Research On Low Voltage Series Arc Fault Identification Method Based On CBAM-SCNN

Arc fault is an important cause of electrical fire.Effective detection can ensure the normal operation of the line and reliable operation of electrical equipment.In the detection of low voltage series arc fault,the traditional recognition algorithm can only detect the arc fault of one or several kinds of load effectively,and the generalization ability is insufficient;For the arc fault of carbonization path,because of its difficulty in obtaining experimental data,there is no suitable detection method to collect the experimental data of such load effectively.Therefore,in order to further pro pose a more generalized arc fault detection method,this paper studies the identification method of low voltage series arc fault.Firstly,according to the relevant standards,the point contact and carbonization path arc fault test platform is designed and built.The active load box is used to simulate different power factor loads in the actual circuit,and the time domain data collected is saved in half cycle.The high pass filter is used to filter out the interference of power frequency information in pra ctice and generate continuous time series.One-dimension continuous time series is coded as two-dimensional characteristic matrix by two-dimensional coding method,so as to establish the basic data set of arc fault.Then,the main influence of data missing on the actual model is analyzed.For the missing carbonization path data set,some of the carbon path data collected are taken as the data set,and the pseudo real data is generated by using the improved generation counter network of Wasserstein distance and gradient penalty factor.The results show that the improved network generated arc fault coding matrix is more diverse,and the generated data can replace the actual arc fault data coding matrix completely.Finally,an improved convolutional neural netw ork is designed to detect the characteristics of the time-domain information coding matrix.The network adopts SELU activation function and Lecun normal initialization method,residual connection,CBAM attention mechanism and adaptive asymmetric convolution kernel method,which improves the convergence speed in the calculation process of convolution neural network,effectively prevents the degradation of neural network,and improves the extraction effect of convolution layer on arc fault characteristics.The accuracy of the method is 99.11%,the miss alarm rate is 1.23%,the false alarm rate is 0.53%.The model has better generalization ability.