A TIADC static-dynamic error joint calibration method and system

Abstract

The present invention relates to the field of ADC calibration technology, and specifically discloses a TIADC static-dynamic error joint calibration method and system. First, an ideal error-free TIADC signal is generated, and the dynamic error and static error of the simulated TIADC are generated to generate an error-containing TIADC signal to construct a data set. Then, a TIADC joint error calibration model is constructed based on a neural network. The TIADC joint error calibration model mainly includes an error extraction network and an error fitting network, which are respectively used to extract static-dynamic error features and fit the error features, effectively learn the local features and long-term statistical characteristics of the high-speed ADC output sequence, and realize the joint calibration of dynamic and static errors through back propagation training. Compared with the traditional calibration method and the existing ML-based calibration method, the present invention expands the error range of TIADC calibration and has better calibration performance.

Description

Technical Field

[0001] The present invention relates to the technical field of ADC calibration, and in particular to a TIADC static-dynamic error joint calibration method and system. Background Art

[0002] An analog-to-digital converter, or A / D converter, or ADC for short, usually refers to an electronic component that converts an analog signal into a digital signal. As the instantaneous bandwidth of systems such as nuclear electronics, ultra-wideband radar, and electromagnetic spectrum reconnaissance expands from tens of megahertz to gigahertz levels, the demand for high-speed, high-precision ADCs has become more urgent. However, under existing process conditions, such ADCs often face serious "speed-precision" constraints. Ideally, a time-interleaved ADC (TIADC) can achieve the effect of doubling the sampling rate and maintaining the same sampling accuracy by using multiple sub-ADCs with lower sampling rates in a time-alternating manner.

[0003] However, in practice, due to factors such as manufacturing discreteness, the performance of TIADC is often affected by a variety of static and dynamic errors. Among them, the static error is independent of the input signal frequency, and mainly includes: gain mismatch error and offset mismatch error. The dynamic error increases with the increase of the input signal frequency, and mainly includes: time mismatch error and clock jitter error. Under the combined effect of these static and dynamic errors, the TIADC output signal will produce false components, resulting in a decrease in its effective number of bits, a decrease in dynamic range, and a decrease in overall performance. Therefore, in response to the above errors, exploring innovative ADC structures and new calibration technologies has become a key issue that needs to be overcome in the field of TIADC design.

[0004] Digital calibration technology is one of the most direct and efficient methods to improve TIADC performance, and has received more widespread attention under the development trend of advanced processes. Traditional digital calibration methods mainly rely on accurate modeling of TIADC errors, and cannot take into account some unknown or unconsidered error factors. In addition, the processes based on different error classification estimates are relatively independent, and calibration is rarely performed based on the comprehensive impact of mixed mismatch, so only partial errors can be calibrated.

[0005] In recent years, inspired by the success of artificial intelligence technology in the computer field, the integrated circuit field, such as TIADC calibration research, has also begun to try to use machine learning (ML), especially deep learning (DL) to improve the performance of TIADC. At present, ML-based research mainly eliminates non-ideal components in the TIADC output signal by constructing a fully connected neural network and fitting the nonlinear relationship between the actual TIADC output and the ideal reference signal. However, although these ML-based methods have made progress in calibrating multiple errors, they focus on exploring the nonlinear fitting ability of neural networks and regard TIADC error calibration as a simple linear regression problem, but ignore its powerful feature extraction performance. Therefore, these methods can only fit some static errors under limited model volume, and cannot be applied to errors with dynamic periodic trends and complex fluctuations, such as clock jitter.

[0006] Therefore, currently, neither the traditional calibration method nor the ML-based TIADC calibration method can cover more comprehensive static and dynamic errors, resulting in the performance of TIADC being limited, especially in broadband applications. Summary of the invention

[0007] The present invention provides a TIADC static-dynamic error joint calibration method and system, and solves the technical problem of how to achieve joint calibration of the static and dynamic errors of the TIADC.

