Multi-Gate Mixture-of-Experts for Predistortion of Power Amplifiers in Satellite Multi-Antenna System

To meet the high data rate and large capacity demands of communications, multi-antenna technology is widely adopted in satellite communication systems, significantly increasing system capacity and spectrum efficiency by transmitting data through multiple parallel data streams. However, the increase in the number of transmission paths introduces significant challenges, such as power amplifier (PA) nonlinearity and inter-antenna crosstalk. Traditional digital predistortion (DPD) techniques typically rely on independent compensation for each PA, which leads to high hardware complexity and scalability issues. Moreover, signal crosstalk distorts the feedback signal used for DPD parameter estimation, thereby degrading the accuracy of the nonlinear modeling. To solve the above-mentioned problems, this paper proposes a novel DPD approach that jointly compensates multiple PAs using a single model, thereby ensuring beam linearity while substantially reducing hardware overhead. Specifically, we design a multi-gate mixture-of-experts-based DPD model to mitigate PA nonlinear distortions and multi-antenna crosstalk. The proposed model formulates the inverse modeling of each PA as a multi-task learning problem, in which shared and task-specific representations are jointly leveraged to extract both common and distinctive nonlinear features, thereby enabling efficient and simultaneous compensation across multiple PAs. Experimental evaluations show that the proposed method outperforms conventional polynomial-based and neural network-based approaches in compensating for nonlinear distortions and suppressing inter-antenna crosstalk. Specifically, in a 4-path antenna array, the proposed model achieves an improvement of 2 dB in adjacent channel power ratio (ACPR) and a 3 dB reduction in normalized mean square error (NMSE).