Resumen de A hybrid radio approach for resource-constrained end-devices in cognitive networks
The number of wireless devices, as well as the data traffic they generate, continues to grow at an unprecedented rate. Furthermore, these devices usually operate in the same frequency bands, substantially increasing their occupancy. As a result, spectrum scarcity is nowadays one of the essential challenges faced by wireless communications service providers. Cognitive radio and edge computing are two of the main paradigms proposed to address this problem by increasing communications efficiency and data processing on end-devices. However, there are fundamental limitations for approaching both paradigms from the perspective of resource-constrained end-devices, which in the coming years will be the largest subsegment within wireless devices.
The main objective of this thesis was to analyze these limitations and propose and assess the technical feasibility of an alternative approach to overcome them. First, the hardware architectures of current end-devices and Software-Defined Radio (SDR) systems were studied. This allowed us to identify the architectural constraints and desirable characteristics of each of them for the challenges that end-devices currently face. Additionally, to delve into the limitations of end-devices, we extended this analysis with an empirical study of the effects of increasing their load of processing and communication tasks. Our results show that increasing this load may result in the appearance of a series of cross-effects between both types of tasks and significantly affect their performance. The findings of this phase of the thesis motivated the search for an alternative solution.
The main contribution of this thesis is a hybrid radio approach for resource-constrained end-devices in cognitive networks. This approach was conceived to simultaneously address the energy-efficiency requirements of these devices and the hardware flexibility demanded by the current challenges of cognitive radio and edge computing. Specifically, we propose to provide these devices with the ability to operate both as a current end-device and as an SDR system, and to dynamically switch their mode of operation, exploiting hardware processing and SDR capabilities only for those actions that strictly require them, and operating as a low-power end-device for the remaining tasks. Possible sporadic uses of SDR operation include: spectrum sensing, network synchronization, dynamic spectrum access, or hardware acceleration of critical tasks.
To assess the technical feasibility of the proposed approach, we designed, implemented and evaluated MIGOU, a low-power hybrid radio experimental platform. The power consumption of this platform was measured in its different modes of operation. These measurements were compared with the corresponding ones of other representative platforms: a resource-constrained end-device, a low-power SDR system, and two widely used high-performance SDR platforms. In addition, the hardware features of all these devices were compared. The results obtained confirm that a state-of-the-art tradeoff between hardware flexibility and energy efficiency is achieved with our hybrid radio approach.
Finally, as a real use case to better understand the multiple challenges that arise when adapting existing techniques to this new type of device, we addressed a widely studied problem in the cognitive radio field from the perspective of a hybrid radio end-device, that is, considering its limitations. We focus on the Automatic Modulation Classification (AMC), as an integral part of intelligent radio systems. Specifically, we conducted a study on how multiple parameters affect the classification accuracy and memory footprint of a reference deep learning model for AMC. From this study, we propose a new solution simpler than the reference one. We trained and tested our solution with over-the-air measurements of real radio signals. Our results show that the proposed solution has a memory footprint of 73.5 kBytes and achieves a classification accuracy of 92.4%, which is an improvement of 51.74% and 8.7% respectively compared to the reference method. These results confirm the feasibility of approaching AMC from the perspective of a resource-constrained hybrid radio end-device, both for its ability to acquire raw radio signals and for the memory footprint of the solution.
