Time offset signal calibration method for high-speed time-interleaved analog-to-digital converter
By constructing a time-bias signal calibration network model with a cascaded feature extraction network and a weighted summator, the problem of decreased calibration performance of high-speed time-interleaved analog-to-digital converters in the high-frequency region is solved, achieving more robust and accurate time-bias signal calibration and improving signal quality.
Patent Information
- Application Number
- CN202610125492.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-15
AI Technical Summary
In existing technologies, the time offset signal calibration effect of high-speed time-interleaved analog-to-digital converters decreases in the high-frequency region, making it difficult to adapt to complex and ever-changing characteristics.
A time-bias signal calibration network model is constructed, which includes a cascaded feature extraction network, a weight generation network, and a weighted summer. A fully connected module is loaded between the feature extraction network and the weighted summer. Through the combination of various calibration mappings and weighted summers, intelligent calibration of the time-bias signal is achieved.
It improves the accuracy and stability of time offset signal calibration, especially maintaining good calibration performance in the high-frequency region, and enhances the performance of spurious-free dynamic range (SFDR) and signal-to-noise ratio (SNDR).
Smart Images
Figure CN122052790A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of analog-to-digital conversion technology and relates to a method for calibrating the time offset signal of a high-speed time-interleaved analog-to-digital converter. Technical Background
[0002] High-speed time-interleaved analog-to-digital converters (ADCs) employ an architecture that uses multiple channels for alternating sampling. By time-interleaving, the sampling task is distributed across multiple channels, thereby improving the overall system sampling rate. Time offset signal refers to the error component in the output signal of a high-speed time-interleaved ADC when there is a phase error between the sampling clocks of different channels. This error severely degrades the dynamic performance of the high-speed time-interleaved ADC, significantly affecting its signal-to-noise ratio (SNDR) and spurious-free dynamic range (SFDR). The clock offset corresponding to the time offset signal is typically on the picosecond scale, and the distortion it introduces intensifies with increasing time offset signal frequency, resulting in high calibration complexity. Given these characteristics, the core of time offset signal calibration lies in achieving stable and effective calibration for time offset signals of different frequencies within the Nyquist frequency range. Currently, the main idea behind time offset signal calibration network models is to extract features and map these features to the calibration result.
[0003] For example, Zhifei Lu et al., in their paper "A Novel NN-based Fast-Convergence Background Calibration for Timing", proposed a novel approach to NN-based fast-convergence background calibration for timing. The paper "Mismatch in TI ADCs" discloses a time-bias signal calibration method. This method uses a feature extraction network to extract features of the time-bias signal to obtain a feature map. A downsampling layer reduces the size of the feature map, resulting in a smaller feature map. A fully connected layer is then used to perform calibration mapping on the feature map to obtain the calibration result, thus improving calibration accuracy. However, its use of a single fully connected layer for feature map calibration mapping makes it difficult to adapt to the complex and variable features of time-bias signals in the high-frequency region, resulting in a certain degree of decrease in calibration performance in the high-frequency region. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of the prior art by proposing a time offset signal calibration method to solve the technical problem of decreased calibration effect in the high-frequency region in the prior art.
[0005] To achieve the above objectives, the technical solution adopted by the present invention includes the following steps:
[0006] (1) Obtain the training sample set and the test sample set:
[0007] For the output of a high-speed time-interleaved analog-to-digital converter Each time-biased signal is fitted, and the fitted ideal signal is used as the label of the corresponding time-biased signal. The training sample set consists of one time-biased signal and its label, and the remaining time-biased signals form the test sample set. ;
[0008] (2) Constructing a time-biased signal calibration network model:
[0009] A time-biased signal calibration network model is constructed, which includes a cascaded feature extraction network, a weight generation network, and a weighted summer, and a fully connected module is loaded between the input of the feature extraction network and the input of the weighted summer.
