Layer-by-layer Visual Analysis and Optimization of Arc Fault Detection Models for Electric Vehicle Charging Devices

Under the guidance of the "double carbon" target and the transformation of the green economy, China's new energy electric vehicle industry has ushered in unprecedented development opportunities. As the battery capacity of electric vehicles continues to increase, fire accidents in electric vehicles and their charging devices are frequent, and the risk of electricity safety is increasing. There is an urgent need to carry out research on electric vehicle and charging fire protection devices to provide protection for the safe and stable operation of the distribution network. This paper proposes a layer-by-layer visualization and optimization method for arc fault detection model based on the interpretive analysis of the network model, and investigates the importance of the signals in each frequency band of the arc current spectrum by selecting the Gradient-weighted Class Activation Mapping (Grad-CAM), and determines the sensitive frequency band of the network model for arc currents from 5 kHz to 60 kHz. The contribution of different convolutional layers in fault feature identification was also investigated using the Shapley Additive Explanation (SHAP) to determine the optimal number of layers for the network model to be 12. The results show that the optimized model has an accuracy of 98.56% for the training set and 97.95% for the validation set compared to the original model, verifying the feasibility of the network model optimization method.