Research On Relevance Vector Machine Soft Sensor Modeling And Its Application

The ultimate goal of almost all the industrial production is to produce high quality products,so when safe production is ensured,quality control is the core of the production process.Although to a certain extent,the emergence of soft sensor technology has compensated the lack of sensors and overcome the disadvantages of off-line detection,the industrial production process requires real-time and accurate data,as a result,the development of soft measurements is put forward higher requirements.Therefore,the improvement of modeling method and the introduction of optimization algorithm are of great significance to simplify the structure of soft measurement model,improve the accuracy of soft modeling and improve the speed of modeling.This paper focuses on two kinds of soft sensor modeling methods based on kernel function—relevance vector machine and fast relevance vector machine,and their related improvements by the introduction of optimization algorithms:1.A soft sensor of RVM model based on optimized mixed kernel function is proposed.In order to simultaneously get better prediction and sparsity,a mixed kernel function and a fitness function which makes a tradeoff upon performances between regression accuracy and sparsity are proposed,and the genetic algorithm is used to automatically optimize the kernel parameter group of RVM.The proposed method is used for a soft sensor modeling in a cleavage-recovery unit of Bisphenol A,and the results show that it can guarantee better sparsity and regression accuracy than the general SVM mixed kernel model and GA-RVM single kernel model.2.Numerous equipment of modern process industry lead to high crafts complexity,and technical indicators from each aspect of the production process are always affected by various factors,so that online measurement of product quality indexes becomes quite a difficult problem.Thus,how to achieve online estimation of product quality index by suitable soft sensor model,how to extract effective features from multiple input variables and how to quickly and effectively determine the model parameters have become a research focus.In order to solve this problem,here KPCA is used to extract features of input variables,and then RVM regression model is established.Consider that the kernel parameters of KPCA and RVM can determine the performance of KPCA and RVM,Harmony Search(HS)algorithm is used to simultaneously search the two best kernel parameters,so a KPCA-RVM model optimized by HS algorithm is ultimately constructed.Simulation results show that,compared to HS-SVM,GA-RVM and HS-RVM,the proposed algorithm has higher prediction accuracy and calculation speed,and has achieved good results.3.It is necessary for the soft measurement model applied in the modern process industry to meet the requirements of large data processing,high accuracy and high real-time performance,so the relevance vector machine(RVM)is replaced by the fast relevance vector machine(FRVM)as the soft measurement regression model,which reduces the computational complexity and the training time;Meanwhile,in order to quickly and accurately determine the kernel parameters of FRVM,an improved Harmony Search Algorithm(HS)method with non-linear changing PAR and new variables selection way for optimizing the kernel parameters of FRVM is proposed.The simulation results show that the improved method proposed in this chapter can effectively solve the problem that the harmony search algorithm is easy to fall into local optimum,and the prediction accuracy and running speed of the proposed method are better than the HS algorithm based on linear-PAR and the HS algorithm based on constant-PAR.