Reducing Parameter Estimation Error of Behavioral Modeling and Digital Predistortion via Transfer Learning for RF Power Amplifiers

Digital predistortion (DPD) has been widely used in linearizing radio frequency (RF) power amplifiers (PAs). However, model coefficients could not always be estimated accurately for a variety of reasons. Several regularization methods have been developed for parameter identification. However, the performance improvement is limited due to the missing information. Fortunately, if parameters from earlier operating conditions are available, they can be employed to enhance the accuracy of DPD in the current state. Despite the fact that many adaptive DPD methods are based on related concepts, they merely use past parameters as initialization for the target task. In this article, we proposed some novel transfer learning-based parameter estimation techniques for PAs operating in time-varying operating configurations. By effectively utilizing the structure knowledge of noncurrent parameters as a priori rather than just initializing them, the estimation error can be significantly decreased. Applying few-sample learning (FSL), for instance, can help to simplify the computational process of parameter extraction, but its robustness is poor. And the experimental results prove that the proposed method is useful for reducing the parameter estimation bias in FSL with negligible extra computational complexity.