Research On Virtual Sample Generation Method In Chemical Process Field Based On GAN

In process industry,data-driven soft sensor models are used to obtain information that is difficult to obtain in real production.Both the quality and quantity of data are critical to the construction of soft sensor model.However,due to the small fluctuations,multiple repetitions,and high collection costs of chemical production process data,the insufficient size,uneven distribution,and lack of diversity of valid samples become important factors leading to the poor accuracy of soft sensor models.In this kind of problem,it is a novel and effective method to construct complete virtual samples by generating virtual input samples and their corresponding output samples and use them to expand the data set.In order to generate more diverse and homogeneous virtual input samples to fill the sparse areas of data and make the constructed models more capable of extracting more potential information from chemical data,this paper proposes a Wasserstein generative adversarial network with gradient penalty(WGAN-GP)based virtual input sample generation method for sparse areas.In the proposed method,KNN is first used to identify the points where the data sparse regions are located and thus construct a hyper-rectangle data sparse region,and then WGAN-GP is used to generate virtual input samples in the data sparse regions based on the uniformity selection method.Then,numerical simulations and actual chemical data sets are used to verify the effectiveness of the proposed method.The experimental results show that the virtual input samples generated by the method conform to the characteristics of the original samples,and increase the coverage and uniformity of the input space,effectively fill the inter-data gap in the sample input space.In order to further improve the accuracy of the soft sensor model,this paper is based on the conditional generative adversarial networks(CGAN),and it is improved to form CGAN with the cycle structure(CS-CGAN)to generate the virtual output samples.The reverse structure of CGAN can generate the output samples by capturing the conditional distribution(|).While CS-CGAN has the property of learning bi-directional conditional distributions and can improve the accuracy of the virtual sample input and output mapping by selecting one mapping with less information loss among multiple input and output mappings through the consistency test by the specificity of its structure,so as to ensure the accuracy of virtual output and the consistency with the original output distribution,and further improve the soft model accuracy.Combining the input generation method with the output generation method can obtain the complete virtual sample generation method.With the validation of numerical simulation and chemical applications,the results show that the proposed virtual sample generation method effectively enhances the accuracy of the soft sensor model,and the enhancement of CS-CGAN is higher than that of reverse CGAN.