Transformer Fault Diagnosis Based on Stacked Contractive Auto-Encoder Net

Abstract

Dissolved gas analysis (DGA) is an effective method for oil-immersed transformer fault diagnosis. This paper proposes a transformer fault diagnosis method based on Stacked Contractive Auto-Encoder Network (SCAEN), which can detect the transformer’s internal fault by using DGA data, including H2, CH4, C2H2, C2H4, C2H6. The network consists of a three-layer stacked contractive auto-encoder (SCAE) and a backpropagation neural network (BPNN) with three hidden layers. A large amount of unlabeled data is used to train to obtain initialization parameters, and then a limited labeled dataset is used to fine-tune and classify the faults of trans-formers. The proposed method is suitable for transformer fault diagnosis scenarios, which contains very limited labeled data. when tested on real DGA dataset, the fault diagnosis accuracy is up to 95.31% by SCAEN, which performs better than other commonly used models such as support vector machine (SVM), BPNN, auto-encoder (AE), contractive auto-encoder (CAE) and SCAE.