Abstract: In response to the problem of inaccurate identification of fault arcs in the current power grid, a fault arc detection method based on multi-channel time-frequency feature extraction and lightweight convolutional neural network is proposed. Firstly, high-speed ADC is used to sample the current signal on the live line, and different band-pass digital filters are used to obtain the target sampling signal. Secondly, different frequency channels are designed to extract multidimensional time-frequency features. Thirdly, multidimensional features are input into convolutional neural networks to mine their relationship with fault arcs. Considering the problem of large amounts of high-speed feature data, a lightweight convolutional neural network based on pruning algorithm is designed to ensure real-time model calculation. Finally, numerical analysis was conducted on different application scenarios. The results show that the proposed method has good superiority and robustness, strong generalization ability, and is particularly suitable for mixed scenarios.