【摘要】 针对现有剩余电流保护装置难以识别触电事故的问题，提出一种基于小波高频分布特征和概率神经网络的触电事故识别新方法.利用S变换对包含触电电流的剩余电流信号进行频谱分析，发现触电时刻高频分量存在幅值突增现象；通过小波多分辨分析提供的多尺度频率窗口提取其各层小波高频分布，利用量纲一化处理后的各层小波高频分布突变量的累计和，量化剩余电流信号前5层高频特征描述触电事故；充分考虑触电事故发生时间的随机性，对所提取的小波特征进行类别划分；构建基于PNN的触电事故识别模型，并按指定步长对定义域内网络平滑参数进行寻优，同时采用均值聚类方法优化网络构.结果表明：触电时刻剩余电流信号500 Hz以上高频段幅值存在明显突变，量纲一化处理后的各层小波高频分布幅值突变量累计和能较好地描述各层小波高频分布对应阶段的幅值突增现象；所建立的PNN网络模型最优平滑参数区间为0.15～0.29,对应的触电事故最佳识别率为95.5%.更多还原

【Abstract】 To solve the problem that the existing residual current protection device was difficult to identify electric shock accidents, a new method of electric shock identification based on wavelet high-frequency distribution characteristics and probabilistic neural network(PNN) was proposed. The S-transform was used to analyze the spectral characteristics of the residual current signal including the moment of electric shock, which was found that the high-frequency component of the residual current signal at the time of electric shock had sudden change in amplitude. The wavelet high-frequency distribution of each layer was extracted through the multi-scale frequency window provided by the wavelet multi-resolution analysis. To describe the electric shock accident, the normalized processing of each layer of wavelet high-frequency distribution mutations was used to accumulate and quantify the high-frequency characteristics of the first 5 layers of the residual current signal. Taking full account of the randomness of the time of electric shock accidents, the extracted wavelet features were divided into categories. A PNN-based electric shock accident recognition model was constructed, and the network smoothing parameters in the defined domain were optimized according to the specified step. The mean clustering method was also used to optimize the network structure. The results show that there is significantly sudden change in the amplitude of the residual current signal above 500 Hz at the time of electric shock. After normalization, the cumulative sum of the amplitude mutations of the wavelet high-frequency distribution of each layer can well describe the amplitude sudden increase in the corresponding stage of the wavelet high-frequency distribution of each layer. The optimal smoothing parameter interval of the established PNN network model is from 0.15 to 0.29, and the best corresponding identification rate of electric shock accidents is 95.5%.更多还原