The number and length of cracks in the friction lining is one of the c ore evaluation criteria to measure the braking performance of high-speed railwa y s.Effective crack detection is of great significance to the safe operation of h igh-speed railways.This paper proposes an improved algorithm based on the back b one network architecture of CSPDarkNet53to realize online automatic detection o f friction plate cracks.Firstly,the dual-path feature extraction network is f u sed to enhance the sensitivity to crack feature detection and effectively improv e the accuracy of friction plate crack detection; Secondly,in the de-redundanc y calculation of prediction box of YOLO detection module,the weighted fusion alg orithm of target box (WBF) is used to reduce the false detection rate.The exper imental results show that compared with the current most representative types of target detection algorithms,the accuracy of the method used in this paper is s ignificantly improved,and the average accuracy is increased by 7.64%.