PD characteristics and defect type identification of typical defects in oil-pressboard insulation

In view of partial discharge caused by internal defects in oil-impregnated transformers, an identification method based on partial discharge signals and patterns from three defective oil-pressboards and deep neural networks is proposed method according to the three signal patterns with excessive partial discharges. Firstly, the characteristics of the partial discharge patterns are obtained and extracted as the input. Then the influence of different parameters of the deep neural network is compared on the identification effect and the optimal deep neural network structure is found. Then trained deep neural network is used to identify the partial discharge type. The results show that the use of Adam algorithm to make the loss function quickly decline and converge, the accuracy of identifying partial discharge signals is higher than the Gradient Descent method and RMSProp algorithm. The loss function can quickly decline and converge with the learning rate of 0.001. The performance of the neural network is optimal without overfitting under the network structure of 64–32–32–32–16. The partial discharge identification rate of the optimized deep neural network is higher than that of the support vector machine and BP neural network.