Dissolved Gas Analysis of Transformer Oil Based on Deep Belief Networks

Dissolved gas analysis (DGA) has been proven an effective method to diagnose the internal fault of transformers for decades. This paper proposes an approach of DGA based on Deep Belief Networks (DBN) which belongs to deep-learning theory. With composed of Restricted Boltzmann Machines (RBMs), DBN can built the mapping relationship between the characteristic gas and the fault types automatically to achieve the accurate diagnosis, namely Pattern Recognition (PR). In this paper, DGA data is divided into three categories, called training data, fine-tuning data and test data. Training data is used for unsupervised-learning to initialize parameters of RBMs in DBN, and fine-tuning data is for supervised-learning with fault modes to optimize parameters. Finally, the test data is used to calculate the recognition rate of transformer fault diagnosis. It comes to conclusion that DBN model proposed shows a promising results of transformer fault diagnosis with high accuracy of 84.87% Compared with Back Propagation Neural Network (BPNN) method, DBN model not only achieves a high recognition rata of transformer fault, but also have strong generalization ability under big data, which could provide a powerful tool for the internal fault diagnosis of transformers.