Fault diagnosis method based on feature selection (FS) and linear local tangent space alignment (LLTSA) is proposed, aiming to solve the problem of non-sensitive features and the high dimension of the feature set. An improved kernel distance measurement feature selection method (IKMD-FS) is proposed, which considers both the distance between classes and the dispersion within class, and the selected sensitive features are weighted by their sensitive-values. The weighted sensitive feature subset is compressed through LLTSA to reduce dimension and get the compressed more sensitive feature subset. Then, the feature subset is fed into weighted k nearest neighbor classifier (WKNNC), whose recognition accuracy is more stable compared with k nearest neighbor classification (KNNC), to recognize the fault type. At last, the validity of the proposed method is verified by the instance of the fault diagnosis of a rolling bearing.