Transformer fault diagnosis based on improved deep coupled dense convolutional neural network

Abstract

The normal operation of the power transformer guarantees the safety and reliability of the power system. However, the data of gas in oil exists the phenomenon of insufficient and unbalanced fault samples and few features. This paper proposes a fault diagnosis model, which contains the adaptive synthetic oversampling (ADASYN), the reconstructed data method, and an improved deep coupled dense convolutional neural network (CDCN). Firstly, the ADASYN expanded the normalized data set. Then, the characteristic gas in the data set is reconstructed to increase the number of features. Finally, the improved CDCN extracts the features from the generated data and obtains the fault status of the power transformer. The IEC TC 10 data and the collected data are used as the test data. The experiment results show the proposed method obtains better performance than the compared algorithms. Primarily, the accuracy of the proposed method acquires 94.05%.