An AI-powered ‘virtual expert’ in a field of sensing research could help advance the development of new technologies, including sophisticated healthcare monitoring techniques, thanks to research taking place at the University of Glasgow and Imperial College London.
Engineers are behind the development of the powerful new artificial intelligence-based tool, known as RFSensingGPT, which can be run on consumer-grade computers.
It utilises a custom-built database of knowledge in radio frequency (RF) sensing to provide highly-reliable technical advice, computer coding expertise, and image analysis which could be used to support the design and development of new devices.
Radio frequency sensing is an emerging technology which uses high-frequency radio waves to explore physical spaces. It applies sophisticated analysis techniques to build up a detailed picture of objects and people as the waves reflect from them, and can resolve images even through walls.
Future healthcare devices, for instance, may use RF sensing to monitor the movements and vital signs of people in their homes or in medical settings without relying on cameras or wearable devices. RF sensing technologies will also be integrated into next-generation 6G communications networks, facilitating real-time monitoring of physical spaces for applications beyond healthcare into transportation, manufacturing, and everyday life.
The development of RFSensingGPT is summarised in a new Early Access paper published in the journal IEEE Transactions on Cognitive Communications and Networking.
The paper’s corresponding author is Professor Qammer H. Abbasi of the University of Glasgow’s James Watt School of Engineering, who is also director of the University’s Communications, Sensing and Imaging Hub.
“In recent years, large language models like ChatGPT have proven that they are capable of providing engaging user experiences, with convincing answers to questions about a wide range of topics,” said Professor Qammer. “However, because they are trained on a broad spectrum of general information, they often provide inaccurate or misleading answers when the conversation moves to more specialised topics, where their training data is less complete.
“Our aim in developing RF SensingGPT was to build a system that could be relied on to provide sophisticated image analysis and accurate, easily-accessible answers to questions from both experts and developers new to the field.”
The engineers used a technique called Retrieval-Augmented Generation, or RAG, to enhance the ability of a large language model to provide specialist answers.
RFSensingGPT draws on more than 80,000 pieces of RF-specific technical information assembled by the team, including technical documents, code repositories and research papers, to provide accurate responses to queries in accessible language. In the team’s testing, it provided access to relevant technical documents nearly 98% of the time in response to user queries – a notable improvement over the 36% rate of correct answers provided by standard AI models tested by the team.
It can also handle the challenging task of accurately interpreting spectrograms – complex visual representations of RF signals which are used by sensing systems to keep track of the movements of objects or people within range of the RF waves.
The researchers trained RFSensingGPT to be able to pick out patterns in the spectrograms produced by sensing signals in three different ranges of the RF spectrum – 24GHz, 77GHz, and Xethru. RFSensingGPT was able to accurately interpret signals showing people sitting, walking, crawling or bending with 93% accuracy.
The team also tested RFSensingGPT’s performance on two computer systems, one a standard Windows desktop using an Intel processor and another fitted with a mid-range Nvidia GPU. In both instances, RFSensingGPT provided fast and accurate results, showing that it is suitable for use even on relatively low-powered computer systems.
Professor Muhammad Imran, head of the James Watt School of Engineering, is a co-author of the paper. He said: “RFSensingGPT could significantly accelerate research and development in healthcare monitoring by making specialised knowledge more accessible to both experts and developers new to the field, helping to democratise RF expertise across industry.
“We’ve shown that it works very well to provide accurate support without the need for complex computing setups, and we hope that it will become widely-adopted by RF sensing researchers. We’re also looking at how the system could be trained to provide similar services for other related communications fields in the years to come, driving innovation in advance AI-driven 6G networks, smart city infrastructure, and healthcare monitoring applications.”
Muhammad Zakir Khan, Yao Ge and Dr Micheal Mollel of the James Watt School of Engineering and Professor Julie McCann of Imperial College London co-authored the paper.
There’s plenty of other editorial on our sister site, Electronic Specifier! Or you can always join in the conversation by commenting below or visiting our LinkedIn page.
