In addressing the issue of rapid loss of sulfurin the lubricating oil in high-speed train traction system gearboxes, it is imperative to monitor the sulfur content in in-service gear oil to determine the optimal oil change timing. The attenuation mechanism of the characteristic absorption peaks of sulfur-containing additives in the gear oil infrared spectra was analyzed. The peak height of the 1 164 cm^-1 characteristic peak with the highest correlation to sulfur content was selected as the input variable for setting a univariate linear regression (LR) model. The optimal number of principal components was determined using the tesrset R² after 5-fold cross-validation, and the partial least squares regression (PLSR) model was constructed. Finally, six characteristic absorption peaks were selected based on the partial least squares regression coefficients, and the reduced feature PLSR (RF-PLSR) model was established after the optimization of the characteristic peaks. The proposed models predicted 39 oil samples in the prediction set. The results indicated that both the univariate linear regression and partial least squares regression models exhibited good predictive capabilities. The predicted R² of the univariate linear regression model was 0.961, the RMSE was 973.1, and the RPD was 5.084. The predicted R² of the PLSR model and the characteristic peak-optimized RF-PLSR model were 0.997 and 0.994, the RMSE was 250.1 and 376.3, and the RPD was 19.780 and 13.149, respectively. This indicated that the RF-PLSR model built after characteristic peak optimization could still maintain high accuracy and prediction ability while reducing model complexity, and the model is more interpretable because the information of irrelevant variables was eliminated. Thus, infrared spectroscopy technology offers a reliable and precise method for detecting sulfur content in gear oil, providing a feasible solution for monitoring the condition of gear oil.