Handheld near infrared spectrometer and machine learning methods applied to the monitoring of multiple process stages in industrial sugar production.

This work aimed to evaluate the performance of a handheld NIR spectrometer in developing calibration models to quantify brix and pol at various stages of an industrial sugar production process. Because of sample variability, collected over two harvesting seasons, NIR measurements were acquired either in transmittance or diffuse reflectance. For modelling purpose, partial least squares (PLS), also combined with variable selection techniques, and support vector machine regression (SVR) were investigated. SVR was applied to handle non-linearities within the data. In general, results illustrated the best performance of SVR, that yielded lower root mean square error of prediction (RMSEP) values for brix and pol for spectra acquisition in transmittance (0.59 and 0.69%w/w) and using diffuse reflectance (1.44 and 2.44%w/w), respectively. Results from models using spectra collected in transmittance were comparable to those reported in other works where benchtop instruments were used, highlighting the cheaper and simpler employment of the portable spectrometer.