Modeling And Application Of Chemical Process Soft Sensing Based On Relevance Vector Machine

Soft sensor technology is a new kind of intelligent online-detection technology which has fast development in industrial process and control field,its core is modeling method.The artificial intelligence modeling method is the most extensive and the fastest-growing soft sensor technology,which mainly contains Artificial Neural Network(ANN)、Support Vector Machine(SVM)、Relevance Vector Machine(RVM)and others.Comparing with ANN、SVM,RVM is more easily to choose kernel function 、 lower computational cost 、 more sparse and so on,however,the predictive performance of RVM still need to improve and it only can be applied in single output system.So the improved RVM methods are developed and applied in the soft sensor of olefin yield in the methanol to olefin(MTO)reaction process.The main research and innovations are as follows:(1)The basic principle of soft sensing technology was summarized,the application of different artificial intelligent algorithm in the soft sensing field was compared,a literature review on application of RVM in the soft sensing technology was introduced,and its future studying direction was discussed.(2)In order to improve RVM model predictive performance,a nonlinear fusion modeling of MRVM model based on AFSA-BPNN was presented to meet the need of the nonlinearity and multiple operating modes features of actual industrial production.The initial parameters of BPNN was optimized by AFSA to accelerate the convergence rate of BP network and achieve global optimization.Then the BPNN was used to nonlinearly combine different sub-models,so that the model performance can be better.Lastly,a nonlinear function was used to test the effectiveness of the proposed modeling method.(3)In order to meet the need of multi-input and multi-output industrial object,a improved RVM modeling method based on mixture kernels was proposed,the traditional RVM was replaced by multi-variable relevance vector machine(MVRVM),and the relevant parameters was optimized by artificial fish school algorithm(AFSA)to avoid man-setting subjectivism.Finally the effectiveness of the presented modeling method was test by the concrete slump test data in the UCI database.(4)The process of the methanol to olefin(MTO)reaction process as the application object,the above two methods was applied to develop soft sensor of olefins yield in the MTO reaction process.The experimental results indicated that the nonlinear fusion modeling of MRVM model based on AFSA-BPNN can finish nonlinear prediction and has better performance.And theMVRVM with mixture kernels optimized by AFSA can predict ethylene and propylene yield simultaneously,and its performance is better.The proposed improved RVM modeling methods and its application in soft sensors are of high research significance and practical value for the soft sensor modeling research and improvement of process industry,and provide an effective method for the quality monitoring of the MTO production process.These methods can be regard as the basis of control and optimization of chemical industry,they are also helpful to improve industrial production,and meet the need of efficiency and low consumption,so as to enhance the competitiveness of traditional industry.