[0010] (3) Iteratively train the time offset signal calibration network model:
[0011] The time-off signal calibration network model is iteratively trained using a training sample set to obtain a well-trained time-off signal calibration network model.
[0012] (4) Obtain the time offset signal calibration results:
[0013] The test sample set is used as input to the trained time-biased signal calibration network model for forward propagation to obtain the time-biased signal calibration result corresponding to the test sample.
[0014] Compared with the prior art, the present invention has the following advantages:
[0015] In the process of training the time-biased signal calibration network model and obtaining the time-biased signal calibration results, the fully connected module performs multiple calibration mappings on the time-biased signal. The weighted summer generates a weight vector based on the weights of the network output and performs a weighted summation on the multiple preliminary calibration results output by the fully connected module. This avoids the shortcomings of existing technologies that use a single fully connected layer to calibrate and map the feature map, which makes it difficult to adapt to the complex and variable time-biased signal features in the high-frequency region. This solves the problem of decreased calibration effect in the high-frequency region. Attached Figure Description
[0016] Figure 1 This is a flowchart illustrating the implementation of the present invention.
[0017] Figure 2 This is a schematic diagram of the time offset signal calibration network in this invention.
[0018] Figure 3 (a) and (b) are calibration effect diagrams of the prior art and the present invention, respectively. Detailed Implementation
[0019] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0020] Reference Figure 1 The present invention includes the following steps:
[0021] Step 1) Obtain the training sample set and the test sample set:
[0022] For the output of a high-speed time-interleaved analog-to-digital converter Each time-biased signal is fitted, and the fitted ideal signal is used as the label of the corresponding time-biased signal. The training sample set consists of one time-biased signal and its label, and the remaining time-biased signals form the test sample set. In this embodiment , .
[0023] Among them The time-biased signals were fitted using the least squares method.
[0024] Step 2) Construct a time-biased signal calibration network model, the structure of which is as follows: Figure 2 As shown:
[0025] A time-biased signal calibration network model is constructed, comprising a cascaded feature extraction network, a weight generation network, and a weighted summer, with a fully connected module loaded between the input of the feature extraction network and the input of the weighted summer. Wherein:
[0026] The feature extraction network consists of stacked convolutional layers, pooling layers, and flattening layers.
[0027] Weight generation network, consisting of stacked fully connected layers and Function layer.
[0028] Fully connected modules, including parallel ones A fully connected layer, in this embodiment .
[0029] Step 3) Iteratively train the time offset signal calibration network model, which is implemented as follows:
[0030] (3a) Initialize the number of iterations to be The maximum number of iterations is , , No. The weights and bias parameters of the network model in the next iteration are respectively , and order In this embodiment ;
[0031] (3b) Randomly select samples from the training sample set In this embodiment, training samples are used as input to the time-biased signal calibration network model for forward propagation. The implementation steps are as follows:
[0032] (3b1) The convolutional layer in the feature extraction network extracts features from each training sample. It extracts multiple local features of the training sample by sliding the convolutional kernel and assembling them into a feature map. The pooling layer pools the extracted feature map, reducing its size while preserving the main feature information. The flattening layer flattens the pooled feature map, unfolding the two-dimensional feature map into a one-dimensional vector in column order to connect to the input of subsequent fully connected layers, thus obtaining... 1 eigenvector;
[0033] (3b2) The fully connected layers in the weight generation network classify each feature vector, and different classes correspond to different calibration strategies. Function layer obtains classification The scores for each category are non-linearly normalized to obtain... The weight vector, the weight vector in the weight vector element values This indicates that the training sample is suitable for using the first... The degree of calibration mapping.
[0034] (3b3) Parallelism in fully connected modules Each fully connected layer performs [processing] on each training sample. A calibration mapping is obtained The preliminary calibration results show that each fully connected layer represents a different calibration mapping. Different fully connected layers are good at handling training samples with different features. This avoids the shortcomings of existing technologies that use a single fully connected layer to calibrate and map feature maps, which makes it difficult to adapt to the complex and variable time-biased signals in high-frequency regions. This solves the problem of decreased calibration performance in high-frequency regions.
