直流故障电弧的特征提取与识别

发布时间：2018-10-14 17:36

【摘要】：直流电弧因为没有周期性和过零点等特性,致使直流故障电弧的检测要比交流故障电弧困难。国内外对直流故障电弧的特征提取缺乏多样化,因此本文旨在丰富直流故障电弧的特征种类,从特征提取和识别两个方面给出特征的评价标准。本文所研究的电弧属于热阴极、气体中、串联、直流、故障电弧,并通过实验的方法采集正常工作和发生故障电弧时的电流数据,选取电弧刚发生时的不稳定阶段作为特征提取的数据基础。对采集的数据从时域、频域、小波、谱分析和混沌分析这几个角度共提取了 17个特征。在时频分析中,从电流变化的角度提取了时域特征4个,从统计学的角度提取了频域特征3个,利用小波变换提取了小波特征3个。在谱分析中,从幅值角度提取功率谱特征3个,从幅值和相位的角度提取高阶谱特征3个。在混沌分析中,基于自相关法计算时间延迟、基于虚假最近邻法计算嵌入维数,利用这两个参数重构电弧电流的相空间,并利用小数据量法提取最大Lyapunov指数特征1个。给出特征提取层面的特征评价指标,即价值量百分比,并计算各特征的指标值大小,其中混沌特征的指标值最大。利用支持向量机(Support Vector Machines,SVM)对故障电弧特征进行识别。选择基于高斯核函数的非线性支持向量机作为最终分类器。利用网格搜索法和3-折交叉验证法对SVM的参数进行寻优,得到的惩罚参数是0.054409,核参数是1。训练出1个总分类器,准确率为99.999%,17个比较分类器,各自准确率都有下降,但整体保持在99.9%以上。给出识别层面的特征评价标准,即对准确率的贡献程度,并计算各特征的指标值,其中混沌特征的指标值最大。结果证明:本文所用特征提取方法和模式识别策略正确;所提两个指标都可以用来评价特征;直流故障电弧的检测问题本质是一个混沌辨识问题。本文所做研究丰富了直流故障电弧的特征种类,并且针对特征有效性给出了评价标准。
[Abstract]:Because DC arc has no periodicity and zero crossing, it is more difficult to detect DC fault arc than AC fault arc. The feature extraction of DC fault arc is lack of diversification at home and abroad, so the purpose of this paper is to enrich the characteristics of DC fault arc, and give the evaluation criteria from two aspects: feature extraction and recognition. The arc studied in this paper belongs to hot cathode, gas, series, DC, fault arc. The current data of normal operation and fault arc are collected by experimental method. The unstable phase of arc is selected as the data base of feature extraction. A total of 17 features were extracted from time domain, frequency domain, wavelet, spectral analysis and chaos analysis. In time-frequency analysis, four features in time domain, three features in frequency domain and three features in wavelet domain are extracted from current variation, statistics and wavelet transform respectively. In spectral analysis, three power spectral features are extracted from amplitude and three higher-order spectral features are extracted from amplitude and phase. In chaos analysis, time delay is calculated based on autocorrelation method, embedded dimension is calculated based on false nearest neighbor method, the phase space of arc current is reconstructed by these two parameters, and one feature of maximum Lyapunov exponent is extracted by the method of small amount of data. The index of feature evaluation in feature extraction level, that is, the percentage of value, is given, and the index value of each feature is calculated, among which the index value of chaotic feature is the largest. Support vector machine (Support Vector Machines,SVM) is used to identify fault arc features. The nonlinear support vector machine based on Gao Si kernel function is chosen as the final classifier. The parameters of SVM are optimized by grid search method and 3-fold cross validation method. The penalty parameter is 0.054409 and the kernel parameter is 1. One general classifier was trained, the accuracy rate was 99.999, and 17 comparative classifiers were all reduced, but the overall accuracy remained above 99.9%. The evaluation criterion of recognition level is given, that is, the degree of contribution to accuracy, and the index value of each feature is calculated, among which the index value of chaotic feature is the largest. The results show that the feature extraction method and pattern recognition strategy used in this paper are correct, the two indexes can be used to evaluate the features, and the problem of DC fault arc detection is essentially a chaotic identification problem. The research in this paper enriches the characteristic types of DC fault arc, and gives the evaluation criteria for feature validity.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TM501.2

【相似文献】