Mitigating Decision Boundary Confusion: A Classifier Prediction-Oriented Domain Adaptation Network

Domain adaptation has been extensively studied and applied in fault diagnosis of rotating machinery to address significant distribution differences between source and target domains. In industrial production, however, decision boundary confusion caused by strong connectivity among similar categories is prevalent. In response to this challenge, a classifier prediction-oriented domain adaptation network (CPDAN) is proposed in this article. Specifically, CPDAN contains two parts. Firstly, a target domain category separation algorithm (TDCSA) is designed to address the problem of decision boundary confusion. TDCSA effectively reduces intercategory correlations by minimizing the inner product between class predictions and their corresponding transposes within the target domain. It enhances feature discriminability, thereby facilitating category separation. Meanwhile, a similar uncertainty mechanism (SUM) is developed to highlight samples with high contributions to category separation. Secondly, to enhance the application effectiveness of TDCSA, a class feature distribution alignment algorithm (CFDAA) oriented to classifier prediction feature reinforcement mechanism (CPFRM) is proposed. CFDAA augments the generated features based on classifier-predicted class information and embeds these new features into a distance regularization term to guide class-level distribution alignment. This effectively captures feature transferability, further strengthening interdomain similarity. Extensive experiments on two bearing datasets have demonstrated the effectiveness and superiority of CPDAN.