A Comparison of AI-Enabled Digital Twins for DSP-based Self-Interference Cancellation in Wideband Full-Duplex Communications

We propose the concept of a “digital twin” - namely, a data-driven simulation model of a full-duplex transceiver chain - to enable real-time computational modeling of the entire radio frequency (RF) system, including non-linearities, self-interference coupling, and other performance metrics that depend on multiple hard-to-model factors including the transmit signals and their waveforms, the peak power level, the instantaneous transmit power level, the transceiver circuit settings, and the proximity of the receiver. A machine-learning approach is explored for obtaining artificial intelligence (AI)-based digital twins of various types, including traditional (shallow) neural networks, deep belief networks, and deep convolutional networks for supervised learning, as well as reinforcement learning approaches. While several related approaches exist in the literature, they have generally been limited to simulations. By contrast, the goal of this work is to study the performance of state-of-art machine learning algorithms in an experimental framework using real-world test data obtained from a recently completed STAR front-end [1] . The proposed STAR front-end has measured bandwidths of better than 500 MHz across al-3 GHz range and provides better than 80 dB analog cancellation of self-interference. A dataset of experimental measurements of this front-end will be collected and applied to a suite of ML algorithms for self-interference cancellation that has recently appeared in the literature. The machine learning algorithms will be evaluated for performance in terms of computational complexity, latency, power consumption, and the reduction of self-interference from the transmitter into the receiver during STAR operation.