p.55
p.63
p.69
p.79
p.87
p.99
p.107
p.119
p.131
A Method for Damage Detection Based on Correlation Characteristic of Acceleration Response
Abstract:
A damage detection method using BP neural network based on a novel damage index, the correlation characteristic of the acceleration response, is proposed, and is evaluated through the FEM simulation and experiment verification. On the basis of achievements in existence, the feasibility of using the correlation characteristic as damage index is validated theoretically. The damage detection for a simple-supported beam using the proposed method was FEM simulated. The results showed that the trained BP neural network can correctly detect the location and extent of damages in both single damage case and multi-damage case. A model test of a reinforced concrete simple-supported beam was performed to verify the validity and efficiency of the damage detection method. From the results of the model test, it is shown that the trained BP neural network can correctly locate the damage mostly detect the extent of damage. A conclusion is given that the novel damage detection method using the correlation characteristic of the acceleration response as damage index is feasible and efficient.
Info:
Periodical:
Pages:
87-98
Citation:
Online since:
January 2013
Authors:
Price:
Сopyright:
© 2013 Trans Tech Publications Ltd. All Rights Reserved
Citation:
004 0.
001 3 0.
005 0.
001 0.
003 4 0.
004 0.
002 0.
002 5 0.
006 0.
006 0.
002 6 0 0 0 -0. 004 0 -0. 002 7 1.
[1] 003 0 -0. 002 0 -0. 001 8 0 -0. 016 1.
[1] 003 0 -0. 003 9 0.
005 0 0 0.
004 10 0.
003 0.
003 1.
999 11 0.
003 0 -0. 003 1.
002 12 0.
005 0.
004 0.
002 13 0.
006 1.
996 0.
001 14 0 -0. 005 0 0 0 -0. 001 15 0.
003 0.
004 0 0 16 0 0 0.
003 0.
001 17 0 -0. 003 0.
001 0 0 18 0 0 0 -0. 004 0 0 19 0 -0. 002 0 -0. 002 0 -0. 002 20 0.
006 0.
004 0.
003 Fig. 6 Single-Damage Sample Fig. 7 Multi-Damage Sample As in Table 1, detecting the damaging location using BP neural network is relatively accurate, though a set of continuous units, like 10# and 11#, are hard to be specified. This inaccuracy could be reduced by decreasing the speed and the error of the detection. Detection of Damage Extent. The damage factors of the post-damage structure in Fig. 4 have been put into the database for training as samples while a 20% and a 30% of the damage have been used as the input data to verify the adopted input method. The detecting results are listed in Table 3. The results show significant accuracy as the maximum error is only 1. 2%. Table 2 Results of the Damage Extent from Single-Damage Samples Via Neural Network Method Number 20%Damage Result Error(%) 30%Damage Result Error(%) 1.
000 -0. 004 -0. 40.
000 -0. 007 -0. 70 2.
000 -0. 001 -0. 10.
000 -0. 002 -0. 20 3.
000 -0. 002 -0. 20.
000 -0. 008 -0. 80 4.
000 -0. 004 -0. 40.
000 -0. 012 -1. 20 5.
000.
002.
20.
000.
008.
80 6.
000.
004.
40.
000.
007.
70 7.
000 -0. 001 -0. 10.
000.
000.
00 8.
000.
000.
00.
000 -0. 002 -0. 20 9.
000.
000.
00.
000.
003.
30 10.
200.
202.
20.
300.
311.
[1] 10 11.
200.
206.
60.
300.
309.
90 12.
000.
000.
00.
000 -0. 001 -0. 10 13.
000.
003.
30.
000.
007.
70 14.
000 -0. 001 -0. 10.
000.
001.
10 15.
000 -0. 001 -0. 10.
000 -0. 007 -0. 70 16.
000 -0. 001 -0. 10.
000 -0. 004 -0. 40 17.
000.
002.
20.
000 -0. 002 -0. 20 18.
000 -0. 002 -0. 20.
000 -0. 005 -0. 50 19.
000 -0. 002 -0. 20.
000 -0. 004 -0. 40 20.
000.
000 -0. 40.
000.
000 -0. 70 Detections of Damage Extent and Localization. Combine the samples in 4. 1 and 4. 2 to detect multiple levels of damages in various locations by repeating the training samples, the results can be seen in Table 3. The repeatedly trained samples show high detection accuracy and can precisely detect the damage level and location. Table 3 Results of the Damage Extent from Multi-Damage Samples Via Neural Network Method Number Damage Result Error(%) Damage Result Error(%) 3 0.
