A class confusion-aware spherical network approach for closed-set domain adaptation in rotating machinery fault diagnosis

Abstract Closed-set domain adaptation is an effective approach to addressing the challenge of significant distribution discrepancies between source and target domains while maintaining consistent class categories in rotating machinery fault diagnosis. However, conventional methods often overlook inter-class confusion within the target domain, leading to ambiguous decision boundaries and limiting adaptation performance. To address these challenges, this paper proposes a class confusion-aware spherical domain adaptation network (CCASDAN), which integrates adversarial training and self-supervised class confusion learning to enhance cross-domain adaptation for fault diagnosis. Specifically, a spherical classifier and discriminator are designed to align global feature distributions across domains using adversarial loss. Furthermore, a dynamic class confusion matrix construction mechanism is introduced to model inter-class confusion relationships based on target domain prediction probabilities and optimize the classifier through self-supervised learning. Additionally, an instance uncertainty weighting mechanism and a class normalization strategy are incorporated to mitigate negative transfer caused by low-confidence instances and class imbalance. Experimental results demonstrate that CCASDAN significantly outperforms existing methods across multiple benchmark datasets, validating its effectiveness in alleviating inter-class confusion and enhancing cross-domain robustness.