Hundreds Million Hertz Broadband Digital Pre-distortion System Design And Implementation Based On ZYNQ SoC

The RF power amplifier(PA)is an important component of the RF frontend,which is the core part of a base station as well.Academia and industry have paid close attention to its linearity and efficiency.In order to improve the linearity and efficiency of the power amplifier,digital predistortion(DPD)technology has become the mainstream solution due to its wide applicability and flexible parameter configuration.With the advent of 5G in mobile communication technology,the demand for information transmission rates has increased,resulting in an increasing signal transmission bandwidth.High-speed and wide-bandwidth mobile communication scenarios present new challenges to traditional DPD systems.High-speed rates mean increased difficulty in real-time processing of signal flows,and wide bandwidth means more complex nonlinear distortion of the power amplifier and increased complexity of coefficient solving algorithms.To address the increased difficulty in real-time processing of highspeed signals,a multi-phase parallel processing approach is adopted to reduce the requirements for the main clock frequency.Based on this,a lowcomplexity dual-loop digital predistortion learning method is proposed to overcome the shortcomings of traditional predistortion learning structures and the difficulties in modeling and extracting parameters caused by broadband signals.The coarse adjustment loop prunes the DPD model using a low-complexity doubly orthogonal matching pursuit algorithm and introduces matrix inversion optimization to transform matrix inversion into matrix multiplication.The optimal number of parameters is evaluated based on the Bayesian information criterion,and the obtained approximate value serves as the initial value to escape from local optima.Then,the system switches to the fine-tuning loop,which uses a Gauss-Newton iteration and Householder rank-revealing reduced QR decomposition with column pivoting to minimize residual errors and achieve online real-time updating of model coefficients,ultimately converging to the global optimum.Based on the proposed method,a broadband DPD system is designed and implemented on a 5G pico microcell base station equipped with ZYNQ SoC as the validation platform.OpenBLAS is compiled and configured as the linear algebra computation library,and optimizations such as NEON instructions and matrix blocking are employed to optimize matrix multiplication.The system was tested using a 100MHz bandwidth 5G NR signal on the experimental platform,and the experiments show that the system can effectively prune model parameters and maintain good linearization performance.The dual-loop learning structure is capable of suppressing system noise to some extent.With the increase in training data,the fine-tuning loop iteration curve stabilizes and converges,indicating that the system has good robustness and error correction capability,enabling real-time tracking and compensation of broadband PA nonlinearity.