A Low-Speed Heavy Load Bearing Fault Diagnosis Method Based on Interpretable Few-Shot Learning

Low-speed heavy-load bearings are widely used in engineering; however, their low-speed characteristics make fault feature extraction challenging, and reproducible fault samples are scarce. Furthermore, the "black-box" nature of deep learning models presents a significant challenge in constructing high-accuracy and interpretable fault diagnosis models when sample data is limited. To address these issues, this paper proposes a lightweight fault diagnosis method for low-speed, heavy-load bearings based on a few-shot learning strategy and interpretability. First, an efficient few-shot learning model, SATNet, is constructed based on the N-shot K-way few-shot learning strategy, incorporating a self-attention mechanism and optimized specifically for the complex fault features of low-speed, heavy-load bearings. Then, the Teager energy operator-based signal enhancement algorithm is applied to preprocess the acoustic emission and vibration signals of the low-speed, heavy-load bearings in both normal and fault conditions. This step extracts the signals' nonlinear and energy features and obtains corresponding time-domain features. The preprocessed acoustic emission and vibration signals are input into the SATNet model for training and testing, accurately classifying faults under different operating conditions. Finally, the SHAP and LIME algorithms are employed to conduct an interpretability analysis of the SATNet model's diagnostic results, thoroughly exploring and revealing the key time-domain features that influence classification decisions and their importance. Experimental results show that the proposed innovative diagnostic method significantly improves both diagnostic accuracy and stability for low-speed, heavy-load bearing fault diagnosis.