In complex industrial systems due to the dramatic increase in the number of sensors, high-dimensional noise and random disturbances are generated, which seriously affect the data continuity and control accuracy of multivariate time series. However, the existing pairs of multivariate time series have the problems of temporal inconsistency over time, deviation of space vectors, and redundancy of spatio-temporal graphical models. In this paper, a new multivariate time series anomaly detection method STGAD is proposed. First, a gating mechanism is introduced to improve the multiscale convolutional network from a high-resolution granularity level to control the process of information interaction between features. Then, two graph structures are designed to eliminate redundant spatio-temporal dependencies, enabling GAT to effectively learn spatio-temporal correlations. In addition, an attention mechanism-based GRU module is proposed to capture the importance of variables over different time windows. Finally, modules for joint optimization prediction and reconstruction. Extensive experiments on three publicly available datasets show that the average F1-score of the proposed model is higher than 0.94, which significantly outperforms other benchmark models on high-dimensional datasets.