Rail Defect Recognition Based on Waveform Subtraction and Rule Base

Abstract

Internal defects in rail may affect the safety of the train travel and deteriorate the infrastructure of rail transit. However, the existing manual defect detection mode has difficulty coping with the increasing mileage and defect detection data. To accomplish the comprehensive detection of internal defects in rail, we propose a waveform subtraction recognition method based on rail defect features in B-scan images. First, rail structure waveforms are detected based on the You Only Look Once (YOLO) v4 network. Then, after removing normal structure waveforms, abnormal waveforms of defect are located. Last, according to the law of B-scan imaging and relationships with structure waveforms, abnormal waveforms are screened and classified into specific types of defect by the rule base. The detection method was tested on a real-world data set, and the test results showed that the accuracy of normal waveform detection was 0.871 and the accuracy of defect detection was 0.755. Furthermore, previous defect detection methods were compared, and results showed that the proposed method is feasible for the comprehensive detection of rail defects.