Data calibration method and device for successive approximation analog-to-digital converter
Data-driven calibration of successive approximation analog-to-digital converters through neural network calibration models solves the problems of design complexity and high manufacturing cost in the prior art, and achieves the performance improvement of analog-to-digital converters with high precision and low noise floor.
Patent Information
- Application Number
- CN202510289088.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-12
- Publication Date
- 2025-06-27
AI Technical Summary
The prior art is difficult to effectively calibrate successive approximation analog-to-digital converters with low design complexity and manufacturing costs, resulting in the performance indicators not reaching the ideal state.
The neural network calibration model is used to calibrate the data of the successive approximation analog-to-digital converter. By building an ideal model and an error model, an ideal data set and a training data set are generated, and a neural network with a single hidden layer fully connected network structure is trained and calibrated.
It effectively suppresses stray signal components and reduces noise floor, significantly improves the accuracy and spurious-free dynamic range of successive approximation analog-to-digital converters, meets the needs of high-precision applications, and reduces the complexity of system design and manufacturing costs.
Smart Images
Figure CN120223076A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of calibration of analog-to-digital converters, and particularly relates to a data calibration method and device for a successive approximation analog-to-digital converter. Background Art
[0002] An analog-to-digital converter (ADC) can convert analog signals in the real world into digital signals in an electronic system and is widely used in various fields of the digital information age. With the rapid development of digital processing technology and the improvement of the performance of digital processing circuits, compared with other types of ADCs, the successive approximation register (SAR) ADC with a relatively high proportion of digital circuits is more widely used and has become a research hotspot in the academic and industrial circles.
[0003] However, in practical applications, the performance of high-precision SAR ADCs is often restricted by various non-ideal factors. Among them, capacitor mismatch and comparator offset are two common non-ideal factors. Capacitor mismatch will cause errors in the conversion process of the SAR ADC, affecting the conversion accuracy; while comparator offset will cause the threshold of the comparator to shift, thereby affecting the accuracy of the SAR ADC. The existence of these non-ideal factors makes the performance indicators (such as accuracy, spurious-free dynamic range, etc.) of the SAR ADC unable to reach the ideal state, restricting its popularization and use in high-precision application scenarios.
[0004] Traditional calibration technology algorithms are relatively complex, and specific calibration algorithms need to be designed for different error factors, which not only increases the design difficulty but also raises the manufacturing cost. Summary of the Invention
[0005] Aiming at the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is: to provide a data calibration method and device that can effectively calibrate a successive approximation analog-to-digital converter on the premise of low design complexity and manufacturing cost.
[0006] To solve the above technical problem, a technical solution adopted by the present invention is: to provide a data calibration method for a successive approximation analog-to-digital converter, including the following steps:
[0007] Obtain an ideal data set and a training data set;
[0008] Construct a neural network calibration model and train the neural network calibration model based on the ideal data set and the training data set;
[0009] Use the trained neural network calibration model to calibrate the data of the successive approximation analog-to-digital converter.
[0010] Further, in the steps of obtaining the ideal data set and the training data set, the following sub-steps are included:
[0011] Construct an ideal model of the successive approximation analog-to-digital converter and generate an ideal data set based on the ideal model of the successive approximation analog-to-digital converter as the target value of the neural network calibration model;
[0012] Introduce non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establish an error model of the successive approximation analog-to-digital converter, and generate a training data set based on the error model of the successive approximation analog-to-digital converter.
[0013] Further, the method of constructing an ideal model of the successive approximation analog-to-digital converter and generating an ideal data set based on the ideal model of the successive approximation analog-to-digital converter specifically includes: constructing an ideal model of the successive approximation analog-to-digital converter, setting the sampling frequency and the analog input frequency based on the coherent sampling theorem, generating ideal successive approximation analog-to-digital converter data, and constituting the ideal data set.
