Multi-Scale CNN-LSTM for Short-Circuit Fault Diagnosis of Shipboard Power System

Shipboard power systems are essential to the safe and stable operation of marine vessels, while short-circuit faults may lead to equipment damage and system interruption under complex onboard operating conditions. To improve fault diagnosis performance in this setting, this study proposes an interpretable short-circuit fault diagnosis framework that combines a multi-scale CNN-LSTM model with Shapley value analysis. Relative changes between pre-fault and fault-state electrical signals are used to construct the input representation, which helps characterize fault-related variations more effectively. The multi-scale convolution branches extract patterns associated with different temporal ranges, and the LSTM layer further models their sequential dependence. Shapley value analysis is introduced to quantify the contribution of voltage- and current-related features, identify the most informative inputs, and support feature screening. Experiments on a Simulink-based shipboard power system dataset show that the proposed method achieves competitive fault diagnosis performance compared with baseline models, including CNN, LSTM, and LightGRU. Under repeated runs, the proposed framework attains an average diagnostic accuracy of 99.03 ± 0.20%, while also maintaining strong precision, recall, and F1-score performance. Under the tested noise conditions, it shows better robustness than the comparison methods. These results indicate that the proposed framework can provide accurate and interpretable fault diagnosis for shipboard power systems.