In the fault period of high-speed train lateral damper, the vibration signal is non-linear and nonstationary, and features extracting is relatively difficult. In order to solve this problem, a method of features extracting based on variational mode decompositi on and multiscale entropy is proposed. The original signal is decomposed into several intrinsic mode function components after being processed by the variational mode decomposition method. Then, the best component is selected by the mutual information index. The feature matrix is constructed through the multiscale entropy of the best component, and removed redundant features using feature evaluation algorithm. The fault type of lateral damper is judged by transforming in the best subset of feature matrix in support vector machine. Experimental results show that the proposed method can extract the feature and judge the fault type of lateral damper effectively, which proves the feasibility of this mechanical fault diagnosis method.