Surface Crack Defect Detection Of High-speed Rail Friction Pad Based On Deep Learning Study

In recent years,high-speed rail has become a popular transportation tool.As the core part of high-speed rail friction plate,timely and effective detection of surface defects is of great significance to driving safety.This paper examines whether there are defects in the surface cracks of high-speed rail friction plates.The main research contents are as follows:(1)In this paper,a data set of surface cracks on high-speed rail friction plates is established.Based on 2,040 high-speed rail friction plate images collected by industrial CCD camera with high resolution,the characteristics and defect categories are analyzed and the data enhancement operation is conducted.Annotate the expanded image to complete the production of the final data set.(2)Aiming at the problem of low efficiency and accuracy in defect detection of surface cracks on high-speed rail friction plates,this paper constructed a YOLO-DUAL defect detection model for surface cracks on high-speed rail friction plates based on YOLO V4 network.On the basis of YOLO v4’s CSPDark Net 53 backbone feature extraction network,an auxiliary feature extraction network composed of a combination of multi-layer pooling and convolutional layers is added.Based on the weighted fusion algorithm of the target frame,the size of the prior frame is recalculated and optimized,and the K-means++ algorithm is selected to cluster the hyperparameters of the prior frame to adapt the morphological characteristics of the target at different scales.The m AP value of the improved YOLO-DUAL network has increased from 60.17% of the YOLO v4 network to 61.58%,and the m Io U has increased from 52.32% to 58.68%,which is an increase of 6.36%.Therefore,the improved YOLO-DUAL network structure can effectively improve the detection performance of surface crack defects of high-speed rail friction plates.(3)This paper carries out lightweight processing on the YOLO-DUAL model,and a lightweight YOLO-MDUAL defect detection model is proposed,and a crack detection system for mobile high-speed rail friction plate crack detection is designed and implemented.The first 16 layers of the Mobile Net v3 network is selected to replace the CSPDark Net 53 backbone feature extraction network in YOLO-DUAL,and the new module is added to the feature extraction stage of the YOLO-MDUAL model.Then,based on the lightweight development framework of wx-tools,the weights and parameters of the trained YOLO-MDUAL model are transplanted and integrated,and the mobile-end high-speed rail friction plate crack detection system is realized.Compared with the YOLO-DUAL model,the improved model reduces the number of parameters from 63 million to 32 million.At the same time,the real-time detection speed of the network is increased by 2.29 times,which can finally meet the real-time application of hardware devices with lower computing power.