[Paper Review] Convolutional Neural Network for Behavioral Modeling and Predistortion of Wideband Power Amplifiers

Xin Hu, Zhijun Liu+9 more|arXiv (Cornell University)|May 20, 2020

Advanced Power Amplifier Design43 references4 citations

TL;DR

This paper proposes a real-valued time-delay convolutional neural network (RVTDCNN) for behavioral modeling and predistortion of wideband power amplifiers (PAs). By leveraging convolutional layers to extract basis functions from I/Q and envelope-dependent signal components, the model efficiently captures strong memory effects with low complexity, achieving performance comparable to deeper neural networks while reducing training burden.

ABSTRACT

In this paper, we propose a novel behavior model for wideband PAs using a real-valued time-delay convolutional neural network (RVTDCNN). The input data of the model are sorted and arranged as the graph composed of the in-phase and quadrature (I/Q) components and envelope-dependent terms of current and past signals. We design a pre-designed filter using the convolutional layer to extract the basis functions required for the PA forward or reverse modeling. The generated rich basis functions are modeled using a simple fully connected layer. Because of the weight sharing characteristics of the convolutional structure, the strong memory effect does not lead to a obvious increase in the complexity of the model. Meanwhile, the extraction effect of the pre-designed filter also reduces the training complexity of the model. The experimental results show that the performance of the RVTDCNN model is almost the same as the NN models and the multilayer NN models.

Motivation & Objective

• To address the challenge of modeling strong memory effects in wideband power amplifiers using data-driven techniques.
• To reduce the complexity of neural network models for PA behavioral modeling and predistortion.
• To improve training efficiency while maintaining high modeling accuracy for wideband signals.
• To leverage convolutional structures for effective basis function extraction from I/Q and envelope-dependent signal components.
• To demonstrate that a lightweight RVTDCNN can achieve performance on par with deeper neural networks.

Proposed method

• The RVTDCNN uses a real-valued time-delay convolutional layer to extract basis functions from the in-phase (I), quadrature (Q), and envelope-dependent components of current and past input signals.
• The input is structured as a graph of I/Q and envelope terms, enabling the network to model memory effects through temporal dependencies.
• A pre-designed filter in the convolutional layer reduces the need for extensive backpropagation, lowering training complexity.
• Weight sharing in the convolutional layer maintains low model complexity despite strong memory effects.
• The extracted features are passed through a simple fully connected layer for final output prediction in forward or reverse modeling.
• The architecture is trained end-to-end for behavioral modeling and predistortion, with performance evaluated on wideband PA data.

Experimental results

Research questions

• RQ1 Can a convolutional neural network effectively model the nonlinear and memory-dependent behavior of wideband power amplifiers?
• RQ2 Does the use of a pre-designed filter in the convolutional layer reduce training complexity without sacrificing modeling accuracy?
• RQ3 How does the RVTDCNN compare in performance and complexity to standard and multilayer neural networks for PA modeling?
• RQ4 Can the RVTDCNN efficiently capture strong memory effects with minimal parameter growth?
• RQ5 Is the RVTDCNN suitable for real-time predistortion applications in wideband communication systems?

Key findings

• The RVTDCNN achieves modeling accuracy comparable to standard and multilayer neural networks, demonstrating strong generalization on wideband PA data.
• The model maintains low complexity due to weight sharing and the pre-designed filter, avoiding the exponential parameter growth typically seen with memory effects.
• The training complexity is significantly reduced compared to conventional deep neural networks, thanks to the filter design and convolutional feature extraction.
• The RVTDCNN effectively captures the nonlinear and memory-dependent behavior of wideband PAs using only I/Q and envelope-dependent signal components.
• The experimental results confirm that the RVTDCNN is a viable and efficient alternative for both behavioral modeling and predistortion in wideband systems.
• The model's performance is robust across various wideband signal conditions, indicating suitability for practical implementation.