An adc light online calibration method based on knowledge distillation

Abstract

The application discloses an ADC light online calibration method based on knowledge distillation, and relates to the technical field of ADC calibration, and comprises the following steps: obtaining standard spectral characteristics and ADC spectral characteristics to obtain spectral residuals; constructing an ADC nonlinear model and dividing the nonlinear order, establishing an order support domain and extracting order spectral residuals; inputting the order spectral residuals into an improved DeepONet model, establishing order sub-operators through an order decomposition mechanism, and performing knowledge distillation based on compensation parameter contribution to obtain light order sub-operators; identifying target drift order according to online order residuals, adjusting the light order sub-operators and generating updated compensation parameters, and performing nonlinear compensation on online sampling sequences. The application can improve the nonlinear order distinguishing ability and reduce the online calibration calculation amount.

Description

Technical Field

[0001] This invention relates to the field of ADC calibration technology, and in particular to a lightweight online ADC calibration method based on knowledge distillation. Background Technology

[0002] Analog-to-digital converters (ADCs) are core devices for converting analog signals to digital signals, and are widely used in signal acquisition scenarios such as communications, radar, electronic measurement, and industrial control. Due to the influence of semiconductor manufacturing processes and the inherent characteristics of devices, ADCs have varying degrees of nonlinear errors. When multiple analog signals of different frequencies are input to an ADC, nonlinear errors will generate harmonic and intermodulation components in the output signal. If some of these harmonic and intermodulation components fall into the effective sampling frequency band, they will cause distortion of the effective signal and reduce the spurious-free dynamic range of the ADC. In multi-channel, large dynamic range signal acquisition scenarios, spurious components may also form cascading interference.

[0003] Existing technologies have employed neural networks for ADC nonlinear calibration. For example, CN116015292B uses a fully connected neural network to learn ADC nonlinear errors and calibrate the output data, while CN119788073B extracts and fuses error features through Volterra convolutional branches of different orders. These methods can improve ADC nonlinear errors, but different nonlinear orders are typically represented or fused within a unified network. When some nonlinear orders drift, it becomes difficult to independently identify and locally adjust specific nonlinear orders.

[0004] Therefore, how to provide a lightweight online calibration method for ADC based on knowledge distillation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a lightweight online calibration method for ADCs based on knowledge distillation. This invention constructs an order support domain and order sub-operators, performs knowledge distillation on the order sub-operators to form lightweight order sub-operators, and combines online order residuals to identify the target drift order and update the lightweight order sub-operators. This solves the problems of insufficient distinguishability of different nonlinear orders and large computational load for overall online model updates in existing technologies, and realizes lightweight online calibration oriented towards the target drift order.

[0006] A lightweight online calibration method for an ADC based on knowledge distillation, according to an embodiment of the present invention, includes the following steps:

[0007] S1. Obtain the standard spectral characteristics of the calibration signal and the ADC spectral characteristics of the ADC to be calibrated to obtain the spectral residual;

[0008] S2. Use Taylor polynomials to construct the ADC nonlinear model, and divide the nonlinear terms in the ADC nonlinear model into multiple nonlinear orders.

[0009] S3. Calculate the harmonic frequencies and intermodulation frequencies of different nonlinear orders based on the input frequency of the calibration signal. Form the order support domain by the frequency positions corresponding to the harmonic frequencies and intermodulation frequencies, and extract the spectral components in the order support domain from the spectral residual to obtain the order spectral residual.

[0010] S4. Input the order spectral residual and the frequency position in the order support domain into the improved DeepONet model, introduce an order decomposition mechanism into the backbone network of the improved DeepONet model, and establish order sub-operators according to different nonlinear orders.

[0011] S5. Generate compensation parameters corresponding to different nonlinear orders based on the order sub-operators, and perform knowledge distillation based on the contribution of the order sub-operators to the compensation parameters to obtain lightweight order sub-operators.

[0012] S6. Obtain the online sampling sequence of the ADC to be calibrated and extract the online spectral features. Obtain the online spectral residual based on the online spectral features and the standard spectral features. Extract the online order residual according to the order support domain. Input the online order residual into the lightweight order sub-operator to obtain the online compensation parameters and identify the target drift order.

[0013] S7. Adjust the lightweight order sub-operator of the target drift order according to the online order residual of the target drift order to obtain the updated compensation parameter. Use the updated compensation parameter to perform nonlinear compensation on the online sampling sequence to obtain the calibration sampling sequence.

[0014] Optionally, S1 specifically includes:

[0015] A high-precision signal source is used to generate single-tone calibration signals and two-tone calibration signals. The single-tone calibration signals and two-tone calibration signals are then connected to a high-precision spectrum analyzer to collect frequency domain amplitude, frequency and spurious distribution to form standard spectrum characteristics.

