Abstract: As a critical transmission component，the operational stability of water-lubricated bearings is essential for the safety and efficiency of ships. To address the potential failures of ship water-lubricated bearings during operation and the unreliability of diagnostic results based on single signals, this study proposes a Transformer-1D CNN model that integrates dynamic and tribological information. The model leverages the self-attention mechanism to effectively capture long-range dependencies and combines the local feature extraction capabilities of convolutional networks. It enables comprehensive analysis of multi-sensor data, such as friction coefficient, temperature, and vibration signals. Additionally, a cross-condition state identification method based on transfer learning is introduced. Experimental data are obtained from a high-performance water-lubricated bearing comprehensive testing platform. The results show that for state identification tasks under fixed conditions, the proposed model achieves an accuracy of 99.97% on the test set. For cross-condition identification tasks, the feature-based transfer learning approach outperforms the parameter-based method，achieving accuracies of 94.91%, 91.38%, and 96.96% in three cross-condition tasks, representing improvements of 17.43%, 18.03%, and 13.59%, respectively, compared to non-transfer learning methods. These findings demonstrate the effectiveness of the proposed Transformer-1D CNN model and its integration with transfer learning in enhancing the accuracy of fault diagnosis for ship water-lubricated bearings, providing a reliable safeguard for ship operational safety.

Key words: water-lubricated bearings, state identification, deep learning, transfer learning