Pseudo-label assisted contrastive learning model for unsupervised open-set domain adaptation in fault diagnosis

Abstract

The operation of mechanical equipment is frequently characterized by complexity and variability, leading to signal domain shifts. This phenomenon underscores the significance of cross-domain fault diagnosis for maintaining the reliability and safety of mechanical systems. Due to the absence of labeled data in many operational contexts, there's a clear need for an unsupervised domain adaptation technique that does not rely on labeled information. Moreover, traditional domain adaptation methods presuppose identical label distributions across source and target domains. Nevertheless, real-world engineering scenarios often present novel fault categories out of distribution, thereby challenging the efficacy of established domain adaption methods. To address these challenges, we proposed a pseudo-label assisted contrastive learning model (PLA-CLM) for Unsupervised Open-set Domain Adaptation. Based on contrastive learning, the proposed model effectively minimizes the discrepancy between samples of identical pseudo-label across domains, while simultaneously integrating distance, density, and entropy to isolate out-of-distribution samples. After training, the model adaptively identifies known faults and detects OOD faults using thresholds calculated based on sample distribution. Experimental results on two datasets demonstrate that our method surpasses existing approaches, ensuring enhanced reliability of mechanical systems’ operation and maintenance.