A soft sensor regression model for complex chemical process based on generative adversarial nets and vine copula

Background: In the increasingly complex industrial process, it is extremely important to measure the key variables that directly affect the timely operation of the entire process. However, some key variables are challenging to be measured by traditional methods, so it is meaningful to use relevant variables to establish a soft sensor regression model to predict them. Usually, a large number of labeled samples are needed for modeling accurately. However, in some complex chemical industries, only a small number of labeled samples can be used to build a regression model and the model's description of the process is therefore inaccurate. Methods: This paper proposes a soft sensor method based on generative adversarial nets and vine copula regression (GAN-VCR) to address this problem. This method uses generative adversarial nets (GAN) to generate a large number of samples with a distribution similar to the labeled samples. For more reasonable sample augment, a sample selection strategy based on the generalized local probability (GLP) index is used to select augmented samples from a large number of generated samples to augment the training samples. Then, the augmented training samples are used to build a vine copula regression model and predict the key variables. Finally, nu-merical and industrial examples are used to prove the effectiveness and practicability of the method. Significant findings: Based on the developed soft sensor, sample augmentation can be effectively performed and the performance can be improved compared to traditional soft sensor methods, and the product quality in practical can be effectively improved.

Automated summary

This paper proposes a soft sensor method based on generative adversarial nets and vine copula regression (GAN-VCR) to predict key variables in complex industrial processes. The method uses generative adversarial nets to generate samples with a similar distribution to the labeled samples and employs a sample selection strategy based on the generalized local probability index to enhance the training samples. Experimental results demonstrate the effectiveness and practicality of the proposed method.