High-speed analog-to-digital converter calibration method and circuit based on margin feedback neural network
By employing a parallel calibration method based on a margin feedback neural network, the problems of nonlinear error, gain error, and capacitance mismatch in analog-to-digital converters are solved, achieving high-speed and high-precision calibration results.
Patent Information
- Application Number
- CN202510938858.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-08
- Publication Date
- 2025-11-07
AI Technical Summary
Existing analog-to-digital converters suffer from nonlinearity errors, gain errors, and capacitance mismatch during calibration. Traditional calibration methods are computationally complex, time-consuming, and have unstable accuracy, making it difficult to meet high-precision requirements.
A calibration method based on margin feedback neural network is adopted. The output results of analog-to-digital converter are calibrated in parallel through a pre-trained calibration network model. By utilizing forward propagation and backward propagation mechanisms, parallel calibration of nonlinear error, gain error and capacitance mismatch error is achieved.
It shortens calibration time, improves calibration accuracy and stability, avoids the accuracy loss in the traditional step-by-step calibration mode, and achieves efficient and high-precision calibration results.
Smart Images
Figure CN120915294A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of ADC calibration, and particularly relates to a high-speed analog-to-digital converter calibration method and circuit based on a residual feedback neural network. BACKGROUND
[0002] In the field of modern electronic technology, mixed architecture pipeline successive approximation type analog-to-digital converters (Pipeline-SAR ADCs), pipeline analog-to-digital converters (Pipeline ADCs) and successive approximation type analog-to-digital converters (SAR ADCs) are widely used due to their unique performance advantages. Pipeline ADCs can realize high-speed and high-precision analog-to-digital conversion by virtue of a multi-stage pipeline structure, and thus occupy an important position in fields such as communication and multimedia that have relatively high requirements on conversion rate; SAR ADCs are favored in low-power application scenarios such as portable electronic devices and sensor interfaces due to their low power consumption, high resolution and simple circuit structure; Pipeline-SAR ADCs combine the advantages of the former two, and have the characteristics of high speed and high precision, and thus can meet complex and diverse application requirements.
[0003] However, these analog-to-digital converters inevitably face problems such as non-linear error, gain error and capacitor mismatch in actual operation. Non-linear error is caused by the non-linear characteristics of internal circuit elements of the ADC, such as the non-linear transmission characteristics of the sample-and-hold circuit, which can cause the conversion result to deviate from the ideal linear relationship, produce harmonic distortion, and reduce the signal-to-noise ratio and effective number of bits of the signal; gain error is usually caused by factors such as inaccurate amplifier gain and inter-stage gain mismatch, which makes the input-output gain of the ADC deviate from the design value, and affects the conversion precision; capacitor mismatch mainly occurs in ADCs based on capacitor arrays, and due to the limitation of manufacturing process, the actual value of the capacitor unit deviates from the ideal value, which can introduce static error and destroy the monotonicity and linearity of the ADC. The existence of these non-ideal factors, like hidden "reefs" in the system, seriously limits the running speed and conversion precision of the analog-to-digital converter, making it difficult to fully exert its performance potential. Therefore, taking effective calibration measures to eliminate the influence of these non-ideal factors has become a key link to improve the performance of the analog-to-digital converter.
[0004] Traditional nonlinear calibration methods usually need to construct an error function first, and then adjust the harmonic coefficients of high-order terms through continuous iteration to minimize the error function, so as to compensate for nonlinear errors. However, this method has high computational complexity and long calibration time. For gain error and capacitance mismatch calibration, existing methods mainly fall into two categories: one is to correct the error by directly adjusting the weight parameters, but this method relies on accurate models and parameter settings, and has poor robustness; the other is to inject a disturbance signal into the system, and then complete the calibration according to the correlation between the disturbance signal and the error. With the development of neural network technology, some error networks that can simultaneously calibrate multiple errors have emerged. These networks mostly use direct input-output technology to achieve calibration function. However, due to the lack of effective response mechanism to environmental changes, temperature fluctuations and process angle differences, the calibration stability is easily affected by external condition changes, and in complex and variable actual application scenarios, it is difficult to ensure the reliability and consistency of the calibration results. In recent years, although the error calibration network based on neural network can handle multiple errors at the same time, it lacks an effective response mechanism to environmental changes, temperature fluctuations and process angle differences. In actual application scenarios, slight changes in environmental temperature and manufacturing process will cause a significant decrease in calibration stability, which cannot guarantee the reliability and consistency of the calibration results, greatly limiting its application range.
