Non-linear Memory System Modeling via Complex Domain Preprocessing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing models for non-linear memory systems, such as power amplifiers, face challenges in accurately modeling their behavior without actual output data, often resulting in poor coefficient quality due to overfitting or inadequate training data configurations.
Innovation Solution
An electronic device and method that preprocess AM-AM and AM-PM data by converting them into complex domains, generating filters through inverse Fourier transforms, and performing convolutions to model non-linear memory systems without requiring actual output data, using Finite Impulse Response (FIR) filters to generate output signals and model the system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If a memory polynomial model is used to model a non-linear memory system, then the system can capture non-linear behavior and memory effects, but the coefficients produced during training may have poor quality due to overfitting, training data quality, or training data configuration
Solution Approach 1:
The patent segments the modeling process into distinct stages: data preprocessing (converting AM-AM and AM-PM data to complex domain), filter generation (converting preprocessed data to time domain), and system modeling (performing convolution to generate output signal). This segmentation allows each stage to be optimized independently, improving overall coefficient quality while maintaining modeling accuracy.
Solution Approach 2:
The patent performs preliminary actions by preprocessing the training data before actual model training. Specifically, it converts AM-AM and AM-PM data into the complex domain, then transforms this preprocessed data into the time domain to generate filters. This preliminary data preparation improves the quality of coefficients produced during subsequent training by providing better-structured input data.
2Reliability
If actual output data from the non-linear memory system is used for training, then the model can be trained with real system behavior, but the process requires physical measurements and actual system output which may not always be available
Solution Approach 1:
The patent creates a virtual copy of the training process by generating synthetic output signals through convolution of input signals with generated filters. Instead of requiring actual physical measurements from the non-linear memory system, the method synthesizes training data that mimics real system behavior, making the training process easier to operate without physical system access.
Solution Approach 2:
The system performs self-service by generating its own training data internally through the filter generation and convolution process. The electronic device creates preprocessed data parts from AM-AM and AM-PM data, converts them to time-domain filters, and uses these filters to generate output signals for training, eliminating the need for external data acquisition from the actual non-linear memory system.
3Productivity
If the model is trained with inadequate training data configuration, then training can proceed with available data, but the coefficients produced will have poor quality
Solution Approach 1:
The patent transforms the parameters of training data by changing the domain representation. It converts AM-AM and PM-AM data from their original polar coordinates into the complex domain, then further transforms this preprocessed data into the time domain. This parameter transformation improves coefficient quality while maintaining training efficiency by creating better-structured training inputs.
Solution Approach 2:
The patent replaces the mechanical process of physical data collection and manual training data configuration with an automated signal processing system. The electronic device automatically performs frequency-to-time domain transformation, filter generation, and convolution operations, substituting manual data preparation with automated computational processes that improve coefficient quality without reducing productivity.
Data Source
AI summary
An electronic device includes a processor that converts AM-AM data and AM-PM data into complex domains for each of a plurality of different amplitudes to acquire a plurality of preprocessed data parts, converts the plurality of preprocessed data parts into a time domain to generate a plurality of filters for each of a plurality of different amplitudes, generates an output signal by performing a convolution on an input signal with the plurality of filters, and models a non-linear memory system using the input signal and the output signal.