Meta-autoencoder-based zero-shot learning for insulation defect diagnosis in gas-insulated switchgear

Abstract

Data-driven methods are the main methods of training models for diagnosis of insulation defects in gas-insulated switchgear (GIS). Because of complicated operating environments, sometimes no target samples are available for training, which leads to insufficient feature learning. Therefore, a meta-autoencoder based zero-shot learning method (MAZL) is proposed for diagnosis of GIS insulation defects. Firstly, visual features of insulation defects signals are extracted by a convolutional neural network. Then, the mapping between visual space and semantic space is learned by semantic autoencoder (SAE). For SAE, triplet loss is adapted to align semantic information, and unknown class semantic constraints are used to alleviate the domain offset. Finally, the nearest-neighbor algorithm is used to achieve zero-sample diagnosis. In model training, episode training is introduced to make MAZL model learn more discriminative semantic prototype, which alleviates classification deviations. The MAZL achieved a diagnostic accuracy for single and multi-source GIS insulation defects of 96.215% and 90.41% without using test class for training, respectively. This provides ideas for diagnosis of new GIS insulation defects.