Research On Virtual Sample Generation Method Based On Gibbs Sampling Algorithm

The popularity of the Internet,the advancement of the Internet of Things,and the continuous improvement of computing power have accelerated the pace of human beings entering the era of big data.Large industries,represented by petrochemicals,have also started to transform digitally and intelligently.Building an accurate and generalizable data model is crucial for business decision makers to conduct transformation studies.Currently,data-driven soft sensor modeling relying on data is widely used for modeling complex industrial processes.However,for the process industry,the stable operation of the production process under modern control systems and advanced devices,it makes the collected data suffer from limited fluctuation range,poor distribution,and insufficient valid information,limiting the accuracy of the data-driven soft sensor model,and thus affects the misjudgment of the production process by the staff and even causes economic losses.There are many methods to address the data quality problem,including Virtual sample generation(VSG)as an efficient and new data augmentation method,it has very important research significance in the field of machine learning and has great application prospects and values in practical applications.In order to generate high quality virtual samples,ensure the correlation and distribution of virtual samples and original samples are similar,and make more obvious advantages in improving modeling accuracy.A Gibbs Sampling Algorithm Based Virtual Sample Generation(GS-VSG)is firstly proposed in this paper.this method starts from the correlation between the input data,makes use of the Gibbs sampling algorithm to complete the virtual input sample generation.Next,uses the generalized regression neural network to predict the virtual output sample,the virtual sample generation process is completed.The validation experiments are conducted on a standard dataset with low-dimensional and the high-dimensional real industrial dataset,and the experimental results show that the GS-VSG method is valid in the low-dimensional space and efficient in the high-dimensional space.The method achieves large-scale augmentation of the original labeled dataset while ensuring the correlation between the original data remains unchanged.In order to solve the problem of inaccurate virtual output samples,the idea of "voting strategy" is introduced in this paper,and performed filtering of the generated virtual samples based on the GS-VSG method,namely GSVSG-Vote method.Since the screening criteria of the virtual samples is not sufficient to achieve accurate selecting,the voting preference strategy is used to select the reliable data and remove the suspicious data with multiple regressors.In order to prove the effectiveness of the improved method,simulation experiments were conducted on two data sets.The results demonstrate that the VSG method with screening strategy can improve the quality of virtual samples,and at the same time,improve the accuracy of the soft sensor model is better than other advanced methods of the same type.