Multivariate time series are ubiquitous in industrial production, power systems, and transportation, exhibiting complex interdependencies between different variables that make anomaly detection a challenging task. Current methods are mostly based on theoretical assumptions, which either limit the model’s generalization ability or fail to effectively handle outliers when processing anomalous data. This paper proposes an anomaly detection model that combines Graph Attention Networks (GAT) with Transformers. The model captures spatial dependencies between sensors through GAT layers and models temporal dependencies using Transformer encoders, thereby improving anomaly detection performance. To enhance the model’s expressive power, trainable weights are introduced in the feature fusion process to perform weighted integration of spatial and temporal features. Experimental results show that the overall performance outperforms baseline methods on four benchmark datasets: SWaT, WADI, PSM, and MSL.