Study Of Soft Sensor Technology And Its Application In The Production Of Polycarbonate

Polycarbonate (PC) is a kind of excellent engineering plastics. Because of its excellent performance, PC has been widely used in various industries, and it has good market prospects. In the production process of PC based on interfacial polycondensation method, some parameters affecting the quality of the products, for example, esterification in the stage of phosgenation, degree of polymerization in the stage of polycondensation and the molecular weight of the products, can not be directly measured by sensor or measured of high cost, sample analysis results have great lag, which is very detrimental to the quality of the products. Soft sensor technology provides convenience for measuring these variables. In view of PC production as the research background, a soft sensor model of PC molecular weight based on MATLAB-BP neural network is devetoped.Firstly, this paper discusses the general situation of the soft sensor technology. Through the analysis of several modeling methods, modeling method based on BP neural network is selected. Then the basic principle of BP neural network and its realisation based on MATLAB are studied. The prediction of BP neural network is applied to the characterization of PC quality index of molecular weight. After the primary factors of the data acquisition and processing through the analysis of the PC reaction mechanism, the3a rule, wavelet analysis and principal component analysis are used to eliminate the abnormal data, improve the data quality, and realize data selection and dimensionality reduction, as the premise for the soft sensor modeling.Then, the processed data are used to build soft sensor model. By comparing the training effect and precision through MATLAB simulation with different hidden layer nodes and improved learning algorithms after the analysis and determination of the input and output layers, neuron numbers and transfer function of every layer, the conclusion is determined that the number of hidden layer nodes is9, and the training algorithm is L-M algorithm.Finally, the soft sensor model is used for actual prediction. The average relative error of7.03%shows good generalization capability.