In the near-infrared online detection of the fermentation process, bubbles are often generated in the fermentation broth due to the need to continuously pass oxygen into the fermentation broth to promote microbial growth and metabolic activities. When the bubbles in the fermentation broth pass in front of the probe, they will interfere with the intensity of the near-infrared (NIR) spectrum. To eliminate the abnormal spectra caused by bubbles collected during the near-infrared online detection of fermentation broth and reduce spectral fluctuations, a normal distribution screening method is proposed in this study. In this study, 600 g of glucose solution with a mass fraction of 10% was prepared, adding 2 g of glucose solution to a reactor containing 600 mL of distilled water every 30 s, stirring well, then calculating and recording the mass fraction of glucose solution in the reactor, and generating bubbles by passing oxygen to the bottom of the reactor, and collecting the NIR spectra of the glucose solution in the reactor by using NIR spectrometer, respectively. After the anomalous spectra affected by the air bubbles were excluded by principal component analysis (PCA) combined with Mahalanobis distance method, Euclidean distance method, isolated forest, and normal distribution screening method, the sample set of spectra was randomly divided into the correction set and the prediction set according to the ratio of 4∶1, and then, after the spectral pre-processing, the glucose concentration prediction model was established for the correction set using the partial least squares method (PLSR) and the prediction set was analyzed by the established PLSR model. The correlation coefficient of the correction set, the root mean square error of the correction set, and the correlation coefficient and root mean square error of the prediction set were compared and analyzed. The results of the constructed model after removing the anomalous spectra affected by bubbles using the four methods are as follows: the correlation coefficient $R_{\text{c}}^2$ of the correction set obtained after removing the anomalous spectra by PCA combined with the Mahalanobis Distance Method is 0.998 208, and the root-mean-square error RMSECV is 0.000 764, and the correlation coefficient $R_{\text{p}}^2$ of the prediction set is 0.997 994, and the root mean square error RMSEP is 0.000 764; The correction set $R_{\text{c}}^2$ obtained after removing the anomalous spectra by the Euclidean distance method is 0.998 628, the root mean square error RMSECV is 0.000 652, the prediction set correlation coefficient $R_{\text{p}}^2$ is 0.998 628, and the root mean square error RMSEP is 0.000 655; the correction set $R_{\text{c}}^2$ obtained after removing the anomalous spectra by the isolated forest method is 0.998 255, the RMSECV is 0.000 739, the prediction set $R_{\text{p}}^2$ is 0.998 132, and the RMSEP is 0.000 740; the correction set $R_{\text{c}}^2$ obtained after the removal of anomalous spectra by the normal distribution screening method is 0.998 641, with a root mean square error RMSECV of 0.000 645, and the prediction set $R_{\text{p}}^2$ is 0.998 628, with a RMSEP of 0.000 636. Comparing the four methods, the normal distribution screening method can effectively reduce the fluctuation of spectral intensity and eliminate abnormal spectra more effectively than other methods.