Research On Soft Sensor Modelling Of Chemical Process Using EWT-MKL Methods

With the rapid development of social economy in recent years,the chemical industry has become an important part of the prosperity of a progressive country.Based on the National Advocacy of green development,the means of modern chemical process monitoring still need to be deepened and expanded.The basic information such as process parameters measured by conventional instruments can not meet the needs of process operation and control.In many practical applications,various information measurement methods need to be combined,in order to exercise effective control over the process.The problems of production process,fault analysis and diagnosis of measurement system and condition monitoring make it urgent to monitor the variables which are difficult to be directly monitored in chemical process.In this paper,Empirical Wavelet Transform（EWT）and Multiple Learning Kernel（MKL）are studied,which are applied to real-time monitoring of chemical process variables to improve product quality and meet the demand of industrial production process for control system,so as to improve the competitiveness of chemical enterprises.The research contents and achievements are as follows:（1）Regression analysis using chemical process data directly will reduce the accuracy and reliability of prediction due to the actual noise and random interference.In order to suppress the influence of interference signals on the model accuracy in chemical process data,empirical mode decomposition（EMD）and empirical wavelet transform（EWT）are used to decompose and reconstruct the test data with Additive white Gaussian noise,the empirical wavelet transform with better performance is selected as the Data pre-processing method to build the soft sensor model,and the multi-core learning method is introduced into the analysis based on the SVM theory,multi-core learning based on semi-infinite programming is realized by MKL Wrapper Algorithm and MKL chunking algorithm,and EWT-MKL method is used to construct soft sensor model.The experimental results show that the proposed method can effectively improve the accuracy of soft-sensing model,and the MKL method can assign the weights of multiple kernel functions through the Kernel Weight Coefficients,which avoids the shortcoming of the traditional kernel learning method,which has only one kernel function,the measurement precision and generalization ability of the soft sensor model are improved（2）The data provided by the Tennessee Eastman simulation platform are used to verify the theoretical analysis of this method.The experimental results show that the proposed method can effectively improve the accuracy of soft-sensing model,and the MKL method can assign the weights of multiple kernel functions through the Kernel Weight Coefficients,which avoids the shortcoming of the traditional kernel learning method,which has only one kernel function,the measurement precision and generalization ability of the soft sensor model are improved.This method is applied to the determination of H₂S and SO₂concentrations in the bottom of a debutane tower and in a Sulfur Recovery Unit（SRU）,the soft-sensing models based on SVM,MKL wrapper,MKL chunking,Simple MKL,EWT-MKL wrapper,EWT-MKL chunking,and EWT-MKL methods are established and used for cross-comparison.The experimental results show that the proposed method can achieve high modeling precision,in the chemical process can be used for the measurement of key components to ensure product quality,but also to achieve the monitoring of harmful gases,environmental protection.At the same time,it avoids the problem that the traditional kernel learning method is too dependent on the kernel function and its parameter selection in soft sensor modeling,and can be widely used in chemical process soft sensor.