[0005] Based on the existing error calibration method, the nonlinear error calibration, gain error calibration and capacitance mismatch calibration methods generally adopt a modular independent calibration mode. In this mode, to complete the calibration of the above three errors, it is necessary to proceed step by step in sequence, and only after completing the calibration of each module can the next step be entered. In this way, the entire calibration process is time-consuming and long, and the calibration speed is greatly limited; and in the process of multiple step-by-step calibration, due to the influence of various uncontrollable factors, the calibration accuracy is easily reduced, and it is difficult to meet the high-precision calibration demand. SUMMARY
[0006] In order to solve the above problems existing in the prior art, the present application provides a high-speed analog-to-digital converter calibration method based on residual feedback neural network. The technical problem to be solved by the present application is solved by the following technical scheme: In a first aspect, the present application provides a high-speed analog-to-digital converter calibration method based on residual feedback neural network, the method comprising: parallel calibrating nonlinear errors, gain errors and capacitance mismatch errors in the conversion results output by the analog-to-digital converter using a pre-trained calibration network model, to obtain a network output; processing the network output and the conversion results output by the analog-to-digital converter using an adder to obtain a calibrated analog-to-digital conversion result; wherein, The pre-trained calibration network model is obtained through iterative training of a preset number of rounds based on forward propagation, residual feedback mechanism and back propagation.
[0007] In an embodiment of the present application, the calibration network model comprises: an input layer, a plurality of one-dimensional convolution layers, a flattening module, a first full connection layer and a second full connection layer connected in sequence, wherein the input layer comprises a reshaping module, an activation function is arranged after each one-dimensional convolution layer in the plurality of one-dimensional convolution layers, and a batch normalization layer is arranged after the first two one-dimensional convolution layers.
[0008] In an embodiment of the present application, the training process of the calibration network model comprises: S01, initializing the hyperparameters of the calibration network model; S02, performing forward propagation on the conversion result output by the analog-to-digital converter to obtain an error result, and performing fitting processing on the conversion result output by the analog-to-digital converter to obtain a label; S03, calculating the mean square error loss according to the error result and the label; S04, updating the hyperparameters of the calibration network model through back propagation according to the mean square error loss; S04, iteratively performing S02-S04 for a plurality of times until the number of iterations reaches a preset training round, and obtaining the trained calibration network model.
[0009] In an embodiment of the present application, the hyperparameters of the calibration network model comprise: learning rate, training round and batch size.
[0010] In an embodiment of the present application, the forward propagation on the conversion result output by the analog-to-digital converter to obtain an error result comprises: preprocessing the conversion result output by the analog-to-digital converter by using the reshaping module in the input layer to obtain a preprocessing result; processing the preprocessing result through the one-dimensional convolution layer, the activation function, the one-dimensional convolution layer and the batch normalization layer in sequence to obtain normalized data; processing the normalized data through the activation function, the one-dimensional convolution layer, the activation function, the one-dimensional convolution layer, the activation function and the flattening module in sequence to obtain feature data; processing the feature data through the first full connection layer and the second full connection layer in sequence to obtain an error result.
[0011] In an embodiment of the present application, the fitting processing on the conversion result output by the analog-to-digital converter to obtain a label comprises: The conversion result output by the analog-to-digital converter is differentiated to obtain the differential result; The analog-to-digital converter is weighted using a weighted search method to complete the calibration of capacitor mismatch and gain bias. The differential result is used to compensate for the nonlinear error of the analog-to-digital converter, resulting in a label without deviation.
[0012] In one embodiment of the present invention, the expression for the mean squared error loss is as follows: ; in, This represents the mean squared error loss. This indicates the number of bits in the output of the digital-to-digital converter. Indicates the first in the conversion result The error result corresponding to the bit Indicates the first in the conversion result The corresponding label.
