Research Of Soft-sensing Method Based On DBN

In process control of petrochemical industry,quality control becomes more and more important.These quality index can't be measured in real time.Thus the requirement of on-line monitoring and implementation of advanced process control cannot be fulfilled.Soft sensor technology has been used to solve this problem.In this paper,soft-sensing method based on DBN(Deep Belief Network)was investigated to improve generalization ability,while avoiding overfitting problems of traditional soft-sensing methods.By introducing kernel extreme learning machine,computation efficiency of soft-sensing method based on DBN was improved.The investigated soft-sensing methods were simulated in a CSTR system and then was applied in a real polypropylene production process.Simulation results show the effectiveness of the proposed soft-sensing method based on DBN.Comparing with soft-sensing method based on BP,SVM and PLS,the generalization ability and the accuracy of the soft-sensing method based on DBN was improved.The main research works are summarized as follows:Soft-sensing method based on BP network,SVM and PLS was studied in a CSTR system and a polypropylene production process..Training data set is selected using odd-even method to contain more information of the process;In addition,the structure of the BP network and SVM were optimized to improve generalization ability and the accuracy.Because of the existence of grade switching in the polypropylene production process,a piecewise soft sensing model was developed.Comparing with the single model without considering production grades,the piecewise soft sensing model is more accurate and more suitable to the industrial process.With consideration of the weakness of soft-sensing method based on BP network,SVM,such as uncontrollable convergence speed and local minima,exponential growth of amount of calculation with increasing of data samples,DBN was introduced into soft sensing methods and then was studied.DBN uses latent variables to express process variables with high correlation,thus DBN has a good expression ability.Soft-sensing simulation was carried in CSTR system and the results show the effectiveness.The soft-sensing method based on DBN was applied in a real polypropylene production process for the first time.Comparing with soft-sensing method based on BP,SVM and PLS,the generalization ability and the accuracy was improved.In order to improve learning speed and generalization ability of DBN based soft-sensing method,kernel extreme learning machine was imported.Then the soft-sensing method based on DBN-KELM was studied.Simulations have been done in CSTR system and polypropylene production process.The results show that the soft-sensing method based on DBN-KELM has promising performance and can improve computation speed and generalization ability.