Semi-supervised sparse feature optimization transfer learning for fault diagnosis under cross-condition and imbalanced data

Abstract Fault diagnosis in intelligent manufacturing faces challenges from cross-condition variations and data imbalances, especially with rare faults. Existing methods typically address these issues separately, yet both often coexist in industrial settings. To tackle these dual challenges, this study proposes a semi-supervised sparse feature optimization transfer learning diagnostic method (SSFOD). This method introduces two strategies: (1) improved enhanced sparse filtering to optimize feature sparsity and improve detection sensitivity for minority class faults, and (2) adaptive resampling maximum mean discrepancy to dynamically adjust data distributions, enhancing model adaptability and generalizability. Experimental results show that SSFOD achieves an average accuracy of 99.3%, significantly outperforming existing methods. This approach effectively addresses the combined challenges of cross-condition and imbalanced data fault diagnosis, advancing the field in complex industrial applications.