[0013] Secondly, the present invention provides a high-speed analog-to-digital converter calibration circuit based on a margin feedback neural network, comprising: The system includes an input data buffer, an RFNN controller, a weight RAM module, an RFNN based on margin feedback, and a data output module; among these, The input data buffer is used to store the conversion result output from the analog-to-digital converter, and waits for the weights to be written and the RDY signal to be effectively enabled; The RFNN controller is used to control the reading of the conversion result output by the analog-to-digital converter and the control of the overall enable and reset signals. It is also used to control the reading of the completed neural network weights and the writing of the trained neural network weights. The weight RAM module is used to read and write neural network weights under the control of the RFNN controller. The margin feedback-based neural network RFNN is used to perform parallel calibration of nonlinearity error, gain error and capacitance mismatch error in the conversion result of the analog-to-digital converter output, and obtain the network output. The data output module is used to process the conversion results of the network output and the analog-to-digital converter output to obtain the calibrated analog-to-digital conversion result.
[0014] The beneficial effects of this invention are: The solution provided by this invention utilizes a pre-trained calibration network model to perform parallel calibration of the nonlinearity error, gain error, and capacitance mismatch error in the conversion result output by the analog-to-digital converter. Compared with the traditional distributed calibration mode, this greatly shortens the calibration time and effectively avoids the accuracy loss caused by distributed calibration, thereby ensuring high accuracy of the calibration results. Compared with the traditional neural network calibration method, this invention can ensure high calibration accuracy in different scenarios by adding a margin feedback mechanism. Attached Figure Description
[0015] Figure 1 This is a schematic diagram illustrating the steps of a high-speed analog-to-digital converter calibration method based on a margin feedback neural network provided in an embodiment of the present invention. Figure 2 This is a schematic diagram illustrating the training and calibration processes in a high-speed analog-to-digital converter calibration method based on a margin feedback neural network, as provided in an embodiment of the present invention. Figure 3 This is an architecture diagram of the calibration network model in a high-speed analog-to-digital converter calibration method based on a margin feedback neural network provided in an embodiment of the present invention. Figure 4 This is a flowchart illustrating a high-speed analog-to-digital converter calibration method based on a margin feedback neural network, as provided in an embodiment of the present invention. Figure 5 This is a schematic diagram of a high-speed analog-to-digital converter calibration circuit based on a margin feedback neural network, provided in an embodiment of the present invention. Detailed Implementation
[0016] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.
[0017] This invention provides a high-speed analog-to-digital converter calibration method and circuit based on a margin feedback neural network.
[0018] Below, we will first introduce a high-speed analog-to-digital converter calibration method based on a margin feedback neural network provided in the embodiments of the present invention.
[0019] like Figure 1 As shown in the embodiment of the present invention, a high-speed analog-to-digital converter calibration method based on a margin feedback neural network may include the following steps: S1, the conversion result of the analog-to-digital converter is forward propagated using a pre-trained calibration network model to obtain the network output; S2, using an adder, processes the conversion results from the network output and the analog-to-digital converter output to obtain the calibrated analog-to-digital conversion result; where, The pre-trained calibration network model is obtained through preset iteration training based on forward propagation, residual feedback mechanism and back propagation.
[0020] Specifically, the training and calibration schematic diagram in the high-speed analog-to-digital converter calibration method based on the residual feedback neural network is as shown in Figure 2 It can be seen that the trained network is obtained through the foreground training method first, and then one forward propagation is performed to complete the calibration. After offline deployment, a lightweight network is constructed to save hardware consumption. The input signal is converted into a digital signal through ADC, and is directly transmitted in one way and sent to the trained calibration network model in the other way to obtain the network output. After addition is completed by the adder, the analog-to-digital conversion result after calibration is finally output, so as to realize the calibration processing of the ADC signal, and the overall system framework of utilizing network training to combine the foreground and background calibration ADC output signal is constituted.
[0021] The calibration network model can include: Figure 3 The input layer, a plurality of one-dimensional convolution layers, a flattening module, a first full connection layer and a second full connection layer connected in sequence; wherein the input layer includes a reshaping module, an activation function is arranged after each one-dimensional convolution layer in the plurality of one-dimensional convolution layers, and a batch normalization layer is arranged after the first two one-dimensional convolution layers.
