Rail Fatigue Crack Classification Based on the CNN-BiTCN-CA Method
摘要
To improve the accuracy of rail fatigue crack signal classification in high-speed railways and reduce the poor noise resistance of traditional methods in complex railway environments, a rail fatigue crack classification method based on feature cross-attention (CA) fusion is proposed, with the CNN-BiTCN-CA classification model established. Convolutional neural network (CNN) and bidirectional temporal convolutional network (BiTCN) are used to extract time-frequency features respectively, and the CA is employed to fuse these time-frequency features, thereby fully extracting features from the original acoustic emission signals. During the model training process, an additional Dropout layer is added between BiTCN and the fully connected layer to suppress the overfitting behavior of the model; L2 weight decay is introduced to constrain the parameter scale. The fully connected layer is utilized to achieve precise classification of rail fatigue crack stages. Experimental studies show that the classification accuracy of rail fatigue cracks using the CNN-BiTCN-CA model is 99.83%, which shows varying degrees of improvement compared to other classification models. The CNN-BiTCN-CA model can deeply extract features from signals and effectively improve the accuracy of rail fatigue crack classification.