A Comparison of CNN and PLSR for Glucose Monitoring Using Mid-Infrared Absorption Spectroscopy

With the development of mid-infrared (MIR) photoelectric devices, spectroscopy has become one important methods for non-invasive detection blood glucose. The region (4000 - 400 cm-1) well-known fingerprint (1200 800 glucose, which clearer characteristic absorption peaks and better specificity. There is a lot molecular information about glucose in MIR. by needs to achieve certain accuracy, quantitative model an factor affecting accuracy detection. In this paper, samples imitation solution containing only mixed are taken as research objects, spectral data collected. full spectrum partial least squares Regression (PLSR) model, SNV + Ctr-PLSR MSC convolutional neural networks (CNN) 3000 900 cm-1 band were constructed. Full PLS CNN 1200 experimental results show that optimal two bands CNN, then correlation coefficient prediction set (Rp) 0.95, root mean square error pre-diction (RMSEP) value 22.10. Rp RMSEP 22.54. promising method, higher than PLSR, especially suitable modeling human complex environment. addition, study provides theoretical practical basis feature selection interpretation.