[0022] The training process of the calibration network model can include: Figure 4 S01, initializing the hyperparameters of the calibration network model; S02, based on the hyperparameters of the calibration network model, performing forward propagation on the conversion result output by the analog-to-digital converter to obtain an error result ; performing fitting processing on the conversion result output by the analog-to-digital converter to obtain a label ; S03, calculating the mean square error loss according to the error result and the label; S04, updating the hyperparameters of the calibration network model through back propagation according to the mean square error loss; S04, iteratively executing S02-S04 for a plurality of times until the preset training round is reached, and obtaining the trained calibration network model.
[0023] In the implementation process, the data collected by the ADC is normalized to map the data to a specific interval (such as [0, 1]), unify the data scale, and improve the stability and convergence speed of model training. The specific single training process is as follows: the preprocessed training data is input into the calibration network model, and the calibration network model is based on the current network structure and parameter setting. The calculation unit composed of neurons in each layer extracts and calculates the features of the input data layer by layer, and then sequentially passes through Figure 3 The method shown in the figure is sequentially transmitted, and finally the network output is obtained. Subsequently, using a preset loss function, the error value between the model calibration result and the label reflects the deviation between the current model prediction and the true situation. Based on the backpropagation algorithm, the error starts from the model output layer and is transmitted in reverse along the network connection. According to the chain rule of derivation, the contribution of each network parameter to the error, i.e. the gradient value, is calculated. The gradient value indicates the direction and amplitude of the model parameter adjustment to reduce the error. Finally, using an optimization algorithm (such as the stochastic gradient descent algorithm or the Adam algorithm), the model parameters (including weights and biases) are updated according to the calculated gradient, so that the model is optimized in the direction of reducing the loss function value and improving the calibration accuracy. Repeat the above single training process, and each time an iteration is completed, the fitting ability of the model to the training data is enhanced, and the prediction accuracy is gradually improved. Continue to iterate the above steps until the number of iterations reaches the preset training rounds, at which time the trained calibration network model is obtained, and its parameters have been optimized and adjusted several times, which can achieve high-precision calibration task for ADC.
[0024] The hyperparameters of the calibration network model can include: learning rate, training rounds, and batch size.
[0025] For S02, the forward propagation of the conversion result output by the analog-to-digital converter to obtain the error result can include: using the reshaping module in the input layer to preprocess the conversion result output by the analog-to-digital converter to obtain a preprocessing result; processing the preprocessing result through a one-dimensional convolution layer, an activation function, a one-dimensional convolution layer, and a batch normalization layer in turn to obtain normalized data; processing the normalized data through an activation function, a one-dimensional convolution layer, an activation function, a one-dimensional convolution layer, an activation function, and a flattening module in turn to obtain feature data; processing the feature data through a first full connection layer and a second full connection layer in turn to obtain an error result.
[0026] In the preprocessing process, the conversion result output by the analog-to-digital converter can be input data with variable dimensions, which is reshaped by a reshaping module to an initial form suitable for subsequent one-dimensional convolution processing, so as to prepare for subsequent convolution operation.
[0027] In the one-dimensional convolution and activation operation, first, one-dimensional convolution is performed, using a one-dimensional convolution kernel with 3, a step of 1, and padding of 1, combining the input channel and output channel dimensions, and then passing through an activation function to introduce the nonlinearity of the network, and then performing batch normalization operation to adjust the data distribution and improve the stability and efficiency of training. Subsequent one-dimensional convolution operations are performed multiple times, and the process is repeated. After two convolutions, normalization operation is performed, and an activation function is connected after each convolution. Nonlinear operation is introduced after each convolution to gradually extract data features and expand feature dimensions and expression capacity.
[0028] In the flattening and full connection processing, after multiple convolutions, the feature data of the network can be input to the last full connection layer through flattening operation. The dimension is changed through the first full connection layer, and the final output result is output through the second full connection layer. The full connection layer is used to realize the mapping from the convolution feature extraction to the final output, and complete the forward propagation calculation process of the entire network.
