Interwoven waves: Enhancing the scalability and robustness of Wi-Fi channel state information for human activity recognition
Jeroen Klein Brinke is a PhD student in the department Pervasive Systems. (Co)Promotors are Prof.dr.ir. P.J.M. Havinga † and dr.ir. A. Chiumento from the faculty of Electrical Engineering, Mathematics and Computer Science.
This PhD dissertation explores the future of unobtrusive sensing in the context of radio wave-based sensing in realistic healthcare situations. Wi-Fi sensing is a new technology that could be employed in human health care for both human activity recognition and physiology monitoring by analyzing the multi-path reflections of the radio waves, enabling true unobtrusive sensing at a larger scale.
To that extent, Wi-Fi-based, or more specifically, channel state information-based, sensing nicely fits in the health care domain due to its unobtrusive, yet widely available nature. Healthcare is an important domain in our current society, as both the number of people and the average age of people are ever-increasing. Therefore, monitoring the well-being and health of family members, friends, and other loved ones is increasingly important to facilitate a safe and comfortable life, preferably living in their known home longer. Current solutions focus on either wearable devices that may be forgotten to be worn, audiovisual technologies that may heavily invade privacy, or expensive and unscalable changes to the infrastructure to facilitate continuous monitoring. On the other hand, channel state information-based sensing allows for distant monitoring of activities, physiology, and social contacts.
However, future sensing systems employing channel state information-based sensing will likely happen at already congested frequency bands in Wi-Fi, and current solutions often only add more congestion by adding random noise to these bands. Additionally, while existing solutions offer high performance, they usually solve niche problems in very specific and idealistic setups. Combining these two makes it hard to justify using the congested spectrum for sensing, as both more devices and sending random noise could result in increased latency and thus increased discomfort (slower internet speed) or risks (medical critical information being sent with a delay). Either of these would offset the benefit of unobtrusive sensing that Wi-Fi could provide.
To harness the potential of Wi-Fi sensing, future sensing solutions should seamlessly be woven into wireless communication; sensing and communication should collaborate and negotiate within the existing network requirements. Therefore, the main focus is on integrating communication and sensing seamlessly within existing infrastructure to allow for true unobtrusive sensing in locations that may otherwise be hard to monitor without generating more wasteful device-based solutions and creating adverse effects on wireless communication.
In order to progress in this direction, this thesis categorizes the relationship between sensing and communication through radio waves into three definitions: parasitic, opportunistic, and mutualistic. In the parasitic model, sensing operations disregard the existing wireless infrastructure, potentially degrading network performance due to additional noise and congestion. On the other hand, the opportunistic approach leverages existing traffic flows for sensing, ensuring that communication is not adversely affected. The most harmonious model, mutualistic sensing, seeks an optimal balance between sensing and communication, enhancing the capabilities of both functions without compromising the other.
Each of these definitions is explored: the possibility of using channel state information in harsh environments or situations that benefit from more data for scalability is explored in parasitic sensing, input-agnostic solutions that allow for sensing in communication networks with realistic packet flows are discussed in commensal sensing, while federated learning and a dynamic consus approach are discussed in mutualistic sensing.
Therefore, the main research objective is:
How can the robustness and scalability of channel state information-based sensing for human activity recognition be enhanced to facilitate their seamless integration into in-home environments through minimal impact on existing infrastructure?
Part I explores parasitic sensing, where the sensing infrastructure is naïve and does not consider the existing wireless communication channels. These solutions heavily pollute the electromagnetic spectrum (especially the 2.4 and 5 GHz bands) due to sending random packets at very short intervals. However, due to their flooded approach, they often allow for high performance. This part explores the data instability and transfer learning of such systems over different days and participants, as well as the opportunities regarding human health care in challenging real-life environments such as the bathroom.
Part II focussed on commensal, or opportunistic, sensing, which is sensing only when there is data being transmitted. In this case, wireless communication may be viewed as the highest priority, meaning sensing only happens if there is a packet flow, resulting in a sensing type that neither harms nor benefits wireless communication. In particular, this part explores the effect and correlation of different transmission rates on pre-trained models to ultimately proposes a technique that allows a model to adapt to changing network parameters, such as varying transmission rates.
Part III moves towards mutualistic sensing, where the requirements of both sensing and communication are balanced, and learning happens across the entire network. The goal here is to optimize both communication (reduced data transmissions, balanced latency) and sensing (achieving the highest performance by selecting the most useful devices). To this extent, the final part explores the complexities of using federated learning in a real-world scenario and proposes a distributed solution that allows for consensus voting among different receivers in a room using existing infrastructure that could reduce training and model complexity and allow for greater scalability and improved robustness.
