The Research And Application Of Soft-measurement Modeling System

Soft-measurement is a new, developing and prospecting industry technology, which is now getting widely used in industry. The modeling and application of soft-measurement essentially is a time-consuming, energy-consuming and complicated work. In the practice of modeling, if lack of special and strong tools, technician have to program every part of soft-measurement. It not only request technician have the ability of comprehending the algorithm correlative, but also take so much time of them that they can't devote to the design and implementing of soft-measurement models. Thus it is helpful to own a set of soft-measurement modeling tool to improve the modeling efficiency. Some soft sensors are provided in the form of software overseas, whereas these soft sensors are expensive very much. Some experts domestic have done a lot of work on theory study and application of soft measurement, however, there are few software provided in the form of merchandise. In addition, according to document, those Soft sensors overseas and domestic emphasize particularly on some certain modeling method, which haven't include other modeling methods or provide only a limited function. It result in users haven't the chance to choose different modeling method.The phenomena mentioned above affect the application and popularization of soft-measurement technology in industry process to some extent. Therefore, it is a significant topic to develop a modeling system that is simple to operate and has sufficient functions. The developing of a soft-measurement modeling system is studied in the thesis. The main features of the system display as follow:(1) Kinds of methods to select secondary variables are provided, such as manual selecting method and PCA method.(2) Sorts of data pretreatment methods are supplied, such as 3 , average filtering and slicing.(3) Kinds of soft-measurement modeling methods are presented, which are mainly divided into two categories: One is regression method concretely including MLR, MSR and PCR, and so on; the other is neural network method including BPN and RBFN. 9 kinds of faster training algorithm of BPN are presented besides the standard steepest descent algorithm. The structure of presenting many kinds of modeling methods in the system make it possible to gain a ideal model by comparing and selecting different modeling methods. The arrangement improved the structure of some soft-measurement systems that emphasize particularly on some special modeling method.(4) The method of integrating PCA with neural network is already reported. However it is not yet seen to integrate the method in soft-measurement modeling tools. The method is actualized in the system.(5) In order to improve the generalization of models, "Early stopping" method is adopted in the system. The method overcomes the disadvantage of some systems that pay attention to fitting effect of models but neglect generalization of models.(6)The result of modeling is displayed and saved in the form of graphs, formulas and statistical figures, which provide standard for users to compare and select models.In order to verify the modeling effect of the system, soft-measurement models of key components of butadiene distillation column in a factory are studied by using the system.In the first chapter of the paper, the achievements on the development of soft sensor are introduced and compared after expatiating on Soft-measurement technique systematic. Subsequently the meaning of developing the system and characteristic of the system are proposed. The structure of the system and some key technology are introduced in the second chapter of the paper. In the following three chapters, the realization of each son system of soft-measurement modeling system is discussed, which are data pretreatment system, modeling system and model correcting system. In the last chapter of this paper, the contents of the paper are summarized and the next tasks are arranged.