Adaptive uncertainty-aware hybrid digital twin architecture for predictive diagnostics of ship power plants

Abstract

Reliable prediction of the technical condition and remaining useful life of ship power plants (SPPs) remains a challenging task because operating conditions continuously change, degradation mechanisms are highly nonlinear, and sensor measurements are inevitably affected by uncertainty. Although hybrid digital twins (HDTs) have significantly expanded the capabilities of predictive maintenance, many existing implementations still employ fixed interaction schemes between physics-based and data-driven models, reducing their adaptability when the reliability of individual information sources varies during operation. The study develops a HDT architecture that combines a deterministic physical model, a CNN–LSTM prediction module, and a cognitive simulation model (CSM) within a single diagnostic framework. Instead of maintaining a fixed balance between analytical and empirical models, the architecture continuously adjusts their contribution according to the estimated uncertainty of the physical model. This makes it possible to retain physical consistency whenever analytical knowledge is reliable while allowing data-driven prediction to dominate when model uncertainty increases. The cognitive simulation component complements numerical prediction by explicitly representing causal relationships between degradation processes and diagnostic conclusions, providing engineering explanations that are difficult to obtain from conventional deep neural networks. Performance of the developed framework was investigated using hardware-in-the-loop (HIL) simulations reproducing representative operating regimes of SPP subsystems, including progressive degradation and stochastic sensor disturbances. Under these conditions, the hybrid architecture reduced remaining useful life prediction error by 32–36% relative to standalone physics-based models and by up to 19% compared with conventional deep-learning solutions. Fault classification reached an FI-score of 0.89 while maintaining stable operation under sensor noise levels of up to 30%. Real-time inference below 20 ms satisfies the requirements of onboard automation systems, whereas asynchronous federated learning enables scalable deployment across distributed fleets without increasing onboard computational load. The combination of adaptive uncertainty-aware model fusion, physics-informed learning, and cognitive reasoning extends the capabilities of current HDT technologies beyond conventional hybrid diagnostic frameworks. Besides improving prediction accuracy and robustness, the architecture enhances transparency of maintenance decisions and demonstrates practical feasibility for intelligent SPPs, with an estimated reduction in annual operating costs of approximately 27%. These results indicate that the developed framework represents a practical step toward reliable and explainable predictive maintenance of next-generation maritime energy systems.