Open set transfer learning for bearing defect recognition based on selective momentum contrast and dual adversarial structure

Abstract

Data-driven rotating machinery fault diagnosis methods gradually enter people’s vision due to their fast and accuracy. However, in practical engineering applications, the unpredictability of machine failure modes and the appearance of new defect types can severely disrupt the diagnostic task with closed set defect recognition data-driven methods. Addressing this concern, this paper delves into the defect transfer diagnosis of rolling bearings in an open-set context and introduces the Momentum Contrastive Dual Adversarial (MCDA) domain adaptive method. This novel approach leverages instance-based difference comparison for unsupervised clustering and implements Negative Samples Selection Optimization (NSSO) via K Nearest Neighbor across all domains. It features intra-module adversarial learning for domain adaptation between source and target domains, and inter-module adversarial learning where encoders and feature extractors across different modules share weights. The methodology prioritizes the separation of private classes during the initial phase of contrastive learning, ensuring the isolation of target private classes. Concurrently, it aligns shared classes in the subsequent phase through semi-supervised training for domain adaptation. Our experiments on the Paderborn Dataset and Railway Freight Wheel-set Bearing (RFWB) Dataset demonstrate the effectiveness of this method.