Integrated Gaussian Process Regression Soft Sensor Modeling Method

In process industry,some key quality variables can only be obtained through off-line analysis in laboratory due to the hardness of online detection,which seriously restricts the monitoring,controlling and optimization level of the process.Therefore,soft sensor technique is introduced to provide real-time estimations of the quality variable by establishing a mathematical model between the the easy-to-measure variables and the difficult-to-measure variable.As a typical local learning algorithm,ensemble learning is widely used in soft sensor development.In order to build the ensemble learning soft sensor model with high performance and strong model framework,this dissertation studies the modeling method of ensemble Gaussian process regression(GPR)soft sensor,and carries out the corresponding simulation experiments and results analysis.The research results are summarized as follows:(1)Traditional ensemble learning soft sensor methods usually ignore the diversity of input variable selection and do not consider ensemble pruning,resulting in high computational complexity and limited prediction performance.For this reason,a soft sensor method of selective ensemble learning based on diverse subspaces is proposed.This method combines bootstrapping resampling and partial mutual information relevance selection to construct a variety of subspaces and build the corresponding GPR base models.Then,the evolutionary multi-objective optimization algorithm is used to achieve ensemble pruning.Finally,the effectiveness of the method is verified in the application of Tennessee Eastman chemical process and penicillin fermentation process.(2)To address the problem that traditional ensemble learning soft sensor models only consider the single perturbation factor of diversity and the model cannot adapt to the time-varying process characteristics,a selective ensemble learning adaptive soft sensor method based on multimodal perturbation is proposed.In this method,input feature perturbation and training sample perturbation are introduced to construct the diverse base model.Then ensemble pruning is carried out by evolutionary multi-objective optimization and stacking ensemble strategy is used to achieve the integration.In addition,in order to prevent the deterioration of model performance caused by time-varying process characteristics,the base model and the ensemble model are updated based on the moving window update strategy.Finally,TE chemical process,penicillin and chlortetracycline fermentation process show the superiority of the method.(3)It is a common phenomenon that abeled samples are lacked and the unlabeled samples are rich in industrial processes.In order to make full use of the process information contained in unlabeled samples,a semi-supervised ensemble learning soft sensor method is proposed in combination with ensemble learning and co-training.This method firstly constructs the diverse subspaces based on evolutionary multi-objective optimization,and builds diverse GPR base models to obtain the pseudo labeling samples.Then,the pseudo labeling samples are selected with high confidence level based on the prediction variance information and update the GPR basis model.The process is repeated until the iteration stop condition is reached.The effectiveness of the method is verified in TE chemical process.