Research on YOLO Model and Its Application in Fault Status Recognition of Freight Trains

Abstract

Compared with other object detection and recognition methods, YOLO is an end-to-end algorithm that integrates target region prediction and target category prediction into a single neural network model. It can realize rapid target detection and recognition with high accuracy, which is more suitable for field application. The network models of YOLOV3 and YOLOV4 are studied, aiming at the application environment of fault state detection in railway transportation and freight. In view of the shortcomings of YOLOV3 algorithm, the Dense block-BC is used to replace Darknet-53 to build the YOLOV3 backbone network. CIoU is used to replace the MSE to calculate the boundary box loss, and Focal loss is used to redesign the loss function to improve the YOLOV3 model. The improved YOLOV3 and YOLOV4 are respectively used to detect the fault state in railway locomotive transportation, and the better detection results are obtained.16,660 pictures are collected directly from the railway transportation site as train set, and other 3,000 pictures are used as test set. The omission factor can be reduced to 4.164%, and the error rate can be decreased to 4.218%.The detection speed of each picture is 0.4s, which can meet the requirements of non-stopping fault state detection in railway transportation.