26 / 2018-08-21 20:07:17

A New Approach of Dissolved Gas Analysis for Oil-immersed Transformer Fault Diagnosis Using Convolutional Neural Networks

dissolved gas analysis; fault diagnosis; deep learning; Convolutional Neural Networks (CNNs).

Insulating oil, as the main insulation material in oil-immersed transformer, would gradually deteriorate in the course of operating subjected to electrical, thermal, mechanical, environmental stresses and so on. When the dielectric strength of the insulation material is decreased to a certain degree, which is no longer sufficient to resist the combined effect of external stress, the fault of overheat or discharge in transformers would occur. These faults will destroy the molecular structure of oil or paper to generate various characteristic gas such as CH4, C2H6, C2H4, C2H2 and etc. In decades, dissolved gas analysis (DGA) has been proven an effective method to diagnose the internal fault of transformers, and researchers have proposed many effective methods for the internal fault diagnosis by DGA, such as three-ratio method, Duval's triangle method and etc. Recently, based on the characteristics of fault diagnosis with equivalent to a mapping from gas to fault modes, a series of new intelligent diagnosis methods are proposed by the development of computer science and artificial intelligence technology, such as Genetic Algorithm (GA), Artificial Neural Network (ANN) and Support Vector Machine (SVM). However, due to the complexity of internal fault of transformers and the uncertainty of the correspondence between fault and gas, the accuracy of various intelligent algorithms is not entirely satisfactory so that actually, operation and maintenance personnel in most cases still use the initial way to have the diagnosis. Fortunately, the development of deep learning theory in recent years provides a new consideration to solve this diagnostic problem. This paper, adopting Convolutional Neural Networks (CNNs) on the basis of deep learning theory for reference and inspiration, presents an approach of DGA for transformer fault diagnosis. The input data of DGA is divided into several parts (local sensory area) which are abstracted by each layer in CNNs to get the characteristics of the data and Using the Back Propagation (BP) algorithm to reduce the network errors so as to achieve the matching of DGA data and fault types. In addition, the influence of the size of convolution kernel, sampling width and the number of pool layers on the capability of the model is also considered in this paper. Moreover, compared with traditional methods, such as three-ratio method, Duval's triangle method, Genetic Algorithm (GA), Artificial Neural Network (ANN) and Support Vector Machine (SVM), CNNs is the first algorithm training multi-layer network structure successfully, which uses the spatial relation of data to reduce the number of algorithm parameters to improve the efficiency of the BP algorithm. It comes to conclusion that the model proposed in this paper shows a promising result with high accuracy of fault diagnosis as 84% under large DGA data, which also avoids the trouble of "dimensional disaster" and "local optimal value" in traditional intelligent methods. And this method based on deep learning theory provides a new opportunity to improve the accuracy of fault diagnosis for oil-immersed transformers.