[0014] Further, in the step of introducing non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establishing an error model of the successive approximation analog-to-digital converter, and generating a training data set based on the error model of the successive approximation analog-to-digital converter, the following sub-steps are included:
[0015] Introduce non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establish an error model of the successive approximation analog-to-digital converter, and the non-ideal factors include capacitor mismatch and / or comparator offset;
[0016] Set the same sampling frequency and analog input frequency as when obtaining the ideal data set, generate error data with non-ideal factors, and constitute the training data set.
[0017] Further, the neural network calibration model adopts a single hidden layer fully connected network structure, including:
[0018] An input layer for receiving external data and transferring the received data to the hidden layer;
[0019] A hidden layer for receiving the data transferred from the input layer, calculating the weighted sum of the data and the weights, and performing a non-linear transformation on the weighted sum data using the Sigmoid activation function;
[0020] An output layer for outputting prediction data.
[0021] Further, the weighted sum is obtained by the following formula:
[0022]
[0023] In formula (1), n represents the number of neurons in the input layer, i represents the i-th neuron in the input layer, j represents the j-th neuron in the hidden layer, and y j represents the weighted sum of the j-th neuron in the hidden layer, and x ij represents the numerical information transmitted from the i-th neuron in the input layer to the j-th neuron in the hidden layer, and w ij represents the synaptic weight between two neurons.
[0024] Further, in the step of training the neural network calibration model based on the ideal data set and the training data set, the following sub-steps are included:
[0025] Input the training data set into the neural network calibration model, perform forward propagation, and obtain the forward network result after weighted calculation by neurons;
[0026] Use the ideal data set as the target value of the neural network calibration model, calculate the mean square error between the forward network result and the target value, and obtain the error value;
[0027] Backpropagate the error value, and use the stochastic gradient descent algorithm to update the network weights;
[0028] Iteratively train until the error value reaches an acceptable range or the set number of iterative training times is reached.
[0029] Further, in the step of introducing non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establishing an error model of the successive approximation analog-to-digital converter, and generating a training data set based on the error model of the successive approximation analog-to-digital converter, a test data set is also generated while generating the training data set;
[0030] After the steps of constructing the neural network calibration model and training the neural network calibration model based on the ideal data set and the training data set, the following steps are further included:
[0031] Input the test data set into the trained neural network calibration model, and obtain the calibrated data through forward propagation;
[0032] Analyze the dynamic performance indexes of the calibrated data, and clarify the improvement degree of the dynamic performance indexes of the calibrated data.
[0033] To solve the above technical problems, another technical solution adopted by the present invention is: to provide a data calibration device for a successive approximation analog-to-digital converter, including:
[0034] A data acquisition module for acquiring an ideal data set and a training data set;
[0035] A model construction and training module, configured to construct a neural network calibration model and train the neural network calibration model based on an ideal data set and a training data set;
[0036] A calibration module, configured to calibrate data of a successive approximation analog-to-digital converter by using the trained neural network calibration model.
[0037] Furthermore, the data acquisition module further includes:
[0038] A first acquisition sub-module, configured to construct an ideal model of a successive approximation analog-to-digital converter and generate an ideal data set based on the ideal model of the successive approximation analog-to-digital converter as a target value of the neural network calibration model;
[0039] A second acquisition sub-module, configured to introduce non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establish an error model of the successive approximation analog-to-digital converter, and generate a training data set based on the error model of the successive approximation analog-to-digital converter.
[0040] The data calibration method and device for a successive approximation analog-to-digital converter of the present invention have at least the following beneficial effects: data-driven calibration of non-ideal factors is performed through a neural network calibration model, effectively suppressing spurious signal components and reducing the noise floor, thereby greatly improving the accuracy and spurious-free dynamic range of the successive approximation analog-to-digital converter, meeting the requirements of high-precision applications, and providing technical support for the engineering application of high-precision successive approximation analog-to-digital converters; a single neural network model is used to correct multiple non-ideal factors simultaneously, without the need to develop specific calibration algorithms for different error sources, breaking through the limitation of the traditional method that requires multiple algorithms to cooperate, reducing the complexity of system design, shortening the development cycle, and the independent verification of the test data set further ensures the generalization performance of the model; the neural network calibration model adopts a single-hidden-layer fully connected network structure, and only a single-layer non-linear transformation is required to complete feature extraction and error compensation. Compared with deep networks, this structure achieves high-precision calibration through weight iteration optimization without using a large amount of training data; there is no need to additionally introduce complex calibration circuits or high-precision components, and process errors such as capacitor mismatch in the manufacturing process can be compensated only through digital background calibration, relaxing the accuracy requirements of the production process and reducing the manufacturing cost. Description of the Drawings
[0041] The drawings described herein are used to provide a further understanding of the present application, and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:
[0042] Figure 1 It is a flowchart of an embodiment of the data calibration method for a successive approximation analog-to-digital converter of the present invention.
