A dual-path architecture based on time series decomposition and degradation correction for remaining useful life prediction of aero-engine

Abstract

Accurately predicting remaining useful life (RUL) is crucial to improving the reliability of aero-engines. Although data-driven approaches have become a powerful tool for RUL prediction, most studies only focus on improving model structure and ignore the characteristics of the data itself. Due to changes in the environment and operating conditions, monitoring data often show non-linearity, non-stationarity and high noise, which affects the accuracy of prediction. Therefore, this paper proposes a dual-path prediction architecture based on time series decomposition and degradation correction (DP-TSDDC) to mine fine-grained information in the data in a targeted manner. First, the time series decomposition is used to extract trend components and non-stationary components from the data. Then, the dual-path architecture is used to select networks in a targeted manner according to the characteristics of each component. For the trend component, the temporal convolutional network is introduced to extract the engine’s long-term degradation features; for the non-stationary component, the Patch idea is used for data preprocessing, the long short-term memory is used within the Patch to capture detail features, and the multi-head attention is used between the Patch to capture global correlation features. Then, based on the gated linear unit and the nonlinear relationship between each feature, the features are enhanced and adaptively fused. At the same time, the degradation correction constraint is established based on physical knowledge to reduce the impact of data noise on the prediction results. Finally, the prediction effect of the DP-TSDDC method is verified by experiments on the CMAPSS and N-CMAPSS datasets.