Abstract: Taking fermentation broth of fermentation process of L-isoleucine as the samples, using partial least square method, we investigated effects of two spectral acquisition methods, 7 spectral pretreatment methods and different spectral band selection on the accuracy of prediction models of 5 amino acids. By establishing the optimal spectral prediction model of the main and byproducts in the fermentation process of L-isoleucine, we determined the optimal spectral information acquisition method, spectral pretreatment method, spectral wavelength range and the number of factors of the model. The results show that the optimal prediction model of the spectrum obtained by reflection scanning is superior to that of transmission scanning. The correlation coefficients (Rc) of the optimal correction model of L-Ile, L-Ala, L-Glu, L-Leu and L-Thr are all greater than 0.96, the root mean square error of cross validation (RMSECV) are 1.760,0.462,0.430,0.259,0.199, and the residual predictive deviation (RPD) are 7.8,6.8,6.3,5.0,6.4, respectively. 
Therefore, the proposed near infrared spectroscopy is stable and feasible for rapid detection of various components in amino acid fermentation broth.

Key words: amino acid detection, nearinfrared spectroscopy, spectral acquisition, chemometrics