[0008] The present invention provides a TIADC static-dynamic error joint calibration method, comprising the steps of:

[0009] S1, constructing an ideal TIADC model and a non-ideal TIADC model, wherein the non-ideal TIADC model simulates various static errors and dynamic errors of the TIADC, and the ideal TIADC model has no errors;

[0010] S2, using the ideal TIADC model to generate an ideal TIADC signal, and using the non-ideal TIADC model to generate an error-containing TIADC signal; constructing a data set after preprocessing the ideal TIADC signal and the error-containing TIADC signal; dividing the data set into a training set and a test set;

[0011] S3, constructing a TIADC joint error calibration model based on neural network for one-time calibration of static and dynamic errors;

[0012] The TIADC joint error calibration model includes an error extraction network, an error fitting network and a signal output module. The error extraction network extracts static error features related to local details from input samples and obtains dynamic error features related to signal frequency through long sample sequences, and inputs the obtained static error features and dynamic error features into the error fitting network; the error fitting network fits the corresponding statistical error according to the static error features and dynamic error features; the signal output module is used to subtract the input sample from the statistical error to obtain a calibrated output signal;

[0013] S4, using the training set to train the TIADC joint error calibration model, wherein minimizing the difference between the output signal and the ideal TIADC signal is a training goal during the training process;

[0014] S5. Use the test set to test the trained TIADC joint error calibration model to determine whether the TIADC joint error calibration model meets the requirements. If so, perform TIADC calibration with the current TIADC joint error calibration model. If not, optimize the TIADC joint error calibration model and retrain and test it.

[0015] Furthermore, in step S1, after adding static and dynamic errors, the sampling signal of the kth channel is expressed as:

[0016] ,

[0017] Among them, V in is the input signal, represents the gain mismatch error of the kth channel, represents the fixed time mismatch error of the kth channel, M represents the number of channels, represents the clock jitter in the kth channel, represents the offset error of the kth channel, n represents the nth cycle, T s represents the sampling period, Represents the output signal of the kth channel in the nth sampling period.

[0018] Furthermore, the error extraction network is composed of multiple cascaded residual networks; each of the residual networks is sequentially provided with a first dilated convolution layer, a first regularization layer, a first ReLU non-linear activation function, a second dilated convolution layer, a second regularization layer, a second ReLU non-linear activation function, and a 1×1 convolution layer; wherein the first dilated convolution layer and the second dilated convolution layer extract local and sequence features along the TIADC output sequence; the 1×1 convolution layer is used to add the input of the residual network and the output of the second ReLU non-linear activation function and input them into the residual network of the next layer.

[0019] Furthermore, the kernel of each residual network is set to 8, and 16 channels are set; the receptive field and kernel length of the first layer residual network are both 8, and the expansion factor is 0; the receptive field of the second layer residual network is twice 8, that is, 16, and the expansion factor is 2; as the residual network is further stacked, the expansion factor and receptive field of the residual network increase exponentially by 2 times.

[0020] Furthermore, the error extraction network is provided with a 7-layer residual network; the acceptance field of the 7th layer residual network is 512 time intervals, and the expansion factor is 32.

[0021] Furthermore, the error fitting network adopts a fully connected neural network, including an input layer, a hidden layer, and an output layer, and a Relu nonlinear activation function is added between the input layer and the hidden layer.

[0022] Furthermore, the input layer is composed of 16 neuron nodes, corresponding to the four channels of TIADC; the hidden layer is composed of 64 neuron nodes; and the output layer is composed of 1 neuron node.

[0023] Furthermore, in step S4, the training process of the TIADC joint error calibration model includes:

[0024] The training data is input into the error extraction network to obtain the mixed feature E feature ;

[0025] The error fitting network is used to fit the mixed feature E feature Perform nonlinear transformation and use activation function to activate and get the corresponding statistical error E fit ;

[0026] The statistical error E fit Subtract the training data to get the calibrated output signal D cali ;

[0027] Calculate the output signal D cali With the real data D gt The loss function between

[0028] Use the back-propagation algorithm to calculate the gradient of the loss function to the model parameters, and update the parameters using the gradient descent algorithm;

[0029] The above training process is repeated until the set number of training rounds is reached or the loss function converges, and finally a trained TIADC joint error calibration model is obtained.

[0030] Furthermore, the gradient descent algorithm adopts the Levenberg-Marquardt algorithm.

[0031] The present invention also provides a TIADC static-dynamic error joint calibration system, the key of which is: a model generation module and a model application module are provided, the model generation model is used for the TIADC static-dynamic error joint calibration method to obtain a trained TIADC joint error calibration model that meets the requirements; the model application module is used to apply the trained TIADC joint error calibration model that meets the requirements in the TIADC calibration task.

