Toward Decision Transformer for Remaining Useful Life Prediction: A Sequence Modeling-Based Approach

Accurate prediction of Remaining Useful Life (RUL) plays a vital role in predictive maintenance, where anticipating equipment degradation can prevent costly downtime and safety-critical failures. This paper explores a Decision Transformer-based architecture for RUL prediction, leveraging self-attention mechanisms to model long-term dependencies in multivariate time-series sensor data. Unlike conventional recurrent and convolutional approaches, the proposed framework adapts the sequence modeling principles of the Decision Transformer, using a bidirectional Transformer encoder with joint RUL regression and maintenance urgency classification, to treat degradation trajectories as structured temporal sequences, enabling the model to capture long-range temporal dependencies without recurrence constraints. A unified preprocessing pipeline incorporating Savitzky-Golay filtering, sensor selection, min-max normalization, and piecewise linear RUL labeling supports consistent and comparable evaluation across studies. Experimental evaluation on the NASA C-MAPSS benchmark dataset demonstrates effectiveness across all four subsets (FD001–FD004), encompassing diverse operating conditions and fault modes. The Decision Transformer achieves better performance compared to the selected deep learning baselines, with RMSE values of 3.12 (FD001), 2.28 (FD002), 3.46 (FD003), and 3.18 (FD004) cycles, and PHM08 scores of 23, 48, 30, and 70, respectively, representing reductions of up to 60% in RMSE and 77% in PHM score compared to recent deep learning methods. The findings demonstrate that transformer-basedsequence modeling offers measurable improvements in RUL prediction accuracy and late-prediction risk reduction, while also highlighting practical challenges related to computational cost, training variability, and interpretability in safety-critical applications.