Improved Virtual Sample Generation Method Using Enhanced Conditional Generative Adversarial Networks with Cycle Structures for Soft Sensors with Limited Data

In the modern chemical industry, only a small number of representative samples can be used to build soft models due to practical factors. However, the accuracy of the soft model built in this case is not sufficient to meet the demand. To overcome this problem, a novel virtual sample generation (VSG) method based on conditional generative adversarial networks (CGANs) with a cycle structure (CS-CGAN) is proposed to augment the sample data sets and enrich the sample diversity. In the proposed method, first, for obtaining the inputs of virtual samples, the Wasserstein GAN with gradient penalty (WGAN-GP) is used to generate new samples x based on the original sample distribution to fill the scarcity regions of the data. Second, the reasonable outputs of the newly generated samples are determined by the CS-CGAN with consistency test. To verify the performance of the proposed new method, numerical simulations and real-world data sets are used. The results show that the proposed new method can effectively generate realistic samples and outperform other methods in improving the performance of soft sensors.