High-speed time-interleaved ADC neural network calibration method based on target optimization
By directly optimizing the signal-to-noise ratio (SNDR) in the time-interleaved analog-to-digital converter and updating the calibration network model using the gradient descent method, the problem of SNDR degradation caused by manufacturing process deviations is solved, achieving high-precision and robust calibration results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIDIAN UNIV
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-15
AI Technical Summary
In existing time-interleaved analog-to-digital converters, gain errors, offset errors, and time deviations caused by manufacturing process variations lead to a deterioration in the signal-to-noise ratio (SNDR). Furthermore, existing calibration methods rely on reference signals, resulting in low calibration accuracy and poor robustness.
A high-speed time-interleaved ADC neural network calibration method based on target optimization is adopted. By directly optimizing the signal-to-noise ratio (SNDR) as the training target, the weights and bias parameters of the calibration network model are updated using the gradient descent method, and a loss function is constructed to maximize SNDR, thereby compensating for mismatch errors in multiple channels.
Simultaneous compensation for mismatch errors of multiple channels was achieved within the Nyquist frequency range, improving calibration accuracy and robustness. This solves the problems of low calibration accuracy and poor robustness caused by the indirect and reference signal dependence of the optimization target in existing technologies.
Smart Images
Figure CN122052791A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of time-interleaved analog-to-digital converter technology, and more specifically to a target-optimized high-speed time-interleaved ADC (Analog to Digital Converter) neural network calibration method. Background Technology
[0002] Time-interleaved analog-to-digital converters (ADCs) employ multi-channel parallel sampling, overcoming the sampling rate bottleneck of single-channel ADCs. However, due to manufacturing process variations, gain errors, offset errors, and timing skew exist between the parallel channels in the system. These mismatch errors introduce additional spurious components into the output spectrum, severely degrading key performance indicators such as the system's signal-to-noise and distortion ratio (SNDR), thus limiting the overall system performance. Zhifei Lu et al., in their paper "A Novel NN-based Fast-Convergence Background Calibration for Timing Mismatch in TI ADCs," disclosed a neural network calibration method for time-interleaved ADCs. This method first obtains a reference signal by fitting the mismatch signal, then trains the neural network with the goal of minimizing the mean square error (MSE) between the neural network output signal and the reference signal, making the neural network output signal approximate the reference signal. Finally, the trained neural network is used to calibrate the time-interleaved ADC.
[0003] However, this calibration strategy is indirect. The ultimate goal of calibration is to optimize SNDR, but minimizing the mean-square error (MSE) with the reference signal is not equivalent to directly maximizing SNDR, resulting in low calibration accuracy. Moreover, existing technologies are highly dependent on the accuracy of the generated reference signal. In practical applications, especially when the input signal is complex and variable, obtaining an accurate reference signal is very difficult, thus exhibiting poor robustness to calibration for different signal types.
[0004] Therefore, how to improve the low calibration accuracy caused by the optimization target and the poor calibration robustness caused by the reliance on the reference signal in the existing technology has become an urgent technical problem to be solved. Summary of the Invention
[0005] To address the aforementioned problems in the existing technology, this invention provides a high-speed time-interleaved ADC neural network calibration method based on target optimization. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a high-speed time-interleaved ADC neural network calibration method based on target optimization, comprising: The training sample set consisting of K training samples is determined based on the original digital sequence output by the time-interleaved analog-to-digital converter, where K is a natural number greater than 0. In the t-th iteration, A training samples are randomly selected from the K training samples, and the i-th training sample is input into the calibration network model obtained after training with the (i-1)-th training sample to obtain the i-th calibrated sequence. The i-th signal-to-noise ratio obtained from the i-th calibrated sequence is used to construct a loss function to update the weights and bias parameters of the calibration network model, and to determine the calibration network model obtained after training with the i-th training sample, until the finally trained calibration network model is obtained, 0 < A < K, 1 ≤ i ≤ A; The test sample is input into the trained calibration network model to obtain the calibration result.
