Abstract: To overcome the challenges presented by current unsupervised anomaly detection models in capturing the temporal and spatial dependencies of multivariate time series data, a spatial- temporal attention- based anomaly detection model, anchored on the Transformer architecture, is developed. The Transformer's encoder was innovatively restructured into two components: a temporal encoder and a spatial encoder, which allowed for the concurrent extraction of dependencies from both time and space dimensions of multivariate time series data. After establishing the anomaly detection model, the POT method was utilized to adaptively determine an anomaly threshold, flagging any data point exceeding this threshold as anomalous. The experimental results on four multivariate time series datasets indicate that the anomaly detection performance of the model has improved by 1.8% to 11.5%, enabling accurate detection and identification of anomalies.