Estimating the Remaining Useful Life (RUL) of aircraft engines plays a vital role in the field of prognostics and health management. In multi-dimensional time series regression tasks, accurately capturing both time series features and sensor features, as well as integrating these two types of features, poses a significant challenge for RUL prediction. The sensor features represent the weights of each sensor on the RUL prediction results. To overcome this challenge, we introduce a hybrid model based on a dual-attention mechanism. Initially, a temporal feature extraction block is applied to map the time-step dimension into a hidden representation space, facilitating the capture of complex temporal dynamics. These patterns are then refined using a multi-head self-attention mechanism. Subsequently, a sensor feature extraction block is applied to capture sensor-specific characteristics. Each sensor sequence is treated as a separate channel, compressed to derive sensor weights, and integrated to form global features that fuse temporal and sensor-level representations. Finally, RUL is estimated via a regression layer. The proposed method is demonstrated to be effective on the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) dataset. Compared with the state-of-the-art CTNet model, the proposed method achieves 7% and 9% gains in RMSE and Score, respectively, on the FD001 dataset.
He et al. (Thu,) studied this question.