Research On Virtual Sample Generation Technology And Its Application To Industrial Modeling

In the era of big data,small sample problems still exist and cannot be ignored.With the development and application of information technology,large amounts of data have been accumulated in the chemical industry.However,due to the low probability of occurrence or repetition of sample data and the high cost of sample data acquisition,the sample data used in the research can be limited.Because the number of samples of small sample data sets is not sufficient,and the distribution of samples is not uniform and uneven,so the training and learning based on small sample is poorly performed,the precision of prediction modeling is relatively low,and the modeling quality is poor.How to effectively solve the training and learning problems of small samples and get a precise and robust predictive model is undoubtedly a problem to be solved urgently in the field of data driven modeling.As an effective solution to small sample problem,it is of great theoretical significance and application value to study and research the topic selection of virtual sample generation technique in small sample problems.Because of the characteristics of incompleteness and imbalance of the small sample data set,currently,there are mainly two approaches in the academic groups to build accurate and robust prediction models based on small samples:the Gray Model theory and machine learning algorithms based modeling methods and modeling with the aid of manmade virtual samples.Therefore,on the basis of previous literature research summary,a novel virtual sample generation(VSG)method based on Monte Carlo method and Particle Swarm Optimization(PSO)algorithm is proposed in this article to improve the precision and accuracy of small sample based predictive models and its prediction performance.Based on the original small sample data,this approach utilizes Monte Carlo method and PSO algorithm to generate and optimize virtual samples that conform to the original small sample,which effectively fills the information interval between original small samples,and improve the accuracy of the Extreme Learning Machine(ELM)prediction model using virtual samples.The effectiveness,practicability and progressiveness of the proposed method is verified through two typical chemical industrial process datasets(PTA system and ethylene production system).The simulation experimental results show that the accuracy of ELM prediction model has been greatly improved.