Industrial Process Soft Measurement Based On Multivariate Statistical Project Method

In many process control situations, it is often difficult to estimate some important process variables due to the limitation of process techniques or measurement techniques. These variables, which are key indictors of process performance, are normally determined by off-line sample analyses in the laboratory or using an on-line analyzer. Product quality values calculated on-line by using indirect but readily variable measurements (known as soft measurements) have been considered as an efficient method to solve this problem. Soft measurement can be used to obtain a regression model between easily obtained measurements and quality variables using statistical techniques, or neural network, etc.Soft measurement based on multivariable statistical project method is a kind of data modelling techniques. The paper systematically elaborates this technique from many aspects. Combining the actual application of industrial gas-fluidized bed polyethylene process with the problems which exist in traditional methods, some new algorithms and solutions are brought forward. Contents in this paper are as following:1. The paper summarizes the definition, content, developing tendency and modeling methods of soft measurement, especially provides complete and historical introduction on multivariate statistical project method.2. Briefly introduces the academic tools of soft measurement based on multivariate statistical project method, such as principle component analysis, principle component regression, partial least squares and neural network.3. Analyzing the characters of the gas-fluidized polyethylene process, the paper propose design of control system. In order to build the product quality soft measurement, we discussed the relation of between process variables and product quality variables. Finally, all modeling variables and data are proposed.4. A new algorithm, finite memory based on recursive PLS, is proposed. The adaptive algorithm is applied to build adaptive soft measurement which have more strong tracking ability and higher precision than traditional model. An application of the method to Slurry-Feed Ceramic Melter and industrial fluidized bed reactor for predicting quality variables is presented to show the effectiveness.5. By using the universal approximation property of neural networks (NN), a three-layer feedforward neural networks is embedded into the framework of standard PLS (partial least squares) modeling method resulting in a nonlinear PLS-NN model. A gradient descent learning algorithm is employed to train the network.Finally, the whole thesis is summarized, and some future research areas are highlighted.