Predictive Health Monitoring of a Naval Propulsion Plant Using Multi-Output Random Forest Regression

Abstract

The reliability of naval propulsion systems is closely related to the degradation of critical gas turbine components, particularly the compressor and turbine. This study proposes a machine-learning-based predictive health monitoring framework for a naval propulsion plant using the Naval Propulsion Plant dataset. The framework was developed from 11,934 observations, 14 operational variables, and 2 degradation targets, namely the compressor decay state coefficient and turbine decay state coefficient. A multi-output Random Forest regressor with 500 trees was employed to simultaneously predict both degradation indicators. The methodology included data preprocessing, exploratory data analysis, model development, performance evaluation, feature importance analysis, and health index formulation. The model achieved an MAE of 0.1899, RMSE of 0.2992, and R² of -0.0484 for the compressor decay state coefficient, while the turbine decay state coefficient achieved an MAE of 0.1604, RMSE of 0.2667, and R² of 0.1814. To improve practical interpretability, the predicted degradation outputs were transformed into a health index and classified into GREEN, YELLOW, and RED conditions. The results showed that 2,387 testing observations were classified as GREEN. The novelty of this study lies in integrating multi-output degradation prediction, feature-based interpretability, and health-index-based condition classification within a single predictive maintenance framework for naval propulsion systems.