Research on nonlinear quantification of Rebaudioside A crystallization process based on near-infrared sensor fusion.

Purposes: Rebaudioside A (RA) is a natural sweetener whose key manufacturing step is crystallization. However, a rapid detection method suitable for this process is currently absent. With miniaturization becoming a trend in industrial applications, near-infrared (NIR) spectroscopy has become a powerful tool for understanding the manufacturing process of natural products. The potential of miniature NIR for quantification of RA content and visualization of the crystallization process was investigated in this study. Methods: The RA crystallization process was simulated in our laboratory. During the crystallization, the NIR spectra were collected using two micro-sensors (1350 ~ 1650 nm and 1550 ~ 1950 nm). To improve the performance of the model, convolution neural network (CNN) with sensor fusion was proposed. Results: The CNN model based on raw spectral fusion showed the best predictive capacity with RMSEP and RPD of 2.895 mg/mL and 3.919, respectively. Conclusions: Overall, the research showed that the CNN model with data fusion was promising for quantifying and visualizing RA crystallization based on NIR. This study promoted the popularization of portable spectrometers by providing a theoretical basis for instrument customization. [ABSTRACT FROM AUTHOR]