Power Amplifier Predistortion With Meta-Learned Polar Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current digital predistortion techniques for power amplifiers in wireless communication systems are inefficient due to high data requirements and slow learning times, making them unsuitable for real-time adaptation to varying operational conditions.
Innovation Solution
A digital predistortion module using two neural networks in the polar domain for amplitude and phase correction, with a meta-learning optimization module that reduces data needs and enables rapid adaptation to changes in the power amplifier's operation, allowing for real-time linearization.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If neural networks are used for digital predistortion to correct power amplifier nonlinearities, then linearization performance is improved, but training data requirements and training time increase significantly
Solution Approach 1:
The patent segments the training process into two distinct phases: offline training that establishes initial weights from comprehensive datasets, and online adaptation that fine-tunes these weights using minimal real-time data. This segmentation allows the system to achieve high linearization performance while minimizing real-time training requirements.
Solution Approach 2:
The patent performs preliminary training offline to establish initial neural network weights before deployment. This preliminary action pre-configures the network with general knowledge about power amplifier characteristics, enabling rapid online adaptation with minimal data and reducing real-time training time significantly.
2Adaptability or versatility
If traditional neural networks are used for digital predistortion, then adaptation to varying operational conditions is possible, but the amount of data required for retraining increases
Solution Approach 1:
The patent implements a dynamic training approach where the neural network transitions from static offline-trained weights to dynamic online adaptation. This allows the system to adapt to varying operational conditions continuously while requiring minimal data updates, as the network only needs to fine-tune pre-existing knowledge rather than learn from scratch.
Solution Approach 2:
The patent changes the training parameter regime by using different learning rates and data quantities for offline versus online phases. Offline training uses comprehensive data with standard learning rates, while online adaptation uses minimal data with higher learning rates for rapid convergence, thereby reducing overall data requirements while maintaining adaptability.
3Manufacturing precision
If digital predistortion module design is made complex to achieve accurate inverse characteristic compensation, then linearization accuracy is improved, but device complexity increases
Solution Approach 1:
The patent replaces complex analytical design methods with neural network-based learning approaches. Instead of manually designing intricate predistortion circuits to achieve accurate inverse characteristics, the system uses neural networks that automatically learn the required compensation through training, simplifying the design process while maintaining or improving accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The invention relates to a device (18) for linearizing, by digital predistortion, a power amplifier, comprising: - a digital predistortion module (30), configured to determine by inference a predistortion in the polar domain to be applied to a signal, and comprising a first neural network (32) and a second neural network (34) respectively configured to correct an amplitude distortion, and a phase distortion, produced by the amplifier; - an optimization module (36) of each of said neural networks (32, 34) configured to implement a meta-learning, using: - a meta-initialization providing beforehand an initialization of the initial weights of each of said neural networks (32, 34); - a meta-adaptation of the initial weights into optimal weights of each of said neural networks (32, 34).