[0029] In S02, the conversion result output by the analog-to-digital converter is fitted to obtain a label, which can include: The conversion result output by the analog-to-digital converter is differentiated to obtain a differential result; The weight search method is used to calibrate the weight of the analog-to-digital converter to complete the capacitance mismatch and gain deviation calibration; The differential result is used to compensate the nonlinearity error of the analog-to-digital converter to obtain a final label without deviation.
[0030] The expression of the mean square error loss calculated according to the error result and the label is as follows: ; Wherein, represents the mean square error loss, represents the number of bits of the conversion result output by the analog-to-digital converter, represents the error result corresponding to the i-th bit in the conversion result, represents the label corresponding to the i-th bit in the conversion result. It can be understood that the flowchart of the high-speed analog-to-digital converter calibration method based on the residual feedback neural network proposed in the embodiment of the application is as follows:
[0031] It can be understood that the flowchart of the high-speed analog-to-digital converter calibration method based on the residual feedback neural network proposed in the embodiment of the application is as follows: Figure 4 As shown, it can be seen that the specific calibration method can be divided into two parts of training and calibration, after the training starts, the hyperparameters of the network need to be initialized first: learning rate, number of rounds and batch size, combined with the conversion results output by the analog-to-digital converter, and the error results are generated in parallel, after the label is obtained through the fitting method, the mean square error loss is calculated, the network parameters are updated through back propagation, if the number of iterations is less than the set cycle, continue to execute the above operation, until the number of rounds meets the requirement, end the training, and the process is optimized through multiple iterations to optimize the model calibration capability. The calibration directly starts from the forward propagation of the neural network, directly calls the calibrated network model, processes the conversion results output by the analog-to-digital converter, and quickly completes the calibration of the conversion results output by the analog-to-digital converter.
[0032] The embodiment of the application innovatively breaks the shackles of the traditional calibration method. Through the unique technical scheme, the traditional step-by-step calibration mode is abandoned, and parallel calibration of multiple errors is realized. This calibration method not only greatly shortens the calibration time and significantly improves the calibration speed, but also effectively avoids the precision loss caused by step-by-step calibration, thereby ensuring the high precision of the calibration result; and completely solves the speed bottleneck problem caused by serial calibration, thereby bringing new breakthroughs and development directions to the calibration technology field.
[0033] In a second aspect, corresponding to the above method embodiment, the embodiment of the application also provides a high-speed analog-to-digital converter calibration circuit based on a residual feedback neural network, as shown in the figure. Figure 5 As shown, the circuit can include: an input data buffer, an RFNN controller, a weight RAM module, a neural network RFNN based on residual feedback, and a data output module; wherein, The input data buffer is used to store the conversion results output from the analog-to-digital converter, and wait for the valid enable of the weight write and RDY signal; The RFNN controller is used to control the reading of the conversion results output by the analog-to-digital converter and the control of the overall enable and reset signal, and is also used to control the reading of the neural network weight and the writing of the trained neural network weight; The weight RAM module is used to read and write the neural network weight under the control of the RFNN controller; The neural network RFNN based on residual feedback is used to perform parallel calibration on the nonlinear error, gain error and capacitance mismatch error in the conversion results output by the analog-to-digital converter to obtain the network output; The data output module is used to process the network output and the conversion results output by the analog-to-digital converter to obtain the calibrated analog-to-digital conversion results.
[0034] It can be understood that the high-speed analog-to-digital converter calibration circuit can include a plurality of key modules to control appropriate data interaction. All the modules are in a way of synchronous clock and asynchronous reset to ensure stable and efficient operation of calibration.
[0035] Specifically, in the neural network RFNN based on the residual feedback, the receptive field control is used for controlling the number of nodes fed to the RFNN. The processing module is used for performing addition and multiplication operations. The ReLU and BN layers are the operations of the activation function and batch normalization. The specific working process or working principle of each module can be referred to the method embodiments of the first aspect, and will not be described here.
[0036] The embodiment of the present application utilizes the pre-trained calibration network model to perform parallel calibration on the nonlinear error, gain error and capacitance mismatch error in the conversion result output by the analog-to-digital converter. Compared with the traditional distributed calibration mode, the calibration time is greatly shortened, and the precision loss caused by the distribution is effectively avoided, so as to ensure the high precision of the calibration result. Compared with the traditional neural network calibration mode, the present application can ensure high calibration precision in different scenarios by adding the residual feedback mechanism.
