Railway switch fault diagnosis based on Multi-heads Channel Self Attention, Residual Connection and Deep CNN
Xirui Chen, Hui Liu, Zhu Duan
Abstract
Abstract A novel switch diagnosis method based on self-attention and residual deep convolutional neural networks (CNNs) is proposed. Because of the imbalanced dataset, the K-means synthetic minority oversampling technique (SMOTE) is applied to balancing the dataset at first. Then, the deep CNN is utilized to extract local features from long power curves, and the residual connection is performed to handle the performance degeneration. In the end, the multi-heads channel self attention focuses on those important local features. The ablation and comparison experiments are applied to verifying the effectiveness of the proposed methods. With the residual connection and multi-heads channel self attention, the proposed method has achieved an impressive accuracy of 99.83%. The t-SNE based visualizations for features of the middle layers enhance the trustworthiness.