High-quality process data are essential for modern manufacturing processes to enable advanced control techniques, fault detection, and predictive maintenance. However, real-world industrial datasets often contain missing values due to sensor failures, power outages, and equipment maintenance. This paper proposes a novel implicit–explicit diffusion model that jointly utilizes both hidden and visible properties for industrial data imputation. The model employs a dual-branch architecture: one branch uses multi-scale dilated causal convolutions to capture hierarchical periodic patterns inherent in industrial time series, while the other branch leverages structured state space (S4) models to learn long-term dependencies. A gated fusion mechanism adaptively combines these complementary representations. Extensive experiments on Debutanizer and Sulfur Recovery Unit (SRU) datasets demonstrate that the proposed method achieves the best root mean squared error (RMSE) across all tested missing rates (20–80%) on both datasets, and exhibits particularly strong advantages in high-missingness scenarios (60–80%), where the proposed multi-scale and long-range modeling capabilities prove most beneficial.
Liu et al. (Tue,) studied this question.