Semi-Supervised Calibration Modeling of Near-Infrared Spectroscopy via Just-In-Time Learning With Data Augmentation for In-Situ Measurement of Multiple Component Contents During the Fermentation Process

For measuring multiple component contents during a fermentation process by the near-infrared (NIR) spectroscopy, a novel spectral calibration method is proposed to improve the real-time measurement accuracy in this paper. To overcome the lack of assay samples for spectral calibration modelling in engineering practice, a data augmentation (DA) approach is introduced to generate virtual samples possessing a similar distribution with the real samples assayed, by designing a conditional variational autoencoder (CVAE) to construct the generative model. Moreover, a semi-supervised robust probabilistic latent variable regression (SSRPLVR) algorithm is developed to properly use both the labeled and unlabeled spectral data collected for calibration modelling, such that the key features of spectral absorbance with respect to the fermentation process components could be effectively extracted to ensure model accuracy. To accommodate for the spectral absorbance nonlinearity under different component contents and process conditions, a data augmentation based just-in-time learning (DA-JITL) strategy is presented to choose the most relevant samples in the training dataset for calibration model building with respect to the moment of measurement during a fermentation process, which could efficiently improve the in-situ measurement accuracy. A numerical case study demonstrates the effectiveness and advantages of the proposed method. Experiments on a 500 ton fermentation tank of monosodium glutamate production validates that the root mean square errors of prediction on glutamic acid concentration and glucose concentration are reduced by 57% and 37%, respectively, compared to the traditional partial-least-squares regression model.