The Study On Variable Selection Methods Of Soft-Sensor Technique

Soft sensor technique has already been widely used and plays a more and more important role in the development of process detection and control system in recent years. Its basic principle is to select a set of secondary variables that are easy to detect and have close relationship with the primary variable according to certain "optimal" criteria. The selected secondary variables are then used to obtain the on-line estimation of the primary variable by constructing some mathematic relationship between these variables. It can be seen from the above definition that the selection of secondary variables is of great importance for building an effective and mature soft sensor model and this dissertation devoted to this problem. The selection of secondary variables is to find a subset of a pre-specified set of independent variables, which can best describe the dependent variables, or to select a subset of the original data which contains a relatively a small number of variables in such a way that the selected subsets of variables retain, as much as possible, the overall multivariate structure of the complete original data. Through variable selection, we can not only simplify the soft sensor model and make it easier to understand, but also greatly reduce the cost of information collection. This dissertation uses principal component analysis and partial least square as major mathematic tools. We use backward regression and genetic algorithm to find the "optimal subset". The main research work conducted in this dissertation summarized as follows:1. A variable selection method is proposed by using PC A and PLS;2. A new method is proposed to select the number of principal component of PLS in PLS regression;3. A variable selection method is proposed based on the combination of genetic algorithm and Bayesian statistics and verified by simulation data;4. A soft sensor model on the concentration of 4-CBA is proposed by using variable selection method based on real industrial data.Finally, several problems for further research and exploration are proposed based on the summary of the research results.