The property of high dimensionality impacts on the process and results in the field of time series data mining, and the traditional methods about principal component analysis have some limitations to represent multivariate time series. Therefore, a feature representation of multivariate time series based on correlation among variables is proposed. The distribution and relationships among variants of every time series are described by the covariance matrix, and principal components are extracted from an integrated covariance matrix by principal component analysis. In this way, the dimensionality of multivariate time series can be reduced and the features can be represented. The experimental results show that the proposed method not only improves the quality of multivariate time series data mining but also efficiently mines on the data with different lengths.