001.
09 0 -0. 002 -0. 18 4 0.
001.
09 0.
001.
08 5 0.
005.
50 0 -0. 001 -0. 11 6 0 -0. 004 -0. 42 0 -0. 001 -0. 08 7 30%Damage.
304.
36 10% Damage.
100.
00 8 0.
002.
18 20% Damage.
200.
00 9 0 -0. 002 -0. 16 20% Damage.
200.
00 10 20% Damage.
199 -0. 11 0.
001.
13 11 30% Damage.
307.
70 0.
000.
05 12 0.
004.
39 0 -0. 001 -0. 06 13 0 -0. 001 -0. 12 0.
000.
02 14 30% Damage.
307.
71 0.
000.
04 15 0.
004.
40 0 -0. 001 -0. 11 16 0 -0. 004 -0. 37 0 -0. 001 -0. 11 17 0 -0. 002 -0. 22 0 -0. 001 -0. 09 18 0.
004.
36 0.
001.
06 5 Experimental study To verify the neural network method for damage detection based on correlation characteristic of acceleration response presented in this paper, a set of concrete simply supported beams has been experimented to model the damage of the structure. With a cross section of 0. 3m×0. 15m and a calculation length of 3. 6m (Showed in Fig. 8), the beam has been constructed with C30 concrete and HRB335 steel bars. A 6-stage static loading case has been loaded on the beam to model different levels of damages. The loading curve and the stiffness in each stage can be seen in Fig. 9 and Table 4. The time history of the acceleration response of 5# makes the curve in Fig. 10, and the power spectrum is in Fig. 11. Meanwhile a finite element model of the experimenting structure has been simulated for establishing the neural network sample database of the damaging factors on various testing spots. The samples have been trained repeatedly for the damage detection. Fig. 8 Layout Drawing of the Reinforced Concrete Beam Table 4 Static Rigidity and Damage Load Grade Static Rigidity (EI) Damage Degree(%) Good.
[7] 4e6 0 1.
[6] 3e6 -14. 44 2.
[5] 5e6 -25. 35 3.
[5] 3e -28. 13 4.
[4] 8e6 -34. 65 5.
[4] 0e6 -46. 09 6.
[1] 7e6 -76. 97 Fig. 9 Force-Displacement Curve of Beam Fig. 10 Time Histories of White Noise Excitation Fig. 11 Power Spectrum of White Noise Excitation See Fig. 12 for the variation of the damage factors on real beam. According to previous analysis, damage occurred between censor4 and censor5 after the 4th stage load. The damage locations can be identified with a repeatedly trained neural network under different load stages. Table 5 shows the detection results under stage 3. The damage locations have been precisely spotted while the error is still not ignorable on detecting the damage level. This can be explained as the inevitable insufficiency of the sample database for all engineering structures as the accuracy of the neural network majorly relies on the number of the neurons. Table 5 Damge Detection of 3rd Degree Number Damage Degree(%) Result Error(%) 1-2.
[5] 68.
[5] 73 -0. 90 2-3.
[9] 53.
[8] 63.
[9] 45 3-4.
[18] 62.
[18] 56.
30 4-5.
[25] 33.
[28] 60 -12. 90 5-6.
[20] 12.
[21] 33 -6. 00 6-7.
[12] 81.
[11] 25.
[12] 15 7-8.
[6] 58.
[6] 35.
[3] 45 Fig. 12 Regression Coefficient under Damage 6 Conclusion A BP neural network method for damage detection with the correlation characteristic of acceleration response as the damage index has been presented in the paper. Results from the derivation on the correlation characteristics of the acceleration response and the modal parameters, the numerical simulation of the damage detection process of a concrete simply supported beam and the damage detection experiment of the simulated beam are all been combined to prove the following conclusion.
[1] Theoretically proved the modal information of the structure is contained in the correlation characteristic of the acceleration response, which reveals the damage of the structure.
[2] The correlation characteristic is sensitive to the location and the level of the damage, thus can be practically used as the damage index during damage detection.
[3] The neural network method for damage detection using the correlation characteristic of acceleration response as the damage index can accurately spot the damage location and the extent, thus can verify the reliability and the effectiveness of the method presented in the paper. Acknowledgements This work was financially supported by the National Natural Science Foundation of China (Grant#50978009 and #51008102), Beijing Municipal Commission of Education Foundation (KZ200910005002). References.