Acoustic Emission Feature Extraction and Classification for Rail Crack Monitoring

Abstract

Crack monitoring of rails aims to identify fatigue cracks in advance in order to ensure a safe and smooth operation of railway system. This study focuses on the rail crack monitoring using acoustic emission (AE) technique in the railway field typically with complex crack conditions and high operational noise. There are mainly three types of AE waves respectively induced by operational noise, crack propagation and impact, which need to be carefully distinguished from each other for the sake of quantitative crack monitoring. Wavelet transform (WT) was applied to represent the features of AE waves in the time-frequency domain. AE waves induced by different mechanisms were found to contain various instantaneous frequency components. A deep convolutional neural network (CNN) with transfer learning was proposed as an automatic feature extractor and classifier to evaluate the WT plots of AE waves. The CNN was trained, validated and tested using AE data collected through field and laboratory tests. The results demonstrated that the proposed methodology performed well in classifying AE waves induced by different mechanisms in the railway field.