Research On Online Measurement And Modeling Method Of Solution Concentration In Cooling Crystallization Process

Cooling crystallization is commonly used for separation or purification of solid substances according to the solubility and temperature,which is widely used in the separation and purification of organic substances in pharmaceutical,fine chemical,food and other industrial processes.The measurement of solution concentration,especially online measurement,is the foundation for the modeling and control of crystallization processes.The measurement accuracy and reliability not only affect the accuracy of the crystallization process kinetic model,but also directly affect the control effects of product quality,yield,crystal purity,and other aspects of the crystallization process.In recent years,with the continuous development of process analytical technology,various online measurement techniques of solution concentration have received widespread attention and significant development.Spectroscopic techniques have become the primary method for online measurement of solution concentration in crystallization processes due to their advantages of real-time in-situ and non-destructive.The accuracy and reliability of solution concentration prediction models have been studied from the perspectives of data processing,variable selection and modeling methods based on the glutamic acid cooling crystallization by Attenuated Total Reflectance-Fourier Transform Infrared(ATR-FTIR)spectrometer.The main research contents are given as follows:(1)This dissertation systematically studies various technologies for solution concentration measurement in crystallization processes,analyzes their measurement principles,advantages and disadvantages,classifies them according to the measurement mode,and systematically summarizes the research progress,application status,and existing problems of various measurement techniques.In particular,the online measurement principles of solution concentration and the chemometrics methods research progress based on spectroscopic technology are expounded in focus.The dissertation also points out the current issues and potential prospects of spectroscopic measurement technologies,and provides directions for the development of solution concentration measurement technologies in crystallization processes.(2)For the nonlinear spectral issues arising from temperature changes,noise,and other factors in online ATR-FTIR spectra,this dissertation investigates the effects of data preprocessing techniques on the calibration model of solution concentration measurement.Four groups of 24 candidate calibration models were established based on the partial least squares regression(PLSR)method,combing variable selection with baseline correction,smoothing,difference reduction and scale transformation preprocessing method.All the candidate calibration models were verified by validation experiments at different concentrations and temperatures,as well as solubility experiments of glutamic acid in water.The results demonstrated that the data preprocessing by combing smoothing with derivatives can effectively eliminate spectral nonlinearities,which also shows a distinct advantage in detecting ATR probe nucleation.(3)For the issues of the difficulties in improving the model accuracy and reliability that used the full-spectra due to redundant variables,the variable selection approach was studied in this dissertation.A novel two-layer variable selection algorithm was designed by combining Fisher combined population analysis(FCPA)and improved binary Cuckoo search(IBCS)algorithm.The model complexity was reduced without affecting model accuracy and reliability.Results demonstrated that the proposed algorithm can greatly eliminate redundant variables(～88%),and the established calibration model presents higher predictive ability and reliability compared with other variable selection algorithms.(4)For the problem of low model accuracy of existing linear modeling methods due to the nonlinearity between solution concentration and spectra affected by solution temperature and optical fiber,this paper investigates the effects of machine learning methods such as Random Forest and Support Vector Machine(SVM)algorithms on model prediction performance from the perspective of modeling methods,and compares them with PLSR modeling methods.The model performance was verified by experiments on glutamic acid cooling crystallization process.Results showed that the calibration model based on SVM has higher prediction accuracy and robustness compared with the conventional PLSR modeling method.Finally,a simple,accurate,and reliable online prediction model of solution concentration by combing data preprocessing,variable selection,and modeling method was established based on ATRFTIR spectroscopy in crystallization process.