A Bearing Fault Diagnosis Method Based on Adaptive Graph Structure Sparse Coding and Bayesian Probabilistic Neural Network

Abstract

Modern complex industrial systems have put forward unprecedented requirements for accuracy and reliability in the early fault identification and precise diagnosis of critical rotating machinery. Although current mainstream diagnostic methods have enhanced recognition performance with deep learning models, they are generally trapped in the "representation bottleneck" and "black-box dilemma". On the one hand, end-to-end learning neglects the explicit modeling of fault physical mechanisms, leading to insufficient interpretable structural priors for feature representation. On the other hand, deterministic models cannot quantify prediction uncertainty, which limits their application in supporting trustworthy decision-making for high-risk scenarios. This study aims to construct a novel diagnostic paradigm integrating high discriminability, strong interpretability, and reliable uncertainty quantification. The core contributions are summarized as follows. First, an Adaptive Graph Structure Sparse Coding (AGSC) method is proposed. It innovatively embeds multi-domain features of vibration signals into a dynamic graph structure space. By jointly optimizing feature representation and graph topology, the non-Euclidean manifold correlations among features are explicitly modeled, realizing discriminative representation learning that fuses physical mechanism inspiration and data-driven learning. Second, a Bayesian Probabilistic Neural Network (BPNN) is developed. Through constructing a hierarchical probabilistic model of kernel widths and a variational inference framework, the empirical parameters of traditional probabilistic neural networks are converted into learnable probability distributions. Thereby, confidence estimation can be synchronously output in classification decisions, which enables the diagnostic results to achieve uncertainty awareness.