An uncertainty relevance metric-based domain adaptation fault diagnosis method to overcome class relevance caused confusion

Unsupervised domain adaptation methods have recently achieved satisfactory results in detecting mechanical faults with slight class relevance. However, in engineering, decision confusion caused by strong class relevance is ubiquitous. As a typical application neglected in most existing studies, the unsupervised domain adaptation scenario with compound faults considers interrelated and cross-influenced fault types under distribution shift, intensifying class confusion and threatening fault diagnosis reliability inevitably. To this challenge, an innovative sample-level distance metric, termed uncertainty relevance (UR), is proposed to overcome class confusion. Specifically, the metric is constructed from the class relevance matrix and uncertainty weighting to measure discrepancies between predictions, whose max–min optimization enhances discriminability and more tradeoffs on multiclass information. Combined with the metric, a novel gradual inference domain adaptation method is developed, whose backbone, termed gradual inference, consists of a multilayer extractor and multiple classifiers, structurally achieving prediction diversity. Functionally, optimizing UR among multiple classifiers enables class-level domain adaptation to reduce class confusion, simultaneously treating classifiers as domain discriminators to construct hierarchical domain adversarial reaches global-level domain adaptation. Moreover, the theoretical risk upper bound is provided by introducing Rademacher complexity. High-precision performance on extensive trials demonstrates the proposed method improves the decision reliability in mechanical fault diagnosis.