电力电容器局部放电模式识别研究

For improving the quality of power capacitor it is important to analyze the influence of material, structure and technology on discharge performance through pattern recognition of partial discharge (PD) within power capacitor. Usually the acoustic emissive signals of partial discharge are detected for specimen of large capacitance. Studying the time and frequency domain characteristics of ultrasonic signals caused by PD and studying the pattern recognition by using artificial neural network have also essential scientific significance for fault diagnosis theory.A computer-based acoustic emissive signal-detecting device is developed for measuring PD within power capacitors. The studying results show that the frequency band should be in the range of 20kHz ~ 500kHz and the total gain of amplifiers should be no less than 60dB.Five types of PD models were designed to represent typical phenomena of PD (gas cavity discharge, oil gap discharge, discharge along oil-paper surface, discharge caused by metal impurity, surface discharge of bushing) in power capacitors and a lot of acoustic signals were obtained through model experiments. The signal duration of gas cavity discharge is very short, about 1 ms, and the signal durations of other models are in the order of millisecond, so gas cavity discharge is easy to be differentiated from other discharge patterns. There are differences among the wave shapes and frequency spectra of acoustic signals caused by various model discharges, therefore the pattern discriminations could be performed according to the time domain and frequency domain characteristics of acoustic signals.The methods used to extract the feature vector from acoustic signals were studied. The investigated 9 kinds of feature extraction methods (time domain data suppression, frequency domain data suppression, auto-regression function, covariance method, power spectrum evaluation, time domain and frequency domain characteristics, time-frequency pattern, AR parameter model, united features method) process corresponding recognition effects for different PD patterns and the united features method, time domain and frequency domain characteristics method, AR parameter model and power spectrum evaluation method are better than the others. During PD pattern recognition the gas cavity discharge is firstly discriminated based on the time duration of acoustic signals. Then the combinational neural network (CNN) is used as a tool to gradually recognize the other four kinds of discharge patterns. The feature vectors extracted by different methods are used as the input vectors of the different sub-networks, it makes the CNN more efficiency and the recognition rates could be more than 98%.

识别电力电容器的局部放电模式，可分析材料、结构、工艺对放电特性的影响，对提高电容器质量和电力系统的运行可靠性具有重要意义。对大电容量试品常检测放电的超声信号，研究放电声信号的频谱；研究以声信号为基础的三维谱图，提取灰度图形矩特征，进而用神经网络识别放电模式；不仅具有重大经济效益，而且具有重大的故障诊断理论学术意义。.