农业工程学报
農業工程學報
농업공정학보
2014年
23期
238-245
,共8页
韩晓慧%杜松怀%苏娟%关海鸥%邵利敏
韓曉慧%杜鬆懷%囌娟%關海鷗%邵利敏
한효혜%두송부%소연%관해구%소리민
优化%电流%信号检测%总泄漏电流%触电电流%最小二乘支持向量机%网格搜索%交叉验证
優化%電流%信號檢測%總洩漏電流%觸電電流%最小二乘支持嚮量機%網格搜索%交扠驗證
우화%전류%신호검측%총설루전류%촉전전류%최소이승지지향량궤%망격수색%교차험증
optimization%electric currents%signal detection%total leakage currents%electric shock currents%least square-support vector machine%grid search%cross validation
针对如何从低压电网总泄漏电流中检测出生物体触电电流信号的难题，提出了一种基于网格搜索和交叉验证的最小二乘支持向量机的触电电流信号检测方法。首先在剩余电流动作保护装置触电物理试验系统平台上通过故障录波器获得生物体在3个典型时刻（电源电压最大时刻、电源电压过零时刻及电源电压任意时刻）发生触电过程的总泄漏电流和触电电流波形，并截取触电前1个周期和触电后3个周期共800个采样点的信号数据作为触电试验样本数据；然后将触电试验样本数据进行滤波预处理，预处理后的多个样本采样点的总泄漏电流组合成特征向量输入最小二乘支持向量机（least square-support vector machine，LS-SVM），相应样本采样点的触电电流作为其输出，并通过网格搜索与交叉验证相结合的方法来优化最小二乘支持向量机参数，利用输出最优参数组合对触电电流与总泄漏电流的关系进行训练，从而建立了触电电流的检测模型；最后利用该方法对10组测试样本数据进行了检测，检测结果为：当训练样本数据为20组时，检测均方误差为14.0040，当训练样本数据为40组时，检测均方误差为11.7469，当训练试验数据为65组时，检测均方误差为11.1849。与径向基（radial basis function， RBF）神经网络方法相比，最小二乘支持向量机方法比径向基神经网络方法检测均方误差分别低3.7272、1.9132、0.1556，从而可较准确地从总泄漏电流中检测出生物体触电电流信号，为开发新一代基于生物体触电电流分量而动作的自适应型剩余电流保护装置提供理论依据。
針對如何從低壓電網總洩漏電流中檢測齣生物體觸電電流信號的難題，提齣瞭一種基于網格搜索和交扠驗證的最小二乘支持嚮量機的觸電電流信號檢測方法。首先在剩餘電流動作保護裝置觸電物理試驗繫統平檯上通過故障錄波器穫得生物體在3箇典型時刻（電源電壓最大時刻、電源電壓過零時刻及電源電壓任意時刻）髮生觸電過程的總洩漏電流和觸電電流波形，併截取觸電前1箇週期和觸電後3箇週期共800箇採樣點的信號數據作為觸電試驗樣本數據；然後將觸電試驗樣本數據進行濾波預處理，預處理後的多箇樣本採樣點的總洩漏電流組閤成特徵嚮量輸入最小二乘支持嚮量機（least square-support vector machine，LS-SVM），相應樣本採樣點的觸電電流作為其輸齣，併通過網格搜索與交扠驗證相結閤的方法來優化最小二乘支持嚮量機參數，利用輸齣最優參數組閤對觸電電流與總洩漏電流的關繫進行訓練，從而建立瞭觸電電流的檢測模型；最後利用該方法對10組測試樣本數據進行瞭檢測，檢測結果為：噹訓練樣本數據為20組時，檢測均方誤差為14.0040，噹訓練樣本數據為40組時，檢測均方誤差為11.7469，噹訓練試驗數據為65組時，檢測均方誤差為11.1849。與徑嚮基（radial basis function， RBF）神經網絡方法相比，最小二乘支持嚮量機方法比徑嚮基神經網絡方法檢測均方誤差分彆低3.7272、1.9132、0.1556，從而可較準確地從總洩漏電流中檢測齣生物體觸電電流信號，為開髮新一代基于生物體觸電電流分量而動作的自適應型剩餘電流保護裝置提供理論依據。
침대여하종저압전망총설루전류중검측출생물체촉전전류신호적난제，제출료일충기우망격수색화교차험증적최소이승지지향량궤적촉전전류신호검측방법。수선재잉여전류동작보호장치촉전물리시험계통평태상통과고장록파기획득생물체재3개전형시각（전원전압최대시각、전원전압과령시각급전원전압임의시각）발생촉전과정적총설루전류화촉전전류파형，병절취촉전전1개주기화촉전후3개주기공800개채양점적신호수거작위촉전시험양본수거；연후장촉전시험양본수거진행려파예처리，예처리후적다개양본채양점적총설루전류조합성특정향량수입최소이승지지향량궤（least square-support vector machine，LS-SVM），상응양본채양점적촉전전류작위기수출，병통과망격수색여교차험증상결합적방법래우화최소이승지지향량궤삼수，이용수출최우삼수조합대촉전전류여총설루전류적관계진행훈련，종이건립료촉전전류적검측모형；최후이용해방법대10조측시양본수거진행료검측，검측결과위：당훈련양본수거위20조시，검측균방오차위14.0040，당훈련양본수거위40조시，검측균방오차위11.7469，당훈련시험수거위65조시，검측균방오차위11.1849。여경향기（radial basis function， RBF）신경망락방법상비，최소이승지지향량궤방법비경향기신경망락방법검측균방오차분별저3.7272、1.9132、0.1556，종이가교준학지종총설루전류중검측출생물체촉전전류신호，위개발신일대기우생물체촉전전류분량이동작적자괄응형잉여전류보호장치제공이론의거。