[0035] (3b4) The weighted summer is obtained based on the calibration mapping of the weight vector. The preliminary calibration results are weighted and summed to obtain... Calibration results of time-off signal This step enables the model to intelligently select and fuse the most suitable calibration mapping for each specific training sample, thereby obtaining more robust and accurate calibration results.
[0036] The weighted summer is obtained based on the calibration mapping of the weight vector. The preliminary calibration results are weighted and summed using the following formula:
[0037] ;
[0038] in, Indicates the first The calibration results for each sample Indicates the first Each weighting coefficient Indicates the first Preliminary calibration results.
[0039] (3c) Using the mean squared error loss function, the loss value of the network model is calculated by calibrating each time-biased signal output by the time-biased signal and its corresponding true label. .
[0040] The loss value of the network model The calculation formula is:
[0041] ;
[0042] in, Indicates the first Labels for each sample.
[0043] (3d) Gradient descent method is used, through right , The time-off signal calibration network model for this iteration is then updated.
[0044] The gradient descent method is used to... , The update is performed using the following formulas:
[0045] ;
[0046] ;
[0047] in, , They represent , The update results Indicates the learning rate. express right Take the partial derivative, express right Take the partial derivative.
[0048] (3e) Judgment If it holds true, then obtain the trained time-biased signal calibration network model; otherwise, let... Then proceed with step (3b).
[0049] Step 4) Obtain the time offset signal calibration result:
[0050] The test sample set is used as input to the trained time-biased signal calibration network model for forward propagation to obtain the time-biased signal calibration result corresponding to the test sample.
[0051] The technical effects of this invention will be explained below with reference to simulation experiments:
[0052] 1. Experimental conditions and contents:
[0053] The hardware platform for the simulation experiment was as follows: the arbitrary waveform generator was a Rigol DG70004; the ADC evaluation board was a four-channel interleaved ADC12D1800RFRB with a sampling rate of 3.6 GS / s and a resolution of 12 bits; the processor was an Intel(R) Core i7-8700 CPU; and the graphics card was an NVIDIA GeForce RTX 4060.
[0054] The software platform is WaveVision 5, used for controlling and analyzing the ADC evaluation board. The operating system is Ubuntu 20.04, Python version 3.6, and PyTorch version 1.7.1.
[0055] Experimental Method: A clean signal with the required frequency and amplitude was generated using an arbitrary waveform generator (DG70000) and input to the ADC evaluation board (ADC12D1800RFRB). The ADC evaluation board was controlled using WaveVision 5 software to acquire the input signal at a sampling rate of 3.6 GS / s, obtaining time-biased signal data including clock skew. Under the PyTorch framework, both an existing neural network calibration model and the model proposed in this invention were loaded. The acquired time-biased signal was input into the model to obtain the calibrated data. Under the same experimental conditions, the spurious-free dynamic range (SFDR) and signal-to-noise ratio (SNDR) were calculated for the data before and after calibration using both methods. The comparison results are shown below. Figure 3 As shown in (a) and (b).
[0056] Spurious-free dynamic range (SFDR) describes the ratio of the amplitude of the dominant frequency component of a signal to the amplitude of the maximum spurious component, reflecting the system's ability to suppress spurious signals. Its calculation formula is:
[0057] ;
[0058] in, Main signal power, This represents the maximum stray component.
[0059] Signal-to-noise ratio (SNDR) represents the ratio of signal power to the sum of the power of noise and distortion components, comprehensively reflecting the noise and distortion performance of a system. Its calculation formula is:
[0060] ;
[0061] in, For noise power, This represents the power of the distortion component.
