A multi-prototype structure constrained class-incremental method for Compound Fault Diagnosis of rolling bearings

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Rolling bearing Compound Fault Diagnosis (CFD) is constrained by several factors in industrial environments, including the staged emergence of fault classes, limited sample, and boundary ambiguity among highly correlated fault classes. Traditional diagnostic models are poorly aligned with continuous class-incremental monitoring. To address these problems, a structured multi-prototype diagnosis method is proposed within a class-incremental learning (CIL) framework. Multi-prototype representations are constructed in the embedding space to model the intra-class distribution of compound fault categories. Structural constraints on intra-class compactness, inter-class separation, and historical prototype stability are imposed during prototype updating, limiting old class representation drift within the cumulative class space and refining separability between compound faults and their associated single fault categories. Experimental results on the public and Lanzhou University of Technology (LUT) datasets show that the proposed method performs consistently in continuous class-incremental scenarios. The proposed method offers a practical framework for CFD of rolling bearings in complex industrial environments.