| At present,stevia is more and more favored by food and drug industry because of its high sweetness and low heat.Among them,Rebaudioside A(RA)is better than other stevia glycosides in terms of sweetness and physicochemical properties.What’s more,its characteristics are the most similar to sucrose,making it an effective substitute for sucrose.Therefore,how to obtain high quality RA is the hot and difficult spot that stevia industry needs to break through.The quality of raw stevia leaves and the crystallization and purification process are the key factors limiting the yield and purity of RA at present.However,for the above two key links,there is currently a lack of deep understanding and effective process evaluation and characterization methods.In view of this,this study mainly focused on the identification and quality screening of stevia leaves,and the optimization and quality control of the key process of RA crystallization purification,to break through the bottleneck of production technology and quality monitoring technology.The quality characterization method of stevia leaf screening process and the "visualization" technical system of crystallization process were put forward,so as to ensure that the quality of joint link in the production process is visible,the product quality is uniform and stable,and provide experimental basis and theoretical basis for the improvement of the core competitiveness of domestic stevia production enterprises.Firstly,an evaluation system based on near infrared spectroscopy(NIR)was established to solve the problems of different varieties and quality of stevia leaves.In terms of static modeling,the effects of the resolution of the NIR spectrometer,the number of scans,the particle size of the sample and the drying time on the stability and repeatability of the NIR spectral data were investigated.Then,principal component analysis(PCA)and soft independent modeling of class analogy(SIMICA)and partial least squares discrimination analysis(PLS-DA)were used to establish the strain discrimination model.In terms of dynamic modeling,online quality discrimination of stevia leaves based on conveyor belt transport mode was carried out.The influence of optical path instability was eliminated by using angle spectral conversion method,and then PCA and cluster analysis were used to conduct online quality discrimination of stevia leaves.The results showed that in terms of static modeling,when the samples of different varieties of stevia leaf were crushed through a 100-mesh screen and placed in a drying oven at 55℃ for 10 h.And under the condition of resolution of 16cm-1 and scanning times of 32,the collected NIR spectra had good stability.Using PCA,SIMICA and PLS-DA methods,the qualitative models of high RA type,high STV type,RA type and mixed type of stevia leaves were successfully established.The PLS-DA discriminant model established after the preprocessing of original spectra by multiplicative scatter correction(MSC)was the best,and the recognition rate and rejection rate of the model reached 100%.In terms of dynamic modeling,the influence of optical path could be effectively eliminated by angle spectral conversion,and then PCA could exclude part of the inferior leaves from the 95%control limit.Cluster analysis could also effectively identify the residual branches and withered leaves.This part of the study showed that the NIR spectral analysis method could effectively evaluate the quality of stevia leaves in both static and dynamic aspects,improve the understanding of product quality in the field acquisition and production feeding,strengthen the design of product quality,and provide a basis for quality by design.Secondly,crystallization is another key unit operation that affects the content of RA.In this study,mixed solvent crystallization method was used to optimize the crystallization processes of RA80(The content of RA>80%)and RA98(The content of RA>98%),and the optimal process conditions were determined.Meanwhile,Raman imaging method was used to establish a visual characterization technology for the change of solid phase system during the crystallization process of RA.In terms of the crystallization process,the purification process of RA 80 was firstly investigated,and then the purification process of RA 98 was investigated with the content and transfer rate of RA as the indexes.The type and proportion of the solvent,the amount of the solvent,the time of heating reflux and cooling were investigated by single factor experiment to determine the best crystallization process,and the pilot test was carried out at last.Raman imaging was used to characterize the content trend of RA in the solid phase system during the crystallization process.The results showed that the production process of RA80 was as follows:the total stevia glycosides with 40%RA was added with 8 times amount of ethanol to reflux for 1.5 h,then crystallized for 4 h,filtered and dried.The production process of RA98 was as follows:RA80 was added with 12 times amount of mixed solvent(methanol:ethanol=1:2)to reflux for 2 h,then crystallized for 4 h,filtered and dried.The imaging results of solid samples during the crystallization process showed that the content of RA in the crystals gradually increased with time,and showed a rapid change in the early stage and a gradually stable crystallization process in the later stage.Through the research of this part,two RA purification processes were established,which were simple,easy to operate and transfer on a large scale.At the same time,a visual qualitative characterization method of RA in the production process was established based on Raman spectroscopy.Thirdly,on the basis of process optimization,the introduction of NIR process analysis technology(PAT)can help to better understand the crystallization process and lay the foundation for quality by design.Bench type laboratory NIR instruments are large in volume and high in maintenance cost,while micro NIR instruments are difficult to achieve wide band coverage due to size limitation.Based on this,this paper carried out the fusion design research of the micro NIR sensor,and developed a set of near infrared spectrum analysis software that can realize the simultaneous acquisition of spectra by two sensors.A fusion sensor detection system based on transmission mode was established to meet the