Research On Online Monitoring Method For Solution Concentration In Cooling Crystallization Process Based On ATR-FTIR

On-line monitoring of the crystallization process is of great significance to the safe and stable operation of the process,ensuring product quality and improving the level of industrial control.Solution concentration is the most important variable in the crystallization process,which directly reflects the supersaturation of solution,so its measurement techniques have received extensive attention and research.With the development of in situ measurement techniques,spectroscopy has been widely used to monitor solution concentration in cooling crystallization processes due to its noninvasive and its ability to reflect changes at the molecular level in real-time.In this thesis,the problems of data pre-processing,variable selection,and calibration model building in the process of online monitoring of solution concentration using Attenuated total reflection-Fourier transform infrared spectroscopy are investigated in the context of the glutamate cooling crystallization process.The main research elements are as follows.To address the problem of reduced model prediction accuracy due to interference in the spectral data,an empirical modal decomposition technique is introduced to complete spectral preprocessing.Regardless of the sequence and parameters of the preprocessing steps,this method which is an adaptive preprocessing method based on data drive removes both noise and baselines from the spectrum.In addition,considering the effect of temperature variations on the spectrum during cooling crystallization,a bidimensional empirical modal decomposition method was proposed to simultaneously extract the spectral wavenumber and the absorbance change of sampling time to eliminate the effect of single sampling and full process variations on the spectral data.Comparative experiments demonstrate that the bidimensional empirical modal decomposition pre-processing method can effectively eliminate the interference of spectral data in-process monitoring and significantly improve the accuracy of the concentration calibration model.Due to the high dimension of spectral data and many irrelevant variables,the prediction accuracy of the calibration model is low and the interpretability is poor.Therefore,based on the model population analysis strategy,this thesis puts forward the corresponding improvements in three aspects,namely,sampling method,importance evaluation index,and sub-model modeling algorithm.Firstly,an iterative shrinkage window sampling method is proposed to solve the problems such as poor stability caused by univariate sampling;secondly,an important evaluation index weighted by the stability of regression coefficients and the frequency of selected variables is proposed to change the limitations of a single evaluation;finally,a modeling strategy based on switching between different partial least squares algorithms is proposed,which accelerates the modeling speed of sub-models and improves the computational efficiency.The experimental validation shows that the calibration model established by the improved algorithm for selecting variables has higher accuracy and stability,and the speed of selecting variables is faster,which is of practical application to improve the accuracy and reliability of measuring solution concentration in cooling crystallization process using ATR-FTIR spectroscopy.To address the non-linearity between absorbance and concentration due to the dynamic changes in the crystallization process and the influence of the external environment,an extreme learning machine is introduced in this thesis for the modeling of quantitative concentration analysis.In addition,an adaptive robust extreme learning machine algorithm is proposed to reduce the influence of the possible outliers in the reference concentration provided by off-line laboratory measurements on the extreme learning machine.This algorithm replaces the squared loss in the extreme learning machine by introducing a generic robust loss function and estimates the output weights using iterative weighted least squares.Meanwhile,the Bayesian optimizer is used to achieve adaptive optimization of the loss function and model parameters during the training.Simulation results show that this method can effectively avoid the influence of abnormal concentration value on model updating without additional correction of outof-bounds sample process,which is of certain significance to realize on-line modeling and maintenance of solution monitoring.