Dual-branch remaining useful life prediction based on local pattern learning and functional trajectory reconstruction

Abstract

Remaining useful life (RUL) prediction is a core task in prognostics and health management because it supports maintenance scheduling and downtime reduction. Deep learning methods have improved temporal representation for RUL prediction, but degradation information extracted from fixed-length recent monitoring windows may still be affected by local variance, noise, and stochastic fluctuations. Similar current patterns may also correspond to different future degradation trajectories under heterogeneous degradation modes. To address this limitation, this study proposes a dual-branch RUL prediction framework that introduces functional trajectory reconstruction to complement local degradation-pattern learning. The local branch uses dynamic time warping retrieval and a pattern-segment convolutional network to learn current degradation patterns from historically similar segments. The functional trajectory branch reconstructs discrete observations into functional degradation trajectories and generates a long-horizon lifetime estimate from accumulated degradation history. The two estimates are then reconciled through lifetime-scale Bayesian fusion, which infers their relative weights and quantifies predictive uncertainty through posterior sampling to obtain RUL predictive intervals and reliability-oriented indicators. Case studies on benchmark aero-engine data and industrial machine-tool degradation data show that the proposed framework achieves competitive RUL prediction accuracy and quantifies predictive uncertainty.