CQA On-Line Detection And Steam Regulating Valve Opening Control In The Extraction Process For Oral Preparations Of TCM

With the implementation of China’s strategic plan to "go global" in traditional Chinese medicine(TCM),domestic and foreign drug regulatory agencies have put forward higher requirements for monitoring and control of critical quality attributes(CQA)in the production process of oral preparations of TCM.The extraction process,as one of the main steps in the production of TCM oral preparations,is mainly controlled by fixed process parameters,which has the problem of large batch-to-batch quality variation that seriously affects the stability and uniformity of the final quality of the drug.In response to the problems above,this paper takes the extraction process of oral preparation of TCM as the research object,constructs the dynamic mechanism model of CQA,combines the near infrared(NIR)spectral analysis technology,data analysis algorithm and automation control technology,designs and builds the online NIR detection platform,establishes the online rapid detection model of CQA content,develops the online control system of steam regulating valve based on the correlation model of CQA and key process parameters(CPP),to realize the improvement of CQA content.Aiming at the problem that it is difficult to obtain online NIR spectral data efficiently and stably in the extraction process of oral preparation of TCM,six modules including circulation module,filtration module,constant temperature module,debubbling module,detection module and sampling module are designed to build an online NIR detection platform.Based on the conservation law and mass transfer theory,a mechanism model of the extraction process of xiao’er xiao ji zhike oral liquid(XXZOL)is established to elucidates the changes of the content of octafluorine,vapor pressure and storage volume of the extract during the production process,which provided the theoretical basis for the online data acquisition system.The finite element analysis method is used to test the performance of the flow cell,the key part of the detection module.The simulation results show that:under the pressure of 1.2MPa,the flow cell can slow down the flow velocity of the detection channel and reduce the pressure loss.The experiment results of the extraction process of XXZOL show that the online NIR detection platform can reduce the impurity content,maintain the temperature at 35±2℃,and have the ability to remove air bubbles,which can realize the efficient and stable online NIR spectral data acquisition function.Aiming at the problem that the abnormal online NIR spectral data in the extraction process of XXZOL is difficult to identify and the labeling cost is high,an abnormal spectral data classification algorithm(E-SVM-AL)is proposed to reduce the labeling cost of online NIR spectral data and improve the classification performance of interference spectra by using the online NIR spatial features of unlabeled samples and constructing interference samples to balance the number of different types of samples.The algorithm performance validation experiments are conducted through 2717 sets of online NIR spectral data collected from the online NIR detection platform for the extraction process of XXZO,and the results show that the overall accuracy(OA)of the E-SVM-AL algorithm is 90.4%,the abnormal recognition rate(MR)is 94.5%,the kappa coefficient(k)is 0.735,and the computational cost is 218.3s.Compared with principal component analysis(PCA),support vector machine active learning(SVM-AL)and closest support vector marginal sampling(MS-cSV),E-SVM-AL has the best classification performance.Aiming at the problems of high noise and low key information content in online NIR spectral data,a CQA quality evaluation system for XXZOL is established by using high performance liquid chromatography(HPLC)method,and a database for CQA rapid detection model training is built based on HPLC and NIR data.Based on the variational autoencoder(VAE)and piecewise direct standardization(PDS)loss function,a deep learning framework(VasLine)is constructed to mine the hidden relationship between online and offline NIR spectral data,and establish a fast quantitative analysis model for online NIR spectral data and CQA.The experimental results show that the RMSEs of VasLine for the seven CQAs of cynarin,betaine,chlorogenic acid,allantoin A,naringenin,hesperidin and neohesperidin is 0.0038 mg/mL,0.0017 mg/mL,0.0010 mg/mL,0.0382 mg/mL,0.0808 mg/mL,0.0756 mg/mL and 0.0926 mg/mL,respectively,and the mean value of R2 is 0.9731.Compared with other algorithms,VasLine can significantly improve the accuracy of online NIR spectral data modeling,and can effectively determine the inflection points of seven CQAs in the production process,meeting the demand for online rapid detection of CQAs.Aiming at the problem of large amount of data of process parameters in the extraction process of XXZOL,and the problem that the online NIR spectral labeling samples are few and the labeling cost is high,CALF algorithm is proposed to improve the efficiency of sample query and enhance the version space of labeled samples,which is based on using parameter combination and feature selection methods to establish a CPP database for the extraction process of XXZOL,and combined with active learning and conditional variational autoencoder(CVAE),and the association model of CQA and CPP is established.The experimental results show that the RMSEs of CALF algorithm for the seven CQAs are 0.0028 mg/mL,0.0021 mg/mL,0.0011 mg/mL,0.01209 mg/mL,0.0361 mg/mL,0.0749 mg/mL,0.0443 mg/mL,respectively.Compared with results of the greedy algorithm(Greedy),the query by committee algorithm(QBC),and expected model change maximization algorithm(EMCM),the RMSE results of CALF algorithm are optimal and the prediction error is minimum,which proves the feasibility of applying CALF algorithm to model the production process.Aiming at the problem of large quality differences between batches in the extraction process of oral preparations of TCM,a control system for steam regulating valve opening is developed based on the CALF production process model and extraction termination control strategy,with steam regulating valve as the control object.The online control system is used to carrt out the extraction experiments of XXZOL.The results show that the quality parameters of all batches meet the production requirements,and the fluctuation interval of the final seven CQA contents is reduced by more than 30%compared with the historical production process of 16 batches,and the extraction time is shortened by more than 10%.