A novel method for transformer fault diagnosis based on refined deep residual shrinkage network

Abstract

Abstract This study proposes a novel method to improve the fault identification performance of transformers. First, to couple multiple factors, a high‐dimensional feature map composed of the feature gas concentrations and some associated variables is constructed. Second, the deep residual shrinkage network is revised using the updated alternating direction multiplier, and the newly constructed variable soft thresholding is proposed to eliminate constant deviations. In addition, the fast iterative shrinkage‐thresholding algorithm is adopted, as it can speed up the determination of the threshold. For the output end, the uniform manifold approximation and projection algorithm are adopted to ensure the integrity of the local optimal solution and the global solution. Compared with traditional dissolved gas analysis methods, the novel refined deep residual shrinkage network exhibits superior precision, which is justified through experiments. The results show that the recognition accuracy of the new model is more than 1.3% higher than that of the existing methods. The new method has good scalability in power applications and fault prevention.

Keywords

• fault diagnosis
• learning (artificial intelligence)
• power transformer protection
• transformer oil