In blockchain abnormal transaction detection, supervised graph neural networks (GNNs) are constrained by the scarcity of labeled data, whereas graph contrastive learning confronts the distinctive challenge of topological coupling in blockchain transaction graphs—namely, dense capital flows induce strong inter-node correlations, which directly violate the key instance independence assumption on which contrastive learning relies. To address this, we propose an efficient boundary-aware model Semantic-Decoupling Contrastive Learning Model for Blockchain Abnormal Transaction Detection (SD-ATD). The model employs spectral clustering to uncover transaction pattern clusters, followed by node partitioning into multiple clusters. Subsequently, the central node of each cluster—selected based on topological centrality and semantic representativeness—is used to reconstruct the graph into a capital-flow subgraph, thereby alleviating topological dependencies through semantic decoupling. Then reconstruct each subgraph to obtain the disturbed graph for comparative learning. Extensive node classification and node clustering experiments conducted on five benchmark datasets and four baselines empirically validate the effectiveness of SD-ATD in capturing abnormal transactions. Furthermore, the method outperforms other existing methods. The code is available at https://github.com/guiei/SD_ATD .