A filling method for missing soft measurement data based on a conditional denoising diffusion model

Abstract

In complex industrial processes, incomplete datasets are common due to problems such as different sampling periods and data loss, which reduces the accuracy of industrial soft sensing models. To solve this problem, this paper proposes a missing data generation and filling method based on a conditional denoising diffusion model. First, a missing area detection method based on a binary mark array is used to locate the region of missing data, and a masking mechanism is applied to obtain the accurate location and size of the missing data. Then, the correlation between the original data and the mask matrix is learned with a multi-head self-attention mechanism, and is used as the condition for the original denoising diffusion model to ensure the accuracy of the generated data. Finally, the generated data are filled into the missing areas to construct a complete dataset, with the aim of improving the prediction accuracy of the soft sensor model. The simulation results demonstrate that the proposed imputation method performs exceptionally well in filling missing data. Compared to traditional methods, it significantly enhances the prediction accuracy of the soft sensor model, reducing the mean squared error by approximately 40 %.