Gas Turbine Decay Prediction for Condition-Based Maintenance of Naval Propulsion Systems: An Extrapolative Generalization Assessment and a Cost-based Decision Perspective

As naval maintenance shifts to condition-based maintenance (CBM), estimating gas-turbine degradation from sensor data is essential. Using the public CODLAG frigate dataset, this study moves beyond model comparison and contributes along three axes. First, because the data is a deterministic full-factorial simulator grid, conventional random cross-validation overstates accuracy: six nonlinear models score R²≥0.97, yet all models degrade sharply under extrapolation to unseen operating points or decay levels; repeated runs over five seeds show a multilayer perceptron (MLP) robustly retains a positive R² only under decay-level extrapolation (0.94±0.03 for the compressor decay coefficient), while under operating-point extrapolation even its R² is unstable across seeds (mean -0.08±0.51)—and rankings reshuffle. Second, we link prediction to decisions through cost-based model selection, sensor-portfolio optimization, conformal prediction intervals, and inspection-policy optimization; a policy designed under optimistic (interpolation) error costs about 46.6% more under realistic error. Third, beyond Gaussian noise we assess structural noise, showing MLP is fragile to outliers and missingness while random forests are robust. Significance is confirmed by Friedman and Nemenyi tests. The study recalibrates achievable CBM performance and offers decision-oriented guidance.