The method based on miniature near-infrared spectrometer combined with Y fiber optic probe to detect the protein and fat in milk powder by diffuse reflectance spectroscopy in the wavelength range of 900～1 700 nm was proposed. By selecting the appropriate spectral bands, the correction models of protein and fat were established with partial least squares algorithm using Unscrambler 9.7 Chemometrics software. The determination coefficients R2 of the correction modes are 0.987 and 0.986 for protein and fat respectively, and the root mean square errors RMSEC are 0.385 and 0.419 respectively. Using these correction models to predict the protein and fat contents with 30 sets of forecast sample data, the prediction standard deviation is SEPProtein=0.751 for protein, and is SEPFat=1.109 for fat. The results indicate that these correction models have prediction capability with unknown samples and meet the on line requirements.