[0043] Figure 2 is Figure 1 the flowchart of step S100 in
[0044] Figure 3 the spectrum analysis diagram of the ideal model output signal of a successive approximation analog-to-digital converter
[0045] Figure 4 is Figure 2 the flowchart of step S120 in
[0046] Figure 5 the model structure diagram of a neural network calibration model
[0047] Figure 6 the spectrogram of the test data before calibration
[0048] Figure 7 the spectrogram of the test data after calibration
[0049] Figure 8 the structural block diagram of an embodiment of the data calibration device for the successive approximation analog-to-digital converter of the present invention Specific Embodiments
[0050] The present invention will be further described below with reference to the accompanying drawings.
[0051] The successive approximation analog-to-digital converter (SAR ADC) has been widely used in high-precision fields such as medical electronics and industrial control due to its advantages of high digital circuit ratio and low power consumption. However, as the precision is increased to 16 bits and above, non-ideal factors such as capacitor array mismatch errors and comparator offset voltages generated in the process manufacturing will significantly deteriorate the performance indicators of the ADC. The specific solution of the present invention will be described in detail below with reference to the data calibration example of a 16-bit non-ideal SAR ADC. Of course, it can be understood that this solution is not limited to the SAR ADC with a precision of 16 bits, and is applicable to all SAR ADCs with a precision lower than 16 bits.
[0052] Please refer to Figure 1 , which is the flowchart of an embodiment of the data calibration method for the successive approximation analog-to-digital converter of the present invention. This embodiment specifically includes the following steps:
[0053] S100. Obtain an ideal data set and a training data set.
[0054] Please refer to Figure 2 , this step S100 includes the following sub-steps:
[0055] S110. Obtain an ideal data set.
[0056] Specifically, an ideal model of a successive approximation analog-to-digital converter is constructed, and an ideal data set is generated based on the ideal model of the successive approximation analog-to-digital converter as the target value of the neural network calibration model. During specific operations, an ideal model of a successive approximation analog-to-digital converter is constructed, the sampling frequency and the analog input frequency are set based on the coherent sampling theorem, and ideal successive approximation analog-to-digital converter data is generated to form an ideal data set. Please refer to Figure 3 , which is the spectral analysis diagram of the output signal of the ideal model of the successive approximation analog-to-digital converter, showing the spectrogram of the sine signal of the 16-bit ideal successive approximation analog-to-digital converter model.
[0057] S120. Obtain a training data set.
[0058] Specifically, non-ideal factors are introduced into the ideal model of the successive approximation analog-to-digital converter to establish an error model of the successive approximation analog-to-digital converter, and a training data set is generated based on the error model of the successive approximation analog-to-digital converter. Please refer to Figure 4 , and this step S120 includes the following sub-steps:
[0059] S121. Establish an error model.
[0060] Specifically, non-ideal factors are introduced into the ideal model of the successive approximation analog-to-digital converter to establish an error model of the successive approximation analog-to-digital converter, and the non-ideal factors include capacitor mismatch and / or comparator offset.
[0061] S122. Obtain error data.
[0062] Specifically, the same sampling frequency and analog input frequency as those when obtaining the ideal data set are set to generate error data with non-ideal factors to form a training data set. In this embodiment, in order to test the improvement degree of the dynamic index of the unknown error data after model calibration subsequently, a part of the generated error data will also be divided into a test data set that does not participate in the model training process.
