A Fastener Inspection Method Based on Defective Sample Generation and Deep Convolutional Neural Network

For the safety of railways, well-trained workers are required to check the fastener constantly, which shows the disadvantage of large time cost, huge labor cost and might being dangerous to workers. To address this and achieve automatic detection, an inspection model based on deep convolutional neural network (DCNN) is adopted in this paper. However, the inspection model suffering from the unbalanced training samples of defective vs normal due to defective fasteners are far less than normal fasteners in real railways. To tackle this problem, a novel sample generation method is proposed to generate defective fastener samples using the normal fasteners to realize sample augmentation. The comprehensive experiments are conducted on the collected real fastener samples and generated samples. The experimental results show that our method has good performance for fastener inspection on unbalanced samples and outperforms other state-of-the-art methods.