CTNet: Improving the non-stationary predictive ability of remaining useful life of aero-engine under multiple time-varying operating conditions

Remaining useful life (RUL) estimation has been widely studied in prognostics and health management. Multiple time-varying operating conditions explicitly influence the measured performance parameters of the aero-engine, which results in non-stationarity and impedes the predictive ability of deep learning models. Previous datadriven studies primarily ignore the causality between operating condition descriptors and measured performance parameters, which is inconsistent with practical aero-engine engineering applications. To address this issue, we propose a causality Transformer network (CTNet), which improves the prediction stability of nonstationary series under multiple operating conditions. First, a novel graph structure is used to capture the connectivity relationship between operating condition descriptors and measured sensor nodes. Then the causality attention is proposed to inject additional causality information into the self-attention module without affecting the similarity computations of query and key values within the window. The causality attention can generate distinguishable attention for operating conditions, which properly release the predictive potential for nonstationary series. Extensive experiments are performed on public datasets (CMAPSS, N-CMAPSS), showing that CTNet outperforms existing state-of-the-art methods. We have also conducted experiments on the real-world quick access recorder (QAR) data, which validates the performance of the proposed model in perceiving timevarying operating conditions.