A novel virtual sample generation method based on dual correlation generative adversarial nets for soft sensing application

Abstract

In industrial production processes, due to the limitation of the environment and economy, the soft sensor models faced the challenge of poor prediction effect caused by the small sample. To solve the above challenges, this paper puts forward a novel virtual sample generation method based on dual correlation generative adversarial nets (DCRGAN-VSG). In DCRGAN, Wasserstein GAN (WGAN) is used to generate the process variables, the improved Conditional GAN (CGAN) is used to generate the quality variables, and the mean square error is added to the value function of CGAN to improve the mapping relationship between the process variables and quality variables. In addition, the Pielou index is used to select process variables that can fill the sparse region from the generated samples. Finally, training soft sensors using the dataset containing virtual samples. The benchmark function and industrial case called the purified terephthalic acid (PTA) data are used to verify the virtual sample generation performance of the DCRGAN-VSG. Simulation results show that DCRGAN-VSG improves the accuracy of soft sensing models by approximately 70% in benchmark experiments and by nearly 40% in industrial case studies.