人工神经网络识别放电能力与网络结构关系的研究

Study on Relationships between Discharge Recognition Ability of Artificial Neural Network and Its Structure

用蒙特卡洛方法得到了局部放电模拟试验数据,数据的多少可根据需要任意选择,除统计分散性外,具有良好的重复性。用这些数据研究了前馈人工神经网络识别放电三维谱图(放电量q、放电发生相位、放电重复率 n)的能力。为减少放电分散性对网络识别能力的影响,讨论了采集放电数据的必要工频周期数。为选择较好的网络结构,研究了输入层、隐含层神经元数对网络识别能力的影响。最后分析了网络对放电识别具有的一定的智能性。

Simulated partial discharge (PD) experiment data were achieved through Monte-Carlo method.These data could be produced as much as necessary and besides the statistical dispersity the repetitiveness of them was enough.The ability of recognizing 3-dimensional PD pattern (discharge quantity q,occurring phase and pulse repetition rate n) by feed forward artificial neural network (ANN)was studied.For reducing the effect of the dispersity of PD data,the necessary power frequency cycle of data acquisition was discussed. In order to choose the better ANN structure,the effects of neuron numbers of input layer and hidden layer on the ability of recognizing PD of ANN were investigated.At last the certain intelligence of ANN in recognizing discharge was analyzed.