System behaviors could be recorded as multivariate time series by smart sensors. It is challenging to interpret such high-dimensional dataset with a one-dimensional temporal state sequence. In this paper, we propose a new approach DBSE (Distribution-based States Extraction) which is based on statistical and clustering analysis. State extraction problem could be resolved through using distribution parameters to describe each subsequence in multivariate time series and applying clustering method to extract states based on distribution similarity in order to obtain a temporal state sequence and information of each state. We validate DBSE and demonstrate how to use DBSE in real-world by extracting state information from a wearable sensor dataset (PAMAP2_Dataset). By comparing DBSE with TICC (Toeplitz Inverse Covariance-based Clustering) and FCM (Fuzzy C-means Clustering), the new approach is more accurate and effective. Moreover, DBSE is also expected to facilitate future behavior analysis.