Study On Soft Sensor Based On Stacked Neural Networks

Melt index is the most important indicator of quality in the production process of polypropylene. Because of the constraints of process equipment and technical, the melt index is difficult to measure online. If we directly measure the melt index of the polypropylene product samples, that will result the measurement delay, melt index can not be controlled effectively online. Because the melt index is difficult to measure online, model of melt index prediction based on stacked neural networks was established.This article mainly includes the following:(1) Above all, introduced the characteristics of the polymerization process and problem of polymerization process study. Discussed the feasibility and significance of soft sensor. Introduced the basic idea of soft sensor and its implementation steps, generally described the situation of various soft sensor modeling method and its application in the industrial field.(2) Analyzed some commonly artificial neural network algorithm, including the RBF neural network and BP neural network. Because of the defect of establishing a nonlinear model base on artificial neural network, introduced combined forecasting method. Detailedly described the status quo, classification and combination coefficient of combined forecasting. Finally, proposed stacked neural networks based on neural network and combined forecasting methods, and to minimizing the maximum absolute error as the criteria for solving the combination weight. Proposed the combined neural network modeling based on minimizing the maximum absolute error.(3) Integrated analysed the catalyst system of Spheripol process,homopolymerization mechanism and production process, determine auxiliary variables of impacting melt index. Secondly, preprocessed sample data, including different date exclusion, random error processing and data normalization. Then, determined the number of BP network hidden layer and nodes based on experience piece-try method. Finally, build combined neural network model, and through comparing combined neural network with a single BP neural network, proved the effectiveness of combined neural network.