Rail Defect Detection Method Based on Recurrent Neural Network
Qinhua Xu, Qinjun Zhao, Gang Yu, Liguo Wang, Tao Shen
Abstract
At present, there are some problems of low efficiency, heavy workload and low accuracy in the manual detection data of rail defect detection vehicle. According to the characteristics of channel distribution, digital combination and time sequence of ultrasonic B-scan data, a method of rail defect data recognition based on recurrent neural network is established. In this method, the B-scan image of the rail defect detection vehicle is processed into a time series language, which is identified and classified by the recurrent neural network model. Taking the screw hole crack as an example, the validity of the method is verified by using the data of rail defect calibration line. This method solves the problem that the type of rail damage is complex and changeable, and the characteristics of rail defect are difficult to extract manually. Under the requirements of manual analysis, it greatly improves the efficiency of rail defect identification and reduces the workload of flaw detection personnel.