Feature Extraction And Adaptive Detection Of Arc Fault

Arc fault detection plays an important role in ensuring safe and reliable operation of electrical equipment and avoiding electrical fires.In AC electrical system,due to the diversity and complexity of arc bring great challenges to arc fault detection,arc fault detection is faced with at least two major problems:feature extraction of current and adaptive arc fault detection.Therefore,from the aspects of the feature extraction of current and adaptive arc fault detection,our research work is as follows:1.Traditionally,feature extraction for arc fault utilizes signal analysis tools,such as fast Fourier transform(FFT),discrete wavelet transforms(DWT)to manually extract features of current.However,studies have shown that traditional methods of feature extraction are subjective in a certain sense,which limits the expression ability of original current.Based on the task of time series modeling,this paper applies deep learning to feature extraction of current,and temporal convolutional networks(TCN)is proposed for features extraction of current.Then a series of optimizations based on the modern convolutional network technology is applied to improve the performance of the TCN.Finally,experiments show that the TCN can contribute to extract salient and discriminated features of current.2.Most of the existing deep networks for time series fall into the category of time-domain methods,and completely ignore frequency information of time series.Based on the standard discrete wavelet decomposition,this paper further applies a novel neural network structure called multilevel wavelet decomposition networks(MWDN)to extract features of current,which seamlessly embeds the wavelet transform into the deep learning framework so that all parameters can be trained.Then,on the basis of the TCN and MWDN,a multi-frequency temporal convolutional networks(MF-TCN)is elaborately designed to capture the frequency information of current.Finally,the proposed MF-TCN is demonstrated in the feature extraction of current,which shows that MWDN inherits the decomposition ability of discrete wavelet decomposition while has the learning ability of deep neural network.3.In order to deal with the drawbacks of current,such as tremendous data volumes,the imbalance of normal and abnormal data,concept drift and so on,adaptive Gaussian mixture model(GMM)is proposed to detect arc fault.Firstly,this paper combines GMM with Bayesian inference to improve the performance of online arc fault detection.Then,an adaptive strategy using improved EM algorithm is proposed to detect arc fault while collect historical streaming data to dynamically update the model.Finally,the adaptive GMM method is demonstrated in arc fault detection.The experiments demonstrate that the adaptive GMM for arc fault detection can be automatically updated according to the data stream,and achieve high recall rate while keep a low false alarm rate.Meanwhile,the designed architecture and analysis approach of arc fault detection reveal its effectiveness and generality in the feature extraction and anomaly detection for other time series.