Granularity knowledge-sharing supervised contrastive learning framework for long-tailed fault diagnosis of rotating machinery

Abstract

The long-tailed distribution of monitoring data poses challenges for deep learning-based fault diagnosis (FD). Recent efforts utilizing supervised contrastive learning (SCL) and reweighted loss have made progress, but have overlooked two key issues: 1) prevailing random undersampling introduces sample influence bias and suboptimal model learning; and 2) focusing only on improving the FD average accuracy compromises fundamental fault judgement (FJ), heightening missed-detective and false-alarm risks unsuitable for real-world deployment. To fill these research gaps, this paper proposes a granularity knowledge-sharing SCL (GKSSCL) framework for long-tailed FD, encompassing GKS supervised contrasting and GKS classification stages. In the former, normal data are clustered into multiple fine-grained subclasses that are similar in size to the fault categories for balanced contrasting. Moreover, a mixed-granularity contrastive loss facilitates knowledge sharing across granularities. In the latter, FJ and FD tasks were concurrently trained through a knowledge graph-based adaptive sharing strategy. Experiments on two fault datasets showed that the GKSSCL can effectively harness all normal data, eliminate sample influence bias, and enhance FD precision without sacrificing FJ reliability.