Semi-Supervised Class Adaptive Prototype Network for Cross-Working Rolling Bearing Fault Diagnosis Under Limited Samples

Deep transfer learning-based fault diagnosis methods are prone to encounter challenges in cross-domain knowledge transfer, particularly when facing limited labeled samples and the diagnostic performance depends on the amount of source domain data and sample distribution. For this purpose, a semi-supervised class adaptive prototype network (SSCAPN) is proposed, which utilizes a meta-learning paradigm to enable cross-working condition rolling bearing fault diagnosis under limited sample scenarios. Firstly, a dualscale feature fusion extraction module is proposed, which captures both local detail features and global trend features, significantly enhancing the model’s fault diagnosis capability cross-working conditions. Secondly, a similarity measure is utilized to generate pseudo-labels for prototypes derived from unlabeled samples. To reduce the misuse of pseudo-labels, a confidence-weighted approach is used to adaptively reduce the discrepancy between class prototypes and cluster centers. Moreover, a class boundary adaptive enhancement module is proposed, which adaptively expands the boundary between different classes by introducing boundary value and gradient descent methods, significantly improving the classification accuracy of the model in the boundary region. Finally, the experiments conducted by using the Jiangnan University rolling bearing dataset and the rolling bearing fault test bench dataset indicate that, the proposed SSCAPN method achieves efficient fault diagnosis performance across working conditions under few labeled samples.