[0062] 2. Analysis of experimental results:
[0063] Figure 3(a) shows the calibration effect of the existing neural network calibration model. Before calibration, the SFDR and SNDR of the signal decrease with increasing frequency; after calibration, although both indicators improve, they still show a significant decreasing trend with increasing frequency, especially in the high-frequency region. Figure 3(b) shows the calibration effect of the model proposed in this invention. After calibration, the SFDR is stably improved to about 85 dB, and the SNDR is stably improved to about 68 dB, with no performance degradation across the entire frequency range, and the calibration effect remains stable.
Claims
1. A method for calibrating the time offset signal of a high-speed time-interleaved analog-to-digital converter, characterized in that, Includes the following steps: (1) Obtain the training sample set and the test sample set: For the output of a high-speed time-interleaved analog-to-digital converter Each time-biased signal is fitted, and the fitted ideal signal is used as the label for the corresponding time-biased signal. The training sample set consists of one time-biased signal and its label, and the remaining time-biased signals form the test sample set. ; (2) Constructing a time-biased signal calibration network model: A time-biased signal calibration network model is constructed, which includes a cascaded feature extraction network, a weight generation network, and a weighted summer, and a fully connected module is loaded between the input of the feature extraction network and the input of the weighted summer. (3) Iteratively train the time offset signal calibration network model: The time-off signal calibration network model is iteratively trained using a training sample set to obtain a well-trained time-off signal calibration network model. (4) Obtain the time offset signal calibration results: The test sample set is used as input to the trained time-biased signal calibration network model for forward propagation to obtain the time-biased signal calibration result corresponding to the test sample.
2. The method according to claim 1, characterized in that, The time-biased signal calibration network model described in step (2) includes a feature extraction network comprising stacked convolutional layers, pooling layers, and flattening layers; and a weight generation network comprising stacked fully connected layers and... Function layer; fully connected modules include parallel... A fully connected layer, .
3. The method according to claim 1, characterized in that, The iterative training of the time offset signal calibration network model described in step (3) is implemented as follows: (3a) Initialize the number of iterations to be The maximum number of iterations is , , No. The weights and bias parameters of the network model in the next iteration are respectively , and order ; (3b) Randomly select samples from the training sample set The training samples are used as input to the time-biased signal calibration network model for forward propagation to obtain... Calibration results of the time offset signal; (3c) Using the mean squared error loss function, the loss value of the network model is calculated by calibrating each time-biased signal output by the time-biased signal and its corresponding true label. ; (3d) Gradient descent method is used, through right , The time offset signal calibration network model for this iteration is then updated. (3e) Judgment If it holds true, then obtain the trained time-biased signal calibration network model; otherwise, let... Then proceed with step (3b).
4. The method according to claim 3, characterized in that, The forward propagation described in step (3b) is implemented as follows: (3b1) In the feature extraction network, the convolutional layer extracts features for each training sample, the pooling layer pools the extracted feature maps, and the flattening layer flattens the pooled feature maps to obtain... 1 eigenvector; (3b2) The fully connected layers in the weight generation network classify each feature vector. Function layer obtains classification The scores for each category are non-linearly normalized to obtain... One weight vector; (3b3) Parallelism in fully connected modules Each fully connected layer performs [processing] on each training sample. A calibration mapping is obtained Preliminary calibration results; (3b4) The weighted summer is obtained by calibrating the mapping for each training sample based on each weight vector. The preliminary calibration results are weighted and summed to obtain... Calibration results of time-off signal .
5. The method according to claim 4, characterized in that, The weighted summer described in step (3b4) is obtained based on the calibration mapping of the weight vector pairs. The preliminary calibration results are weighted and summed using the following formula: ； in, Indicates the first The calibration results for each sample , They represent the first A preliminary calibration result and its weighting coefficients.
6. The method according to claim 4, characterized in that, The loss value of the network model described in step (3c) The calculation formula is: ； in, express The corresponding tags.
7. The method according to claim 3, characterized in that, The gradient descent method described in step (3d) is used for... , The update is performed using the following formulas: ； ； in, , They represent , The update results Indicates the learning rate. express right Take the partial derivative, express right Take the partial derivative.