[0032] The present invention provides a TIADC static-dynamic error joint calibration method and system, which first generates an ideal error-free TIADC signal and simulates the dynamic error and static error of the TIADC to generate an error-containing TIADC signal to construct a data set, and then constructs a TIADC joint error calibration model based on a neural network. The TIADC joint error calibration model mainly includes an error extraction network and an error fitting network, which are respectively used to extract static-dynamic error features and fit the error features, effectively learn the local features and long-term statistical characteristics of the high-speed ADC output sequence, and realize the joint calibration of dynamic and static errors through back propagation training. Compared with traditional calibration methods and existing ML-based calibration methods, the present invention expands the error range of TIADC calibration and has better calibration performance; the present invention has a larger scope of application and has a good calibration effect under full-band signal input. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 is a schematic diagram of a TIADC static-dynamic error joint calibration method provided by an embodiment of the present invention;

[0034] Figure 2 is a flow chart of training a TIADC joint error calibration model provided by an embodiment of the present invention;

[0035] Figure 3 is a structural diagram of an error extraction network provided by an embodiment of the present invention;

[0036] Figure 4 is a signal FFT spectrum diagram before and after TIADC calibration provided by an embodiment of the present invention;

[0037] Figure 5 It is a comparison diagram of SNDR before and after TIADC calibration under full-band input provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0038] The following specifically illustrates the implementation mode of the present invention in conjunction with the accompanying drawings. The embodiments are provided for illustrative purposes only and cannot be understood as limiting the present invention. The accompanying drawings are provided for reference and illustration only and do not constitute a limitation on the scope of patent protection of the present invention, because many changes can be made to the present invention without departing from the spirit and scope of the present invention.

[0039] The embodiment of the present invention first provides a TIADC static-dynamic error joint calibration method, such as Figure 1 As shown in the schematic diagram, the method includes the steps of:

[0040] S1. Construct an ideal TIADC model and a non-ideal TIADC model, wherein the non-ideal TIADC model simulates various static errors and dynamic errors of TIADC, and the ideal TIADC model has no errors;

[0041] S2, using the ideal TIADC model to generate an ideal TIADC signal, and using the non-ideal TIADC model to generate an error-containing TIADC signal; constructing a data set after preprocessing the ideal TIADC signal and the error-containing TIADC signal; dividing the data set into a training set and a test set;

[0042] S3. Construct a TIADC joint error calibration model based on a neural network for one-time calibration of static and dynamic errors. Different from the traditional fully connected structure for static error fitting, the TIADC joint error calibration model includes an error extraction network, an error fitting network and a signal output module. The error extraction network extracts static error features related to local details from the input samples and obtains dynamic error features related to signal frequency through long sample sequences, and inputs the obtained static error features and dynamic error features into the error fitting network. The error fitting network fits the corresponding statistical error based on the static error features and dynamic error features. The signal output module is used to subtract the input samples from the statistical error to obtain the calibrated output signal.

[0043] S4, using the training set to train the TIADC joint error calibration model, wherein minimizing the difference between the output signal and the ideal TIADC signal is the training goal during the training process;

[0044] S5. Use the test set to test the trained TIADC joint error calibration model to determine whether the TIADC joint error calibration model meets the requirements. If so, calibrate the TIADC with the current TIADC joint error calibration model. If not, optimize the TIADC joint error calibration model and retrain and test it.

[0045] The method first generates an ideal error-free TIADC signal and simulates the dynamic error and static error of TIADC to generate an error-containing TIADC signal to construct a data set, and then constructs a TIADC joint error calibration model. The TIADC joint error calibration model mainly includes an error extraction network and an error fitting network, which are used to extract static-dynamic error features and fit the error features, respectively, effectively learn the local features and long-term statistical characteristics of the high-speed ADC output sequence, and through back-propagation training, realize the joint calibration of dynamic and static errors. The present invention abandons the error classification calibration idea of ​​traditional high-speed ADCs, and adopts a more comprehensive and flexible data-driven method, which can not only cover a larger error range, but also better adapt to the error calibration needs in different scenarios.

[0046] The following describes each step in more detail.

