Interpreting network knowledge with attention mechanism for bearing fault diagnosis

• Published in: Applied soft computing Vol. 97; p. 106829
• Main Authors: Yang, Zhi-bo, Zhang, Jun-peng, Zhao, Zhi-bin, Zhai, Zhi, Chen, Xue-feng
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01.12.2020
• Subjects: Attention mechanism; Bearing fault diagnosis; Interpretability; Bearing fault diagnosis; Attention mechanism; Interpretability
• ISSN: 1568-4946; 1872-9681
• DOI: 10.1016/j.asoc.2020.106829

Condition monitoring and fault diagnosis of bearings play important roles in production safety and limiting the cost of maintenance on a reasonable level. Nowadays, artificial intelligence and machine learning make fault diagnosis gradually become intelligent, and data-driven intelligent algorithms are receiving more and more attention. However, many methods use the existing deep learning models directly for the analysis of mechanical vibration signals, which is still lack of interpretability to researchers. In this paper, a method based on multilayer bidirectional gated recurrent units with attention mechanism is proposed to access the interpretability of neural networks in fault diagnosis, which combines the convolution neural network, gated recurrent unit, and the attention mechanism. Based on the attention mechanism, the attention distribution of input segments is visualized and thus the interpretability of neural networks can be further presented. Experimental validations and comparisons are conducted on bearings. The results present that the proposed model is effective for localizing the discriminative information from the input data, which provides a tool for better understanding the feature extraction process in neural networks, especially for mechanical vibration signals. •The attention mechanism is introduced in bearing fault diagnosis.•A deep learning-based bearing fault diagnosis model via attention is proposed.•The results of the model have clearer interpretability compared with other models.•Experiments on bearing datasets verified the effectiveness of the model.