Robust Marine Propulsion Performance Forecasting Using Robust Heterogeneous Graph Transformer Network Model

Abstract

Precise assessing of marine engine performance under dynamic load circumstances is significant for modern marine vessels to ensure fuel efficiency, operational dependability and preventative maintenance. In this paper a Robust Heterogeneous Graph Transformer Network (HGTN) is proposed for predicting marine engine performance under dynamic load variations. The raw data incorporates load-dependent indicators, engine log histories, and raw high-frequency sensor readings. The heterogeneous graph nodes characterize engine subsystems, operational parameters; edges encode their correlations and dynamic connections. The HGTN uses multi-type attention and adaptive meta-path learning to capture nonlinear effects arising from fuel injection variability, turbocharger dynamics, thermal fluctuations, and shaft load changes. The proposed HGTN technique accurately predicts brake power, SFOC, exhaust temperature, and vibration signatures, attaining an MSE of 0.0052, RMSE of 0.0719, MAE of 0.032, and R² of 98%. Results confirm the structure efficiency in identifying performance degradation and improving engine tuning strategies.