An Interpretable Self-Guided Learning Model With Knowledge Distillation for Intelligent Fault Diagnosis of Rotating Machinery

Abstract

Neural networks are widely applied in fault diagnosis of rotating machinery due to their powerful feature extraction and classification capabilities. However, their inherent black-box nature and reliance on predefined signal processing methods limit interpretability and adaptability in complex industrial scenarios. Knowledge distillation (KD) offers an effective approach to transfer knowledge from complex models to lightweight models while preserving the original performance of the model, but KD highly requires pretrained complex models. This article proposed a self-guided learning model (SGLM) that integrates adaptive feature extraction with knowledge transfer mechanisms, achieving both high diagnostic accuracy and physical interpretability. Specifically, the proposed SGLM employs learnable wavelet kernel functions to dynamically decompose raw vibration signals into multilevel subbands, adaptively capturing critical features for fault diagnosis. Further, the proposed SGLM eliminates dependence on external complex models by partitioning the network into hierarchical subsections, where knowledge from deeper layers can guide shallow layers. Experimental results on two datasets demonstrate the superior performance of SGLM, achieving 99.50% accuracy on the bearing dataset and 99.67% accuracy on the planetary gearbox dataset. The interpretability of SGLM is proven through three interpretability mechanisms. Meanwhile, SGLM’s effectiveness and practicality are validated via ablation, cross-validation, and efficiency analysis.