For instance, an increase in living and consumption level has significantly led to an increase in demand for food safety and quality of milk and its products. The quality of milk affects the production and consumption of dairy products. In order to ensure the quality of dairy products, methods and procedures have been developed to detect various milk adulterants in the collection, storage and production procedure. Most current analytical methods, such as chemical and instrumental analysis, are targeted detection methods, which require pre-treatment steps designed for adulterants and are cumbersome and time-consuming. In this paper, we proposed a non-targeted method based on mid-infrared spectroscopy developed for the identification of abnormal milk samples. The natural raw milk samples were collected from six pastures of the Mengniu company, and abnormal milk samples were prepared by adding multiple adulterants. Then the mid-infrared spectrum was measured and pre-processed with smoothing, multiple scattering correction, baseline correction and normalization. In order to improve the accuracy and robustness of models, Three different variable selection methods were implemented, such as uninformative variables elimination (MC-UVE), uninformative variables elimination-successive projections algorithm (UVE-SPA) and competitive adaptive reweighted sampling(CARS). Then, two classification algorithms, partial least squares discriminant analysis(PLS-DA) and support vector machine (SVM), were employed and compared in the discrimination models. The results indicated that SVM is the better classification algorithm achieving higher identifying accuracy, and CARS method screening performs better with PLS-DA and SVM classification models. The accuracy of the -SVM-CARS discrimination model achieved 97.84% and 94.55% for validation and prediction, respectively. The variables screened by the CARS method were mainly concentrated in the regions where the spectral features of the anomalous milk samples were more obvious. Further analysis of the misclassified sample showed that the model combination could more accurately identify the abnormal milk samples. These results demonstrate that abnormal milk can be identified successfully using mid-infrared spectroscopy with discriminant analysis, suggesting our techniques to provide an efficient and practical reference for milk adulteration and on-line detection of the production process.