Rolling bearing remaining useful life (RUL) prediction still faces major challenges, including the difficulty of reliably locating the initial degradation point under noisy and long stationary conditions, the lack of discriminative multi-scale features that capture both transient and steady-state degradation patterns, and the limited ability of existing models to fuse heterogeneous temporal information. To address these issues, this study proposes a framework integrating robust regression with a dynamic 3σ-based threshold (RRDT) for first prediction time (FPT) identification and a dual-branch coordinated deep temporal network (DCDTN) for RUL estimation. By accurately detecting the degradation onset amid noise, RRDT provides a more trustworthy starting point for RUL modeling. On the modeling side, DCDTN comprises two feature pathways: a wavelet scattering transform (WST) branch whose coefficients are adaptively refined by the scattering feature adaptive processor (SFAP) unit to suppress redundancy, and a raw-signal branch that employs multi-scale convolutions and a depthwise-separable feed-forward network based temporal convolutional network (FN-TCN) to mine multi-scale deep features. A coordinated temporal fusion module combining Peephole BiConvLSTM and efficient multi-scale attention (EMA) further enhances the representation of complex degradation dynamics, producing the final RUL estimate. Experiments on the XJTU-SY and PHM2012 datasets demonstrate that the proposed method achieves significantly higher prediction accuracy than conventional RUL models. Overall, the framework effectively addresses key obstacles in bearing RUL prediction and shows strong potential for industrial deployment in equipment health monitoring and early-warning scenarios.
扫码关注我们
求助内容:
应助结果提醒方式:
