<p>Anomaly detection for multivariate time series (MTS) data plays a crucial role in industrial process. In MTS data, feature relationships measure the relationships between different time series. However, most existing anomaly detection models for MTS data focus on the temporal dependence, ignoring feature relationships. Meanwhile, contrastive learning are used for MTS data, but common data augmentation may lead to the destruction of temporal characteristics. To address these issues, we propose MMGCL, an anomaly detection model with Multi-view Multi-grained Graph Contrastive Learning. Firstly, to simultaneously capture temporal dependence and feature relationships, the Multi-View Graph with Masked-Attention (MVGMA) is proposed. MVGMA applies a symmetric structure, facilitating the mutual reinforcement of information between views. Secondly, to avoid temporal characteristic destruction caused by unreasonable data augmentation, Multi-Grained Contrastive Learning (MGCL) is proposed. MGCL treats the two views’ results as augmentations. It designs instance contrastive learning and context contrastive learning, separately aiming at relationships between windows and relationships inside a single window. Experimental results demonstrate that our model outperforms the majority of existing MTS anomaly detection methods, achieving an average anomaly detection F1 score exceeding 96.8%.</p>