An extended canonical variates analysis (ECVA) method dealing with multicollinear data of infrared spectrum and a singular within-group covariance matrix is proposed for wine origin discrimination. The wine-origin is classified by infrared spectrum and the chemometrics methods. Comparing the classification results of the k-nearest neighbor algorithm (KNN), partial least-squares discriminant analysis (PLS-DA), and principal component analysis (PCA) methods, the ECVA-KNN method reaches the optimal correctness of 97.73%. The experiment results prove that the ECVA method is fit for wine origin discrimination and can be effectively applied to the qualitative analysis of the collinear infrared spectrum.