Acoustic emission wave classification for rail crack monitoring based on synchrosqueezed wavelet transform and multi-branch convolutional neural network

This study focuses on the acoustic emission wave classification for the sake of more accurate and comprehensive rail crack monitoring in the field typically with complex cracking conditions, high-operational noise, and mass data. There are mainly three types of acoustic emission waves induced by operational noise, impact, and crack propagation, respectively. Synchrosqueezed wavelet transform was introduced to represent intrinsic characteristics of acoustic emission waves more clearly in the time-frequency domain, where acoustic emission waves induced by different mechanisms were found to show various patterns of energy distribution. Then, a multi-branch convolutional neural network model with two branches was developed to automatically classify the three types of acoustic emission waves by taking into account their synchrosqueezed wavelet transform plots in various time-frequency scales. Training, validation, and test data sets were constructed using acoustic emission waves collected through a series of field and laboratory tests with various noise levels and loading conditions. The transfer learning was used to train the model faster, and the Bayesian optimization algorithm was applied to tune the hyperparameters. Finally, the multi-branch convolutional neural network model achieved higher accuracy and robustness than the traditional convolutional neural network model of single branch in identifying different acoustic emission mechanisms. The proposed acoustic emission wave classification method based on synchrosqueezed wavelet transform and multi-branch convolutional neural network is able to detect not only surface rail cracks, where both impact-induced and crack propagation-induced acoustic emission waves would be identified, but also internal rail cracks where only crack propagation-induced acoustic emission waves would be captured.

本研究聚焦于声发射波分类，以便在现场（通常具有复杂的开裂条件、高运行噪声和大量数据）更准确和全面地监测钢轨裂纹。主要有三种分别由运行噪声、冲击和裂纹扩展诱发的声发射波。引入同步挤压小波变换以在时频域更清晰地呈现声发射波的内在特征，发现由不同机制诱发的声发射波呈现出各种能量分布模式。然后，开发了一个具有两个分支的多分支卷积神经网络模型，通过考虑它们在不同时频尺度下的同步挤压小波变换图来自动对这三种声发射波进行分类。利用通过一系列具有不同噪声水平和加载条件的现场和实验室测试所收集的声发射波构建训练集、验证集和测试集。使用迁移学习来更快地训练模型，并应用贝叶斯优化算法来调整超参数。最后，多分支卷积神经网络模型在识别不同声发射机制方面比单分支的传统卷积神经网络模型具有更高的准确性和鲁棒性。所提出的基于同步挤压小波变换和多分支卷积神经网络的声发射波分类方法不仅能够检测钢轨表面裂纹（在此处冲击诱发的和声发射波以及裂纹扩展诱发的声发射波都能被识别），还能检测钢轨内部裂纹（在此处只有裂纹扩展诱发的声发射波会被捕捉）。