The Challenge
As the radio frequency (RF) spectrum becomes more congested it is increasingly difficult to detect specific signals or anomalies within RF clutter. AI offers the potential of automating this task to convert raw data into useful information to better understand the environment. This can benefit applications such as spectrum management, cognitive radio and signals intelligence.
Could an AI classifier be trained on synthetic data to perform this task?
The Approach
There are various possible ways in which an RF signal could be transformed into a suitable form for an AI model to process. In this project, the signal magnitude as a function of time produced by a single antenna was converted into a spectrogram. A spectrogram is a 2D representation that captures both time and frequency information.
Since a spectrogram is effectively an image, existing image-based object detection AI models (e.g. mask R-CNN) could be used as the starting point for RF classification. However, as they have not been trained to classify RF signals, it was first necessary to retrain the AI model using new data. To avoid the need to collect large quantities of real data an RF model was used to generate synthetic data of different RF signal types with varied environmental effects applied. This training set consisted of 1000 examples each of six different signal types (e.g. FM, OOK, MFSK, MPSK, DMR and LoRa) with an additional 100 examples of each for testing. This process is known as transfer learning as it allows an AI model to be initially trained for one application on a large quantity of data but then be retrained on a smaller quantity of data for a different task.

The Outcome
Once this AI model had been retrained to recognise RF signals it was used to process new spectrograms for detection and classification. This worked effectively on simulated data and it successfully detected and classified signals that were overlapping in time and/or frequency. The confusion matrix shows good classification performance across all signals with the exception of Multiple Phase-Shift Keying (MPSK) which, if detected, was sometimes confused with On-Off Keying (OOK) or Multiple Frequency-Shift Keying (MFSK). This was due to the occasional similarity of the signals within the analysis window and could be resolved with further development.


As part of the performance assessment, some measurements of real transmitters were taken and analysed. Despite no real devices being included in the training set the AI model was able to detect and classify most RF signals. Those it was not successful against looked significantly different in the spectrogram in comparison to the training data.

A key outcome from this work was demonstrating that retraining an AI model using synthetic data can still produce a model that can operate on real measured signals. This reduces the requirement for resource intensive real data collection to produce a training set allowing rapid retraining of a model for new applications. A more limited quantity of real data can then be used to enhance the performance in what is effectively a second transfer learning step.
“Retraining an image-based object detection AI model on synthetic RF data allows varied RF applications to be automated, though boosting the training set with some real data will ground the model to produce optimal performance.”
Dr Damien Clarke, Head of Data Science
Ready to automate RF signal classification?

Plextek develops AI-based RF signal classification systems, from algorithm development through to embedded real-time processing solutions. Our transfer learning approach enables efficient training using synthetic data whilst achieving strong real-world performance across spectrum management, ew, and communications monitoring applications.
Get in touch with us to discuss your RF signal classification requirements.



































































