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

Area of Science:

• Electrical Engineering
• Marine Engineering
• Artificial Intelligence

Background:

• Shipboard power systems are critical for vessel operation, but short-circuit faults pose significant risks.
• Existing fault diagnosis methods struggle with the complexity and dynamic nature of onboard conditions.

Purpose of the Study:

• To develop an accurate and interpretable framework for short-circuit fault diagnosis in shipboard power systems.
• To enhance the reliability and safety of marine electrical infrastructure.

Main Methods:

• A novel framework integrating a multi-scale Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model with Shapley value analysis.
• Utilizing relative changes in electrical signals for input representation.
• Employing multi-scale convolutions for temporal pattern extraction and LSTM for sequential modeling.
• Applying Shapley value analysis for feature contribution quantification and screening.

Main Results:

• Achieved an average diagnostic accuracy of 99.03 ± 0.20% in experiments on a Simulink-based shipboard power system dataset.
• Demonstrated competitive performance against baseline models (CNN, LSTM, LightGRU) in precision, recall, and F1-score.
• Exhibited superior robustness under noise conditions compared to existing methods.

Conclusions:

• The proposed framework offers accurate and interpretable short-circuit fault diagnosis for shipboard power systems.
• The integration of CNN-LSTM and Shapley values enhances diagnostic performance and provides insights into feature importance.
• This approach contributes to improved safety and stability of marine power systems.