In electric vehicles, the electrical contact points within the main circuit are susceptible to series arc faults as a result of poor contact and other factors, which poses a significant threat to the safety of vehicle occupants. First, this study constructed an experimental platform centered on the Geely Emgrand EV450 electric vehicle and undertook experimental investigations on arc faults in electric vehicles. With the power supply terminal voltage serving as the research focus, a one-dimensional convolution based arc fault detection model was developed. However, traditional deep learning models encounter two core challenges in practical applications. First, the models have poor interpretability, making it difficult to clarify the key basis for fault detection. Second, they have a large number of parameters and high computational complexity, making it difficult to meet the strict requirements for real-flortime and lightweight fault detection in electric vehicles. To address these issues, this study adopts a network structure search strategy integrating multi-objective optimization. which incorporates accuracy, interpretability indicators, and floating point operations into the search objectives of the network structure search. Through multidimensional trade-offs, adaptive optimization of the network structure is achieved, which effectively improves the initial performance of the model. Subsequently, a feature channel merging strategy was developed by integrating dynamic time warping, particle swarm optimization, and simulated annealing algorithms. Among these methods, Dynamic time warping can measure the similarity of outputs from different channels; Particle swarm optimization, with its global search capability, quickly locates potential optimal channel merging combinations; and Simulated annealing further enhances the rationality and effectiveness of channel merging. Using this strategy, an accurate, interpretable, and lightweight arc fault detection model for electric vehicles has been successfully developed. Finally, generalization analysis and comparative analysis with other detection methods confirm that the model exhibits excellent performance in electric vehicle arc fault detection.