TIADC Calibration Using Reference-Guided Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Time-interleaved Analog to Digital Converters (TIADCs) suffer from mismatches in sub-ADCs due to deviations in sampling times, gain, and phase-shift, leading to distorted output data, which existing calibration methods struggle to accurately model and correct.
Innovation Solution
A two-phase calibration process involving conventional calibration techniques followed by machine learning-based neural networks to approximate non-linear functions, using a reference channel as ground truth for subchannel-specific neural network training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If multiple sub-ADCs are used to increase sampling rate, then productivity is improved, but mismatches in gain, phase, and timing cause distortion that worsens measurement precision
Solution Approach 1:
The patent implements feedback by using the output of a reference sub-ADC to train neural networks that correct the outputs of other sub-ADCs. The neural networks are trained using the reference channel output as ground truth, creating a closed-loop system where measurement errors are continuously corrected based on feedback from the reference channel, thereby resolving the distortion caused by sub-ADC mismatches while maintaining high sampling rates.
Solution Approach 2:
The patent introduces neural networks as an intermediary between the sub-ADCs and the final output. These neural networks act as mediators that process and correct the outputs from multiple sub-ADCs, using the reference channel as a mediator for training. This intermediary layer enables the system to achieve high sampling rates while maintaining measurement precision by correcting mismatches before final data output.
2Device complexity
If conventional calibration methods are used, then device complexity is reduced, but they fail to accurately model and correct non-linear mismatches, worsening measurement precision
Solution Approach 1:
The patent replaces conventional calibration methods with machine learning-based neural networks. Instead of using traditional calibration techniques that rely on simplified linear models, the system substitutes these with neural networks capable of modeling complex non-linear relationships. This substitution enables accurate correction of non-linear mismatches while keeping the overall device complexity manageable through software-based solutions.
Solution Approach 2:
The patent changes the calibration approach from fixed parameter adjustment to dynamic parameter optimization through neural network training. The neural networks learn optimal correction parameters by training on reference channel data, allowing the system to adapt to non-linear mismatches. This parameter change enables precise correction of gain, phase, and timing errors that conventional methods cannot handle, significantly improving measurement precision.
3Measurement precision
If neural networks are trained for each subchannel, then measurement precision is improved, but device complexity and training time increase
Solution Approach 1:
The patent applies segmentation by training separate neural networks for each subchannel instead of using a single unified calibration system. Each subchannel-specific neural network is trained independently using the reference channel output as ground truth. This segmentation allows each network to specialize in correcting its specific subchannel's mismatches, improving measurement precision while managing complexity through modular, independent training processes.
Data Source
AI summary
Methods and systems and method for calibrating a time-interleaved Analog to Digital Converter (TIADC) are disclosed. In an example, a method for calibrating a TIADC involves a first calibration phase that involves adjusting for at least one of offset error, gain error, and time skew of M subchannels of the TIADC, and a second calibration phase that is implemented after the first calibration phase and that involves training a neural network to approximate a non-linear function of each of the M subchannels of the TIADC using an output from a reference channel as a ground truth.