[0063] S200. Construct and train a neural network calibration model.
[0064] Specifically, a neural network calibration model is constructed and the neural network calibration model is trained based on the ideal data set and the training data set. Through a large number of experiments, it is found that in this embodiment, the neural network calibration model adopting a single hidden layer fully connected network structure can not only achieve a high calibration accuracy without a large amount of training data, but also reduce energy consumption and improve calculation efficiency. Please refer to Figure 5, the neural network calibration model includes an input layer, a hidden layer, and an output layer. Among them, the input layer is used to receive external data and transfer the received data to the hidden layer. The hidden layer is used to receive the data transferred from the input layer and calculate the weighted sum of the data and the weights, and use the Sigmoid activation function to perform a non-linear transformation on the weighted sum data. The output layer is used to output prediction data. The number of neurons in the input layer is the same as the number of input data, and the number of neurons in the output layer matches the amount of prediction data. Each neuron outputs a prediction data. In this embodiment, the number of neurons in the input layer is the same as the amount of data of the successive approximation analog-to-digital converter with error in one sampling period, and the output layer outputs the calibrated data set.
[0065] The calculation formula of the weighted sum is as follows:
[0066]
[0067] Among them, n represents the number of neurons in the input layer, i represents the i-th neuron in the input layer, j represents the j-th neuron in the hidden layer, and y j represents the weighted sum of the j-th neuron in the hidden layer, and x ij represents the numerical information passed from the i-th neuron in the input layer to the j-th neuron in the hidden layer, and w ij represents the synaptic weight between two neurons.
[0068] When the hidden layer performs a non-linear transformation on the weighted sum data, the selected Sigmoid activation function formula is F(x) = 1 / (1 + e -x ).
[0069] A method for training a neural network calibration model based on an ideal data set and a training data set specifically includes: inputting the training data set into the neural network calibration model for forward propagation, and obtaining a forward network result after weighted calculation by neurons; using the ideal data set as the target value of the neural network calibration model, calculating the mean square error between the forward network result and the target value to obtain an error value; backpropagating the error value and using the stochastic gradient descent algorithm to update the network weights; iteratively training until the error value reaches an acceptable range or reaches the set number of iterative training times.
[0070] In order to be able to intuitively display the calibration effect of the trained neural network calibration model, as a preferred embodiment, it further includes the following steps: inputting the test data set into the trained neural network calibration model, obtaining calibrated data through forward propagation, analyzing the dynamic performance indicators of the calibrated data, and clarifying the improvement degree of the dynamic performance indicators of the calibrated data. Please refer to Figure 6 and Figure 7 , Figure 6 is the spectrogram of the test data with error,Figure 7 It is the spectrogram of the test data after the neural network calibration. By comparing the spectrograms before and after calibration, it can be seen that the noise floor is significantly reduced after calibration, and the spurious signal components are suppressed to the noise floor. The dynamic performance index of the data after calibration is significantly improved. The signal-to-noise ratio is increased by 31 dB, the spurious-free dynamic range is increased by 35 dB, and the number of effective bits is increased by 5.5 bits.
[0071] S300. Calibrate the data of the successive approximation analog-to-digital converter.
[0072] Specifically, use the trained neural network calibration model to calibrate the data of the successive approximation analog-to-digital converter.
[0073] Please refer to Figure 8 , which is the data calibration device for the successive approximation analog-to-digital converter of this embodiment, and is used to implement the data calibration method for the successive approximation analog-to-digital converter as described in the above embodiment. Specifically, the data calibration device for the successive approximation analog-to-digital converter of this embodiment includes a data acquisition module 100, a model construction and training module 200, and a calibration module 300. Among them:
[0074] The data acquisition module 100 is used to acquire an ideal data set and a training data set. The data acquisition module 100 further includes a first acquisition sub-module 110 and a second acquisition sub-module 120. The first acquisition sub-module 110 is used to construct an ideal model of the successive approximation analog-to-digital converter and generate an ideal data set based on the ideal model of the successive approximation analog-to-digital converter as the target value of the neural network calibration model. The second acquisition sub-module 120 is used to introduce non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establish an error model of the successive approximation analog-to-digital converter, and generate a training data set based on the error model of the successive approximation analog-to-digital converter. In this embodiment, in order to test the improvement degree of the dynamic index of the unknown error data after being calibrated by the neural network calibration model, the second acquisition sub-module is further used to acquire a test data set.
