AI-Driven Predictive Maintenance and Performance Optimization of Marine Diesel Engines Using Multivariate Machine Learning
Abstract
Marine diesel engines operate under severe and continuously varying load conditions, and any unplanned breakdown of the main propulsion engine leads to heavy economic loss, delayed cargo delivery, and, in the worst case, safety hazards to the crew and vessel. Traditional time-based maintenance schedules are often conservative and cannot capture the actual health condition of the engine. In the present work, an artificial-intelligence-driven predictive maintenance framework is developed and experimentally validated on a medium-speed, four-stroke marine diesel engine test-bed of 450 kW rated power. A multivariate sensor suite consisting of cylinder combustion pressure, individual cylinder exhaust gas temperature, fuel injection rail pressure, lubricating oil pressure and temperature, cooling water temperature, crankcase vibration, engine speed, and load was employed to acquire run-to-failure data for five health states, namely healthy condition, fuel injector nozzle wear, exhaust valve leakage, clogged fuel filter, and piston ring wear. Six machine learning algorithms—random forest, XGBoost, support vector machine, artificial neural network, k-nearest neighbours, and a logistic regression baseline—were trained and compared for multi-class fault classification, while gradient-boosted regression and random forest regression were evaluated for remaining useful life (RUL) estimation. The random forest classifier achieved a classification accuracy of 96.8% with an F1-score of 0.967, and the XGBoost regressor predicted the RUL with a root mean square error of 18.4 operating hours and a coefficient of determination of 0.962. Feature importance analysis revealed that individual cylinder exhaust gas temperature deviation, combustion pressure peak, and high-frequency vibration energy were the most discriminative predictors of engine degradation. The proposed framework successfully detected developing faults up to 140 operating hours before functional failure, demonstrating its suitability for onboard condition monitoring and for the performance optimization of marine diesel engines.