How to efficiently detect, trace and interpret anomalies or attack behaviours in massive high-dimensional temporal data has become an urgent challenge. Especially when dealing with unknown temporal relationships, how to construct correlations between data points is also a major challenge. In recent years, the development of Graph Neural Networks (GNNs) has provided a new perspective for deep learning models and brought potential breakthroughs in model interpretability. However, GNNs still have certain limitations in spatio-temporal feature fusion, graph adjacency matrix construction, and model interpretability. Therefore, this paper proposes a method to adaptively update unknown graph relationships using Bayesian networks, which effectively addresses the problem of unknown node relationships between sensors. At the same time, in order to improve the model’s ability in spatio-temporal feature fusion and to highlight the role of key nodes, this paper introduces a node weighting processing mechanism to select certain statistical features as node embeddings that reflect anomalies. The experimental results show that the method proposed in this paper has been validated on two sensor datasets, SWAT and WADI. The results show that, compared to the baseline methods, the proposed model can not only detect anomalies more accurately, but also better capture the correlation between nodes and perform in-depth analysis of the causes of anomalies. Through this series of improvements, the performance and interpretability of our anomaly detection method has been significantly enhanced.