[0006] In one embodiment of the present invention, determining a training sample set including K training samples based on the original digital sequence output by the time-interleaved analog-to-digital converter includes: Obtain the raw digital sequence output by the time-interleaved analog-to-digital converter; Each of the original number sequences is normalized to obtain a normalized number sequence, and the normalized number sequence is used as a training sample set. K normalized number sequences are randomly selected from the training sample set as training samples. In one embodiment of the present invention, during the t-th iteration, A training samples are randomly selected from the K training samples, and the i-th training sample is input into the calibration network model obtained after training with the (i-1)-th training sample to obtain the i-th calibrated sequence. A loss function is constructed using the i-th signal-to-noise ratio obtained from the i-th calibrated sequence to update the weights and bias parameters of the calibration network model, thereby determining the calibration network model obtained after training with the i-th training sample, until the finally trained calibration network model is obtained, including: Step 2.1: In the t-th iteration, randomly select A training samples from the K training samples; Step 2.2: Input the i-th training sample into the calibration network model obtained after training with the (i-1)-th training sample to obtain the i-th calibrated sequence; Step 2.3: With maximizing the signal-to-noise ratio of the i-th calibrated sequence as the direct optimization objective, construct the i-th loss function based on the i-th signal-to-noise ratio; Step 2.4: Using gradient descent, the gradient of the i-th loss function with respect to the weights and bias parameters is calculated and backpropagated to update the parameters of the calibration network model, thus obtaining the calibration network model trained on the i-th training sample. Step 2.5: Repeat steps 2.1 to 2.4 until the calibration network model obtained after training with the Ath training sample is obtained, and use the calibration network model obtained after training with the Ath training sample as the calibration network model after the tth iteration update. Step 2.6: Repeat steps 2.1 to 2.5 until... The trained calibration network model is obtained, where T is the maximum number of iterations.
[0007] In one embodiment of the present invention, step 2.3 includes: Step 2.31: Perform a Fast Fourier Transform on the i-th calibrated sequence to obtain the i-th spectrum; Step 2.32: Locate the dominant frequency component using the i-th spectrum and determine the first power of the dominant frequency component; Step 2.33: Calculate the sum of the power of the other frequency components in the i-th spectrum except for the main frequency component, to obtain the second power; Step 2.34: Obtain the i-th signal-to-noise distortion ratio based on the first power and the second power; Step 2.35: Obtain the i-th loss function based on the i-th signal-to-noise ratio. In one embodiment of the present invention, step 2.32 includes: Determine the peak value of the i-th spectrum, and determine a preset threshold value based on the peak value, wherein the preset threshold value is the difference between the peak value and a preset value; The frequency corresponding to the point in the i-th spectrum whose amplitude is above the preset threshold is taken as the main frequency component; The first power is obtained based on the amplitude corresponding to the main frequency component. In one embodiment of the present invention, the loss function is expressed as:
[0008] in, For the i-th loss function, Let be the signal-to-noise ratio (SNR) of the i-th element.
[0009] In one embodiment of the present invention, the update formulas for the weights and the biases are respectively expressed as:
[0010]
[0011] in, The weights of the calibration network model obtained after training with the (i-1)th training sample The update results The bias of the calibration network model obtained after training with the (i-1)th training sample. The update results For learning rate, Represents the loss function Weights gradient, Represents the loss function For bias The gradient. In one embodiment of the present invention, the calibration network model includes an input layer, a hidden layer, and an output layer cascaded in sequence, wherein: The input layer is used to receive the i-th training sample, perform a linear transformation on the i-th training sample to map it to a high-dimensional space, and then perform non-linear activation on the i-th training sample after the linear transformation to obtain high-dimensional features. The hidden layer is used to perform multi-level linear transformations on the high-dimensional features to obtain deep features; The output layer performs a linear transformation on the deep features to map them to the target dimension, and then constrains the deep features of the target dimension to the range of [-1, 1] to obtain the i-th calibrated sequence. In one embodiment of the present invention, the input layer includes a first fully connected layer and a... Function layer, the first fully connected layer includes 1 node Each node is connected to the Function layer; The hidden layer includes L layers of second fully connected layers, each of which includes... One node; The output layer includes a third fully connected layer and a Tanh function layer. The third fully connected layer includes... 1 node Each node is connected to the Tanh function layer. = .
