Convolutional Neural Network for Behavioral Modeling and Predistortion of Wideband Power Amplifiers

Area of Science:

• Electrical Engineering
• Computer Science
• Signal Processing

Background:

• Traditional neural network (NN) models for power amplifier (PA) behavior modeling suffer from high complexity.
• Existing models struggle to efficiently capture the memory effects in wideband PAs.

Purpose of the Study:

• To propose a novel, simplified behavior model for wideband PAs.
• To reduce the computational complexity and coefficient number of PA models.

Main Methods:

• A real-valued time-delay convolutional neural network (RVTDCNN) was developed.
• Input data was structured as a graph of in-phase/quadrature components and envelope-dependent terms.
• A convolutional layer was used to extract basis functions for PA modeling.
• A fully connected layer integrated these functions into the final model.

Main Results:

• The RVTDCNN model achieved performance comparable to existing NN and multilayer NN models.
• The proposed model significantly reduced the number of coefficients and computational complexity.
• These benefits were particularly evident with wider signal bandwidths and increased PA memory effects.

Conclusions:

• The RVTDCNN offers an efficient and effective solution for wideband PA behavior modeling.
• Its architecture inherently manages complexity associated with memory effects.
• This model presents a significant advancement for PA design and simulation.