Virtual Sample Generation Using Conditional Adversarial Network with Latent Spaces as Noise Inputs

Data-driven soft sensors are widely applied in the chemical industry. However, due to the production conditions and resource limitations, the soft sensor models faced the challenge of poor prediction effect caused by the small sample. To solve this challenge, this paper puts forward a novel virtual sample generation (VSG) method based on conditional generative adversarial networks with latent spaces as noise inputs (CGAN-LSNI). In the CGAN-LSNI, we use the latent space of the Variational Auto-Encoder (VAE) as the noise input to the CGAN. In addition, utilizing the samples generated by VAE and the original samples to co-train the conditional generative adversarial nets (CGAN) by semi-supervised learning method. This method aims to improve generalization ability of the CGAN predictions. To better apply the numerical data obtained from sampling in the chemical industry, we incorporate the mean square error into the existing loss function of CGAN. The benchmark function and industrial purified terephthalic acid (PTA) data are used to verify the performance of the proposed CGAN-LSNI. The results show that the CGAN-LSNI has better prediction accuracy compared to the related methods of generating virtual samples.