High Speed Railway Bolt Detection Based on Deep Learning
Download as PDF
With the development of railway network, the safety maintenance and guarantee of transportation infrastructure is becoming more and more important. Under the high frequency vibration of train running, the bolt is easy to produce defects. At present, railway works departments collect track structure images by means of comprehensive inspection vehicles and manual trolley, but they rely on manual playback to find structural defects due to their limited ability of automatic identification of defects. Considering the Similarity of the inner and outer defective-free bolts of each rail platform, this work proposed a fast location algorithm of bolts based on Structural Similarity (SSIM). The Bidirectional Generative Adversarial Network (BiGAN) is selected and trained to identify the defect bolts, and is applied to 4.25 GB images of 25 km lines. The results show that the method can effectively identify the defective bolts, the recall rates of missing bolt and stampeding plate are 98.62% and 97.93% respectively, the recall rate of defective samples is 93.45%, the accuracy rate is 94.20%, and the misjudgment rate of normal samples is 5.13%.
Deep learning; Bolt defect detection; Adversarial generative network