New Generative End-to-End RUL Prognostics of Bearings Based on Pretrained Conditional Temporal Diffusion Network

Accurate remaining useful life (RUL) prognostics of bearings is beneficial to prolong service life and reliability maintenance of mechanical system. Inspired by non-equilibrium thermodynamics, a novel generative end-to-end RUL prognostics method, i.e., the pretrained conditional temporal diffusion network (PCTDiffNet), is proposed for rolling bearings. Firstly, the RUL ratio lines are designed with the physical parameterized sequence, then feature extraction, dimension reduction and feature transformation are conducted in pretrained conditional encoder module. Furtherly, new U-Net residual block in diffusion model with condition vector is designed, and the differential loss function is constructed to avoid the issues of pattern collapse and convergence. Eventually, following the diffusion points convergence of the reverse denoising process in PCTDiffNet, the RUL of bearing can be estimated. Two run-to-failure datasets are employed to verify the superiority of the proposed method compared with state-of-the-art benchmarks.