Abstract: Automatic dependent surveillance-broadcast （ADS-B） is the primary technology in air traffic surveillance. However， its reliance on open wireless transmission and lack of encryption render it vulnerable to interference， leading to data anomalies， which can compromise air traffic safety， airspace management， and operational efficiency. Although machine and deep learning models have been widely applied in ADS-B anomaly detection， most existing methods fail to adequately extract and enhance multi-feature correlations in the data， resulting in low recognition accuracy for certain anomaly types. To address these limitations， this study designed a model based on graph attention network， temporal convolutional network （TCN）， and Transformer architecture to detect anomalies in ADS-B time-series data and expand the anomaly dataset. The model represents the ADS-B data in a graph structure， leverages its spatio-temporal correlation characteristics， and achieves deep feature extraction and anomaly detection of multi-dimensional parameters through the joint optimization of reconstruction and prediction outputs. In threshold configuration， the Peak Over Threshold setting method was adopted. This method utilizes the Streaming Peak Over Threshold probability model dynamically determining the anomaly classification threshold for enhanced adaptability. On the expanded dataset， the proposed model achieved an average F1-score of 0.941 in anomaly detection， which is a 13% improvement on average compared to that of classical models such as long short-term memory and TCN. The findings demonstrate that representing ADS-B data as a graph effectively enhances the anomaly detection rate. This study provides valuable insights for aircraft status monitoring in complex airspace environments and serves as a useful reference for civil aviation traffic control and radio management.