requirements of the liquid phase system samples in the crystallization process of RA.Firstly,the hardware construction of the micro NIR sensor detection system was completed,and then the corresponding near infrared spectrum analysis software was developed.The software wass realized based on Python,and the function was designed according to the user’s requirements.Pyserial module completed the underlying communication,and PyQT5 completed the writing of the interactive interface.The main functional modules of the software included data acquisition,data analysis and data storage.Through the research of this part,a set of micro NIR sensor fusion detection system was established.It can quickly complete the development of customized software for special collection purposes according to different application scenarios,and can detect by any combination of instruments covering different bands according to the characteristic spectrum of the collected object.It has the advantages of portability,multi-scene application and flexible transformation of collection bands,etc.,which lays a foundation for subsequent spectral data fusion analysis,and also establishes a theoretical foundation for instrument miniaturization and online detection and analysis.Fourthly,aiming at the change rule of RA in liquid phase during crystallization process,the research of micro NIR sensor fusion based on linear modeling method was carried out.The liquid phase system was sampled and filtered at intervals during the crystallization process,and the NIR spectra were collected using the spectral sampling system established in Chapter 4.The linear correction model of RA was established by NIR data non-fusion method,low-level fusion method and intermediate fusion method.The best linear correction model of RA was determined by comparing the modeling results to realize the analysis of RA crystallization process.As it turned out,the root mean squared error of calibration(RMSEC),root mean squared error of cross validation(RMSECV),root mean squared error of prediction(RMSEP),(?)of PLS model established by sensor 1 were 4.417mg/mL,8.990mg/mL and 3.440mg/mL,0.8719,0.4628,0.9081 respectively,and relative percent deviation(RPD)were 3.298.the RMSEC,RMSECV,RMSEP,(?)of PLS model established by sensor 2 were 3.868 mg/mL,6.356 mg/mL,3.220 mg/mL,0.9018,0.7315 and 0.9195 respectively,and RPD was 3.524.It is proved that both sensor 1 and 2 can be used for the quality characterization of liquid phase system during RA crystallization.The results of the fusion data study show that compared with the results of sensor 1,the RMSEP of the original spectral low-level fusion PLS decreased by 21.90%,(?)and RPD increased by 3.94%and 28.05%respectively;compared with sensor 2,RMSEP decreased by 16.55%,(?)and RPD increased by 2.65%and 19.84%respectively.The intermediate data fusion model also shows better prediction performance.Compared with the single PLS model of sensor 1 and sensor 2,the RMSEP of the SO-PLS model decreased by 10.95%and 4.86%,(?)increased by 2.41%and 1.14%,and RPD increased by 12.31%and 5.11%respectively;the RMSEP of So-CovSel model decreased by 10.46%and 4.33%,(?)increased by 2.41%and 1.14%,RPD increased by 11.67%and 4.51%respectively.Compared with the PLS model,the sequential model of sensor 1 decreased by 2.01%in RMSEP and increased by 0.19%in(?);the sequential model of sensor 2 decreased by 4.33%in RMSEP and increased by 1.14%in(?),and the RPD of the sequential model has only a very small increase.In conclusion,the modeling effect of the linear model based on data fusion is better than that of the non-fused single sensor model,indicating that the data fusion technology can effectively improve the prediction accuracy of the model,in which the original spectral low-level fusion PLS is the best linear characterization model for RA quantitative.In this part,the best linear characterization model for the quantitative characterization of RA in liquid phase system during crystallization process was established.The rapid determination of RA content in liquid phase system can help to understand the crystallization process and determine the end point of crystallization scientifically,and also provide a modeling idea for the realization of multi-sensor information fusion and complementation.Finally,the convolutional neural network(CNN),an artificial intelligent(AI)algorithm,was used to describe the crystallization process non-linearly,in order to establish a new modeling algorithm based on Al.In this part,one-dimensional CNN was used to realize quantitative analysis of RA,and the modeling research of low-level fusion,outer product fusion and PCA fusion of original spectrum between micro sensors was explored by using CNN feature extraction.The results showed that the quantitative model of RA established by CNN had better predictive ability than the PLS model.The RMSEP of CNN model of sensor 1 was 4.96%lower than that of linear PLS model,while(?)and RPD increased by 2.41%and 5.06%respectively;the RMSEP of CNN model of sensor 2 was 1.68%lower than that of PLS model,and(?)and RPD increased by 1.36%and 10.61%respectively.In the study of data fusion,the original spectral low-level fusion CNN model with the best modeling effect had a prominent performance compared with the single sensor PLS model.compared with the PLS model of sensor 1 and sensor 2,RMSEP of the fusion model decreased by 20.97%and 15.56%,(?)increased by 3.73%and 2.45%,and RPD increased by 37.66%and 28.83%respectively.The study in this part shows that the CNN non-linear model based on AI can better describe the crystallization process,which lays a foundation for the study of non-linear crystallization kinetics.To sum up,this paper studies the operation of the bottleneck unit that limits the quality of RA.In terms of raw materials,the static and dynamic quality identification systems of stevia leaves were established by NIR spectroscopy.In terms of crystallization and purification,a micro NIR fusion sensor detection system was designed.NIR technology and Raman spectroscopy combined with data fusion method and AI modeling method were used to visualize the quality of solid phase and liquid phase in the crystallization process of RA.The research of this paper strengthens the understanding of product quality formation,and also lays the foundation for quality from design. |