[0075] The model construction and training module 200 is used to construct a neural network calibration model and train the neural network calibration model based on the ideal data set and the training data set obtained by the data acquisition module 100. Through a large number of experiments, it is found that in this embodiment, the neural network calibration model adopting a single hidden layer fully connected network structure can not only achieve high calibration accuracy without a large amount of training data, but also reduce energy consumption and improve computational efficiency. The neural network calibration model adopts a single hidden layer fully connected network structure, including an input layer, a hidden layer, and an output layer. The input layer is used to receive external data and transfer the received data to the hidden layer. The hidden layer is used to receive the data transferred by the input layer and calculate the weighted sum of the data and the weights, and use the Sigmoid activation function to perform a non-linear transformation on the weighted sum data. The output layer is used to output prediction data.
[0076] The model construction and training module 200 is also used to input the training data set into the neural network calibration model for forward propagation, and obtain the forward network result after weighted calculation by neurons; then use the ideal data set as the target value of the neural network calibration model, calculate the mean square error between the forward network result and the target value to obtain an error value; then backpropagate the error value and use the stochastic gradient descent algorithm to update the network weights; iterate the training until the error value reaches an acceptable range or reaches the set number of iterative training times.
[0077] The calibration module 300 is used to calibrate the data of the successive approximation analog-to-digital converter by using the neural network calibration model trained by the model construction and training module 200.
[0078] The present invention performs data-driven calibration on non-ideal factors through a neural network calibration model, effectively suppressing spurious signal components and reducing the noise floor, thereby greatly improving the accuracy and spurious-free dynamic range of the successive approximation analog-to-digital converter, meeting the requirements of high-precision applications, and providing technical support for the engineering application of high-precision successive approximation analog-to-digital converters; using a single neural network model to correct multiple non-ideal factors simultaneously, without the need to develop specific calibration algorithms for different error sources, breaking through the limitation of the traditional method that requires multiple algorithms to cooperate, reducing the complexity of system design, shortening the development cycle, and the independent verification of the test data set further ensures the generalization performance of the model; the neural network calibration model adopts a single hidden layer fully connected network structure, and only a single-layer non-linear transformation is required to complete feature extraction and error compensation. Compared with deep networks, this structure can achieve high-precision calibration through weight iteration optimization without using a large amount of training data; there is no need to additionally introduce complex calibration circuits or high-precision components, and only digital back-end calibration can be used to compensate process errors such as capacitance mismatch during the manufacturing process, relaxing the accuracy requirements of the production process and reducing the manufacturing cost.
[0079] The above content only expresses the preferred embodiments of the present invention. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention patent shall be subject to the appended claims.
Claims
1. A method for calibrating data of a successive approximation analog-to-digital converter, characterized in that: The following steps are involved: Obtain ideal data sets and training data sets; Constructing a neural network calibration model and training the neural network calibration model based on the ideal data set and the training data set; The trained neural network calibration model is used to calibrate the data of the successive approximation analog-to-digital converter.
2. The data calibration method of the successive approximation analog-to-digital converter according to claim 1, characterized in that: The steps of obtaining an ideal data set and a training data set include the following sub-steps: Constructing an ideal model of a successive approximation analog-to-digital converter and generating an ideal data set based on the ideal model of the successive approximation analog-to-digital converter as a target value of the neural network calibration model; Non-ideal factors are introduced into the ideal model of the successive approximation analog-to-digital converter, an error model of the successive approximation analog-to-digital converter is established, and a training data set is generated based on the error model of the successive approximation analog-to-digital converter.
