Train Bearing Fault Diagnosis Based on Convolution Temporal-spatial Mutual Fusion Network

Abstract

Objective The diverse and complex operating environment of axle box bearings makes it challenging for bearing fault diagnostic methods to achieve satisfactory results with the single sensor. To address this issue, research is conducted on train bearing fault diagnosis based on convolutional temporal-spatial mutual fusion network. Method A multi-sensor fusion approach is adopted, applying the proposed CTS-MFN (a convolutional temporal-spatial mutual fusion network) for bearing fault diagnosis upon horizontal and vertical vibration datasets. The ECA module (efficient channel attention), LSTM (long short-term memory network), and similarity distance constraints are introduced to the convolutional autoencoder, enabling the model to extract temporal-spatial attention features that include inter-modal interaction information. An MLP (multilayer perceptron) is then used to fuse and infer temporal-spatial features from each modality. Result & Conclusion Comparative experiments, ablation studies, and generalization performance analysis demonstrate the effectiveness of the proposed model.

Keywords

• train
• bearing fault diagnosis
• convolutional autoencoder
• data fusion
• long short-term memory network