Series Arc Fault Identification Method Based on Wavelet Transform and Feature Values Decomposition Fusion DNN

In low voltage distribution systems, AC series arcing is highly random and the fault characteristics are influenced by the type of loads. As the variety of loads connected to the system increases, a standard for uniform detection between varieties of loads is difficult to find. For such issues, a neural network algorithm based on wavelet analysis and feature value decomposition is proposed. Firstly, the DB5 wavelet base is used to decompose the collected data into 5 layers of wavelet and the wavelet coefficients of the first layer are extracted for the con-struction of the Hankel matrix. The first feature extraction and data compression are completed. The matrix is decomposed by Eigenvalue Decomposition (EVD) and used to construct the eigenvector sigma, and the second feature extraction and data compression are completed. The mean value d1, the root mean square value d2 and standard deviation value d3 of sigma are extracted, the third feature extraction and data compression are completed. The three values are taken as input to the neural network to train the Deep Neural Network (DNN) fault detection model, and the compressed input scale is only 1 x 3. The DNN fault detection model built by this paper is characterized by low complexity and high timeliness. The complexity of the model is reduced in terms of driving data and network structure, and the model can be trained in just 33 s. The model is extremely time efficient because it monitors faults in terms of current half-cycles. The experimental results show that, within the permissible range of GB14287.4-2014 standard, the model has good detection effect on both training set loads and non-training set loads, and the overall recognition rate can reach 98.7%.