Bolt Detection of Key Component for High-speed Trains Based on Deep Learning

Abstract

Safety inspection is a crucial task for high-speed trains. Conventional methods based on computer vision are insufficient in achieving this task by using template matching, the main reason is that it is difficult to extract image features, especially for the images of high-speed trains whose image edges are difficult to detect. In this paper, deep learning is introduced to key component detection. Deep learning method is not suitable when samples are rare, and as for our task, there are few abnormal samples because key component rarely become abnormal. However, abnormal detection can be transformed to the problem of bolt detection; it will be abnormal if bolts are lost. From this paper, Faster R-CNN and YOLO are adopted to detect bolts. Considering the purpose of the abnormal detection of bolts is to confirm whether the bolts are lost, and the location information of bolts contained in the extracted information of target detection is redundant. Therefore, a bolt number detection network based on ResNet is proposed to examine how many bolts on the key component. The number of bolts determine whether the bolts are lost. It shows that, this bolt number detection network achieves the task well.