Short-circuit fault diagnosis for shipboard power grid based on multi-scale depthwise separable convolution and lightweight channel attention

Objectives To address the multiscale characteristics of short-circuit fault signals in shipboard power systems and the high complexity of existing diagnostic models, a lightweight fault diagnosis method is proposed. Methods Firstly, a short-circuit fault simulation model is built in Matlab/Simulink. Secondly, an MDSC-LCA diagnosis network integrating a multi-scale depthwise separable convolution (MDSC) module and a lightweight channel attention (LCA) module is designed, where the MDSC module extracts multiscale fault features and the LCA module enhances key discriminative information. Finally, the network is used to accurately identify normal conditions and ten types of short-circuit faults. Results Under 11 operating conditions, the proposed model achieves 99.51% accuracy and a 99.37% macro-F1 score. Compared with typical benchmark models, it improves diagnostic performance while significantly reducing parameters and computational cost, and still maintains 95.61% accuracy under 10 dB noise. Conclusions The proposed method balances diagnostic accuracy, model lightweight design and noise robustness, providing a technical reference for short-circuit fault diagnosis of shipboard power systems.