Study On Power Transformer Fault Diagnosis Technology Based On DGA And Partial Discharge Detection

Power transformer is the most important equipment in power system, the transformer fault diagnosis is helpful to ensure safety and stability of the system. Early before the transformer fault happening, it will come with some special phenomenon such as electric pulse, local heating and gas decomposition. The data of transformer condition can be got through quantification and measurement, and the type of transformer fault will be identified by analyzing the features extracted from these data. This paper mainly studies on the data acquisition, signal (image) processing, data analysis, pattern recognition and etc..In the research of the transformer fault diagnosis method based on the result from transformer oil online monitor, mathematical morphology is applied to filtering the output signal of chromatograph, then the voltage is extracted through image processing, and voltage-gas concentration nonlinear regression is achieved by using support vector machine. To overcome difficulty of SVM kernel parameter selection, a kernel parameter optimization method and a multi kernel multiclass SVM is designed to identify the type of power transformer fault. The results of experiments indicate that the arithmetic of this paper has high classification accuracy with fast convergence speed and small sample requirement, which proves its effectiveness and usefulness.In the research of the transformer fault diagnosis method based on the result from transformer partial discharge detection, adaptive morphological filter is designed to solve period noise disturbing, and the normalized pattern spectrum is used to feature extraction from single discharge waveform, then the type of power transformer fault is identified by SVM. The results of integrated filter experiments indicates that adaptive morphological filter have obvious inhibiting effect both on periodic interference and white noise, which overcoming the traditional morphological filter’s statistical deviation. The performance of discharge type recognition is compared between the multi class SVM and BP neural network, the experiments showed that the multi class SVM achieve higher classification accuracy and more suitable for small sample learning.Finally, have a brief summary of the research in this paper, and point out the issues that need to be further discussed.