Enhancing non-stationary feature learning for remaining useful life prediction of aero-engine under multiple operating conditions

Remaining useful life (RUL) estimation has been widely concerned, given its significant role in prognostics and health management of industry systems. This paper focuses on the non-stationarity of real-world aero-engine sensors data under multiple operating conditions. Previous methods accomplish the normalization operation to attenuate the inherent non-stationarity of raw series for better predictability, which can be less instructive for bursty RUL prediction tasks. To tackle the limitations of direct normalization, we propose a difference transformer network (DFormer). First, we present a series decomposition module to extract predictable components from different series, which can meticulously perceive the non-stationary information of each component. Then the difference attention is proposed to approximately obtain attention without normalization to maintain the non-stationarity of the time series. Furthermore, DFormer consists of two complementary parts: the encoder for incorporating the non-stationary information of the original series and the decoder for stable distribution of time series. We have conducted extensive experiments on the CMAPSS and N-CMAPSS datasets. The results show that DFormer can significantly improve prediction performance compared to existing state-of-the-art methods. In addition, the non-stationary feature learning performance of the model is validated on QAR engineering data, demonstrating the capability of practical engineering applications.