The anonymity of Bitcoin, while protecting user privacy, also provides a covert channel for illegal fund flows. Coin mixing technology enhances privacy protection by blurring transaction paths, but it also increases the difficulty for law enforcement agencies to track money laundering, ransomware and other criminal activities. Traditional rule-based detection methods rely on fixed patterns and are difficult to cope with the rapid iteration of coin mixing technology, resulting in low recall rates and insufficient generalization ability. To address this issue, this paper proposes a detection model based on Spatio-Temporal Graph Attention Network (ST-GAT), which builds a 30-layer predecessor and 10-layer successor temporal graph structure of the target transaction, and combines the global temporal attention of the transformer architecture with the spatial topological attention of GAT to effectively capture the patterns of fund convergence and dispersion in coin mixing transactions. Experiments show that ST-GAT outperforms traditional GCN and LSTM-TC models in terms of precision, recall and F1 value on the evaluation set, verifying its advantages in complex transaction graph analysis.