A fast and gentle conditional diffusion model for a missing data generation method customized for industrial soft sensor

Abstract Due to sensor failures, interruptions to data transmission, and other factors affecting industrial processes, whole segments may be missing from a dataset, which can reduce the accuracy of an established downstream data-driven model. Existing methods usually treat the issues of filling missing data and building downstream model independently, and do not fully consider the requirements of the downstream tasks, resulting in insufficient filling accuracy for the missing data. In view of this, a fast and gentle conditional diffusion model is proposed in this paper. The main contributions of this paper are as follows. (i) We put forward a basic framework for customized missing data filling to meet the specific needs of downstream tasks, and our target of improving the prediction accuracy of the downstream soft sensing model is achieved. (ii) To meet the demand for lightweight models for industrial applications, a fast conditional diffusion model is proposed. Using a random step sampling strategy in the reverse process effectively accelerates the training speed of the model and reduces computational complexity. (iii) In order to ensure that the sensor detection process has basic qualities such as independence and objectivity, and to reduce the interference of downstream tasks in filling the missing data of the sensor, a gentle feedback strategy is designed. Validation on two datasets from a nickel smelting system and a combined cycle power plant shows that the proposed method is feasible, and is superior to alternative methods in terms of solving the problem of whole segments missing from industrial data.