Multi-Scaled Lightweight Neural Network Based ARC Fault Diagnosis: Feature Combination and Classification

As electrification progresses, series arc faults have become a vital problem in power systems. To quickly and accurately classify the arc fault, this paper proposes a multiscaled lightweight neural network that incorporates multi-scale features from original signals into high dimensions to improve classification performance. The proposed methodology unfolds: First, the original signals are inputted into multi-layer convolutional neural networks with different kernel sizes, in which the residual structure is added for further feature extraction. Then the multi-scaled features are extracted automatically by the network and combined into a higher-dimensional latent space. The latent space helps models find the necessary detailed features. At last, the combined features are utilized as the key factors of the classification layers for a more accurate arc fault classifier. The superiority of MSNN is demonstrated in analytical experiments by comparing it with other deep learning methods and conventional methods. The occlusion experiments and ablation experiments highlight the importance of multi-scale feature filters and hyper-parameters in MSNN. The robustness and scalability of MSNN are also discussed in ablation experiments.