV) of E values in undamaged state.
Table 6
Probability of damage existence (%) for different combination of uncertainties in training and testing data
(b) Uncertainties in training data = 0.5% (frequencies) and 5.0% (mode shapes)
0 0 60.5 31.0 60.2 64.2 62.0 49.2
0.5 5.0 99.9 3.2 8.2 99.0 7.0 10.1
1.0 10.0 98.6 11.9 45.0 74.6 41.2 43.2
2.0 15.0 45.3 53.3 39.0 79.8 22.1 4.2
5 40 12.2 49.2 20.3 18.1 11.3 32.1
(c) Uncertainties in training data = 1.0% (frequencies) and 10.0% (mode shapes)
0 0 31.2 24.3 27.0 52.1 60.3 37.4
0.5 5.0 78.1 18.5 43.8 98.3 40.5 6.0
1.0 10.0 100.0 17.3 3.4 96.0 17.4 4.1
2.0 15.0 75.4 22.9 77.0 46.8 40.8 5.2
5 40 25.2 11.2 23.2 50.7 46.1 32.1
(d) Uncertainties in training data = 2.0% (frequencies) and 15.0% (mode shapes)
0 0 28.1 46.9 27.4 25.1 2.1 43.0
0.5 5.0 35.4 28.1 35.2 59.1 54.1 3.2
1.0 10.0 76.8 24.1 46.1 73.9 29.2 32.1
2.0 15.0 92.1 4.0 5.8 100 11.1 3.3
5 40 37.4 4.0 33.4 23.4 31.9 15.7
5. Experimental example was 32 and the mass density was 2.55 × 103 kg/m3 . The slab
rested on wooden planks placed over three steel UB sections.
A two-span reinforced concrete slab with dimensions of Two point loads were applied to each span using hydraulic
6400 mm × 800 mm × 100 mm was tested. The concrete grade jacks as shown in Fig. 9(a) to induce damage. For the dynamic
N. Bakhary et al. / Engineering Structures 29 (2007) 2806–2815 2813
test, a 5.4 kg impact hammer and nine sensors (accelerometers) and 3, the loads at left span were applied first, followed by at
were used. The sensors were placed at 27 points on the slab. the right span.
Fig. 9(b) illustrates the sensor and impact locations. In this For the purpose of damage detection, the slab is divided
study only the vertical bending mode shapes of the middle slab into seven segments, as shown in Fig. 10. The crack patterns
observed during the experiment for the three damage states are
are taken as the input of the ANN models. Three different load
shown in Fig. 11(a)–(c). At level 1, some cracks are clearly
levels were applied by the jacks to produce three damage states.
seen at the middle of the left span, while no cracks observed
The load levels were (i) 18 kN at the left span (level 1); (ii) in another segment. When the load at level 2 was applied, the
18 kN each at both the left and right span (level 2); and (iii) cracks in the left span seemed to be more obvious, and a group
32 kN each at both the left and right span (level 3). At levels 2 of small cracks appeared at the bottom of right span. There were
2814 N. Bakhary et al. / Engineering Structures 29 (2007) 2806–2815
Fig. 13. Probability of damages existence for slab using statistical ANN.
and measurement noise are assumed to have normal distribu- [6] Yeung WT, Smith JW. Damage detection in bridges using neural networks
tion and zero means. The accuracy of the statistical approach for pattern recognition of vibration signatures. Engineering Structures
2005;27(5):685–98.
was proved using Monte Carlo simulation. Using this method,
[7] Sahin M, Shenoi RA. Quantification and localisation of damage in beam-
the probability of damage existence can be estimated. The nu- like structures by using artificial neural networks with experimental
merical and experimental results demonstrated that, compared validation. Engineering Structures 2003;25(14):1785–802.
with the normal ANN approach, the statistical ANN approach [8] Yun C-B, Bahng EY. Substructural identification using neural networks.
gives more reliable identification of structural damage. Computers & Structures 2000;77(1):41–52.
[9] Lee JJ, et al. Neural network-based damage detection for bridges
considering errors in baseline finite elements models. Journal of Sound
Acknowledgements and Vibration 2005;280(3–5):555–78.
[10] Matsuoka K. Noise injection into inputs in back-propagation learning.
The authors would like to acknowledge the partial financial IEEE Transaction of System, Man and Cybernatics 1992;22(3):436–40.
[11] Papadopoulos L, Garcia E. Structural damage identification: A
support from the Australian Research Council and Main Roads
probabilistic approach. AIAA Journal 1998;39(11):2137–45.
Western Australia under grant number LP0453783. Financial [12] Rosenblueth E. Point estimates for probability moments. Proceedings of
support from the Universiti Teknologi Malaysia for the first the National Acedemy of Science 1975;72(10):3812–4.
author to pursue a Ph.D. study in Australia is also gratefully [13] Bakhary N, Hao H, Deeks A. Vibration based damage detection
acknowledged. using artificial neural network with consideration of uncertainties.
In: Proceeding of structural faults and repair, structural faults + repair
2006. Edinburgh: Engineering Technics Press; 2006.
References [14] Prechelt L. Early stopping — but when? In: Orr GB, Muller OR, editors.
Neural networks: Tricks of the trade. Berlin (Germany): Springer-Verlag
[1] Cawley P, Adams RD. The location of defects in structures from Telos; 1999. p. 57–69.
measurements of natural frequencies. Journal of Strain Analysis 1979; [15] Xia Y, et al. Damage identification of structures with uncertain frequency
14(2):49–57. and mode shape data. Earthquake Engineering and Structural Dynamics
[2] Wu X, Ghaboussi J, Garrett JH. Use of neural networks in detection of 2002;31(5):1053–66.
structural damage. Computers & Structures 1992;42(4):649–59. [16] Helton JC, Davis FJ. Latin hypercube sampling and the propagation of
[3] Pandey PC, Barai SV. Multilayer perceptron in damage detection of uncertainty in analyses of complex system. Reliability Engineering and
bridge structures. Computers & Structures 1995;54(4):597–608. System Safety 2003;81:23–69.
[4] Zhao J, Ivan JN, DeWolf JT. Structural damage detection using artificial [17] Maren AJ, Jones D, Franklin S. Configuring and optimizing the back-
neural network. Journal of Infrastructure Systems 1998;4(3):93–101. propagation network. In: Maren A, Harston C, Pap R, editors. Handbook
[5] Zapico JL, Worden K, Molina FJ. Vibration-based damage assessment of neural computing applications. San Diego (CA): Academic press; 1990.
in steel frames using neural networks. Journal of Smart Materials and [18] Bishop CM. Neural network for pattern recognition. Oxford: Oxford
Structures 2001;10:553–9. University Press; 1995.