Physics-aware dynamic spectral modeling integrated with weakly supervised few-shot learning for planetary gearbox fault diagnosis
Peng Chen, Jia Gao, Yaqiang Jin, Chengning Zhou, Junyu Qi, Changbo He
Abstract
Abstract Intelligent fault diagnosis based on deep learning has shown promising results in industrial applications, yet the requirement for large labeled datasets remains a significant limitation in real-world deployments. This paper proposes a novel physics-aware dynamic spectral modeling integrated with weakly supervised few-shot learning (PADSM-WSFL) framework for fault diagnosis in planetary gearboxes. The key innovations include: (1) integration of physics-based modeling with deep learning to enhance feature extraction, (2) a unique combination of weakly supervised and few-shot learning that effectively utilizes abundant unlabeled data while requiring only extremely limited labeled samples, and (3) a graph-based feature extraction module that captures complex fault patterns. The framework consists of three main components: a physics-aware dynamic spectral modeling approach, a graph construct module for feature extraction, and the integration of weakly supervised learning with few-shot learning models. Experimental validation on two machinery fault diagnosis datasets demonstrates that PADSM-WSFL achieves superior robustness and generalization capabilities compared to state-of-the-art methods, providing an effective solution to the critical challenge of limited labeled data in industrial fault diagnosis.