Research On Detection Method Of Series Arc Fault In Low Voltage Electrical Fire Based On Deep Learning

In household low voltage AC system,the current effective value when series arc fault occurs overlaps with which of normal operation to a great extent,so it has strong concealment.The occurrence location and time are random,so it’s especially difficult to detect series arc fault.Thus series arc fault seriously threatens the safety of household power supply and distribution system.Aiming at the problem that series arc fault is difficult to detected,in order to find a universal detection method which can overcome the influence of load characteristics on the detection results,a detection method based on deep learning is presented.Combine with the characters of current signal,deep learning networks are constructed which are trained and tested by the current waveforms data which are collected by experiments under normal operation and series arc fault.The feasibility of this method is investigated and the detection accuracy is improved.Arc fault simulation device is made,experiment circuit is connected and data acquisition environment is set up.The current waveform data of linear load(including inductance coil,incandescent lamp and the series connection of inductance coil and incandescent lamp,they are used to investigate the influence of impedance angle on series arc fault detection)and nonlinear load(a television)in normal operation and series arc fault state were collected,totaling 9600 groups.Two hundred groups of normal and arc fault current data sets are randomly selected from every load as test data,and test sets are made for four kinds of loads respectively to test the detection effect of deep learning network on a single load.All the test data are mixed randomly as a holistic test set to test the overall detection effect of deep learning network on all loads.The remaining 8000 groups are used as training data to produce a training set for training the constructed deep neural network.An AlexNet is built to detect series arc fault.The detection accuracy is increased by modifying AlexNet from the perspectives of network structure and activation function.The overall detection accuracy is increased to 93.875% and the detection accuracy of single load is increased to 96.5%.Two kinds of ResNet are built to detect series arc fault.The detection accuracy of single load is increased to 99.75% and the overall detection accuracy is increased to 97.125%.Compared with AlexNet and its modified models,ResNets reduce the number of training epochs and parameters but detection accuracy increased.In order to achieve the goal of reducing or avoiding electric fire by forecasting and warning the series arc fault,this paper adopts the deep learning algorithm to detect the line current signal,and realizes the identification of the series arc fault.The current waveform signals of normal operation and series arc fault are collected by experiments.The deep learning networks are constructed and modified.The networks are trained by the gained current waveform signals and the detection effect is tested.The results show that the ResNet method can achieve a better detection effect for low voltage series arc fault.