Soft Sensor Modeling Based On Orthogonal Least Squares

In petrochemical process, there are many important process variables that cannot be measured on-line. Such as purity, dry point of distillation product, distillation tray efficiency, and polypropylene melt index, etc. Soft sensor provides an effective way to solve these problems. With the rapid development of modern industry, the soft sensor technique has a widely applicable prospect and has been one of the most focuses in process instrument and measurement technique. This paper researches on Orthogonal Least Squares (OLS) and establishes a soft sensing model of polypropylene melt index by using this method. The main work is described as follows:Some kinds of traditional methods of soft sensor modeling methods are discussed and summarized in details in this paper, while the characteristic of each approach is described respectively. A new OLS method based on Orthogonal Signal Correction (OSC) is proposed to reduce the noise information which is uncorrelated with output variables, and it is used for soft sensor modeling in industrial process. In order to improve model explanatory ability and generalization, Orthogonal Signal Correction is applied to preprocess OLS model to reduce the noise information which is uncorrelated with output variables. The application results of actual data show that OLS based on Orthogonal Signal Correction can predict the polypropylene melt index more accurately than PLS and OLS.With regard to the limitations of single kernel in kernel modeling methods, an algorithm of orthogonal least squares (OLS) based mixtures of kernels is proposed and it is used for soft sensor modeling in industrial process. Mixtures of kernels have properties of local and global kernel ones. They are used in OLS method to replace the single kernel which can improve the generalization ability and nonlinearity of model. The selection of kernel parameter has great influence on model. Particle swarm optimization (PSO) algorithm is used in optimizing kernel parameters. The application results of actual data show that mixtures of kernels OLS can predict the polypropylene melt index more accurately than PLS and OLS.