Research On Soft Sensor Modeling Method Based On Model Population Analysis

In recent years,due to the requirements of industrial production efficiency and quality,soft sensor has become an important research field.Then there are more and more soft-sensor algorithms.However,a single soft-sensor algorithm usually considers only one characteristic of the industrial process,and ignores other characteristics in practical process.So,one method under some condition with good performance is likely to be badly in other condition.To some degree,ensemble learning can solve this problem.This thesis mainly adopted the model population analysis and the Bayesian algorithm,and introduces the basic ideas and implementation method.In order to solve the limitations of a single value index and a single soft sensor algorithm for model comparison,an ensemble learning system is established under the framework of model population analysis,and designs the correspond GUI platform.These include:1.In the outlier detection,Monte-Carlo cross-validation algorithm based on model population analysis is used to detect outliers.The statistical result indicates that the prediction accuracy of sub-model algorithms and Bayesian algorithm is improved after outlier detection.2.Some soft sensor algorithms,which are applicable to various industrial processes,are chosen to build sub-models,and the results of them are integrated by Bayesian algorithm.The statistical result indicates that the prediction accuracy of Bayesian algorithm is better than any other sub-model algorithm.3.In the model comparison,the influence of the choice of training datasets,which has an effect on the result of model comparison,can be eliminated through producing multiple training sub-datasets by model population analysis.Meanwhile,the diversity of datasets can be improved,and the way of model comparison in statistics is discussed.4.A MATLAB-GUI platform based on the ensemble learning system in this thesis is developed,which could deal with online monitoring.Finally,a summarization of the full text is given and some prospects for further research are discussed.