First Principle And Data-driven Based Soft Sensors For Key Process Parameters In Batch Sugar Crystallization Process

The formation and growth of crystals in sugar process are performed in complicated conditions constructed by temperature,vacuum,concentration,mother liquor supersaturation,liquor purity and other thermodynamics.The regular operation of sugar crystallization is to control the mother liquor supersatruation.However,the dynamic changes in purity will pose great effect on the crystallization rate and mother liquor supersaturation.Also,sugar crystallization is a nonlinear and unstable process involved a series of physical dynamic changes,including nucleation,growth and aggregation.The inadequate knowledge of the process mechanism and its dynamics will result in a variety of issues,such as unsatisfied prediction of crystal size distribution(CSD),varied crystal sizes and low sugar production.Hence,a deep understanding of thermodynamics and dynamics is demanded for a better control of sugar crystallization.Hardware sensor for mother liquor supersaturation,purity,crystal content and CSD is still unrealistic at present.To resolve this problem,this thesis conducted a wide range research on soft sensor for sugar crystallization process and its key methods,which proceeded as follows:(1)A twin support vector machine(TSVR)based soft sensor was proposed to predict the supersaturation and purity of mother liquor.Its performance had proved to be superior.(2)Empirical risk minimization(ERM)principle in the data-driven model was likely to yield over-fitting problems.Aimed to solve this issue,the structural risk minimization(SRM)principle was introduced to replace the ERM,where the complexity of model was described by regularization term.The square kernel matrix of the model was replaced with rectangular to reduce the model computation time and to obtain better generalization,with the model performance remained uncompromised.In the meanwhile,to identify the importance of each sample,the weight coefficient was introduced to assign different penalties to the sample set,which would avoid over-fitting problem at some extent.(3)From the perspectives of energy balance,population balance and mass balance of sugar process first principle model,the crystallization and its dynamics were deeply discussed.A hybrid model constructed by first principle and data-driven model was built to describe the nucleation and growth rate of crystals,and complex aggregation in particular.Based on the knowledge of crystallization process,crystal content and CSD were estimated with high accuracy,which provided basics for automatic control for sugar process.(4)Targeted at the drawbacks of current automatic control system for sugar crystallization,this thesis put forward an overall scheme for a multiple intelligent techniques integrated automatic monitoring system.Its work included the upper computer design of the system and the implementation of data-driven and hybrid model based soft sensor components.In this way,an automatic intelligent system capable of on-line monitoring mother liquor purity,supersaturation,crystal content and CSD was proposed.