A Fuzzy Confusion Matrix-Based Self-Supervised Learning Method to Mitigate Class Confusion for Partial Transfer Fault Diagnosis

Domain adaptation (DA) has significantly advanced the field of fault diagnosis. Partial DA (PDA), which addresses situations where the label space of the target domain is a subset of that in the source domain, has garnered significant interest. However, existing PDA methods insufficiently address class confusion in the presence of severe domain asymmetry. In such cases, the challenge of crossing distributional discrepancies increases, leading to a higher number of samples at classification boundaries and consequently hindering the DA process. This article proposes a novel self-supervised learning method based on a fuzzy confusion matrix (FCM) for partial transfer fault diagnosis to tackle this issue. This approach effectively reduces class confusion and enhances the precise matching of feature distributions between the two domains. Specifically, the FCM is introduced to quantify the certainty and uncertainty of classification decisions, derived exclusively from the classifier's output probabilities rather than true labels. Leveraging the FCM, a batch confusion penalization is formulated to mitigate class confusion within self-supervised learning. Additionally, a class-level weighting mechanism is designed to suppress the negative transfer induced by source private classes. Comprehensive cross-operating-condition and cross-component experiments are conducted to demonstrate the effectiveness and superiority of the proposed method.