A novel virtual sample generation method based on a modified conditional Wasserstein GAN to address the small sample size problem in soft sensing. (May 2022)

• Record Type: Journal Article
• Authors: He, Yan-Lin; Li, Xing-Yuan; Ma, Jia-Hui; Lu, Shan; Zhu, Qun-Xiong
• Is Part Of: Journal of process control. Volume 113(2022)
• Page Start: 18
• Page End: 28
• Publication Date: 2022-05
• Subjects: Soft sensor -- Virtual sample generation -- Conditional Wasserstein generative adversarial networks -- Deep neural networks -- Purified terephthalic acid; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2022.03.008
• Languages: English
• ISSNs: 0959-1524

Abstract

Abstract: In the modern chemical industry process, soft sensing has been widely used. However, the lack of valid and sufficient data has made it difficult to apply advanced soft sensor modeling methods to realistic scenarios. In this paper, a novel soft sensing method based on deep learning is proposed to handle the problem of small data. Aiming at handling the issue of small sample size, a novel virtual sample generation method embedding a deep neural network as a regressor into conditional Wasserstein generative adversarial networks with gradient penalty (rCWGAN) is presented. In rCWGAN, conditional variables are introduced to make the training supervised and a dual training algorithm is specially designed. With the advanced structure and the designed training algorithm, rCWGAN has powerful sample generation capabilities and can well predict quality variables. Finally, an experiment on the purified terephthalic acid (PTA) solvent system is carried out for the validation of the presented rCWGAN. The simulation results indicate that the presented rCWGAN has good sample approximation ability and acceptable prediction accuracy with the small sample size task. Highlights: Novel virtual sample generation method with rCWGAN is proposed. WGAN-GP is combined with condition variables to generate label samples. The training method of rCWGAN is modified to obtain good virtual samples. Simulation results confirm the superior effectiveness of the proposed rCWGAN.