Laser & Optoelectronics Progress, Volume. 58, Issue 22, 2228008(2021)
Laser Ultrasonic Surface Defect Recognition Based on Optimized BP Neural Network
Figures & Tables(17)
Fig. 2. Flow chart of defect depth quantitative identification model based on PSO-BP neural network
Fig. 5. Simulation results of ultrasonic field distributions at different moments of ultrasonic waves excited by laser in aluminum material without surface defects at different moments. (a) t=0.2 μs; (b) t=0.8 μs; (c) t=1.2 μs; (d) t=1.8 μs
Fig. 6. Simulation results of ultrasonic wave field distribution at different moments of ultrasonic wave excited by laser in aluminum material with surface defects. (a) t=0.2 μs; (b) t=0.8 μs; (c) t=2.2 μs; (d) t=2.5 μs
Fig. 7. Transmission time domain signals of different depth defect response. (a) Depth is 0.1 mm; (b) depth is 0.3 mm; (c) depth is 0.5 mm; (d) depth is 0.7 mm
Fig. 8. Transmission frequency domain signals of different height defect response. (a) Depth is 0.1 mm; (b) depth is 0.3 mm; (c) depth is 0.5 mm; (d) depth is 0.7 mm
Fig. 9. Fitness change curve and convergence curves of neural network model. (a) Fitness curve; (b) convergence curves
Fig. 10. Regression results R of BP neural network model. (a) Training set; (b) validation set; (c) test set; (d) total
Fig. 11. Regression results R of PSO-BP neural network model. (a) Training set; (b) validation set; (c) test set; (d) total
Table 1. Thermodynamic parameters of aluminum material in model
View tableTable 1. Thermodynamic parameters of aluminum material in model
Parameter Unit Value Specific heat capacity J·kg-1·K-1 900 Thermal conductivity W·m-1·K-1 238 Thermal expansion coefficient K-1 2.3×10-5 Density kg·m-3 2700 Young’s modulus Pa 7×1010 Poisson’s ratio - 0.33
Table 2. Feature vectors for different defect depths (part)
View tableTable 2. Feature vectors for different defect depths (part)
Defect depth /mm Time domain peak Center frequency 3 dB bandwidth fH fL 0.1 -2.146560 2.21960 3.9740 0.2326 4.2066 0.2 -1.740667 1.64385 3.0609 0.2234 3.0643 0.3 -1.358564 0.66580 1.5584 0.2166 1.1150 0.4 -1.010040 0.84985 0.2665 0.2166 1.4831 0.5 -0.782374 0.48620 0.5574 0.2075 0.7649 0.6 -0.645267 0.46910 0.5574 0.1904 0.7478 0.7 -0.547497 0.45940 0.5825 0.2736 0.8561 0.8 -0.495937 0.44630 0.5506 0.1710 0.7216 0.9 -0.473476 0.43375 0.5461 0.1607 0.7068 1.0 -0.460262 0.40750 0.5450 0.1150 0.7000
Table 3. Test results of BP neural network model and PSO-BP neural network model
View tableTable 3. Test results of BP neural network model and PSO-BP neural network model
Actual depth /mm 0.1 0.3 0.5 0.7 1 Predicted depth of BP neural network 0.1134 0.2516 0.5597 0.5915 0.9089 Absolute error 0.0134 -0.0484 0.0597 -0.1085 0.0911 Predicted depth of PSO-BP 0.0971 0.2901 0.4566 0.6620 0.9471 Absolute error -0.0029 -0.0099 -0.0434 -0.0380 -0.0529
Table 4. Test results for different defect depth intervals
View tableTable 4. Test results for different defect depth intervals
Defect depth /mm 0.1 0.3 0.5 0.7 1.0 Relative error /% 2.86 3.30 8.69 5.11 5.29 Defect depth /mm 2.6 2.7 2.8 2.9 3.0 Relative error /% 10.00 7.67 4.68 7.95 9.60
Table 5. Comparison of relative depth errors of different algorithms%
View tableTable 5. Comparison of relative depth errors of different algorithms%
Algorithm BP neural network PSO-BP RBF SVM Depth is 0.1 mm 13.40 2.86 11.60 8.32 Depth is 0.3 mm 16.10 3.30 8.30 7.89 Depth is 0.5 mm 11.90 8.69 10.50 9.55 Depth is 0.7 mm 15.50 5.11 9.76 10.08 Depth is 1.0 mm 9.11 5.29 11.70 10.63 Average error 13.20 5.05 10.37 9.24
Tools
Get Citation
Copy Citation Text
Chao Chen, Xingyuan Zhang, Siye Lu. Laser Ultrasonic Surface Defect Recognition Based on Optimized BP Neural Network[J]. Laser & Optoelectronics Progress, 2021, 58(22): 2228008
Paper Information
Category: Remote Sensing and Sensors
Received: Jan. 8, 2021
Accepted: Feb. 4, 2021
Published Online: Nov. 10, 2021
The Author Email: Xingyuan Zhang (zhyy_yuan@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20