Multivariate time series data classification has recently attracted interests from both industry and academia, as sensors used in various industries produce a lot of multivariate time series data. Having a lot of features, feature selection from those time series is essential to efficiently construct a classifier. In this paper, we propose a feature selection method to efficiently select features from the multivariate time series data considering variation. The candidate feature set is too large to efficiently select features and there are some feature redundancies. The proposed method can efficiently resolve these problems, and is validated by real datasets obtained from UCI Machine Learning Repository. Experiments show that the proposed method outperforms the typical feature selection methods in terms of accuracy and precision.