In view of the complex diversity and short time persistence of surveillance video anomaly， a weakly supervised video abnormal detection method was introduced to detect anomalies by only using video-level tags， and an anomaly regression network VLARNet based on Variational AutoEncoder （VAE） and Long Short-Term Memory （LSTM） network was proposed as an anomaly detection framework to effectively capture the temporal dependencies in time series data， eliminate redundant information and retain key information in the data. Anomaly detection was considered as a regression problem by VLARNet. To learn detection features， a Triplet-Centered Loss for Anomaly Score Regression （TCLASR） was designed and combined with Dynamic Multiple Instance Learning loss （DMIL） to further improve the discrimination ability of features. The DMIL widened the inter-class distance between abnormal instances and normal instances， but it also widened the intra-class distance. The TCLASR made the distances between the instances in the same class and the center closer and the distances between instances in different classes and the center farther. The proposed VLARNet was comprehensively tested on ShanghaiTech and CUHK Avenue datasets. Experimental results show that VLARNet can effectively utilize various information in video data， achieving Area Under receiver operating characteristic Curve （AUC） of 94.64% and 93.00% respectively on the two datasets， which is significantly better than those of the comparison algorithms.