A neural network approach for wireless spectrum anomaly detection in 5G-unlicensed network

The 5G New Radio Unlicensed (5G-U) technology has enabled manufacturing enterprises to deploy their own private industrial networks, making anomaly event detection more necessary for maintaining wireless communication quality. However, existing statistical analysis algorithms cannot efficiently and accurately detect various kinds of anomaly events caused by the complex industrial environment. These events include electromagnetic interference as well as contention between cross-technology devices for unlicensed spectrum resources. To improve the efficiency, we design a classification algorithm that uses feature extraction in the frequency domain and a convolutional neural network model to detect various kinds of anomaly events (e.g., loose antennas and co-channel interference). We prototyped Slade (Spectrum Learning for Anomaly Detection), an anomaly detection system for industrial 5G networks. To evaluate the system, we collect wireless spectrum data with two industrial 5G-U terminals. Our evaluation on the dataset shows that our methodology can accurately detect different anomaly events, with an accuracy of 97.6% and a recall of 97.1%.