Anomaly detection of spectrum in wireless communication via deep auto-encoders

Anomaly detection is a typical task in many fields, as well as spectrum monitoring in wireless communication. Anomaly detection task of spectrum in wireless communication is quite different from other anomaly detection tasks, mainly reflected in two aspects: (a) the variety of anomaly types makes it impossible to get the label of abnormal data. (b) the complexity and the quantity of the electromagnetic environment data increase the difficulty of manual feature extraction. Therefore, a novelty learning model is expected to deal with the task of anomaly detection of spectrum in wireless communication. In this paper, we apply the deep-structure auto-encoder neural networks to detect the anomalies of spectrum, and the time–frequency diagram is acted as the feature of the learning model. Meanwhile, a threshold is used to distinguish the anomalies from the normal data. Finally, we evaluate the performance of our models with different number of hidden layers by our experiments. The results of numerical experiments demonstrate that a model with a deeper architecture achieves relatively better performance in our spectrum anomaly detection task.