Current radio communication systems that adopt amplitude and phase modulations demand high linearity and high efficiency. The cascade connection between digital baseband pre-distorter (DPD) and power amplifier (PA) can be a cost-effective solution to guarantee the required linearity without compromising the efficiency. In the design of a DPD for a single band PA, direct learning can be used to extract the pre-inverse parameters or, alternatively, indirect learning can be employed by exchanging the position of the system during the identification procedure to avoid the necessity of a PA model within a closed-loop process. The performance of direct learning is substantially dependent on the accuracy of the behavioral model that replaces the PA. Furthermore, in a practical environment where only an approximation to the inverse is achieved, the linearization capability of the indirect learning is affected by shifting the post-inverse placed after the PA to a pre-inverse located before the PA. For concurrent dual-band PAs, an additional advantage of the indirect approach is that the post-inverse identifications for each band are completely independent of each other. In an authors’ previous work, a comparative analysis between the two learning architectures applied to the linearization of concurrent dual-band PAs was performed based on DPDs modeled by polynomials with memory. This work contributions are the extension of such comparative analysis to DPDs modeled by artificial neural networks, the development of complex-valued three-layer perceptrons suitable for concurrent-dual band DPDs and the introduction of a modified indirect approach to improve the accuracy of previous direct and indirect learnings. Spectre-RF transient simulations are performed in the circuit-under-test described by a wideband 130 nm CMOS PA concurrently stimulated by 2.4 GHz Wi-Fi and 3.5 GHz LTE signals. Reported simulation results show that, in a comparison with the previous direct and indirect learnings with similar output mean powers of about 60 mW, the modified indirect approach provides a superior linearity performance. The modified indirect learning reduces the error vector magnitude (EVM) metric to 0.87% and 1.13% for Wi-Fi and LTE bands, respectively, whereas the indirect and direct learnings achieve EVM equal to or larger than 1.05% and 1.49% for Wi-Fi and LTE bands, respectively.

当前采用幅度和相位调制的无线电通信系统要求高线性度和高效率。数字基带预失真器（DPD）与功率放大器（PA）之间的级联连接可以是一种经济高效的解决方案，可确保所需的线性度而不会影响效率。在针对单频段PA的DPD设计中，可以使用直接学习来提取逆参数，或者，可以通过在识别过程中交换系统位置来使用间接学习，以避免不必要的闭环过程中的PA模型。直接学习的表现基本上取决于替代PA的行为模型的准确性。此外，在实际环境中，只能获得近似的倒数，间接学习的线性化能力会受到以下影响：将置于PA之后的逆后方偏移到位于PA之前的逆前方。对于并发的双频段功率放大器，间接方法的另一个优点是，每个频段的逆后标识完全彼此独立。在作者先前的工作中，基于带记忆多项式的DPD建模，对应用于并行双频PA线性化的两种学习体系结构进行了比较分析。这项工作的贡献是将此类比较分析扩展到了由人工神经网络建模的DPD，开发适用于并发双频DPD的复值三层感知器，并引入改进的间接方法以提高以前的直接和间接学习的准确性。Spectre-RF瞬态仿真是在被测电路中执行的，该电路由2.4 GHz Wi-Fi和3.5 GHz LTE信号同时激发的宽带130 nm CMOS PA构成。报告的仿真结果表明，与以前的直接和间接学习方法（具有大约60 mW的类似输出平均功率）相比，改进的间接方法提供了出色的线性性能。修改后的间接学习将Wi-Fi和LTE频段的误差矢量幅度（EVM）指标分别降低至0.87％和1.13％，而间接学习和直接学习实现的EVM等于或大于1.05％和1。