3. The data calibration method of the successive approximation analog-to-digital converter according to claim 2, characterized in that: The method for constructing an ideal model of a successive approximation analog-to-digital converter and generating an ideal data set based on the ideal model of the successive approximation analog-to-digital converter specifically includes: constructing an ideal model of the successive approximation analog-to-digital converter, setting the sampling frequency and the analog input frequency based on the coherent sampling theorem, generating ideal successive approximation analog-to-digital converter data, and forming an ideal data set.
4. The data calibration method of the successive approximation analog-to-digital converter according to claim 3, characterized in that: The steps of introducing non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establishing an error model of the successive approximation analog-to-digital converter, and generating a training data set based on the error model of the successive approximation analog-to-digital converter include the following sub-steps: Introducing non-ideal factors into the ideal model of the successive approximation analog-to-digital converter to establish an error model of the successive approximation analog-to-digital converter, wherein the non-ideal factors include capacitor mismatch and / or comparator offset; The sampling frequency and analog input frequency are set to be the same as those used to obtain the ideal data set, and error data with non-ideal factors are generated to form a training data set.
5. The data calibration method of the successive approximation analog-to-digital converter according to claim 1, characterized in that: The neural network calibration model adopts a single hidden layer fully connected network structure, including: The input layer is used to receive external data and pass the received data to the hidden layer; The hidden layer is used to receive the data transmitted by the input layer and calculate the weighted sum of the data and the weight, and use the Sigmoid activation function to perform nonlinear transformation on the weighted sum data; The output layer is used to output predicted data.
6. The data calibration method of the successive approximation analog-to-digital converter according to claim 5, characterized in that: The weighted sum is obtained by the following formula: In formula (1), n represents the number of neurons in the input layer, i represents the i-th input layer neuron, j represents the j-th hidden layer neuron, and y j represents the weighted sum of the jth hidden layer neurons, x ij represents the numerical information transmitted from the i-th input layer neuron to the j-th hidden layer neuron, w ij represents the synaptic weight between two neurons.
7. The data calibration method of the successive approximation analog-to-digital converter according to claim 1, characterized in that: The step of training the neural network calibration model based on the ideal data set and the training data set includes the following sub-steps: Inputting the training data set into the neural network calibration model, performing feedforward transmission, and obtaining a feedforward network result after neuron weighted calculation; Taking the ideal data set as the target value of the neural network calibration model, calculating the mean square error between the feedforward network result and the target value to obtain an error value; Back propagating the error value, and updating the network weights using a stochastic gradient descent algorithm; Iterate the training until the error value reaches an acceptable range or the set number of iterative training is reached.
8. The data calibration method of the successive approximation analog-to-digital converter according to claim 7, characterized in that: In the step of introducing non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establishing an error model of the successive approximation analog-to-digital converter, and generating a training data set based on the error model of the successive approximation analog-to-digital converter, a test data set is also generated while generating the training data set; After the steps of constructing a neural network calibration model and training the neural network calibration model based on an ideal data set and a training data set, the following steps are also included: Inputting the test data set into the trained neural network calibration model to obtain calibrated data through feedforward transfer; Analyze the dynamic performance indicators of the calibrated data and clarify the degree of improvement of the dynamic performance indicators of the calibrated data.
9. A data calibration device for a successive approximation analog-to-digital converter, characterized in that: include: A data acquisition module is used to obtain an ideal data set and a training data set; A model building and training module, used to build a neural network calibration model and train the neural network calibration model based on an ideal data set and a training data set; The calibration module is used to calibrate the data of the successive approximation analog-to-digital converter using the trained neural network calibration model.
10. The data calibration device of the successive approximation analog-to-digital converter according to claim 9, characterized in that: The data acquisition module also includes: A first acquisition submodule is used to construct an ideal model of a successive approximation analog-to-digital converter and generate an ideal data set based on the ideal model of the successive approximation analog-to-digital converter as a target value of the neural network calibration model; The second acquisition submodule is used to introduce non-ideal factors into the ideal model of the successive approximation analog-to-digital converter, establish an error model of the successive approximation analog-to-digital converter, and generate a training data set based on the error model of the successive approximation analog-to-digital converter.