[0047] (1) Step S1: Constructing TIADC error model and ideal model

[0048] In step S1, an ideal model and a non-ideal model of TIADC need to be constructed. The ideal model outputs a TIADC sequence without error, namely, the Ground Truth, and the non-ideal model outputs a TIADC sequence with a composite error (Multi-Error, including dynamic error and static error) added.

[0049] like Figure 1 As shown in the figure, this example uses the Simulink modeling tool to build a behavioral model of a 4000Msps four-channel TIADC. Each channel has a sub-ADC, and various dynamic errors and static errors are added to the channel. Finally, the outputs of the four channels are integrated (MUX) to obtain the error-containing TIADC signal, that is, the TIADC sequence V containing dynamic errors and static errors. out . Specifically, the static error of TIADC mainly includes the gain mismatch error and bias mismatch error between channels. The dynamic error of TIADC mainly adds the time mismatch error between channels. In addition, the clock jitter in TIADC will introduce obvious errors in the process of broadband signal input. The present invention introduces the clock jitter error into the TIADC error model for the first time and performs joint calibration. Therefore, after adding the above static and dynamic errors, the mathematical expression of the kth channel sampling signal can be expressed as:

[0050] ,

[0051] Among them, V in is the input signal, represents the gain mismatch error of the kth channel, represents the fixed time mismatch error of the kth channel, M represents the number of channels, represents the clock jitter in the kth channel, represents the offset error of the kth channel, n represents the nth cycle, T s represents the sampling period, Represents the output signal of the kth channel in the nth sampling period.

[0052] (2) Step S2: Constructing the dataset

[0053] In step S2, a corresponding data set needs to be constructed for training and testing the model.

[0054] The data set of the neural network directly determines the performance and generalization ability of the neural network model, so selecting and preparing the data set is a crucial part of the neural network model training. The neural network data set used in this example contains two parts, one is the output with error, and the other is the ideal output of TIADC.

[0055] The output with error is generated by the TIADC error model constructed in step S1, and an output signal containing error is obtained. The ideal output in the training set is obtained by sampling the output signal of the ideal model. In this example, the sampled quantized output of the four channels is taken as a data sample.

[0056] The sampled data also needs to be preprocessed to ensure the correspondence between the true signal and the error signal at time, including screening and normalization. Specifically, the true signal and error signal that cannot correspond at time are screened out so that the true signal and the error signal are aligned at time, and the filtered data is normalized.

[0057] This embodiment constructs a data set containing 100,000 data samples to train and test the network, where the training set occupies the majority so that the model can fully learn the characteristics of the data. The test set is used to evaluate the final performance of the model. The ratio of the number of training sets to test sets is 10:1.

[0058] (3) Step S3: Constructing the TIADC joint error calibration model

[0059] like Figure 2 As shown in the figure, the TIADC joint error calibration model constructed in this example includes an error extraction network (TCN), an error fitting network (FCNN) and a signal output module. The input sample (ADC output signal with error Vout ) Enter the error extraction network to obtain the error feature E feature (including static error characteristics and dynamic error characteristics) are input into the error fitting network, and the error fitting network fits the corresponding statistical error E according to the error characteristics fit , the signal output module is used to convert the input sample V out and statistical error E fit Subtract to get the calibrated output signal D cali .

[0060] The error extraction network (TCN) in this example uses Figure 3 The network structure shown in FIG. 1 is composed of multiple cascaded residual networks. In order to extract and calibrate the static and dynamic errors in the TIADC sequence, a residual network is sequentially provided with the first dilated convolution layer, the first regularization layer, the first ReLU nonlinear activation function, the second dilated convolution layer, the second regularization layer, the second ReLU nonlinear activation function, and a 1×1 convolution layer (1×1Conv), that is, a residual network has two dilated convolution layers, two regularization layers, a ReLU nonlinear activation function, and a 1×1 convolution layer (1×1Conv). Among them, the dilated convolution layer extracts local and sequence features along the TIADC output sequence, the regularization layer accelerates the convergence of the network, improves the generalization ability, and better handles the problem of gradient disappearance and explosion, while the activation function introduces nonlinear error fitting ability to the network. The 1×1 convolution layer is used to add the input of the residual network and the output of the second ReLU nonlinear activation function.

