基于实值时间卷积神经网络的功放预失真研究

为了实现传输速率高达千兆比特每秒(Gbps)的目标,5G通信系统需要更宽的传输带宽和更高的调制度,这些对射频功放的线性度提出了更加苛刻的要求。必须对功放的非线性进行线性化。文中构建了一种基于实值时间卷积神经网络(Real-Valued Temporal Convolutional Networks,RVTCN)模型的数字预失真器。RVTCN模型利用扩大因果卷积(Dilated Causal Convolution, DCC)提取功放的当前时序信息,把记忆信息存储在残差块(Residual Block,RB)中,不断获取时序特征并保存于网络中。为了验证RVTCN线性化的性能,文中采用了100 MHz带宽的5G NR信号,对中心频率3.5 GHz的Doherty功放进行了预失真线性化实验验证。实验结果表明:该RVTCN模型具有射频功放的动态非线性行为建模能力,其归一化均方误差可达-40 d B;RVTCN预失真器对测试功放的相邻信道功率比(ACPR)改善可达19.5 d B左右。

Research on Power Amplifier Predistortion Based on Real-valued Temporal Convolutional Neural Networks

In order to achieve transmission rates of up to gigabits per second (Gbps), the fifth-generation mobile communication system (5G) requires wider transmission bandwidth and higher modulation systems. However, these make the linearity requirement of the RF power amplifier more serious. Therefore, the linearization of the amplifier is imperative. In this paper, a digital predistortion based on the real-valued temporal convolutional networks (RVTCN) model is constructed. The RVTCN model uses dilated causal convolution (DCC) to extract current time series information, stores the memory information in the residual block (RB), and continuously acquires the timing features and saves them in the network. In this paper, the 100 MHz bandwidth of the 5G NR signal is used to verify the predistortion linearization experiment of the 3. 5 GHz Doherty amplifier. Experimental results show that the proposed RVTCN model has the ability to model the dynamic nonlinear behavior of RF power amplifiers, of which normalized mean squared error can reach -40 dB. The adjacent channel power ratio (ACPR) of the test amplifier with the digital predistortion built based on the RVTCN can be improved by 19. 5 dB.