Chemical Processes Monitoring Based On Gene Expression Programing

Chemical process monitoring is of key importance for the safety of chemical process system. With the development of control systems in chemical process, vast operating data are collected from the process and stored in real-time database. These data are high dimensional, non-linear, correlated and redundant. In order to monitor the status effectively from these data, large numbers of monitoring technologies based on kernel have been broadly used (e.g. SVM, KPCA, KPLS, KFDA). Kernel-based method is to map data sets from input space to a higher-dimensional kernel feature space. Then the original nonlinear problem is changed into a linear problem, so as to realize nonlinear feature extraction and classification. Currently Polynomial, Gaussian, and Sigmoid kernel functions are often used. Although the method based kernel functions can extract nonlinear features, it is not good to reduce dimension. Furthermore, the selection of kernel functions and parameters is the task of key importance in kernel-based technology, which has a significant impact on monitoring results.Firstly, feature extraction method based on gene expression programming is proposed in this paper to reduce dimension. GEP-PCA algorithm is used to deal with nonlinear data. Through the function sets mapping, the nonlinear relationship between variables are changed into the same dimension of the linear relationship. And then they are treated with PCA method to reduce the dimension. In order to calculate the error of the method namely degree of information retention, the original data should be reconstructed. This paper we adopt neural network method to reconstruct data. The right GEP transfer functions and feature vectors are obtained by adjusting the function group via fitness values so as to achieve the purpose of dimension reduction.In this paper, the state monitoring method based on GEP-kernel for chemical process is proposed to simplify the selection of kernel functions and parameters. Two suitable kernel functions will be gotten automatically by GEP, and then feature extraction is done using GEP-kernel PCA and classification using GEP-kernel SVM. The kernel functions are adjusted by fitness function which is the reciprocal value of the sum of absolute classification error. Finally, the KGEP-PCA-SVM model is obtained. The method is applied to heavy furfural refining and ketone-benzol dewaxing processes of lube oil in a petrochemical plant. And both of the monitoring accuracy rates are more than 90%.