[0012] In one embodiment of the present invention, the sample to be tested is input into the trained calibration network model to obtain calibration results, including: Obtain the raw digital sequence output by the time-interleaved analog-to-digital converter to be tested; The original digital sequence output by the time-interleaved analog-to-digital converter to be tested is normalized, and the normalization result is used as the test sample and input into the trained calibration network model to obtain the calibration result. Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes a time-interleaved analog-to-digital converter calibration method based on direct optimization of SNDR. This method uses maximizing SNDR as the training objective of the neural network, without the need for fitting and generating or relying on reference signals. It achieves simultaneous compensation for mismatch errors of multiple channels across the entire Nyquist frequency range, solving the technical problems of low calibration accuracy caused by indirect optimization objectives and poor calibration robustness caused by reliance on reference signals in existing technologies.
[0013] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0014] Figure 1 This is a flowchart illustrating a high-speed time-interleaved ADC neural network calibration method based on target optimization provided by the present invention; Figure 2 This is a structural diagram of a calibration network model provided by the present invention; Figure 3 This is a schematic diagram of a high-speed time ADC structure provided by the present invention; Figure 4 This is a schematic diagram of a sampling clock deviation provided by the present invention; Figure 5 This is a schematic diagram of the calibration effect provided by the present invention. Detailed Implementation
[0015] 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.
[0016] Please see Figure 1 , Figure 1 This is a flowchart illustrating a high-speed time-interleaved ADC neural network calibration method based on target optimization provided by the present invention. The present invention provides a high-speed time-interleaved ADC neural network calibration method based on target optimization, which includes: Step 1: Determine a training sample set consisting of K training samples based on the original digital sequence output by the time-interleaved analog-to-digital converter, where K is a natural number greater than 0.
[0017] Step 1.1: Obtain the raw digital sequence output by the time-interleaved analog-to-digital converter.
[0018] Specifically, the time-interleaved analog-to-digital converter samples the input analog signal and outputs... A sequence of original numbers.
[0019] Step 1.2: Normalize each original number sequence to obtain a normalized number sequence, and use the normalized number sequence as a training sample set. Randomly select K normalized number sequences from the training sample set as training samples.
[0020] Specifically, for Each of the original number sequences is normalized to obtain a normalized number sequence. Then, K normalized number sequences are randomly selected from all the normalized number sequences as training samples, and the rest... A normalized sequence of numbers was used as the test sample.
[0021] Here, the formula for normalization is:
[0022] in, For the k-th digit code in the normalized digit sequence, It is the k-th digit code in the original digit sequence. It is the largest numeric code in the original numeric sequence. It is the smallest numeric code in the original numeric sequence. . Optionally, In this embodiment , .
[0023] Step 2: In the t-th iteration, randomly select A training samples from the K training samples, and input the i-th training sample into the calibration network model obtained after training with the (i-1)-th training sample to obtain the i-th calibrated sequence. Then, construct a loss function based on the i-th signal-to-noise ratio obtained from the i-th calibrated sequence to update the weights and bias parameters of the calibration network model, and determine the calibration network model obtained after training with the i-th training sample. Continue until the finally trained calibration network model is obtained, where 0 < A < K, 1 ≤ i ≤ A.
