Research On VSG Method Based On AANN Noise Injection And Its Modeling Application

In the traditional petrochemical production process,a reliable mathematical model is often established to effectively describe the relationship between input and output of the feed,and it is convenient for the operator to control the production process more accurately.However,the main problem that currently exists in petrochemical processes is the problem of small samples.Small sample data usually has the characteristics of small number of samples,uneven data distribution,and incomplete data.This results in low accuracy and large errors in the mathematical models established.The overall characteristics of the sample space cannot be fully characterized.Facing the above small sample problem,this paper proposes a method of generating virtual samples based on the noise-injected Auto-Associative Neural Network(AANN),which effectively fills the uneven data,thereby alleviating the small sample data band.The problem of uneven distribution of data from the original data is solved,but the traditional chemical process modeling is slow and low in accuracy.Therefore,combining the incremental method and the Extreme Learning Machine(ELM)principle,The data is used for rapid modeling,and to solve the problem of insufficient precision in the modeling of chemical processes,this paper proposes a multi-activation function extreme learning machine(LV-MAFELM),which uses multiple non-linear functions as the activation functions of hidden layer nodes,and uses principal component analysis The method extracts the main component of the hidden layer output,effectively eliminates the redundant part of the hidden layer output,and solves the problem that the number of nodes in the hidden layer is difficult to select,thereby further optimizing the extreme learning machine network model and improving the modeling accuracy.This paper uses Gaussian noise-added Auto-Associative neural network and industrial ethylene production process data to verify the validity of the method of generating virtual samples.In this paper,terephthalic acid production process data are used to verify the multi-activation function extreme learning machine Stability and applicability.The experimental results show that the virtual sample generation method proposed in this paper can effectively solve the problem of uneven sample distribution caused by small sample data.The LV-MAFELM proposed in this paper can further improve the modeling accuracy based on the traditional ELM rapid modeling.