Infrared and Laser Engineering, Volume. 51, Issue 3, 20210227(2022)
Research on vibrothermography detection and recognition method of metal fatigue cracks based on CNN
Figures & Tables(15)
Fig. 1. Schematic diagram of ultrasonic infrared thermal imaging detection system
Fig. 6. Structure diagram of convolutional neural network designed in this article
Table 1. Crack and optical measurement length of 15 kinds of metal specimens
View tableView in ArticleTable 1. Crack and optical measurement length of 15 kinds of metal specimens
Number of test piece Crack length/μm 01 5374.71 02 5477.40 03 5624.33 04 6570.00 05 6629.00 06 7275.00 07 7507.79 08 7930.00 09 8537.50 10 9143.00 11 9301.36 12 9453.00 13 3474.50 14 3898.49 15 0
Table 2. Description of network model parameters designed in this article
View tableView in ArticleTable 2. Description of network model parameters designed in this article
Layer Description Layer Description input 224×224×3, images with "zerocenter" normallization conv_3 32 3×3×16 convolutions with stride[1 1] and padding[1 1 1 1] conv_1 8 5×5×3 convolutions with stride [1 1] and padding[0 0 0 0] relu_3 Relu relu_1 ReLu maxpool_3 2×2 max pooling with stride [2 2] and padding[0 0 0 0] crossnorm_1 Cross channel normaillization with 5 channels per element fc_1 512 fully connected layer maxpool_1 2×2 max pooling with stride [2 2] and padding [0 0 0 0] relu_4 ReLU conv_2 16 3×3×8 convolutions with stride [1 1] and padding[2 2 2 2] dropout 50% dropout relu_2 Relu fc_2 15 fully connected layer crossnorm_2 Cross channel normaillization with 5 channels per element Softmax Softmax maxpool_2 2×2 max pooling with stride [2 2] and padding "same" classoutput crossentropyex
Table 3. Results of different batch size recognition rate
View tableView in ArticleTable 3. Results of different batch size recognition rate
Batch size Accuracy Time/s 32 99.3% 296 64 100% 206 128 95.4% 188
Table 4. Metal plate specimen and optical measurement of crack size
View tableView in ArticleTable 4. Metal plate specimen and optical measurement of crack size
Number of test piece Crack length/μm Number of test piece Crack length/μm A 9453.00 F 6577.41 B 9301.36 G 6629.00 C 9143.00 H 6740.50 D 8537.50 I 6983.00 E 8014.54 J 7275.00
Table 5. Crack recognition and classification by different algorithms
View tableView in ArticleTable 5. Crack recognition and classification by different algorithms
Algorithm Accuracy Time/s CNN designed in this article 100% 206 Alexnet 99.6% 236 Googlenet 98.9% 326 SVM 95.3% 1154
Tools
Get Citation
Copy Citation Text
Li Lin, Xin Liu, Junzhen Zhu, Fuzhou Feng. Research on vibrothermography detection and recognition method of metal fatigue cracks based on CNN[J]. Infrared and Laser Engineering, 2022, 51(3): 20210227
Paper Information
Category: Image processing
Received: Apr. 6, 2021
Accepted: --
Published Online: Apr. 8, 2022
The Author Email: Feng Fuzhou (fengfuzhou@tsinghua.org.cn)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20