In electric vehicle circuit systems, DC series arc faults frequently occur at loose contact points or damaged line connections, leading to hazards such as fires and explosions. To study series arc faults in electric vehicles, an experimental platform specifically designed for electric vehicle arc faults was established. The causes and patterns of changes in the main current waveforms under various operating conditions were analyzed in detail. Given the stringent real-time requirements for arc fault detection, this study employed a lightweight convolution operation known as depthwise separable convolution to develop an arc fault detection network. This network achieved the detection of arc faults and the identification of fault lines in electric vehicles. To address the limitations of depthwise separable convolution in feature extraction in low-dimensional spaces, this study made improvements and proposed a convolutional operation with superior feature expression: group separable convolution. Ultimately, a progressive ladder structure was implemented, where the number of convolutional kernels within each group of the group separable convolution gradually decreases from the shallow layers to the deeper layers of the network. This approach streamlined and optimized network architecture while ensuring detection accuracy. Further enhancements involved optimizing the convolution kernel size within the detection model and integrating a lightweight attention mechanism into the architecture. A dynamic learning rate adjustment strategy was also applied during the model′s training process. Through these optimization measures, both operational efficiency and detection accuracy were systematically improved. The model achieved a detection accuracy rate of 96. 76% and exhibited good generalization and anti-interference capabilities.