Hybrid DPD for Power Amplifiers With Faster NN Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing digital pre-distortion (DPD) schemes for power amplifiers (PAs) in high-data-rate communication systems, such as 5G, face challenges with high implementation complexity and long training times, particularly in schemes based on generalized memory polynomials (GMP) and artificial-intelligence neural networks (ANN), which hinder optimal hyperparameter identification.
Innovation Solution
An electronic device employs a hybrid DPD scheme combining a generalized memory polynomial (GMP) scheme with a neural network (NN) scheme to identify weight values and hyperparameters, correcting non-linear data in PA input using both schemes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a DPD scheme based on the GMP scheme is used, then signal linearity is maintained, but implementation complexity increases exponentially
Solution Approach 1:
The patent combines the GMP scheme and ANN scheme into a hybrid DPD system where both schemes operate together. The GMP scheme handles the mathematical modeling of PA non-linearity while the ANN scheme learns optimal hyperparameters from training data, merging the strengths of both approaches to maintain signal linearity while managing implementation complexity
Solution Approach 2:
The patent changes the parameters of the ANN scheme by using the GMP scheme's weight values as input features for training the ANN. This parameter transformation allows the system to leverage the structured mathematical model of GMP while allowing the ANN to adaptively learn optimal configurations, thereby reducing the exponential complexity growth
2Reliability
If a DPD scheme based on the ANN scheme is used, then signal linearity is maintained, but training time increases
Solution Approach 1:
The patent performs preliminary action by using the GMP scheme to calculate weight values that serve as pre-processed features for the ANN training process. This preliminary calculation provides the ANN with meaningful starting points and structured data, reducing the time required for the ANN to converge during training while still achieving optimal signal linearity
Solution Approach 2:
The patent implements feedback by using the PA's input and output data to continuously refine both the GMP weight values and ANN hyperparameters. This feedback mechanism allows the system to learn from actual PA behavior and adjust parameters efficiently, reducing training time while maintaining high signal linearity performance
3Measurement precision
If the GMP scheme precision is increased, then signal linearity improves, but computation resources increase exponentially
Solution Approach 1:
The patent introduces the ANN scheme as an intermediary that bridges the GMP mathematical model and the final DPD output. The ANN learns to map GMP weight values to optimal hyperparameters, allowing the system to achieve high precision signal linearity correction without directly computing the full exponential complexity of high-precision GMP models, thus reducing computation resource usage
Data Source
AI summary
An electronic device includes a power amplifier (PA), one or more processors coupled with the PA, and memory storing instructions. The instructions cause the electronic device to monitor input data of the PA and output data of the PA, identify a weight value for a first digital pre-distortion (DPD) scheme, identify a hyperparameter for the second DPD scheme, based on estimated input data of the PA estimated based on a second DPD scheme which is based on a neural network (NN) scheme and estimated input data of the PA estimated based on the first DPD scheme, and the input data of the PA, correct first non-linear data included in the input data of the PA based on the weight value and the first DPD scheme, and correct second non-linear data included in the input data of the PA based on the hyperparameter and the second DPD scheme.