Equipment remaining useful life prediction via dynamic label recalibration and gated multi-scale U-Net framework

Abstract Accurate prediction of industrial equipment remaining useful life (RUL) requires reliable degradation labels and models capable of capturing multi-scale temporal dynamics and interactions among heterogeneous sensors. In practical environments with variable operating conditions, strong noise, and heterogeneous degradation behaviors, widely used single-threshold labeling schemes induce label shift and phase misalignment across units, while conventional fusion architectures operate at fixed temporal scales and treat sensor channels largely independently, limiting robustness and prediction accuracy. To address these issues, this paper proposes an adaptive RUL prediction framework, AD-temporal convolutional network-MSCDIM, which couples an unsupervised dynamic label recalibration strategy with a multi-scale temporal–sensor modeling backbone. First, the proposed MsTD-finite impulse response (FIR)-Rec method performs parallel multi-scale time-delay embedding and zero-phase FIR filtering, combined with kernel principal component analysis based slope calibration, to identify unit-wise failure turning points and reconstruct phase-consistent RUL labels. Building on these refined targets, a gated multi-scale U-Net architecture with a Multi-Scale Channel-Aware U-module adaptively fuses global temporal context and channel-wise information, while a dual-branch adaptive dilated causal temporal convolutional decoder jointly captures long-term degradation trends and local fluctuations. Experiments on the Commercial Modular Aero-Propulsion System Simulation (C-MAPSS) benchmark and the more realistic, noise-contaminated N-CMAPSS dataset show that the proposed framework achieves overall better performance to representative baseline methods in terms of root mean square error and the score, and ablation studies confirm the effectiveness and reusability of MsTD-FIR-Rec and multi-scale channel-aware U-module under multi-condition, noisy industrial scenarios.