There are “high protein”, “high fat”, and other characteristics of milk in the market. In order to realize the nondestructive and rapid grading of super quality milk, high-protein characteristic milk, high-fat characteristic milk and ordinary milk, 5 121 milk samples from 10 pastures in Hebei Province in different months (January, March to October) were collected. Then the mid-infrared spectroscopy data were collected, the protein and fat content in milk were measured, the somatic cell number was measured, and the mid-infrared spectrum model of milk quality grading was established. Firstly, milk spectral analysis was carried out, and redundant bands were removed. Finally, the sensitive band combinations of 9 925~1 597 and 1 712~3 024 cm^-1 were selected as the full spectrum to establish the model. In order to improve the prediction accuracy and efficiency of the model, six spectral pre-processing methods were used to improve the signal-to-noise ratio of the original spectrum, including Standard normal variable transform (SNV), multiple scattering correction (MSC), the first derivative and second derivative, first difference and second-order difference. Comparing the effects of different pretreatment methods by establishing naive Bayes model (NB) and random forest model (RF), the second-order difference obtained the best prediction accuracy. The testing set accuracy was 92.11% and 96.87%, respectively. So second-order difference was identified as the best pretreatment method for further analysis. In order to simplify the models, UVE (Uninformative variable elimination), CARS (Competitive adaptive reweighed sampling), SCARS (Stability Competitive adaptive reweighted sampling) were utilized to extract the characteristic variables from the pre-processed spectrum by second-order difference method. Then, the NB and RF models were established based on the full spectral data and the selected characteristic variable data. The results showed that SCARS was the best feature extraction algorithm for the NB model, and the accuracy rates of the training set and the testing set were 94.45% and 93.94%, respectively.UVE-SCARS is the best feature extraction algorithm of the RF model, and the accuracy of the training set and test set are 99.86% and 96.48%, respectively. In conclusion, the second-order difference-UVE-CARS-RF model established based on Fourier transform the mid-infrared spectroscopy technology can realize the rapid and non-destructive prediction of classification of 4 kinds of milk. Through the establishment of mid-infrared spectrum model, the combination of milk protein, milk fat content and somatic cell number is the first time for direct classification and identification, which is unprecedented in previous studies. In applying the model, we only need to input the obtained milk mid-infrared spectral data into the model to output the prediction category, which has practical application value in the milk industry.