Research On The Transfer Method Of The Near-infrared Quantitative Model Of The Water Content Of The Small-scale Test In The Fluidized Bed Granulation Process To The Pilot Test

Near infrared(NIR)spectroscopy,as a process analysis technology,has the advantages of fast,non-destructive,easy operation,and in-line real-time monitoring.It has been widely used in online quality control and quantitative analysis of traditional Chinese medicine preparation processes.However,the NIR quantitative model established by the limited calibration samples has a certain scope of application.When the spectral information variation of the test set is large and exceeds the spectral information range of the calibration set,the NIR quantitative model is difficult to be directly applied to the test samples,and the model transfer methods are needed to accurately predict the unknown variant samples.In this paper,the applicability and model transfer methods of the NIR quantitative model established by the fluidized bed granulation process samples were investigated at the small test and the pilot test.The main research contents of this paper were as follows:1.This paper collected NIR spectrum and moisture content data of six batches of small test and pilot test in the fluidized bed granulation process of Yixintong tablets.The mtohod of Multivariate Statistical Process Control(MSPC)was used to explore how to preliminarily judge the applicability of NIR quantitative models in different batches and different scales from the perspective of spectrum.One,two,and three small test batches were used as modeling batches M1,M2,and M3,respectively,to establish the small test NIR quantitative models with different applicable scopes.MSPC method was used to observe the degree of deviation from the model in the figures of PCA and Hotelling T2 of the test batches to predict the applicability of the model.The results showed that when the number of modeling batches were little,the spectrum information of most other small test batches exceeds the 95%control limits of the M1 PCl-PC2 scores scatter plot and the Hotelling T2 plot,which indicated that M1 NIR quantitative model might not apply to other small test batches.When three small test batches were used as modeling batches,the spectrum information of the other small test batches were within the control limits of the M3 PCA scores scatter plot and the Hotelling T2 plot.It was concluded that the M3 NIR quantitative model could be applied to other small test batches.The results of the NIR quantitative analysis showed that the relative standard errors of prediction(RSEP)values of the other small test batches predicted by the M3 NIR quantitative model were 6.25%,4.17%and 4.62%reduced from 43.48%,28.41%and 3.93%.predicted by M1 model.The MSPC analysis results of the six pilot batches with the M3 PCA model showed that the spectral information of the six pilot batches were within the control limits of the M3 PCA scores scatter plot and the Hotelling T2 plot.The RSEP values of the six test batches predicted by the M3 NIR quantitative model were 5.10%,4.77%,5.01%,4.95%,8.19%,and 5.04%,which all met the prediction accuracy requirements of NIR quantitative analysis.The above results showed that when the applicable range of model was small,it was difficult for the NIR quantitative model to accurately predict other batches with large spectral variations.When three batches were used to build the NIR quantitative model,the small or pilot NIR quantitative models could be applied to the other batches at the same scale.Morever,the NIR quantitative model of small-scale trial can be applied to the pilot batches,which means that the spectral information of small-scale trial batches and pilot batches was essentially the same.When the scope of modeling information was large enough,it was feasible for NIR quantitative model of small-scale trial to predict the pilot batches across scale,and MSPC analysis could effectively reflect the applicability of NIR quantitative model for test batches from the spectra perspective.2.For the test sets whose spectral information were beyond the applicable range of the model,this study proposed the model transfer method of "Directed Direct Orthogonal Signal Correction Combined with Slope/Bias Correction"(DDOSC-SBC).This method is aimed at the problem that the DOSC orthogonal projection parameters of the modeling set may not be suitable for the test set.The DDOSC orthogonal parameters suitable for the test set were obtained through a small number of representative samples in the test set in order to correct the accidental errors in the test set spectrum more instructively and effectively.Then correct the residual systematic errors in the modeling set and the test set spectrum with the SBC method,so as to realize the accurate prediction of the NIR quantitative model for the test set with large variation.This paper collected the NIR spectrum and moisture content data from the six batches of small test(tf,df,sh1,sh2,stl and st2)and pilot test(TF,DF,SH1,SH2,ST1 and ST2)of fluidized bed granulation process of dextrin,and increase the samples variability at different scales and batches by adjusting the formulation parameters and changing the sampling environment to explore the effectiveness of the NIR quantitative model cross-scale transfer method.The RSEP values of the shl and stl predicted by the original small trial NIR quantitative model established by the tf,df and st2 batches were 22.88%and 13.43%,respectively.After using the DOSC model transfer method,the RSEP values of shl and st1 batches were reduced to 5.45%and 7.24%,which indicated that the DOSC method can effectively remove the interference information from the spectrum.The RSEP of the SH2 batch predicted by the original pilot NIR quantitative model established by the pilot TF,DF and ST1 batches was 73.60%.After using the DOSC method and DOSC-SBC,the RSEP values of the SH2 batch was 85.33%and 20.42%,respectively.Eight representative samples were selected from SH2 through the KS method combined with the PCA scores plot.the RSEP values of SH2 after applying DDOSC and DDOSC-SBC model transfer methods were 48.41%and 6.21%,respectively.Both the correlation graphs of the predicted values and reference values with the DOSC and DDOSC methods showed that the residuals of the predicted values were biased to one side,which indicated that there were systematic errors remaining in the spectrum.In addition to,the predicted values with DOSC method were not linear with the reference values.The disordered samples distribution reflected that in addition to systematic errors in the spectrum,large accidental errors remained.So,the DOSC-SBC method did not achieve good results.However,the prediction values obtained by the DDOSC method have a good linear relationship with the reference values.So,after combining the SBC to correct the initial prediction values,the accuracy of the prediction result has been improved.Using small trial NIR quantitative models to predict cross-scale pilot test sets B(SH1 and ST1)and C(TF,DF,and ST2),their RSEP values were 24.76%,and 16.28%,respectively.Using DDOSC-SBC,Model Update(MU)and MU-DOSC model transfer methods,respectively,the RSEP values of test set B decresed to 4.85%,6.58%and 3.68%,the RSEP values of test set C decresed to 5.23%,4.90%and 14.90%.While the MU-DOSC and DDOSC-SBC methods showed similar results.However,the MU method needed to be re-modeled.And the effect of the MU method was greatly affected by the original model set samples.When there were many original modeling samples,and the newly added representative samples were too few,the correction effect of the model will not good.When the variation of the newly added representative sample spectrum was too large,it might be considered as the abnormal samples and reduce the model performance.The DDOSC-SBC method does not need to change the model,which can well avoid such problems.In summary,the NIR quantitative model of the small-scale preparation process can be used in pilot scale batches across scales,and the applicability of the model can be preliminarily judged by the MSPC method to provide a basis for the precise application of the NIR quantitative model.When the spectral variation of the test set is too large,the DDOSC-SBC method can achieve accurate prediction between batches of the same scale or cross-scale by the NIR quantitative model through the correction of the spectrum and the predictive values.It provides a method for the process quality control of traditional Chinese medicine preparations and can promote the practicality of NIR spectrum technology in the process of solid preparation of traditional Chinese medicine.