HSCPA: a hyperspherical contrastive prototype adaptation network for few-shot cross-domain bearing fault diagnosis

Abstract

Few-shot cross-domain bearing fault diagnosis faces two critical challenges: the scarcity of labeled training samples and significant distribution shifts between source and target domains, which severely degrade the generalization capability of diagnostic models. To address these issues, this study proposes a novel hyperspherical contrastive prototype adaptation (HSCPA) framework. The framework encompasses three key innovations. First, a hyperspherical feature extractor projects raw features onto a unit hypersphere, enforcing geometric constraints—unit normalization, ordered structure, and uniform distribution—to simultaneously enhance inter-class discriminability and intra-class compactness. Second, a contrastive prototype network is developed, integrating four complementary loss functions: intra-class compactness loss, query-prototype contrast loss, prototype uniformity loss, and prototype correlation loss. This design effectively mitigates prototype deviation and overcomes the limitations of conventional metric learning in few-shot scenarios. Third, a cross-domain feature alignment strategy based on Kullback–Leibler divergence is introduced to minimize inter-domain distribution discrepancy while promoting flat minima convergence for robust generalization. Extensive experiments under varying working conditions and equipment configurations demonstrate that the proposed HSCPA framework achieves state-of-the-art diagnostic accuracies of 98.81 and 94.67% in cross-condition and cross-machine tasks, respectively. Compared with eight benchmark methods, HSCPA exhibits superior performance, validating its efficacy for bearing fault diagnosis under data scarcity and domain shift conditions.