Learnable-embedding Waveletformer: signal process and data dual-driven interpretable network for rotating machinery fault diagnosis

Abstract Deep learning networks have achieved significant success in the bearing and gear fault diagnosis. However, most of the existing fault diagnosis methods based on deep learning are purely data-driven models, which leads to high dependence on the quantity and quality of the dataset and poor physical interpretability. To tackle the issues, a signal process and data dual-driven interpretable network, termed learnable-embedding Waveletformer (LEWF) is proposed for rotating machinery fault diagnosis. In LEWF, a learnable-embedding layer with adaptive filter function is designed, where the filtering weights in the neural operator are trained adaptively as part of the network to process the vibration signal without expert expertise. The fault feature components are enhanced and irrelevant interference components are suppress adaptively based on frequency distribution. In parallel, the Waveletformer encoder incorporating a multi-head wavelet attention mechanism is introduced to capture impulse fault features across multiple scales, improving the feature extraction performance of the Transformer encoder. Experimental results on bearing and gearbox fault datasets show that the proposed method achieves high classification accuracy under fewer training samples and performs robustness to the noise. The weight visualizations are supplied to interpret the mechanism of the signal process technique and wavelet attention mechanism in the network. Furthermore, the signal process technique embedding leads to a simplified network structure and presents a substantial reduction in parameters and training time.