[0024] Specifically, in this embodiment, the maximum number of iterations is set to 1. The initial number of iterations is (Initial time) Therefore, in the t-th iteration, A training samples are randomly selected from the K training samples to train the calibration network model obtained after the (t-1)-th iteration. The A training samples are then sequentially input into the calibration network model for training. The first training sample input is the calibration network model obtained after the (t-1)-th iteration. The calibration network model obtained after the (t-1)-th iteration outputs the first calibrated sequence. Then, based on the first calibrated sequence, the first loss function for the t-th iteration is constructed to update the weights and bias parameters of the calibration network model, thus obtaining the calibration network model trained with the first training sample. The network model is first trained with the first training sample, and then the second training sample is input into the calibration network model obtained after training with the first training sample. The calibration network model outputs the second calibrated sequence. Then, the second loss function is constructed based on the second calibrated sequence to update the weights and bias parameters of the calibration network model, thereby determining the calibration network model obtained after training with the second training sample. This process is repeated until the calibration network model obtained after training with the Ath training sample is determined. This calibration network model is the calibration network model obtained after training with the tth iteration. After T rounds of training, the finally trained calibration network model is obtained.
[0025] This embodiment takes maximizing the signal-to-noise ratio of the calibrated sequence as the direct optimization objective, constructs a loss function based on SNDR, and iteratively updates the weight parameters and bias parameters in the calibration network model through the gradient backpropagation algorithm until the model converges, thus obtaining a trained calibration network model.
[0026] Step 2.1: In the t-th iteration, randomly select A training samples from the K training samples.
[0027] For example, .
[0028] Step 2.2: Input the i-th training sample into the calibration network model obtained after training with the (i-1)-th training sample to obtain the i-th calibrated sequence.
[0029] Specifically, for the first training sample, its input is the calibration network model obtained after training for the (t-1)th iteration, and the calibration network model obtained after training for the (t-1)th iteration outputs the first calibrated sequence; for the ith training sample, its input is the calibration network model obtained after training for the (i-1)th training sample, and the calibration network model obtained after training for the (i-1)th training sample outputs the ith calibrated sequence.
[0030] In an optional embodiment, the calibration network model includes an input layer, a hidden layer, and an output layer cascaded in sequence, wherein: The input layer is used to receive the i-th training sample and perform a linear transformation on the i-th training sample to map it to a high-dimensional space. Then, a non-linear activation is performed on the i-th training sample after the linear transformation to obtain high-dimensional features. Hidden layers are used to perform multi-level linear transformations on high-dimensional features to obtain deep features; The output layer performs a linear transformation on the deep features to map them to the target dimension, and then constrains the deep features of the target dimension to the range of [-1, 1] to obtain the i-th calibrated sequence, for example, the target dimension is 8.
[0031] Further, please see Figure 2 The input layer consists of a first fully connected layer and a... Function layer, the first fully connected layer includes 1 node Each node is connected Function layer, for example, ; The hidden layer consists of L cascaded second fully connected layers, each second fully connected layer including... For example, a node. , ; The output layer consists of a third fully connected layer and a Tanh function layer. The third fully connected layer includes... 1 node Each node is connected to the Tanh function layer. = ,For example, .
[0032] Step 2.3: With maximizing the signal-to-noise ratio of the i-th calibrated sequence as the direct optimization objective, construct the i-th loss function based on the i-th signal-to-noise ratio.
[0033] Step 2.31: Perform a Fast Fourier Transform on the i-th calibrated sequence to obtain the i-th spectrum, where the horizontal axis represents frequency and the vertical axis represents amplitude. Step 2.32: Use the i-th spectrum to locate the main frequency component and determine the first power of the main frequency component.
[0034] Step 2.321: Determine the peak value of the i-th spectrum and determine a preset threshold value based on the peak value. The preset threshold value is the difference between the peak value and the preset value.
[0035] Here, the preset threshold is expressed as Q=F-30, where Q is the preset threshold and F is the peak value. For example, if the preset value is 30, then Q=F-30dB.
[0036] Step 2.322: Take the frequency corresponding to the point in the i-th spectrum whose amplitude is above the preset threshold as the main frequency component.
[0037] Step 2.323: Obtain the first power based on the amplitude corresponding to the main frequency component.
[0038] Specifically, the square of the amplitude corresponding to each main frequency component is calculated to obtain the power of that main frequency component, and the power of all main frequency components is added together to obtain the first power.