[0037] It should be noted that in the description of the present application, it should be understood that the terms "first", "second" are only for the purpose of description, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features limited by "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise specifically limited.
[0038] Each embodiment in the specification is described in a related manner, and the same and similar parts between each embodiment can be referred to each other, and each embodiment mainly explains the difference from other embodiments. Especially, for the circuit embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the related parts can refer to the part of the method embodiment.
[0039] The above only describes the preferred embodiments of the present application, and is not used to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.
Claims
1. A method for calibrating a high-speed analog-to-digital converter based on a residual feedback neural network, characterized in that, The application relates to an analog-to-digital converter (ADC) calibration method and device. The method comprises the following steps: The pre-trained calibration network model is based on a forward propagation, a residual feedback mechanism and a back propagation, and is obtained through iterative training of a preset number of rounds. The calibration network model comprises:
2. The calibration method of high-speed analog-to-digital converter based on residual feedback neural network according to claim 1, characterized in that, The input layer comprises a reshaping module, and an activation function is arranged after each one-dimensional convolution layer, and a batch normalization layer is arranged after the first two one-dimensional convolution layers. The training process of the calibration network model comprises the following steps:
3. The method of claim 2, wherein, S01, initializing the hyperparameters of the calibration network model; S02, performing forward propagation on the conversion result of the analog-to-digital converter to obtain an error result, and performing fitting processing on the conversion result of the analog-to-digital converter to obtain a label; S03, calculating the mean square error loss according to the error result and the label; S04, updating the hyperparameters of the calibration network model through back propagation according to the mean square error loss; S04, iteratively executing S02-S04 for a plurality of times until the number of iterations reaches a preset training round, and obtaining the trained calibration network model. The hyperparameters of the calibration network model comprise:
4. The method of claim 3, wherein, Learning rate, training round and batch size. The method comprises the following steps:
5. The method of claim 3, wherein, The pre-processing result is obtained by using the reshaping module in the input layer to pre-process the conversion result of the analog-to-digital converter; The normalized data is obtained by sequentially processing the pre-processing result through the one-dimensional convolution layer, the activation function, the one-dimensional convolution layer and the batch normalization layer; The feature data is obtained by sequentially processing the normalized data through the activation function, the one-dimensional convolution layer, the activation function, the one-dimensional convolution layer, the activation function and the flattening module; The error result is obtained by sequentially processing the feature data through the first full connection layer and the second full connection layer. The method comprises the following steps:
6. The method of claim 3, wherein the method further comprises: The differential result is obtained by performing differential processing on the conversion result of the analog-to-digital converter; The weight calibration of the analog-to-digital converter is completed by using a weight search method to calibrate the capacitance mismatch and the gain deviation; The final label without deviation is obtained by using the differential result to compensate the nonlinear error of the analog-to-digital converter. The expression of the mean square error loss is as follows:
7. The method of claim 3, wherein the method further comprises: The application relates to an analog-to-digital converter (ADC) calibration method and device. ； wherein, represents a mean square error loss, represents the number of bits of the conversion result output by the quantizer, represents the error result corresponding to the bit of the conversion result, represents the label corresponding to the bit of the conversion result.
8. A residue feedback neural network based high speed analog-to-digital converter calibration circuit, comprising: The method comprises the following steps: The input data buffer is used for storing the conversion result output from the analog-to-digital converter, and waiting for the effective enablement of the weight write and the RDY signal; The RFNN controller is configured to control reading of conversion results output by the analog-to-digital converter and control of overall enable and reset signals, and is further configured to control reading of neural network weights and writing of trained neural network weights. The weight RAM module is configured to read and write the neural network weights under control of the RFNN controller. The neural network RFNN based on residual feedback is configured to perform parallel calibration on nonlinear error, gain error and capacitance mismatch error in the conversion results output by the analog-to-digital converter to obtain network output. The data output module is configured to process the network output and the conversion results output by the analog-to-digital converter to obtain calibrated analog-to-digital conversion results.