In this paper, Attenuated Total Reflection Fourier Transformed Infrared Spectroscopy (ATR-FTIR) combined with the multivariate calibration model was used to realize the indirect measurement of the concentration of two main substrates (glucose and sodium glutamate) during the fermentation process of γ-polyglutamic acid (γ-PGA), which could provide feedback information for the fermentation process. The frequent baseline drift phenomenon in the spectrum measurement will seriously affect the performance of the subsequent multivariate calibration model, and it is necessary to use the baseline calibration algorithm to preprocess the spectrum. Most of the popular baseline correction algorithms are based on the Whittaker Smoother (WS) smoothing algorithm. And use integer-order differentials with limited expressive power to constrain the fitted baseline. Because of the poor adaptability of integer-order differential in the existing baseline correction algorithms, we use more flexible fractional-order differentials to constrain the baseline and then propose a baseline correction algorithm based on fractional-order, which realizes the extension of the integral order baseline correction. 5 batches of γ-PGA fermentation experiments were carried out, and the ATR-FTIR spectra of different batches and all batches were subjected to fractional baseline correction respectively; subsequently, the prediction accuracy of each model was improved to some extent. The experimental results show that only in batch 2 the baseline correction effect based on the integer-order is the best; the orders to obtain the best baseline correction effect for other batches were all fractional-order. Italso reflects that the constraint of the fractional-order derivative (including the integer-order derivative) on the baseline is reasonable. At the same time, it is found that the overall baseline correction effect of all batches is far worse than that of a single batch. The reason may be that the baseline of the spectra for each fermentation batch is different. Different orders need to be selected for different batches to achieve the best effect of baseline correction. In addition, the background spectrum was acquired with distilled water as the reference before measuring each γ-PGA fermentation sample. Anegative water peak thus inevitably appears in the wavenumber range of 3 100~3 600 cm^-1 and forms harmful interference signals; the fractional baseline-corrected spectra show that the fractional-order baseline correction algorithm regards the negative water peak as the baseline and eliminates it to a certain extent. In summary, the fractional-order baseline correction algorithm expands the application range of the traditional integer-order baseline correction algorithm and provides a new solution to eliminate negative water peaks in the ATR spectra with water as the background spectrum.