基于广义LSTM神经网络的宽带射频功放数字预失真线性化

在无线通信系统中,射频功放的非线性是信号失真与频谱增生的主要原因,尤其是对于采用64QAM、256QAM等高峰均功率比的复杂调制系统,对射频功放线性度的要求越来越高;然而宽带射频功放中存在的强记忆效应严重地降低了基于传统非线性模型的数字预失真器的线性化性能。文章提出广义长短期记忆(LSTM)神经网络模型,通过输入的时序特性,从时间轴上进行模型迭代,利用LSTM模型独特的长短时序结构以更好地表征宽带射频功放的记忆效应,同时引入时间超前项以构建广义的LSTM模型,进一步增强其动态非线性建模能力。在不同超参数下的建模结果表明,该模型的归一化均方误差(NMSE)指标可达-42.2895 dB。最后,使用20 MHz带宽的4载波WCDMA信号,对中心频率1900 MHz的50 W Doherty功放进行预失真线性化实验验证。实验结果证实了基于广义LSTM神经网络模型的数字预失真器可以使互调分量降低达23.27 dB,大大优于记忆多项式等传统非线性模型的非线性校正性能。

Digital Pre-distortion Linearization Based on Generalized LSTM Neural Networks for Wideband RF Power Amplifiers

In the wireless communication system, the nonlinearity of radio frequency power amplifier is the principal reason for signal distortion and spectrum regrowth. Especially for the complex modulation system with high peak average power ratio such as 64QAM and 256QAM, the linearity of radio frequency power amplifier is strictly required and a critical specification. However, the strong memory effect in the wideband radio frequency power amplifier seriously reduces the linearization performance of the digital predistorter based on the traditional nonlinear models. In this paper, a generalized long short-term memory (LSTM) neural network is presented, in which by using the input sequence features, the model iteration from the time axis, and the length of the unique temporal structure in the LSTM model, the memory effect of wideband radio frequency power amplifier can be represented in better accuracy. At the same time,introducing time ahead to build the generalized LSTM neural network model, the dynamic nonlinear modeling capabilities can be further enhanced. The modeling results under different hyper-parameters illustrate that the NMSE of the model can be as high as -42.2895 dB. Finally, using a 4 carrier WCDMA signal with 20 MHz bandwidth as the test signal, a 50 W Doherty power amplifier with 1900 MHz central frequency is utilized to make experimental validation for predistortion linearization. The experimental results confirm that the digital predistorter based on the generalized LSTM neural network model can reduce the intermodulation component by 23.27 dB, which is much better than the nonlinear correction performance of the traditional nonlinear model such as memory polynomials.