Research On Arc Fault Detection Method Based On Wavelet Transform And Deep Neural Network

With the improvement of living standards,there are more and more electrical equipment in the low-voltage ac power distribution system.While bringing convenience to life,electrical fire accidents occur from time to time.Series arc fault caused by aging damage of insulation layer and poor electrical contact of line or equipment is the main cause of electrical fire.If series arc fault in line or equipment cannot be timely cut off,the fault may further develop and spread.In serious cases,large area power failure,fire,explosion and other malignant events may happen,resulting in unpredictable losses.In order to explore a comprehensive and well-established solution,series fault arc detection method based on wavelet transform and deep neural network is proposed in this paper.With reference to the standard of GB14287.4-2014,a series arc fault generation experimental platform is built to collect 21600 current signals,which contain normal current and fault current of series arc fault under pure resistive load,pure inductive load,resistive and inductive load and nonlinear load.Combined with the wavelet transform theory,effects of different wavelet basis functions and decomposition levels on fault identification are compared and analyzed.The sym8 wavelet basis function is selected to perform three-layer decomposition on collected current signals,coefficients of each layer of reconstructed signal are obtained as fault feature matrices.All fault characteristic matrices were divided into training set and test set at a ratio of 5:1.The theory of deep neural network is studied,and a deep neural network model for series arc fault identification is built.Fault feature matrices extracted by wavelet transform is input to the model for training,and the parameters of model are optimized from three aspects: activation function,learning rate and batch size.The test results show that fault identification accuracy of method which based on wavelet feature and deep neural network is 99.19%.This method can distinguish greatly between normal current and fault current.400 electric fan test samples outside the data set verify that the model has strong generalization ability.In order to effectively identify each type of normal samples and fault samples,data sets were re-labeled and divided into 8 categories.The re-labeled training set is sent to train deep neural network,and the test accuracy is 81.78%.The multi-classification capability of model is improved by increasing the number of network layers.After comparing identification effect,it is found that the identification effect of 11-layer network is better than networks with 16-layer and 19-layer.Its overall accuracy is91.58%.By adding dropout and local response normalization,the overall accuracy of11-layer model is improved to 93.03%.The research results prove that the improved model has an ideal effect of series arc fault detection.It not only can identify whether the sample is in normal state or fault state,but also can identify which kind of load causing it.