Feature Analysis And Pattern Recognition Of Low-voltage Arc Fault

With the continuous increase of social electricity consumption,electrical fires occur frequently,threatening people's lives and properties.The arc fault is the main cause of electrical fires.When arc fault occurs,only a few amps of current can produce local high temperature,and the conventional electrical protection device couldn't work.Arc fault detection devices(AFDD)are electrical protection devices which can cut off circuits after recognizing the arc fault.The key technology of AFDD is to recognize the arc fault accurately without cutting off circuits by mistake.This paper mainly studies the features of the arc fault and the classifier which is used to recognize the arc fault.The main work is as follows.experiments are conducted according to GB/T 31143-2014 standard.The experiments include series arc fault experiments,parallel arc fault experiments,suppressive load shielding experiments,and load normal working experiments.The current data are recorded to build the dataset.The features are extracted according to time domain and frequency domain characteristics of the arc fault.The statistical results show that both the load type and the circuit state will affect the value of each feature.Several feature selection algorithms,which include information gain,gain ratio,and Relief,are used to evaluate the features.The top 14 features are selected as the feature subset to train classifiers.This paper uses the classifier to recognize the state of each power frequency cycle.If the number of fault cycles in the time window is greater than the set threshold,AFDD will cut off the circuit.When the classifier predicts the test sample,there are two types of errors.One is false positive(FP),another is false negative(FN).The potential risk of FP is higher than that of FN.This paper introduces the cost-sensitive learning theory,and designs a cost-sensitive neural network based on Bagging and cost-sensitive decision making.The test results show that the cost-sensitive neural network with reasonable cost matrix values could reduce the number of FP,meanwhile has a high accurate rate.This paper uses feature selection algorithms to evaluate and select features,and introduces cost-sensitive learning to reduce the number of FP when recognize the arc fault.These methods have reference values and application values to the arc fault detection technology of our country.