More will be better: Multi-source-free aggregation adaptation with confidence calibration for fault diagnosis

Considering storage, transmission, privacy, and other limitations of massive monitoring data, source-free adaptation diagnosis techniques have been increasingly required. Although some ingenious methods are presented, two issues remain unexplored. Firstly, the existing methods focus on single-source domain settings and ignore the universal and realistic multi-domain scenarios. Second, existing approaches only evaluate performance with accuracy and lack confidence calibration, resulting in incredible decisions. In light of the above issues, a novel diagnosis method named Multi-Source free Aggregation Adaptation with Confidence Calibration (MSA₂C₂) is proposed in this work. To begin with, a dynamic weight learning strategy is developed to aggregate multi-source domain knowledge. Subsequently, a confident self-training approach is presented to guide the target model learning, employing a combination of dual pseudo-label assignment and data augmentation strategies to maximize the utility of unlabeled target data. Furthermore, a joint feature alignment technique is further incorporated to address target intra-domain distribution shifts. Finally, an unsupervised post-hoc confidence calibration mechanism is developed for more credible diagnosis results. Extensive experiments on three diverse datasets demonstrate the effectiveness and superiority of our method in terms of both diagnosis accuracy and confidence calibration performance.