Search

Based on the characteristics of SF6 decomposition products, machine learning is becoming a popular solution to construct the insulation fault diagnosis model for gas insulated switchgear (GIS). However, a main challenge is to establish a model compatible for diverse operation conditions, due to: 1) the distribution difference between the laboratory and field data, and 2) limited labeled field samples. Inspired by heterogeneous domain adaptation (HDA), a novel weighted heterogeneous domain adaptation (WHDA) network is proposed, which could integrate the latent knowledge of source and target domains with different feature spaces. During the feature integration, the fault samples from each domain are projected into a common subspace with unified dimensions and aligned distributions based on maximum mean discrepancy (MMD). The sample weights are learned to balance the importance of each sample. Besides, the weighted pseudo-label strategy is presented to enhance the reliability of predicted labels of unlabeled data. At last, the optimal diagnosis model is obtained by evaluating the loss of the domain discriminator and classifier. The algorithm experiments are performed in supervised and semi-supervised scenarios using laboratory and field fault samples. The results validate the superiority of the proposed WHDA on GIS insulation fault diagnosis against feature extraction and other HDA methods.