Transfer learning-enabled digital twin for online iteration of marine engine prediction models

Abstract

The lifecycle management of marine engine performance is a complex and dynamic process, facing numerous challenges in real-time monitoring and predictive maintenance. This study proposes a novel methodology that integrates transfer learning and engine performance degradation laws to enable the online iteration of the digital twin model for marine engines throughout their entire lifecycle. Initially, a deep learning neural network that integrates Snake Optimization, Convolutional Neural Networks, and Bidirectional Long Short-Term Memory networks (SO-CNN-Bi-LSTM) is employed to develop a performance prediction model for the engine. Then, based on engine wear and bench test data, a nonlinear fitting relationship between running time, wear, and performance degradation is derived, allowing for the generation of a dynamically updated dataset of engine performance parameters. Finally, the digital twin model is iteratively optimized using a feature-model joint migration framework, achieving high-precision lifecycle performance predictions. Experimental results demonstrate that, under maximum wear conditions, the prediction errors for various performance parameters are significantly lower than those before the iteration, validating the effectiveness of the model in performance prediction. This approach establishes a solid theoretical foundation for the comprehensive operation and maintenance management of marine engines throughout their entire lifecycle.