[0061] The kernel of each residual network is set to 8, and 16 channels are set. The receptive field and kernel length of the first layer residual network are both 8, and the dilation factor is 0. The receptive field of the second layer residual network is twice 8, which is 16, and the dilation factor is 2. As the residual networks are stacked, the dilation factor and receptive field of the residual network increase exponentially by 2 times. For example, the receptive field of the third layer residual network is twice 16, which is 32, and the dilation factor is 4. In this embodiment, the network depth is set to 7 (there are 7 residual networks in total), the receptive field of the 7th layer residual network is 512 time intervals, and the dilation factor is 32 to fully capture local features and long-term features.

[0062] The error fitting network in this example uses a fully connected neural network (FCNN), and its structural diagram is as follows Figure 2As shown inside, it includes an input layer, a hidden layer, and an output layer. A Relu nonlinear activation function is added between the input layer and the hidden layer, so that the neural network can learn and represent more complex nonlinear function relationships, improving the model's expressiveness and performance. Neurons in adjacent network layers of the neural network are interconnected, where the input layer consists of 16 neuron nodes, corresponding to the four channels of TIADC, the hidden layer consists of 64 neuron nodes, and the output layer consists of 1 neuron node. This structure can be simplified to [16 64 1]. It is worth noting that since the error extraction network is introduced before the error fitting network, the number of neurons required for the FCNN model in this example is greatly reduced. Compared with other works, it not only has a more comprehensive error coverage range, but also has a smaller model size.

[0063] (4) Step S4: Training and testing the TIADC joint error calibration model

[0064] The training process of a neural network usually includes forward propagation, loss calculation, back propagation, and parameter update steps. The training data is input into the neural network for forward propagation, the output result is calculated and the loss function is calculated with the real label, the gradient of the loss function to the model parameters is calculated using the back propagation algorithm, and the parameters are updated through optimization algorithms such as gradient descent. This process is repeated until a certain number of training rounds is reached or the loss function converges, and finally a trained neural network model is obtained.

[0065] Combining the above process, the training process of the TIADC joint error calibration model is as follows: Figure 2 As shown, the steps include:

[0066] The training data is input into the error extraction network to obtain the mixed feature E feature ;

[0067] The error fitting network is used to fit the mixed feature E feature Perform nonlinear transformation and use activation function to activate and get the corresponding statistical error E fit ;

[0068] The statistical error E fit Subtract the training data to get the calibrated output signal D cali ;

[0069] Calculate the output signal D cali With the real data D gt The loss function between

[0070] Use the back-propagation algorithm to calculate the gradient of the loss function to the model parameters, and update the parameters using the gradient descent algorithm;

[0071] The above training process is repeated until the set number of training rounds is reached or the loss function converges, and finally a trained TIADC joint error calibration model is obtained.

[0072] The training data is input into the input layer of the error extraction network (TCN). TCN gradually extracts local features and long-term statistical features from the time series data of the training data by stacking convolutional layers with causality and possible expansion. The initial convolutional layer focuses on the immediate and short-term changes of the signal, extracting local features such as spikes and periodic patterns. As the depth of the network increases, the dilated convolution helps expand the receptive field, enabling TCN to capture longer-term trends and changing statistical features of periodic patterns. These multi-level and multi-scale feature extractions provide a rich information foundation for subsequent TIADC output analysis, signal processing and prediction tasks. Subsequently, the error fitting network receives the mixed feature E output by the error extraction network. feature ,FCNN performs nonlinear transformation on feature vectors through multiple layers of fully connected layers, and uses activation functions to enhance its representation ability.

[0073] During training, FCNN calculates the difference between the output and the true label according to the loss function, using the mean squared error (MSE) loss function (N vs. D cali and D gt The smaller the loss function value, the closer the model's prediction result is to the true value.

