Wireless communication and radar systems have been traditionally designed independently from each other, even though both radio-frequency technologies use electromagnetic waves as their main resource. Against this background, the paradigm of integrated sensing and communication (ISAC) pursues the fusion of radar sensing and communication for a more efficient use of hardware, energy, and spectral resources. In addition to this, ISAC will expectedly support new services in telecommunications, transportation, industry, agriculture, smart home, or healthcare. This thesis provides a state-of-the-art overview of ISAC in mobile networks, while addressing some open challenges concerning technology integration and application. At the integration level, the fundamental problem of self-interference is tackled in the thesis via beamforming. In particular, analog and hybrid beamforming architectures are studied with regard to self-interference mitigation to enable monostatic ISAC, i.e., co-located communication transmission and sensing reception. For the analog case, target detection is identified as the key sensing task, which leads to low-sidelobe requirements. Moreover, current beam codebook specifications from the fifth generation (5G) standard are taken as a basis to investigate self-interference reduction formally, through analytically derived bounds on analog and mixed-signal distortion. The designed beam codebooks present substantially improved sensing quality with little impact on communication, hence paving the way for perceptive next-generation mobile networks. Additionally, the thesis considers machine learning (ML) for radar sensing and ISAC applications, focusing on the paradigmatic task of human activity recognition. Firstly, a rule-based activity sensing algorithm is demonstrated within a sub-THz ISAC testbed. Alternatively, activity classification is further explored using frequency-modulated continuous wave (FMCW) radar and deep learning. Different neural-network architectures are evaluated in this context, as well as various techniques for cross-domain learning and domain adaptation across radar configurations. Conclusively, cross-domain learning is envisioned to exploit radar datasets for ML training and deployment in novel ISAC systems, e.g., at experimental sub-THz frequencies, while alleviating the burden of time- and resource-consuming data collection procedures.
Hernangómez, Rodrigo (Wed,) studied this question.