Study On Model Building And Model Transfer In The Application Of Near Infrared Spectroscopy To Online Quality Control Of Chinese Herbal Liquid Tonic

Near infrared spectroscopy(NIRS)is a widely used process analytical technology that can be applied to all areas of quality control in the production of Chinese herbal medicines,largely due to its fast,efficient,real-time,no pretreatment,no sample destruction characteristics.However,A NIR spectrum is greatly affected by the equipment and application environment.When the detection environment or the equipment changes,the predictive performance of the original calibration model may deteriorate under new conditions.In order to address this challenge,the concept of model transfer was proposed by previous researchers.By correcting the spectral data or the original model,the model can still provide good predictive performance under the new conditions.The majority of previous model transfer methods relies on the availability of standard samples for which spectra were taken using both the master and the slave spectrometers.In this work,a new model transfer approach was proposed that makes use of virtual reference samples in applications where no one to one corresponding spectra for the same sample are available.Application of the approach to the monitoring and control of three quality indexes of a Chinese herbal medicine demonstrates the effectiveness of the mew method.The main contents of this paper are as follows:Firstly,taking a Chinese herbal medicine oral liquid product produced by an enterprise as the research object,and taking the model that has already been in use in guiding production as the original model,the online transfer model is studied.As it is difficult to obtain one-to-one standard sample in industrial production,polysaccharide,an important quality index of oral liquid,has no obvious characteristic absorption peak in the near infrared spectrum region,and the transfer method using the standard sample-free model has low transfer precision and large workload.Therefore,this paper combines the sample without standard sample with the model transfer method with standard sample,and uses virtual standard sample to transfer the model online through the model transfer method with standard sample.Under the condition that the master and slave machine samples are unified,for the standard standard samples that cannot be obtained in one-to-one correspondence,the master spectrum whose error does not exceed 3% after one-to-one correspondence with the slave machine samples is used as the virtual standard sample for model transfer with standard samples.The model transfer method is a widely used direct standardization method,and combines principal component dimension reduction to improve model transfer efficiency and transfer performance.In the transfer process,the transfer results of the virtual standard sample and the standard standard sample are compared.The results show that it is feasible to transfer the model with standard samples by using virtual standard samples in production,and the relative error of the prediction of the transfer model meets the error limit required by industry.It provides a new way of thinking for the model transfer process in which it is difficult to obtain one-to-one standard sample in industry,and realizes the model sharing between different instruments and accessories.Secondly,the capability for generalisation of the model transfer method is studied.The master and slave instruments of this application are different not only in spectral detection equipment and accessories,but also in the spectral resolution and number of spectral variables.The spectral variables of the Chinese herbal medicine oral liquid product are 1557 for the host instrument and 778 for the slave machine.The spectral variables are different,which raised challenges to the generality of the model transfer approach.In this work,the spectra of the host computer each with 1557 variables are compressed into spectra each has 778 variables;and the host instrument model was reconstructed used the the compressed spectra each of 778 variables.In this way,the reconstructed host instrument model was used as the original model for model transfer operation.After model transfer the error of prediction for the new instrument for three quality indicators met error tolerance.