AI-Based Anomaly Detection for Industrial 5G Networks by Distributed SDR Measurements

The analysis of the physical layer in the frequency spectrum is subject to vigorous research in the last couple of years. From localization tasks to anomaly detection, the research is starting to incorporate more artificial intelligence-based solutions. In case of anomaly detection, the looming problem is that there are many different and unquantifiable types of anomalous effects. Hence trying to find a model that predicts anomaly itself is not feasible. Therefore, we propose an approach of successfully detecting all known signals in a given signal range, which implicitly leads to finding possible anomalies. This is done by collecting Power Spectral Density waterfall diagrams and segmenting them with a Convolutional Neural Network named U-Net. The results are compared against a knowledge base of the scanned bandwidth and an informed decision on the validity of the observed signal is made. We are able to provide a working concept for the distributed monitoring and stress test system STING for detecting anomalies in private 5G networks. The system is able to achieve an accuracy up to 90%, while providing a false negative rate of 2.37%. We aim to supply full coverage of a given industrial workplace, through the distribution of software defined radios over the STING-system itself and thus are able to detect anomalies over the complete industrial facility in the future.