Semi-Supervised Contrastive Transfer Learning Network for Fault Diagnosis of Cross-Working Condition Rolling Bearings

Résumé fourni par la source

The methods based on feature transfer learning, which solve the problem of cross-condition rolling bearing fault diagnosis without learning the discriminative features between different fault classes, making the existence of fuzzy classification boundaries, resulting in the performance of diagnosis model being limited. Aiming at the problem, this paper proposes a cross-condition rolling bearing fault diagnosis method based on the semi-supervised contrastive transfer learning network (SSCTLN). In SSCTLN, the local maximum mean discrepancy (LMMD) is adopted for the transfer learning of cross-domain features, and the in-domain semi-supervised contrastive learning (SSCL) method is designed, which guides the learning of discriminative features between different classes, to eliminate the fuzzy classification boundaries by the supervision of class information, and also facilitates the transfer learning of cross-domain features. Meanwhile, the negative impact from poor-quality target domain pseudo-labels on SSCL and feature transfer learning is reduced, by dynamically limiting the relevant contrastive learning loss and transfer learning loss gains, and introducing domain adversarial. Finally, the effectiveness of SSCTLN is verified by the transfer fault diagnosis experiments on the Paderborn University (PU) dataset. The experiment results show that SSCTLN improves the overall average accuracy by 22.43% compared to DSAN on six cross-condition diagnosis tasks, and is superior to other popular feature transfer learning methods.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.