This paper proposes STDiff, a novel deep learning method for handling missing values in industrial systems. Unlike existing methods that focus on pattern completion within fixed time windows, STDiff learns the system's state evolution to gradually generate missing values. This approach is suitable for industrial systems that experience dynamic changes, abnormalities, and long-term missingness due to control actions. STDiff utilizes a conditional denoising diffusion model with causal bias to generate missing values based on recently known states and relevant control or environmental inputs. Experimental results demonstrate that STDiff achieves lower error rates than existing methods on public wastewater treatment datasets and real-world industrial datasets, particularly for long-term missing data. While existing window-based models flatten or oversmooth data, STDiff generates dynamically valid time series.
