Principal component analysis and relevant techniques are often used to represent feature for multivariate time series. However, they have some limitations for feature representation, and the dimension reduction results impact on the similarity measure accuracy. Therefore, from the perspective of shape features, a crucial shape feature dimension reduction technique for multivariate time series is proposed. In train datasets, central for multi variate time series of each category is obtained through dynamic time warping, and the importance degree of crucial feature properties components is found according to shape feature. In this way, the dimensionality of multi-time series can be reduced and the original data can be represented by the crucial shape feature. The experiments results show that the proposed method is superior because it can obtain a better classification effect and effective dimension reduction.