Abstract
Tracking the nonlinear behavior of an RF power amplifier (PA) is challenging. To tackle this problem, we build a connection between residual learning and the PA nonlinearity, and propose a novel residual neural network structure, referred to as the residual real-valued time-delay neural network (R2TDNN). Instead of learning the whole behavior of the PA, the R2TDNN focuses on learning its nonlinear behavior by adding identity shortcut connections between the input and output layer. In particular, we apply the R2TDNN to digital predistortion and measure experimental results on a real PA. Compared with neural networks recently proposed by Liu et al. and Wang et al., the R2TDNN achieves the best linearization performance in terms of normalized mean square error and adjacent channel power ratio with less or similar computational complexity. Furthermore, the R2TDNN exhibits significantly faster training speed and lower training error.
Introduction
In modern wireless communications, Power Amplifiers (PAs) are critical components that must operate near their saturation point to maximize energy efficiency. However, this proximity to saturation introduces significant nonlinear behavior, leading to signal distortion and spectral regrowth—interference that leaks into adjacent frequency channels. Digital Predistortion (DPD) is the standard technique used to mitigate these effects by applying an inverse distortion to the signal before it reaches the amplifier. While traditional mathematical models like the Volterra series are effective, they often struggle with the severe nonlinearities and memory effects found in high-bandwidth 5G signals.
The R2TDNN Architecture
To address these challenges, we propose the Residual Real-valued Time-delay Neural Network (R2TDNN). This architecture builds upon the standard Multilayer Perceptron (MLP) but introduces two critical innovations: residual learning and real-valued time-delays.
By incorporating identity shortcut connections, the network is not required to map the entire input-output relationship. Instead, it focuses exclusively on learning the residual—the complex nonlinear distortion component. This allows the hidden layers to dedicate their full capacity to modeling the nonlinearities. Additionally, the network handles complex RF signals by splitting them into In-phase (I) and Quadrature (Q) components and utilizes tapped delay lines to account for the PA's memory effects.
Experimental Evaluation
The R2TDNN was evaluated using the RF WebLab remote experimental setup, featuring a Gallium Nitride (GaN) Power Amplifier. The test signal was a 10 MHz bandwidth OFDM signal, typical of 5G, sampled at 200 MHz. A dataset of $10^6$ samples was used for training and evaluation using the Indirect Learning Architecture (ILA), where a post-distorter is identified and then mirrored as a pre-distorter.
Results and Discussion
The R2TDNN was compared against the Generalized Memory Polynomial (GMP) and other state-of-the-art neural networks, including the RVTDNN and ARVTDNN.
As shown in the performance analysis, the R2TDNN achieved a Normalized Mean Square Error (NMSE) of -33.7 dB, outperforming the standard RVTDNN (-31.5 dB). It also achieved an Adjacent Channel Power Ratio (ACPR) of -42.5 dBc. A key advantage of the R2TDNN is its efficiency; once the computational budget exceeds approximately 500 FLOPs, it provides a significantly more accurate inverse model than any other tested architecture. Furthermore, the residual structure enables faster convergence during training, reaching a lower error floor in fewer epochs.
Conclusion
The R2TDNN offers a robust and efficient solution for DPD in 5G systems. By focusing on learning only the nonlinear residuals of the power amplifier, it achieves high linearization accuracy with manageable computational complexity. Future work will explore scaling the model to wider bandwidths (e.g., 100 MHz or 400 MHz) and optimizing hardware implementations for real-time operation on FPGAs or ASICs.
Summary updated