[0074] By calculating the gradient of the loss function to the model parameters (weights and biases), the chain rule is used to back-propagate the gradient from the output layer to the input layer. The gradient will gradually be transmitted and accumulated as the layers propagate, and finally the gradient of the parameters of each layer is calculated. The optimization algorithm Levenberg-Marquardt gradient descent method is used to update the parameters of the neural network according to the parameter gradient calculated by back-propagation. The Levenberg-Marquardt algorithm shows unique advantages in solving nonlinear least squares optimization problems compared with other optimization algorithms such as gradient descent method and Newton method. It combines the fast convergence of the Gauss-Newton method and the global convergence of the gradient descent method. By introducing a damping term to balance the local search step size and global stability, it effectively avoids the problem of matrix singularity or excessive step size that the Gauss-Newton method may encounter when it is far away from the optimal solution. In addition, the Levenberg-Marquardt algorithm also has good numerical stability and fast convergence speed, which is particularly suitable for complex nonlinear model parameter estimation tasks that require accurate solutions and have high requirements for computational efficiency.

[0075] During the network training process, the optimization algorithm continuously changes the parameter values ​​in the neural network to reduce the loss function value. By repeating this process, the optimization algorithm can make the neural network gradually converge to the local optimal solution or the global optimal solution, thereby improving the calibration performance of the network.

[0076] (5) Step S5: Model testing

[0077] After the neural network training is completed, its performance needs to be tested. The testing process of the present invention is: use 16384 continuous TIADC outputs with dynamic and static errors to input into the trained neural network for calibration, perform FFT processing on the calibrated sampling values, obtain the spectrum information diagram of the signal, and observe the changes in its various indicators. The present invention selects the signal-to-noise distortion ratio (SNDR) and spurious-free dynamic range (SFDR) indicators for analysis, and compares the SNDR and SFDR of the calibrated signal with the SNDR and SFDR of the original signal to determine whether the calibration performance of the neural network meets the target requirements.

[0078] Figure 4 The following is the FFT spectrum diagram of the signal before and after TIADC calibration. Compared with before calibration, the SFDR of the calibrated signal is improved from 41.47dB to 85.47dB, and the SNDR is improved from 37.67dB to 74.52dB. Figure 5 The SNDR comparison chart of the method proposed in this example before and after TIADC calibration under full-band input shows that in the full-band, the average SNDR after calibration is about 70dB, and the average SNDR before calibration is about 38dB. The method proposed in the present invention achieves excellent calibration effect while covering more comprehensive error conditions.

[0079] from Figure 4 and Figure 5 It can be seen from the test results that a TIADC static-dynamic error joint calibration method and system provided in an embodiment of the present invention can simultaneously eliminate the static error and dynamic error of the TIADC, improve the output performance of the TIADC, and broaden the application scenarios of the TIADC, especially in the field of wide-band signal applications.

[0080] (6) Model optimization

[0081] The performance of the trained neural network is tested. If the test results meet the target requirements, it means that the performance of the neural network meets the requirements for TIADC error calibration. If the calibration results are not ideal, the neural network needs to be optimized accordingly. The network structure, hyperparameters, data processing, etc. can be improved. By adjusting the network structure, such as increasing the number of layers and adjusting the number of neurons, the representation and learning capabilities of the network can be optimized. By adjusting hyperparameters, such as learning rate and batch size, the training process can be optimized. Comprehensively adjusting these aspects can improve the performance and generalization ability of the neural network model, making it more suitable for practical application scenarios.

[0082] Based on the above-mentioned TIADC static-dynamic error joint calibration method, this embodiment also provides a TIADC static-dynamic error joint calibration system, which is provided with a model generation module and a model application module. The model generation module is used to execute the above-mentioned TIADC static-dynamic error joint calibration method to obtain a trained TIADC joint error calibration model that meets the requirements; the model application module is used to apply the trained TIADC joint error calibration model that meets the requirements in the TIADC calibration task.

[0083] In summary, the TIADC static-dynamic error joint calibration method and system provided in the embodiments of the present invention are highly flexible, easy to implement, and powerful. They can eliminate the adverse effects caused by TIADC dynamic and static errors, cover a wider error range, improve the performance of TIADC output, and expand the application scenarios of TIADC.

[0084] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the protection scope of the present invention.

