An attention-augmented BiLSTM-LSTM framework with adaptive noise for turbofan engine remaining useful life prediction

Abstract Unplanned failures of aircraft engines pose severe threats to flight safety and result in substantial economic losses, making accurate prognostic maintenance critical for the aviation industry. To address the challenges in degradation modeling and improve reliability, this study proposes a novel deep temporal modeling framework that features architectural decoupling and dynamic regularization for the remaining useful life (RUL) prediction of turbofan engines. Specifically, a bidirectional long short-term memory is employed as a streamlined encoder to strictly capture bidirectional temporal contexts, while a multi-head attention mechanism is dedicated to feature re-weighting, thereby enhancing the representation of critical degradation segments. Subsequently, an LSTM-based predictor is used to model degradation trends and estimate RUL. Furthermore, to bolster the model’s robustness against time-varying industrial noise, an adaptive Gaussian noise perturbation strategy is introduced. Experimental results on the C-MAPSS turbofan engine dataset demonstrate that the proposed method outperforms several representative baselines in terms of prediction accuracy; furthermore, extended validation on the XJTU-SY rolling bearing dataset confirms its superior generalization performance, validating its potential for effective engineering applications.