An Unsupervised Spectrum Anomaly Detection Method Integrating Self-Attention Mechanism

Spectrum anomaly detection (SAD) is an important part of spectrum management. This article proposes an unsupervised SAD method integrated with a self-attention mechanism to address unknown anomaly detection in complex electromagnetic environments. The proposed framework establishes reliable detection capabilities for unseen anomaly types through exclusive use of normal spectrum samples during training. By leveraging the self-attention mechanism of the transformer architecture, the model effectively captures global contextual dependencies within spectrum data and overcomes the inherent limitations of conventional neural networks in long-range context modeling. Experiments based on simulated dataset demonstrate that the proposed method exhibits good detection stability and strong generalization capability.