Feature Self-Enhancement Compound Dimension Interpretable Wavelet Network and Its Application in Rotating Component Fault Diagnosis Under Varying Speed

For rotating components’ fault diagnosis, the existing interpretable networks mostly focus on rotors at constant speeds, neglecting interpretable networks for fault diagnosis in rotors with variable speeds. In industrial settings, rotors mostly operate at variable speeds, leading to challenges in identifying rotor health due to spectral aliasing in faulty rotors under variable speeds. This article proposes a hybrid 1-D and 2-D neural network incorporating a feature enhancement layer and wavelet transformation feature self-enhanced compound dimension wavelet network (FSCDWN). FSCDWN transforms 1-D feature maps generated by wavelet convolution kernels into 2-D maps, facilitating better analysis of interband relationships and easier feature extraction from variable frequency signals. By enhancing fault features before the wavelet convolution layer, downstream networks can more easily extract fault features. Through experiments, the convergence speed of FSCDWN is significantly higher than that of the control group, with accuracy mostly around 90%, markedly surpassing the control group’s 80% and 60%. Based on the principles of signal processing, the feature maps of the network are visualized. FSCDWN enhances and retains the fault characteristics in the variable speed signals, while the significance of the feature maps of the control group network is unclear. This demonstrates the interpretability of FSCDWN and explains the reason for its high accuracy.