An Instant Learning Soft Sensor Modeling Method Based On Evolutionary Optimization And Diversity Similarity

In process industry,some key quality variables can only be obtained through off-line analysis in laboratory due to the lack of effective online sensors,which has become a major obstacle to facilitating efficient process monitoring,controlling and optimization.Soft sensor technique has been recognized as an effective tool to provide real-time estimations of difficult-to-measure variables.The core of a soft sensor is to establish a mathematical model between the easy-to-measure variables and the difficult-to-measure variable to realize online prediction of the quality variables.As a typical data-driven soft sensor modeling method,just-in-time(JIT)learning has gained growing popularity due to its capability of dealing with nonlinearity,time-varying behavior,multi-mode and multi-phase process characteristics.Therefore,this dissertation focuses on JIT learning modeling framework,aiming to develop high-performance JIT soft sensors by exploiting diverse similarity measures and using evolutionary optimization algorithms.The main contributions are summarized as follows:(1)Traditional JIT learning soft sensor methods mainly focus on using single similarity measure,which may lead to failure in dealing with complicated process characteristics.To tackle this problem,a novel JIT learning soft sensor method using mixture weighted similarity(MWS)is proposed.The MWS similarity measure is defined by combining multiple weighted Euclidean distance(WED)based similarity measures.In addition,input variable selection and the MWS parameters are optimized simultaneously by employing the mixed integer genetic algorithms.(2)WED similarity measure can effectively improve the predicition performance of JIT learning soft sensors by considering the importance of the input variables to the output variable.However,it is difficult to obtain a set of optimal weights assigned to input variables.Therefore,an ensemble JIT learning soft sensor method using regression coefficients as the weights of WED similarity measures is proposed.Firstly,a set of diverse training subsets are created by repeatedly performing random subspace construction and Gaussian mixture models clustering.Then,the weights of input variables are determined by using PLS regression,thus allowing definition of a set of diverse WED similarity measures.Finally,the ensemble strategy is used to combine all individual JIT models for providing the final prediction.(3)The difficulty in building a high-performance ensemble JIT learning soft sensor model is to obtain a set of base JIT learning models satisfactory diversity and accuracy.Thus,a novel ensemble JIT learning soft sensor modeling framework is proposed through perturbating similarity measures.This method generates diverse WED similarity using evolutionary multi-objective optimization,and then constructs a set of JIT learning base models which satisfying diversity and accuracy objectives.Then the stacking strategy is used to combine the base models to obtain a high-performance ensemble JIT learning soft sensor model.(4)To achieve an effective balance between the complexity and accuracy of JIT model and consider the differences and advantages of different similarity criteria,an ensemble JIT learning soft sensor method based on triple-modal perturbation is proposed.Firstly,various similarity measures are used to construct JIT learning models,for each of which the model structure and input variable selection are determined through evolutionary multi-objective optimization by minimizing the model complexity and maximizing the prediction accuracy.Then,all base JIT models are conbined through selective ensemble learning,which achieves ensemble size reduction while maintaining or even improving prediction accuracy of JIT models.