A Partial Domain Adaptation Scheme Based on Dual-Weighted Adversarial Network for Bearing Fault Diagnosis

Domain adaptation techniques have achieved significant results in fault diagnosis, but their performance often depends on the assumption that the source and target domains have the same label set. This study focuses on a more realistic diagnostic scenario, where the label set of the target domain is only a subset of the source domain label set, which is more common in practical industrial applications. In this paper, a dual-weighted adversarial adaptation network for bearing fault diagnosis is proposed. This method introduces a class-level weight evaluation strategy for source domain samples that quantifies the uncertainty of each category, dynamically adjusting class weights to ensure the model focuses more on shared categories. Furthermore, a sample-level weight evaluation mechanism is introduced, which evaluates the transferability for target domain samples through an evaluation model, mitigating the negative impact of samples with low-quality during the current domain adaptation training stage. The effectiveness of the method is verified by experiments on the bearing fault dataset of Soochow University.