The Dual-Encoder Transformer for Prediction of Aero-Engine Remaining Useful Life with Uncertainty Quantification

Remaining Useful Life (RUL) prediction plays a crucial role for the Prognostics and Health Management (PHM) of aero-engines. However, existing deep learning methods for RUL prediction usually focus on overall prediction accuracy while neglecting the case-wise prediction reliability. This paper proposes a Dual-encoder Transformer (DT) model with the capability to quantify prediction uncertainty. The proposed method utilizes temporal and sensor encoders to handle temporal dependencies and sensor feature dependencies in multi-dimensional time series data, respectively. It adaptively learns more important feature information through the Gated Fusion Unit (GFU) and extends the model with variational inference to achieve effective uncertainty estimation. The proposed model is validated on the widely used NASA C-MAPSS dataset. Compared to the alternative models, the DT model demonstrates either superior or similar overall prediction accuracy. Moreover, the DT model effectively quantifies uncertainty, enabling reliable RUL predictions.