A hybrid semi-supervised approach for remaining useful life estimation using masked attention-based representations and actor–critic sliding-mode residual correction

The implementation of predictive maintenance necessitates the development of models that possess the capability to accurately estimate the remaining useful life (RUL) of industrial assets under variable operating conditions, sensor noise, and limited availability of labeled data. This problem is especially salient in scenarios where there are substantial volumes of monitoring signals but a paucity of complete trajectories to failure. To address this limitation, this work proposes a semi-supervised hybrid approach for estimating RUL in turbofan engines using the C-MAPSS dataset. The architecture integrates three complementary stages: self-supervised pre-training via masked reconstruction and temporal consistency, supervised prediction based on attention mechanisms, and an actor-critic residual correction guided by a reward inspired by sliding mode control. The primary innovation of the method is its formulation of RUL estimation not only as a regression problem but also as a hybrid scheme of representation, prediction, and robust residual control. The findings indicate that there has been a 18.21%, 30.77%, 7.54%, and 23.53% reduction in root mean square error (RMSE) for FD001, FD002, FD003, and FD004, respectively, when compared to the most advanced contemporary methodologies. Furthermore, a decline in critical errors associated with late predictions was demonstrated, with the score decreasing by 12.07%, 59.70%, 11.54%, and 57.54%, respectively. The statistical analysis confirmed the model’s high explanatory power, with R 2 values ranging from 0.9029 to 0.9403 and strong Pearson correlations ranging from 0.9331 to 0.9703. These findings support the precision, low bias, and predictive stability of the proposed approach.