Real-time anomaly detection for on-orbit satellites is crucial in the early identification of faults to prevent further anomaly expansion. However, current anomaly detection methods for on-orbit satellites face challenges such as low sensitivity to anomalies, high rates of false negatives, and a lack of criteria for assessing anomaly severity. The Spatio-Temporal Causal Graph (STCG) describes the intrinsic temporal and spatial relationships among telemetry parameters. By extracting temporal and spatial features, it facilitates enhanced modeling of normal data and improves the model’s sensitivity to anomalies. Moreover, according to the principle of anomaly propagation, the STCG can be utilized to identify false negatives. Therefore, we propose a framework of on-orbit satellite hierarchical anomaly detection using causal structure learning (OSHAD-CSL). This framework introduces a graph autoencoder for STCG learning and develops a method for dynamically setting detection thresholds. Additionally, the framework puts forward a false negative identification method based on the learned STCG. Lastly, the framework presents an anomaly severity judgment criterion, enabling the determination of the affected system level by the detected anomaly. To evaluate the effectiveness of the proposed framework, we validate the OSHAD-CSL on three public datasets and a dataset from a satellite’s attitude control system. The results demonstrate that the proposed method outperforms the baselines, with an F1-score (a metric used for evaluating the performance of anomaly detection) of 80% for anomaly detection and an accuracy of 83% for anomaly severity judgment. These findings highlight the efficacy of the proposed method in improving the performance of anomaly detection and severity assessment.