Currently, residual current operated protective devices (RCDs) have a wide range of application in low voltage power grids and play an important role in preventing electric shock hazards and avoiding fire disasters caused by ground fault. But, the stocking current of the animals and human beings has no relationship with the setting value of action current from the protection devices, and the root mean square (RMS) value of residual current detected is considered the current value to determine if the protector acts or not. Theoretical analysis and operation experiences indicate that such criterion is unavailable in identifying the shocking current signals of the animals and human beings from the summation leakage current .Thus, in order to identify the electric shock signal from the summation leakage current automatically and accurately, intelligent information processing techniques are adopted and identification method based on least square-support vector machine (LS-SVM) with grid search and cross validation optimization are proposed. Firstly, through the experiments simulating various scenarios of rabbits electric shocking on the electric shock experiment platform of residual current operated protective devices(RCDs), signal data of 800 sample points before the one cycle and after three cycles of electric shock are used as electric shock sample data obtained by fault recorder to get the leakage current and electric shock current waveform on the electric shock process of the power supply voltage at maximum time, zero time, and any time. Secondly, the above sample data needed to be filtered to reduce the impact of the non-stationary for noise data. Then, the leakage currents of sampling points are combined into a high dimensional feature vector which is input into LS-SVM and the corresponding electric current of sampling point is employed as output of LS-SVM. The relation between input and output is trained by applying grid search and cross validation to determine the optimal parameters of LS-SVM automatically, and the ideal model of electric shock signal is obtained. Thirdly, a total of 75 groups of sample data are used as the research object. Among them 10 groups of sample data are used as testing samples, with the experimental results showing that when 20 groups of sample data are used as training samples, the identification mean square error is 14.0040. When 40 groups of sample data are used as training samples, the identification mean square error is 11.7469. When 65 groups of sample data are used as training samples, the identification mean square error is 11.1849. In comparison with the electric shock current identification method proposed method recently, the radial basis function (RBF) neural network method, the proposed method has the lower identification error. Consequently, it can identify the shocking current signals of the animals and human beings from the summation leakage current more accurately and provide a more reliable theoretical basis for developing new generations of adaptive residual current protection devices, which is based on the electric current component of the animals and human beings causing its action.
