In this paper, EdgeConvFormer is introduced as a novel approach to unsupervised anomaly detection in multivariate time series, combining the strengths of graph convolutions and Transformers within a hierarchical structure. The model utilizes Time2Vec to generate temporal embeddings and subsequently constructs dynamic graphs from point embeddings in the 2D sensor-time space, effectively extracting the most significant spatiotemporal relationships. A novel parallel sensor-specific attention mechanism has been developed to enhance the model’s ability to exploit sensor-specific temporal dependencies, significantly boosting its detection capabilities. Extensive experiments have been conducted across six benchmark datasets for anomaly detection, including the expansive Exathlon dataset. The experimental results demonstrate that EdgeConvFormer outperforms state-of-the-art methods across various evaluation metrics.