Signal classification at discrete frequencies using machine learning

Incidents such as the 2018 shut down of Gatwick Airport due to a small Unmanned Aerial System (UAS) airfield incursion, have shown that we don’t have routine and consistent detection and classification methods in place to recognise unwanted signals in an airspace. Today, incidents of this nature are taking place around the world regularly. The first stage in mitigating a threat is to know whether a threat is present. This thesis focuses on the detection and classification of Global Navigation Satellite Systems (GNSS) jamming radio frequency (RF) signal types and small commercially available UAS RF signals using machine learning for early warning systems. RF signals can be computationally heavy and sometimes sensitive to collect. With neural networks requiring a lot of information to train from scratch, the thesis explores the use of transfer learning from the object detection field to lessen this burden by using graphical representations of the signal in the frequency and time domain. The thesis shows that utilising the benefits of transfer learning with both supervised and unsupervised learning and graphical signal representations, can provide high accuracy detection and classification, down to the fidelity of whether a small UAS is flying or stationary. By treating the classification of RF signals as an image classification problem, this thesis has shown that transfer learning through CNN feature extraction reduces the need for large datasets while still providing high accuracy results. CNN feature extraction and transfer learning was also shown to improve accuracy as a precursor to unsupervised learning but at a cost of time, while raw images provided a good overall solution for timely clustering. Lastly the thesis has shown that the implementation of machine learning models using a raspberry pi and software defined radio (SDR) provides a viable option for low cost early warning systems.