Claims

1. A TIADC static-dynamic error joint calibration method, characterized in that: Includes steps: S1, constructing an ideal TIADC model and a non-ideal TIADC model, wherein the non-ideal TIADC model simulates various static errors and dynamic errors of the TIADC, and the ideal TIADC model has no errors; In step S1, after adding static and dynamic errors, the sampling signal of the kth channel is expressed as: ， Among them, V in is the input signal, represents the gain mismatch error of the kth channel, represents the fixed time mismatch error of the kth channel, M represents the number of channels, represents the clock jitter in the kth channel, represents the offset error of the kth channel, n represents the nth cycle, T s represents the sampling period, Represents the output signal of the kth channel in the nth sampling period; S2, using the ideal TIADC model to generate an ideal TIADC signal, and using the non-ideal TIADC model to generate an error-containing TIADC signal; constructing a data set after preprocessing the ideal TIADC signal and the error-containing TIADC signal; dividing the data set into a training set and a test set; S3, constructing a TIADC joint error calibration model based on neural network for one-time calibration of static and dynamic errors; The TIADC joint error calibration model includes an error extraction network, an error fitting network and a signal output module. The error extraction network extracts static error features related to local details from input samples and obtains dynamic error features related to signal frequency through long sample sequences, and inputs the obtained static error features and dynamic error features into the error fitting network; the error fitting network fits the corresponding statistical error according to the static error features and dynamic error features; the signal output module is used to subtract the input sample from the statistical error to obtain a calibrated output signal; The error extraction network is composed of a plurality of cascaded residual networks; each of the residual networks is sequentially provided with a first dilated convolution layer, a first regularization layer, a first ReLU nonlinear activation function, a second dilated convolution layer, a second regularization layer, a second ReLU nonlinear activation function, and a 1×1 convolution layer; wherein the first dilated convolution layer and the second dilated convolution layer extract local and sequence features along the TIADC output sequence; the 1×1 convolution layer is used to add the input of the residual network and the output of the second ReLU nonlinear activation function and then input the sum into the residual network of the next layer; The kernel of each residual network is set to 8, and 16 channels are set; the receptive field and kernel length of the first layer of residual network are both 8, and the expansion factor is 0; the receptive field of the second layer of residual network is twice 8, that is, 16, and the expansion factor is 2; with the further stacking of residual networks, the expansion factor and receptive field of the residual network increase exponentially at a rate of 2; the error extraction network is set with 7 layers of residual networks; the receptive field of the 7th layer of residual network is 512 time intervals, and the expansion factor is 64; S4, using the training set to train the TIADC joint error calibration model, wherein minimizing the difference between the output signal and the ideal TIADC signal is a training goal during the training process; S5. Use the test set to test the trained TIADC joint error calibration model to determine whether the TIADC joint error calibration model meets the requirements. If so, perform TIADC calibration with the current TIADC joint error calibration model. If not, optimize the TIADC joint error calibration model and retrain and test it.

2. A TIADC static-dynamic error joint calibration method according to claim 1, characterized in that: The error fitting network adopts a fully connected neural network, including an input layer, a hidden layer, and an output layer, and a Relu nonlinear activation function is added between the input layer and the hidden layer.

3. A TIADC static-dynamic error joint calibration method according to claim 2, characterized in that: The input layer is composed of 16 neuron nodes, corresponding to the four channels of TIADC; the hidden layer is composed of 64 neuron nodes; and the output layer is composed of 1 neuron node.

4. A TIADC static-dynamic error joint calibration method according to claim 3, characterized in that: In step S4, the training process of the TIADC joint error calibration model includes: The training data is input into the error extraction network to obtain the mixed feature E feature ; The error fitting network is used to fit the mixed feature E feature Perform nonlinear transformation and use activation function to activate and get the corresponding statistical error E fit ; The statistical error E fit Subtract the training data to get the calibrated output signal D cali ; Calculate the output signal D cali With the real data D gt The loss function between Use the back-propagation algorithm to calculate the gradient of the loss function to the model parameters, and update the parameters using the gradient descent algorithm; The above training process is repeated until the set number of training rounds is reached or the loss function converges, and finally a trained TIADC joint error calibration model is obtained.

5. A TIADC static-dynamic error joint calibration method according to claim 4, characterized in that: The gradient descent algorithm adopts the Levenberg-Marquardt algorithm.

6. A TIADC static-dynamic error joint calibration system, characterized in that: A model generation module and a model application module are provided, wherein the model generation module is used to execute a TIADC static-dynamic error joint calibration method according to any one of claims 1 to 5 to obtain a trained TIADC joint error calibration model that meets the requirements; and the model application module is used to apply the trained TIADC joint error calibration model that meets the requirements to a TIADC calibration task.

Images