Universal domain adaptation with navigator-guided strong alignment and dynamic threshold for rotating machinery fault...

Universal domain adaptation (UniDA) is a promising approach for fault diagnosis in rotating machinery where target domain fault classes are uncertain, but suffers from two key limitations: (1) Strong alignment strategies cause negative transfer by misaligning outlier classes, and (2) manual thresholding for unknown-class rejection is inefficient and lacks adaptability. To address these issues, this paper proposes NSADT, a UniDA method featuring Navigator-guided Strong Alignment and a Dynamic Threshold. To mitigate negative transfer, a navigator-guided module is designed to assess sample-wise class certainty and domain affiliation, thereby focusing alignment on known shared-classes. For adaptive unknown private-class rejection, NSADT introduces a novel dynamic threshold mechanism that leverages the discrepancy in feature stability between shared and private classes during a self-supervised reconstruction (SSR) task. Specifically, the method quantifies the feature embedding shift by comparing the Prototype Matching Embedding Distances from the initial SSR task with that from a secondary SSR task. This shift creates a bimodal distribution, allowing for the adaptive determination of an optimal rejection threshold for each class via clustering. Extensive UniDA experiments on self-built gearbox dataset and two public bearing datasets demonstrate that NSADT significantly outperforms existing methods.