[0039] Step 2.33: Calculate the sum of the power of the other frequency components in the i-th spectrum except for the main frequency component, and obtain the second power.
[0040] Specifically, the square of the amplitude of each frequency component other than the main frequency component in the i-th spectrum is calculated to obtain the power of each other frequency component. These powers are then added together to obtain the second power, which is the noise and distortion power.
[0041] Step 2.34: Obtain the i-th signal-to-noise ratio based on the first power and the second power.
[0042] Here, the formula for calculating the signal-to-noise ratio is:
[0043] in, The signal-to-noise ratio (SNR) For first power, This is the second power. Step 2.35: Obtain the i-th loss function based on the i-th signal-to-noise ratio.
[0044] Here, the loss function is expressed as:
[0045] in, For the i-th loss function, Let be the signal-to-noise ratio (SNR) of the i-th element.
[0046] Step 2.4: Using gradient descent, the gradient of the i-th loss function with respect to the weights and bias parameters is calculated and backpropagated to update the parameters of the calibration network model, thus obtaining the calibration network model trained with the i-th training sample.
[0047] Here, the gradient descent method is used to update the parameters, and the update formulas for the weights and the bias parameters are expressed as follows:
[0048]
[0049] in, The weights of the calibration network model obtained after training with the (i-1)th training sample The update results The bias of the calibration network model obtained after training with the (i-1)th training sample. The update results For learning rate, Represents the loss function Weights gradient, Represents the loss function For bias The gradient.
[0050] Step 2.5: Repeat steps 2.1 to 2.4 until the calibration network model obtained after training with the Ath training sample is obtained, and use the calibration network model obtained after training with the Ath training sample as the calibration network model after the tth iteration update. Step 2.6: Repeat steps 2.1 to 2.5 until... This yields a well-trained calibration network model.
[0051] Specifically, judge If it holds true, then obtain the trained calibration network model; otherwise, let... Continue training.
[0052] For example, .
[0053] Step 3: Input the test sample into the trained calibration network model to obtain the calibration result.
[0054] Step 3.1: Obtain the raw digital sequence output by the time-interleaved analog-to-digital converter to be tested.
[0055] Specifically, the original digital sequence output by the time-interleaved analog-to-digital converter to be tested can be the original digital sequence corresponding to the test sample in the training sample set obtained in step 1, or it can be the original digital sequence output by the time-interleaved analog-to-digital converter that actually needs to be tested.
[0056] Step 3.2: Normalize the original digital sequence output by the time-interleaved analog-to-digital converter to be tested, and use the normalized result as the test sample to input into the trained calibration network model to obtain the calibration result.
[0057] Specifically, the original digital sequence output by the time-interleaved analog-to-digital converter to be tested is normalized according to the normalization method in step 1, and the normalization result is input into the trained calibration network model to obtain the calibration result.
[0058] Figure 3This is an 8-channel high-speed time-interleaved ADC architecture, where each sub-ADC is controlled by a clock signal of a different phase (Clk0, Clk1…Clk8). Through alternating sampling by the eight sub-ADCs, the overall effective sampling rate of the system is increased to eight times that of a single ADC. The digital signals output from each sub-ADC are input to a selector (MUX) in time sequence. The MUX arranges the data from each channel according to the sampling time sequence, integrating them into a continuous high-speed digital output stream, fully preserving the time-domain information of the original signal.
[0059] Figure 4 This demonstrates the sampling clocks of all sub-ADCs under non-ideal conditions. The sampling clock phases of each sub-ADC deviate, leading to non-uniform sampling, mismatch, and a significant impact on high-frequency signals (close to the Nyquist frequency).
[0060] Figure 5 Figures (a) and (b) in the figure show the calibration effect of the proposed calibration algorithm on the single-tone signal and wideband signal output of a 12-bit 3.6GS / s 8-channel high-speed time-interleaved ADC, respectively. The frequency points shown in the figures are from the test set. It can be seen that for the single-tone signal, spurious signals caused by mismatch, gain mismatch, and time deviation are significantly suppressed. For complex wideband signals, existing neural network calibration models are limited by fitting the wideband signal, while the proposed algorithm can calibrate it without fitting, effectively compensating for mismatch errors.
