Abstract: Multivariate time series (MTS) anomaly detection aims to identify abnormal patterns in time series data that do not conform to general rules. Existing methods often neglect the importance of different variables at each moment, and the long-term stable and short-term dynamic associations among variables. Moreover, these methods make effects to balance long-term and local temporal dependencies when extracting features from data, leading to insufficient feature extraction and inaccurate anomaly detection results. To address this, we propose the method for MTS anomaly detection by integrating adaptive graph and multi-level attention. Firstly, we learn the long-term stable and short-term dynamic associations between variables using node embeddings and node inputs, respectively. Secondly, a graph loss mechanism is added to guide the learning of long-term stable associations. Then, we design multi-level attention mechanisms: the variable-level attention mechanism is used to capture the key variables at the current time step. The sequence-level weighted attention mechanism overcome the potential global dispersion issue in traditional self-attention mechanisms, effectively capturing both long-term and local temporal dependencies in time series. Thus, the learned normal patterns of MTS could be reconstructed. Finally, experiments on several real datasets show that the average F1-score of the proposed method is high than 0.96, which significantly outperforms the state-of-the-art methods.