Advanced Machine Learning Techniques for Self-Interference Cancellation in Full-Duplex Radios

Abstract

In-band full-duplex systems allow for more efficient use of temporal and spectral resources by transmitting and receiving information at the same time and on the same frequency. However, this creates a strong self-interference signal at the receiver, making the use of self-interference cancellation critical. Recently, neural networks have been used to perform digital self-interference with lower computational complexity compared to a traditional polynomial model. In this paper, we examine the use of advanced neural networks, such as recurrent and complex-valued neural networks, and we perform an in-depth network architecture exploration. Our neural network architecture exploration reveals that complex-valued neural networks can significantly reduce both the number of floating-point operations and parameters compared to a polynomial model, whereas the real-valued networks only reduce the number of floating-point operations. For example, at a digital self-interference cancellation of 44.51 dB, a complex-valued neural network requires 33.7 % fewer floating-point operations and 26.9 % fewer parameters compared to the polynomial model. 500

Hybrid-Layers Neural Network Architectures for Modeling the Self-Interference in Full-Duplex Systems

“…For the CV-TDNN [30], using the previous implementation assumptions for the CV additions and multiplications, the number of FLOPs and parameters can be given by…”

“…In the following experiments, the public dataset employed in [29] and [30] is utilized to train and verify the proposed NN architectures. The dataset is produced using a realistic FD testbed, which generates a 10 MHz quadrature phase-shift keying modulated-orthogonal frequency division multiplexing (OFDM) signal with an average transmit power of 10 dBm.…”

“…The first part is used for training and consists of 90% of samples, whereas the second is employed for verification and contains the rest 10%. The employed hardware test-bed provides 53 dB passive analog RF cancellation using physical separation between the transmit and receive antennas; hence, in this work, no further active analog cancellation techniques are employed since the use of digital cancellation after passive analog suppression is adequate to bring the SI signal's power down to the receiver's noise level [29], [30]. A summary of the parameters employed to generate the aforementioned dataset is presented in Table . I.…”

“…More specifically, the first attempt of using NNs for canceling the SI has been reported in [29], where a real-valued time delay NN (RV-TDNN) 1 is introduced to model the SI signal with computational complexity lower than the polynomial-based canceler. In [30], a recurrent NN (RNN) and a complex-valued TDNN (CV-TDNN) have been investigated for SI mitigation; it is shown that the CV-TDNN has excellent modeling capabilities to approximate the SI with lower computational complexity than the polynomial and RNN-based cancelers. In [31], [32], the hardware design of the polynomial and NN-based cancelers introduced in [29] have been provided.…”

Low-Complex Digital Cancellation of Transmitter Harmonics in LTE-A/5G Transceivers

“…Most concepts are based on polynomial PA models with and without memory. The estimation of the nonlinearity coefficients and the leakage path is implemented either using classical estimation approaches, such as batch least-squares solutions and the recursive least squares (RLS) [16], [17], or adaptive learning methods, like least-mean squares (LMS) algorithms combined with basis function orthogonalization [18], spline adaptive filters [19] and neural networks [20]. For many of these concepts, the computational complexity during training is high.…”