Near-Infrared Spectroscopy for Online Glucose Detection in Fermentation Processes: Transflectance/Transmission Sensor Evaluation and Modeling Optimization

This study employed near-infrared (NIR) spectroscopy for real-time spectral acquisition of fermentation broth in lab-scale bioreactors, comparing the performance of transflectance and transmission sensors through glucose modeling and prediction while optimizing modeling approaches. The results demonstrated superior adaptability of transflectance sensors in fermentation environments: in conventional fermentation, glucose models exhibited lower errors (RMSEC = 4.087 g/L, RMSEV = 9.829 g/L) compared to transmission sensors (RMSEC = 5.972 g/L, RMSEV = 10.904 g/L), with significantly higher predictive performance (RPD = 3.735 vs. 2.369), indicating enhanced fitting accuracy and stability. In complex natural media containing peptone and yeast extract, transmission sensor performance deteriorated dramatically due to turbidity interference (R2cal = 0.134), whereas transflectance sensors maintained robust performance (R2cal = 0.993), confirming their adaptability to complex matrices. Regarding modeling strategies, the 1550–1700 nm spectral region demonstrated optimal feature extraction capability (RMSEC = 3.269 g/L, R2cal = 0.987). Basic preprocessing methods such as the moving average smoothing method have become the preferred preprocessing methods, as they strike a balance between calibration and prediction performance. Outlier removal analysis revealed that moderate elimination of 12 high-error samples (accounting for 30% of the total 39 samples) reduced RMSEC to 1.441 g/L and improved R2cv to 0.996, optimizing model performance; however, excessive removal of outlier samples degraded model capability, necessitating judicious sample selection. For fixed total sample sizes, calibration sets comprising 70–80% of samples yielded more reliable predictions. In conclusion, transflectance sensors demonstrate superior compatibility with multicomponent fermentation systems. Combined with wavelength selection, moving average preprocessing, and rational sample removal and partitioning strategies, this approach provides an effective solution for NIR-based online glucose monitoring.