Research On Partial Discharge Characteristics Based On UHF Detection Method

Currently on-site equipment through actual operation Statistical analysis showed that the main reason for insulation fault is still safe and stable operation of the affected equipment. Ultra high frequency(Ultra High Frequency, UHF) partial discharge detection method is now in a popular method that receives electromagnetic signals radiated partial discharge process within the UHF band UHF antennas by partial discharge fault detection. By building a partial discharge test system, based on a study of a large number of experimental UHF PD frequency characteristics corresponding to different models for defects in the 3-meter anechoic chamber environment and analyze the characteristics of each, depending on the model. On this basis, through the support vector machine(Support Vector Machine, SVM) and relevance vector machine(Relevance Vector Machine, RVM) and other artificial intelligence learning algorithms for pattern recognition defect types. Through the classification results of previous experimental data were analyzed and found artificial neural networks(Artificial Neural Network, referred ANN) slow convergence and easy to fall into local optimum problems affecting its accuracy, SVM kernel function must satisfy the requirements of Messi inherent limitations theorem, etc., need to calculate the kernel function between all samples, the computational complexity will increase dramatically as the number of samples increases, the model sparsity limited. Compared with the first two, relevance vector machine(Relevance vector machine, RVM) low computational complexity, not only can get a probability based on sparse solution kernel output, and test time is shorter, more suitable for online testing.In this paper, support vector machine and relevance vector machine is used to study partial discharge characteristics by binary classification with more than three dichotomous classification of four typical PD classification model, experimental results show that, SVM can be good four typical model for partial discharge distinguish and obtain appropriate recognition accuracy. Compared with SVM, RVM while obtaining recognition accuracy, you can also get more sparse classification model with probabilistic output value, and has a shorter time and better predict the generalization performance.