Design of fault diagnosis for medium-voltage DC full-electric propulsion systems in vessels based on CNN–BiLSTM–Transformer

Medium-voltage DC full-electric propulsion technology in vessels offers high energy density, small space occupation, and flexible system layouts, making it the main direction for the development of integrated power systems in ships worldwide. However, medium-voltage DC full-electric propulsion systems operate in harsh environments with high temperatures, humidity, limited space, and extreme power shock changes. These conditions lead to complex system structures, high component coupling, and frequent operating state changes, increasing the probability and impact of failures and causing propagation of failures, thereby endangering the stable operation of the overall vessel system.

A fault diagnosis technology based on convolutional neural network (CNN)–bidirectional long short-term memory (BiLSTM)–Transformer is developed. To improve the effectiveness of fault feature extraction, a CNN-based single-point feature cascade network is constructed to achieve in-depth extraction of spatial details of fault signals at a specific moment. Furthermore, to improve the accuracy of fault diagnosis, a BiLSTM-based multipoint feature dependency network is designed to effectively learn the feature dependency relationship of fault signals between multiple time points through gate control and bidirectional temporal learning mechanisms. Next, a Transformer-based sequence-feature parallel processing network is established and a self-attention mechanism is used to characterize contextual relationships in fault evolution, improving diagnostic accuracy. Multihead attention design enables parallel processing of input sequence features, enhancing real-time fault diagnosis performance.

Comparative experiments with commonly used diagnostic models, including GRU, LSTM, BiLSTM, and CNN, were conducted, analyzing performance indicators such as average diagnostic accuracy, recall, F1 score, and diagnostic time. The proposed method demonstrated high diagnostic accuracy and the shortest diagnostic time.

Compared with existing mainstream fault diagnosis methods, the proposed method can achieve excellent diagnostic performance across various fault modes, providing strong technical support for the safe operation of medium-voltage DC full-electric propulsion systems in vessels.