BootBoGs-TS: Bootstrap-Based Hyperparameter Optimization for LSTM Models in Remaining Useful Life Prediction

Neural networks, particularly Long Short-Term Memory (LSTM) models, have become prominent for Remaining Useful Life (RUL) prediction due to their ability to capture temporal degradation patterns from multivariate sensor data. However, their predictive performance is highly sensitive to hyperparameter configurations, and conventional optimization methods such as Grid Search or Bayesian Optimization often become inefficient or unstable when applied to high-dimensional, sequential data. To address this, we introduce BootBoGs-TS, a hybrid hyperparameter optimization framework for LSTM-based RUL prediction from time series. BootBoGs-TS extends our previously proposed BootBoGs algorithm, originally developed for classification tasks, by integrating a unit-level Bootstrap that resamples complete engine trajectories to preserve long-term degradation integrity with an out-of-bag (OOB) evaluation scheme, and introducing a performance-based filtering step. This filtering step retains only stable configurations whose validation scores lie within a median ± k$$\cdot$$MAD interval, where MAD denotes the median absolute deviation computed from the empirical distribution of OOB S-scores. This procedure discards unstable or extreme hyperparameter settings and thereby reduces the search space in a statistically principled manner. The proposed framework sequentially combines Bayesian Optimization to define the initial hyperparameter ranges, bootstrap to explore variability in model performance while preserving temporal dependencies, and reduced-grid search for final refinement. We evaluate BootBoGs-TS on the C-MAPSS benchmark dataset. Experiments are conducted on the four C-MAPSS subsets (FD001–FD004). Detailed comparative experiments and ablation studies are performed on the FD004 subset, which represents the most complex scenario. Results show that BootBoGs-TS improves both efficiency and stability, achieving lower RMSE and S-score values compared to standard baselines.