Hybrid and deep learning architectures for predictive maintenance: Evaluating LSTM, and attention-based LSTM-XGBoost on turbofan engine RUL

Abstract

Accurate prediction of a machines Remaining Useful Life (RUL) underpins modern, costeffective predictive-maintenance programmes. This paper proposes a two-stage hybrid pipeline that couples sequence learning with tree-based residual modelling. In stage 1, 50-cycle windows of NASA C-MAPSS sensor data (FD001 and FD004 subsets) are processed by a bi-layer Long Short-Term Memory (LSTM) network equipped with an attention mechanism; attention weights highlight degradation-relevant time steps and yield a compact, interpretable context vector. In stage 2, this vector is concatenated with four statistical descriptors (mean, standard deviation, minimum, maximum) of each window and passed to an extreme gradient-boosted decision-tree regressor (XGBoost) tuned via grid search. Identical preprocessing and earlystopping schedules are applied to a baseline LSTM for fair comparison. The attention-LSTM–XGBoost model lowers Mean Absolute Error (MAE) by 9.8 % on FD001 and 7.4 % on the more challenging FD004, and reduces Root Mean Squared Error (RMSE) by 8.1 % and 5.6 %, respectively, relative to the baseline. Gains on FD004 demonstrate robustness to multiple fault modes and six operating regimes. By combining temporal attention with gradient-boosted residual fitting, the proposed architecture delivers state-of-the-art accuracy while retaining feature-level interpretability, an asset for safety-critical maintenance planning.