[0061] This invention, in training the calibration network model of a high-speed time-interleaved ADC, abandons the traditional roundabout strategy of "fitting the reference signal" and directly optimizes it by maximizing the core performance indicator SNDR of the system. By directly maximizing SNDR through gradient descent, the neural network can adaptively learn the optimal calibration mapping path, achieving end-to-end optimal calibration performance and thus improving calibration accuracy.
[0062] This invention does not require the generation and reliance on a reference signal. Therefore, when the signal is complex and variable, there is no need to consider the problem of inconsistent accuracy of the generated reference signal, thus improving the calibration robustness for different signals.
[0063] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.
[0064] Although the invention has been described herein in conjunction with various embodiments, those skilled in the art will understand and implement other variations of the disclosed embodiments by reviewing the accompanying drawings, disclosure, and appended claims in carrying out the claimed invention. In the claims, the word "comprising" does not exclude other components or steps, and "a" or "an" does not exclude a plurality. A single processor or other unit can implement several functions listed in the claims. While different dependent claims may recite certain measures, this does not mean that these measures cannot be combined to produce good results.
[0065] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, any modifications made without departing from the inventive concept should be considered within the scope of protection of the present invention.
Claims
1. A high-speed time-interleaved ADC neural network calibration method based on target optimization, characterized in that, include: The raw digital sequence output from the time-interleaved analog-to-digital converter is determined based on... K A training sample set of n training samples, where... K It is a natural number greater than 0; In the t In the next iteration, from the K Randomly selected from training samples A The training samples mentioned above, and the first training sample i The training sample is input after the first training sample... i In the calibration network model obtained after training with -1 training samples, the first... i The calibrated sequence, and using the sequence according to the first... i The calibrated sequence obtained from the first i The signal-to-noise ratio is used to construct a loss function to update the weights and bias parameters of the calibrated network model, determining the loss function after the first signal-to-noise ratio is reached. i The calibration network model is obtained after training on a number of training samples, until the final trained calibration network model is obtained, 0 < A < K ,1≤ i ≤ A ; The test sample is input into the trained calibration network model to obtain the calibration result.
2. The high-speed time-interleaved ADC neural network calibration method according to claim 1, characterized in that, The raw digital sequence output from the time-interleaved analog-to-digital converter is determined based on... K The training sample set for each training sample includes: Obtain the raw digital sequence output by the time-interleaved analog-to-digital converter; Each of the original number sequences is normalized to obtain a normalized number sequence, and the normalized number sequence is used as a training sample set. Then, random selections are made from the training sample set. K The normalized sequence of numbers is used as the training sample.
3. The high-speed time-interleaved ADC neural network calibration method according to claim 1, characterized in that, In the t In the next iteration, from the K Randomly selected from training samples A The training samples mentioned above, and the first training sample i The training sample is input after the first training sample... i In the calibration network model obtained after training with -1 training samples, the first... i The calibrated sequence, and using the sequence according to the first... i The calibrated sequence obtained from the first i The signal-to-noise ratio is used to construct a loss function to update the weights and bias parameters of the calibrated network model, determining the loss function after the first signal-to-noise ratio is reached. i The calibration network model obtained after training on a number of training samples, until the final trained calibration network model is obtained, includes: Step 2.1, in the... t In the next iteration, from the K Randomly selected from training samples A The training samples mentioned above; Step 2.2, the first i The training sample is input after the first training sample... i In the calibration network model obtained after training with -1 training samples, the first... i A calibrated sequence; Step 2.3, to maximize the first i The signal-to-noise ratio of the calibrated sequence is the direct optimization objective, and a structure based on the ... i The first of the aforementioned signal-to-noise distortion ratios i A loss function; Step 2.4: Using the gradient descent method, calculate the... i The gradient of the loss function with respect to the weights and bias parameters is calculated and backpropagated to update the parameters of the calibrated network model, resulting in the loss function after the first iteration. i The calibration network model obtained after training on a training sample; Step 2.5: Repeat steps 2.1 to 2.4 until the result is obtained. A The calibration network model obtained after training on the training samples will be used to train the calibrated network model, and then... A The calibration network model obtained after training on the training samples is used as the first... t The calibration network model updated in the next iteration; Step 2.6: Repeat steps 2.1 to 2.5 until... The trained calibration network model is obtained, wherein, T This represents the maximum number of iterations.
