多传感器信息融合的铁路扣件缺陷检测方法

The defect detection of railway fastener based on machine vision has replaced manual detection and improved the detection efficiency. However, the problems including few samples of defective railway fastener, the difficulty of sample labeling and the great influence of light conditions on 2D image detection are still the main obstacles in defect detection of railway fastener. Therefore, a defective railway fastener detection method combining multi-sensor information is proposed. The sample of railway fastener is quickly and efficiently collected by structural light equipment. An adaptive positioning and segmentation method based on depth map was designed to automatically and accurately locate and segment railway fastener in depth map, and the positioning and segmentation parameters were integrated to realize the positioning and segmentation of railway fastener on the intensive map. To solve the problem of unbalanced samples and low efficiency of sample labeling, DCGAN is used to expand the number of samples of defective railway fastener. In addition, a depth neural network with a total of 18 layers of 8 residual blocks was designed based on ResNet network to complete the detection of three types of defects: missing nut, broken spring strip and missing spring strip of railway fastener, and samples of WJ-7 type railway fastener were collected for experiment to verify the proposed method. The results showed that the average accuracy of defect detection of fastener was 97.6%, meeting the needs of engineering applications.