Interpretability study of gearbox fault diagnosis based on MWK-GraphSage modeling

Deep learning has exhibited remarkable accuracy in fault diagnosis. However, its opaque nature complicates the interpretation of decision-making processes, thereby constraining its dependability and practicality in industrial settings. This paper introduces an interpretable graph neural network model named Multi-Wavelet Kernel Graph Sample and Aggregate Network (MWK-GraphSage) to address three limitations in current approaches: limited modeling capacity for non-Euclidean data, insufficient disentanglement of multi-fault characteristics, and one-dimensional interpretability assessment. The proposed model features a multi-wavelet kernel convolutional layer (MWKC) that integrates four wavelet basis functions (Sine, Cauchy, Bump, and Laplace) to extract multi-scale time–frequency attributes, thereby enhancing the capability to disentangle features. An Adaptive Wavelet Weight (AWW) allocation mechanism is devised to dynamically enhance feature fusion performance. Furthermore, various visualization techniques are employed to facilitate multi-dimensional interpretability analysis. Experimental results on the Southeast University gearbox dataset demonstrate that MWK-GraphSage achieves a fault diagnosis accuracy of 99%, surpassing comparative models such as ResGAT and WaveletCNN.