STARNet: Stacked Transfer-Aware for Robust Remaining Useful Life Prediction for C-MAPSS Multi-Regime Engines

Predictive Maintenance (PdM) plays a critical role in reducing unplanned downtime and extending asset longevity in industrial systems like aircraft engines, manufacturing equipment, and energy infrastructure. Proper estimation of the Remaining Useful Life (RUL) of parts allows for preventive maintenance scheduling and increased operational efficiency, safety, and cost savings. However, the development of robust RUL prediction models remains challenging due to high variability in operating conditions and fault modes. Existing deep learning models tend to demonstrate exceptionally good performance on individual datasets but are unable to generalize across diverse regimes without retraining, which makes them less reliable and scalable for real-world deployment. To overcome existing constraints, this study proposes STARNet, a robust RUL prediction framework that incorporates a multi-scale Conv–BiLSTM with attention pooling, window statistics fusion, and mixture-of-experts encoding. The framework incorporates phase-aware normalization and delta features to improve adaptability across multi-regime subsets (FD002/FD004) of the NASA CMAPSS dataset. Furthermore, self-supervised pretraining and targeted transfer learning between subsets are employed to improve representation quality and reduce overfitting, while Monte Carlo dropout provides reliable uncertainty-aware inference. A multi-seed ensemble with nonlinear stacking ensures stable and accurate predictions across all four subsets. The model is thoroughly tested using a multi-metric approach consisting of Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Coefficient of Determination (R²) and NASA Score. In this study, STARNet, a robust RUL prediction network that illustrates steady performance gains across all subsets of C-MAPSS, with average RMSE and score improvements of 5.91% and 13.23%, respectively, has been proposed over existing state-of-the-art models. The findings consistently surpass traditional and cutting-edge deep learning benchmarks, confirming the efficacy, generality, and deployment suitability of the developed framework for real-world predictive maintenance tasks.