A cyber-resilient framework for data-driven ship propulsion power prediction

The digitalization of maritime operations has accelerated the adoption of data-driven models for propulsion power prediction and fuel optimization. However, the increasing reliance on real-time sensor streams introduces critical vulnerabilities to cyber–physical threats, where sensor faults or data manipulation can compromise ship control systems. Existing models often operate under the assumption that input data is trustworthy, making them fragile when exposed to operational anomalies. To address these challenges, this study proposes a Cyber-Resilient Framework designed to maintain trustworthy operations even when input data is compromised. Instead of relying on standard filtering, the proposed architecture secures the modeling pipeline by enforcing physics-based constraints to reject impossible values and utilizing density-based clustering (DBSCAN) to isolate operational outliers. These verified patterns are then processed by an Artificial Neural Network (ANN) optimized for robust estimation. The framework was validated using high-frequency (5 Hz) operational data collected from a post-Panamax containership over a two-year period. The system was trained on Europe–Asia trades and subsequently tested on unseen Asia–America routes to evaluate its generalization capability. The results reveal a critical trade-off where Gradient Boosting achieves high accuracy on familiar routes but fails to generalize to new environments, exhibiting deviations exceeding 20%. In contrast, the ANN maintains superior stability and physical consistency across different domains. This confirms that combining physical validity checks with robust learning algorithms provides a necessary safeguard mechanism for secure ship operations.