A novel domain adaptive network guided by expert experience and active learning for field gas insulated switchgear insulation defect diagnosis

Le résumé fourni par la source

Abstract Despite advances in few-shot learning for gas-insulated switchgear (GIS) insulation defect diagnosis, challenges remain due to domain gaps between laboratory and field conditions as well as the presence of additional novel fault categories in the target domain. To overcome these issues, this paper proposes an active learning domain adaptive network (ALDAN) guided by expert knowledge. ALDAN features a collaborative framework that combines autoencoder-based partial discharge signal reconstruction with prior feature prediction, enabling the extraction of generalized diagnostic features. By embedding expert prior knowledge, the framework reduces dependence on extensive labeled datasets. Then a source-free open-set domain adaptation strategy is introduced, driven by active learning, further enhances model adaptability by leveraging a pre-trained source model and a few expert-labeled, high-value target samples for effective knowledge transfer. Additionally, a locally diverse active selection mechanism based on local label consistency is employed to identify both representative and novel-class samples to improve inter-class discrimination. Experiments on real-world GIS datasets show that ALDAN achieves 93.26% accuracy with only 6% labeled samples, highlighting its effectiveness in data-scarce and complex field scenarios.

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