Parallel time-series mixer network enhanced by slicing procedure and attention mechanism for remaining useful life prediction

Abstract

The prediction of an aero-engine’s Remaining Useful Life (RUL) is a critical task in Prognostics and Health Management (PHM). While deep learning has shown promise, existing models often struggle with the variable-length time-series sequences common in aero-engine monitoring and fail to effectively capture complex temporal patterns and cross-variate information. This paper presents an advanced approach for predicting the RUL of aero-engines through the proposal of the CoSO-pTSMixer-SGA network. The network includes three key blocks: the Concentrating and Slicing Operator (CoSO), the parallel Time-Series Mixer (pTSMixer), and the Scalable Global Attention (SGA), designed to handle variable-length data flexibly and enhance feature extraction. Extensive experiments on the C-MAPSS dataset demonstrate that CoSO-pTSMixer-SGA achieves state-of-the-art performance, with a 6.4% reduction in RMSE and a 3.4% reduction in Score compared to other leading methods. The network is particularly effective under complex operating conditions, outperforming others by up to 10.4% in key datasets’ RMSE. Ablation studies validate the contributions of each element, and a novel RMSE-60 metric is introduced for a more targeted evaluation. The CoSO-pTSMixer-SGA network offers a flexible and precise solution for real-world RUL estimation tasks.