Dissolved Gas Analysis for Power Transformer Fault Diagnosis Based on Deep Zero-Shot Learning
Leixiao Lei, Yigang He, Zhikai Xing
Abstract
Rapid and accurate fault diagnosis methods of power transformers are essential to ensure power systems’ safe and stable operation. However, the problem of limited fault data and missing data leads to inadequate feature learning, mapping offset, and low accuracy of fault diagnosis. To solve this problem, this study presents a deep zero-shot learning (DZSL) model to diagnose the unseen class fault for the dissolved gas analysis (DGA). First, a specific attribute matrix is presented to build a relationship between fault states and the attribute. Then, the channel-space-time attention network extracts the significant features from the dissolved gas in oil. The multiscale deep residual contraction networks (MDRNs) learn the connection between the attribute matrix and the main features. Finally, the cosine similarity comparison method obtains the transformer fault status. The performance of the proposed model is verified by the measured data of dissolved gas in the oil of the power transformer. The experimental results show that the model has a better performance than other compared methods.