Cross-working condition fault diagnosis via multi-classifier discrepancy-driven learning and confidence pseudo-label self-supervision

The maximum classifier discrepancy domain adaptation (DA) method based on a dual-classifier structure has achieved some success in fault diagnosis of cross-working condition, yet its reliance on the dual-classifier structure may fail to effectively represent the data characteristics of the target domain and may suffers from insufficient classification boundary discrimination. To address this, this paper extends the aforementioned structure to multiple classifiers and proposes a multi-classifier discrepancy (MuCD)-driven DA approach, which enhances the representation of target domain data characteristics, through exploiting the discrepancies among multiple classifier pairs. Additionally, given the uncertainty of multi-classifier outputs, the harmonic mean is calculated and a screening threshold is set. Combined with the self-supervised learning losses for target domain, a confidence pseudo-label strategy-guided self-supervised learning (CPSL) mechanism is designed, to guide and enhance discriminative learning among different fault classes in target domain, thereby improving the model’s diagnosis performance. Furthermore, by specifically integrating CPSL into the MuCD framework, a cross-working condition fault diagnosis method based on multi-classifier discrepancy-driven learning enhanced by confidence pseudo-label self-supervision (CPSL-MuCD) is proposed. Finally, comparative experiments and ablation studies on bearing fault, gear fault, and bearing-gear compound fault datasets verify the effectiveness and generalizability of the proposed CPSL-MuCD.