Modeling And Application Of Near-infrared Spectral Detection For Drying And Fermentation Processes

Near-infrared(NIR)spectroscopy has been increasingly applied in industrial batch manufacturing processes such as granules drying and biological fermentation processes in recent years,owing to its advantages of nondestructive detection,fast analysis,and no pretreatment on samples.However,the existing NIR spectral analysis methods are mainly based on taking samples for off-line measurement,but could not be used for on-line measurement or may cause notable measurement errors,let alone on-line quality monitoring.In this dissertation,batch processes of fluidized bed drying(FBD)granules and glucose fermentation are investigated for studying the online measurement of granule moisture content during the drying process,calibration model building of the main component contents during the glucose fermentation process and online monitoring by using in-situ NIR spectroscopy.The main research contents and contributions are as follows.(1)To address the problem that measuring the granule moisture content during the drying process by NIR spectroscopy is affected by the process operating conditions,a spectral calibration model building method by incorporating the operating conditions is proposed to ensure the accuracy of the real-time on-line measurement.A partial-least-squares(PLS)regression modeling method is presented by taking two important operating conditions,i.e.,FBD chamber temperature and heating energy,together with the in-situ real-time collected NIR spectra as the modeling variables for spectral calibration.In order to eliminate the negative effect of measurement outliers in actual detection,a median-absolute-percentage-error(MdAPE)index is proposed to determine the optimal number of latent variables for PLS modeling.The resulting spectral calibration model could achieve a better fit and prediction of granule moisture content under different operating conditions.The validity of the proposed method and its advantages over the existing off-line measurement methods are verified through experiments on batch FBD processes of silica granules under similar and different operating conditions.(2)Concerning the problem that insufficiently labeled samples could be prepared for NIR spectral calibration model building for some kinds of batch FBD processes of granules,a semisupervised PLS calibration model building method is proposed based on variational inference.Its salient advantage is that all the spectral data collected in real-time measurement(including some spectra corresponding to those samples for offline measurement,named as labeled spectra,and the remaining ones are named as unlabeled spectra)are used for building NIR spectral calibration models based on probability density analysis.To deal with the sparsity of highdimensional NIR spectral input variables with regard to the number of labeled data,the gamma distribution is introduced into the spectral input variables to enable automatic variable selection.Through a numerical case and detection experiments on batch FBD processes of silica granules,it is verified that the proposed method could significantly improve not only on-line measurement accuracy but also the credibility of the training model.(3)Regarding the problem that there often exist inaccurate results and insufficient sample data from off-line measurement for some kinds of batches biological fermentation processes,causing difficulty for building NIR calibration models,a ladder network-based NIR spectral calibration modeling method is proposed.Firstly,a layer-by-layer training strategy is proposed to determine the optimal ladder network model structure and parameters,by minimizing the loss function of the validation set.Then a supervised ladder network model is used to remove those NIR spectral data corresponding to inaccurately labeled samples.Subsequently,a semisupervised ladder network-based NIR spectral calibration model is built based on the accurately labeled spectral data after sieving and a large amount of unlabeled spectral data.A numerical example and experiments on batch glucose fermentation processes are conducted,verifying that the proposed method could not only effectively remove inaccurately labeled spectral data,but also significantly improve the accuracy of the real-time measurement.(4)Concerning the engineering problem of real-time monitoring the endpoint of batch drying process by NIR spectroscopy,a switching model based NIR detection method is proposed to monitor the endpoint of granule FBD processes in real-time.Firstly,a global model has been established for real-time monitoring of the early stage of such a drying process.Then a just-intime local model is constructed for real-time monitoring of the post-drying process and determining the drying endpoint.Hence,not only the difficulty to guarantee accuracy by using a global model to on-line monitor the endpoint of such drying process could be overcome but also the excessive computational cost by using a just-in-time local model to monitor the whole process in real-time could be avoided.In addition,a t-distributed stochastic neighborhood embedding(t-SNE)method is introduced to reduce the input variable dimension of highdimensional NIR spectra,in order to facilitate selecting similar spectral samples in the lowdimensional space for calibration modeling.Experiments on real-time detecting FBD processes of silica granules are performed,verifying that the proposed method could accurately monitor the endpoints of drying processes under similar or different operating conditions.