Improving Remaining Useful Life Prediction with Synthetic Data and Black Box Adversarial Reprogramming

Predictive maintenance is essential in modern manufacturing for improving the reliability and efficiency of machinery. A central challenge lies in accurately estimating the remaining useful life (RUL) of critical components such as ball bearings, especially when real-world labeled data is scarce. This study addresses this challenge by combining physics-informed simulation with a regression-based transfer learning approach. A 5-degree-of-freedom simulation model was developed to replicate the lifecycle of ball bearings under varying operating conditions, generating synthetic run-to-failure vibration data. The synthetic signals were segmented and processed using statistical and time-frequency features, with dimensionality reduction performed via statistical relevance and multicollinearity filtering. Tree-based regression models—Random Forest, Gradient Boosting, and XGBoost—were trained on the synthetic data and optimized via Bayesian hyperparameter tuning. To bridge the domain gap, the Black Box Adversarial Reprogramming (BAR) algorithm was applied to adapt these models for real-world data without retraining. Performance was evaluated across twelve structured transfer tasks using RMSE, MAE, and R 2 metrics. Results showed that BAR-enhanced models consistently outperformed their unadapted counterparts, with the Gradient Boosting Regressor achieving an RMSE of 10.25 and R 2 of 0.87 on the best transfer task. A statistical evaluation over 100 iterations confirmed the effectiveness of the approach, while highlighting its stochastic nature. Overall, the integration of physics-based data generation with BAR-based transfer learning offers a scalable and model-independent solution for accurate RUL prediction under data-constrained conditions.