A joint distribution domain adaptive rolling bearing fault diagnosis model for processing noisy signals

Based on deep domain adaptation (DA) methods, which have been widely applied in cross-condition fault diagnosis, traditional DA methods often struggle to extract domaininvariant features under noisy signals and overlook the influence of the quality of pseudo-labels generated in the target domain during class alignment. Therefore, this paper proposes a Joint Distribution-based Robust Cross-domain Adaptation (JACDA) model for fault diagnosis with noisy signals. Firstly, an feature extraction module is proposed to extract transferable features from the source and target domains under noisy signals. Secondly, an adaptive threshold pseudo-label strategy is introduced to address the issue of lacking labels and low label quality in the target domain. Additionally, a discriminative discrepancy module is proposed, utilizing target domain pseudo-labels obtained through an adaptive threshold to compute MK-MMD inter-class distances with source domain data, encouraging tight clustering of features from the same category and separation of features from different categories, thereby enhancing feature discriminability. Experimental results demonstrate the model's strong transferability in cross-domain bearing fault diagnosis.