Interpretable mechanical fault diagnosis based on a multi-scale complex attention network
Abstract
Deep learning-based methods could provide an end-to-end solution that can pass over the accumulation of expert knowledge in feature engineering, which makes the combination of deep learning and fault diagnosis a hot research topic. However, these studies focus more on the way to change the shape of the acquisition data to meet the input requirement of existing deep learning methods. This practice causes three problems: blind hyper-parameter search wastes time and computational resources; altering the data shape destroys the raw time–frequency structure, making results less credible; it amounts to modeling for its own sake, yielding no improvement in our understanding of fault diagnosis. To address these problems, an intelligent diagnosis of mechanical faults method based on the Multi-Scale Complex Attention Network (MSCA-Net) is developed. In experiments, the proposed method achieves higher accuracy than several traditional deep learning methods, demonstrating its advantage. Score-CAM visualization explains the model’s decision, confirming its credibility. Furthermore, the information captured by the model proves consistent with domain expert knowledge, thereby providing a basis for advancing the understanding of unknown faults.