A two-phase soft sensor modeling framework for quality prediction in industrial processes with missing data

Abstract

Quality prediction plays an essential role in modern process industries to improve product quality, ensure production stability, and increase economic efficiency. The high-dimensional, nonlinear, and dynamic characteristics of process variables make it hard for the traditional soft sensor modeling methods. Meanwhile, high-quality data is particularly scarce due to the economic and technological constraints, making it easy to overfit and crash of the model training. In order to solve those problems, a two-phase soft sensor modeling framework based on time series generative adversarial network (TimeGAN) and minimal gated unit (MGU) is proposed. The first phase is performing data augmentation for small samples by TimeGAN to improve the generalization performance of model training, and the second phase uses MGU for soft sensor modeling with the purpose of quality prediction. Finally, two datasets with different thicknesses for the hot rolling process (HRP) are used to verify the accuracy and validity of the proposed scheme.