Power Amplifier Predistortion Using Dynamic Memory-Tap Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural network-based architectures for power amplifier non-linearity compensation fail to effectively utilize apriori information about the impact of previous samples and require complete retraining for minor changes, leading to high training and inference complexity.
Innovation Solution
Implementing a neural network with separate sub-networks for each memory tap, allowing flexible adaptation to sample chronology and enabling online training of only a fixed and trainable part, reducing training and inference complexity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing neural network-based architectures are used for power amplifier non-linearity compensation, then the system can compensate for non-linear effects, but the training and inference complexity becomes high due to complete retraining requirements and inability to utilize apriori information about previous samples
Solution Approach 1:
The patent divides the neural network into separate sub-networks for each memory tap (current sample and previous samples). Each sub-network independently processes samples from different time points, allowing the system to utilize apriori information about the impact of previous samples while reducing overall complexity through modular structure
Solution Approach 2:
The patent implements dynamic adaptability by allowing different sub-networks to be selectively activated based on the chronology and importance of samples. The system can dynamically adjust which memory taps are used and retrain only the relevant sub-networks rather than complete retraining, reducing training overhead while maintaining compensation performance
2Adaptability or versatility
If complete neural network retraining is performed for minor changes, then the system adapts to changes, but the training overhead and time consumption increase significantly
Solution Approach 1:
By segmenting the neural network into independent sub-networks for each memory tap, the system can retrain only the specific sub-networks that need adjustment for minor changes, rather than retraining the complete network. This selective retraining approach maintains adaptability while significantly reducing training time overhead
Solution Approach 2:
The patent applies partial action by retraining only the necessary portion of the neural network (specific sub-networks) rather than the entire network. This partial retraining approach provides sufficient adaptation for minor changes without the excessive time cost of complete retraining
Data Source
AI summary
Embodiments herein disclose a method for controlling a non-linear effect of a power amplifier by an apparatus. The method includes acquiring an input data of the power amplifier of the apparatus and an output data of the power amplifier. Further, the method includes determining an inverse function using a neural network. The inverse function maps normalized output data of the PA to the input data of the PA, where the neural network comprises at least one sub-network for at least one memory tap from a plurality of memory taps in the neural network. Further, the method includes modifying the input data based on the determined inverse function value by dynamically changing a usage of the at least one memory tap from the plurality of memory taps. Further, the method includes compensating the non-linear effect in the output data of the power amplifier.