4. The high-speed time-interleaved ADC neural network calibration method according to claim 3, characterized in that, Step 2.3 includes: Step 2.31, regarding the first... i Perform a Fast Fourier Transform on the calibrated sequence to obtain the ... i One spectrum; Step 2.32, using the aforementioned... i The spectrum is used to locate the dominant frequency component and determine the first power of the dominant frequency component; Step 2.33, calculate the first... i The sum of the powers of all frequency components in the spectrum other than the dominant frequency component is used to obtain the second power; Step 2.34: Obtain the first power and the second power. i The aforementioned signal-to-noise distortion ratio; Step 2.35, according to the... i The signal-to-noise distortion ratio is obtained from the first... i The loss function is described above.
5. The high-speed time-interleaved ADC neural network calibration method according to claim 4, characterized in that, Step 2.32 includes: Determine the first i The peak value of each spectrum is determined, and a preset threshold is determined based on the peak value, wherein the preset threshold is the difference between the peak value and a preset value; The first i The frequency corresponding to the point in the spectrum whose amplitude is above the preset threshold is taken as the main frequency component; The first power is obtained based on the amplitude corresponding to the main frequency component.
6. The high-speed time-interleaved ADC neural network calibration method according to claim 3, characterized in that, The loss function is expressed as: in, For the first i A loss function, For the first i Signal-to-noise ratio.
7. The high-speed time-interleaved ADC neural network calibration method according to claim 6, characterized in that, The update formulas for the weights and the biases are expressed as follows: in, For the first i The weights of the calibration network model obtained after training with -1 training sample The update results For the first i The bias of the calibration network model obtained after training with -1 training sample The update results For learning rate, Represents the loss function Weights gradient, Represents the loss function For bias The gradient.
8. The high-speed time-interleaved ADC neural network calibration method according to claim 1, characterized in that, The calibration network model comprises an input layer, a hidden layer, and an output layer cascaded in sequence, wherein: The input layer is used to receive the first i The training samples mentioned above, and for the training samples mentioned above. i The training samples are linearly transformed to map to a high-dimensional space, and then the linearly transformed samples are... i A high-dimensional feature is obtained by performing nonlinear activation on each training sample. The hidden layer is used to perform multi-level linear transformations on the high-dimensional features to obtain deep features; The output layer performs a linear transformation on the deep features to map them to the target dimension, and then constrains the deep features of the target dimension to the range of [-1, 1] to obtain the first... i A calibrated sequence.
9. The high-speed time-interleaved ADC neural network calibration method according to claim 8, characterized in that, The input layer includes a first fully connected layer and a... Function layer, the first fully connected layer includes 1 node Each node is connected to the Function layer; The hidden layer includes L Each layer of the second fully connected layer includes a second fully connected layer. One node; The output layer includes a third fully connected layer and a Tanh function layer. The third fully connected layer includes... 1 node Each node is connected to the Tanh function layer. = .
10. The high-speed time-interleaved ADC neural network calibration method according to claim 1, characterized in that, The test sample is input into the trained calibration network model to obtain the calibration results, including: Obtain the raw digital sequence output by the time-interleaved analog-to-digital converter to be tested; The original digital sequence output by the time-interleaved analog-to-digital converter to be tested is normalized, and the normalization result is used as the test sample and input into the trained calibration network model to obtain the calibration result.