[0016] The same single-tone calibration signal and two-tone calibration signal are input into the ADC to be calibrated. After sampling and quantization, a digital sampling sequence is obtained. The digital sampling sequence is then subjected to a fast Fourier transform to form the ADC spectral characteristics.

[0017] Based on the frequency position matching standard spectral characteristics and ADC spectral characteristics, the difference calculation is performed on the spectral amplitude at the same frequency position to obtain the spectral residual.

[0018] Optionally, S2 specifically includes:

[0019] The analog input signal of the ADC to be calibrated is used as the independent variable, and the digital output signal of the ADC to be calibrated is used as the dependent variable. A Taylor polynomial is constructed according to constant terms, first-order terms, and second-order and above nonlinear terms to obtain the nonlinear model of the ADC.

[0020] Configure the coefficients of the constant term as offset coefficients, configure the coefficients of the first-order term as linear gain coefficients, and configure the coefficients of nonlinear terms of second order and above as nonlinear coefficients.

[0021] The nonlinear orders are divided according to the power of the nonlinear terms in the ADC nonlinear model. The quadratic terms are divided into second-order nonlinear orders, the cubic terms are divided into third-order nonlinear orders, and the remaining nonlinear terms are divided in ascending order of power, resulting in multiple nonlinear orders.

[0022] Optionally, the step of calculating harmonic frequencies and intermodulation frequencies of different nonlinear orders based on the input frequency of the calibration signal, and forming an order support domain by the frequency positions corresponding to the harmonic frequencies and intermodulation frequencies, specifically involves:

[0023] The input frequencies of single-tone calibration signals and two-tone calibration signals are extracted, and the input frequencies are multiplied according to the nonlinear order to obtain harmonic frequencies of different nonlinear orders.

[0024] Two input frequencies are extracted from the two-tone calibration signal, and the two input frequencies are combined by integer multiples according to the nonlinear order to obtain the intermodulation frequencies of different nonlinear orders.

[0025] Based on the sampling frequency of the ADC to be calibrated, the harmonic frequencies and intermodulation frequencies that are outside the effective sampling frequency band are folded back to the effective sampling frequency band to obtain the frequency positions of the harmonic frequencies and intermodulation frequencies.

[0026] Based on the frequency positions of the harmonic frequencies and intermodulation frequencies gathered according to the nonlinear order, order support domains of different nonlinear orders are formed.

[0027] Optionally, the step of extracting spectral components within the order support domain from the spectral residual to obtain the order spectral residual specifically involves:

[0028] The frequency position in the order support domain is read according to the nonlinear order, and the frequency position is converted into the frequency index in the spectral residual according to the frequency resolution of the ADC spectrum characteristics.

[0029] Based on the frequency index, extract the amplitude residuals at the corresponding frequency position and adjacent frequency positions from the spectral residuals;

[0030] The amplitude residuals extracted from the same nonlinear order are arranged in ascending order of frequency to obtain the order spectrum residuals of different nonlinear orders.

[0031] Optionally, S4 specifically includes:

[0032] The improved DeepONet model includes a branch network and a backbone network;

[0033] The order spectral residuals of the same nonlinear order are arranged into residual vectors in ascending order of frequency. The residual vectors are input into the branch network, and residual features are extracted through multi-layer fully connected operations and nonlinear activation operations. Multiple order base coefficients are generated based on the residual features.

[0034] The frequency position and nonlinear order in the order support domain are input into the backbone network, and an order decomposition mechanism is introduced into the backbone network to generate candidate operator bases based on different nonlinear orders.

[0035] By weighting and combining candidate operator bases of the same nonlinear order based on the order basis coefficients, order suboperators of different nonlinear orders can be established.

[0036] Optionally, the order decomposition mechanism is specifically as follows:

[0037] By comparing the order support domains of different nonlinear orders, frequency positions that exist only in one order support domain are classified as exclusive frequency positions, and frequency positions that exist simultaneously in multiple order support domains are classified as shared frequency positions.

[0038] Nonlinear mapping is performed on the exclusive frequency positions to obtain exclusive frequency features of different nonlinear orders, and shared frequency features of shared frequency positions are extracted according to the nonlinear order.

[0039] Following the increasing order of nonlinearity, the shared frequency characteristics of the previous nonlinear order are projected onto the shared frequency characteristics of the next nonlinear order to obtain the projected components.

[0040] The projection component is subtracted from the shared frequency characteristics of the next nonlinear order to obtain the residual frequency characteristics, and the projection and subtraction are repeated in ascending order of nonlinear order.

[0041] By arranging the exclusive frequency features and residual frequency features according to their frequency positions in the order support domain, candidate operator bases for different nonlinear orders are obtained.

[0042] Optionally, S5 specifically includes:

[0043] The order spectral residual is input into the order sub-operators of different nonlinear orders. The nonlinear coefficients of the same nonlinear order in the ADC nonlinear model are fitted according to the output of the order sub-operators, and the compensation parameters are generated according to the nonlinear coefficients.

[0044] By sequentially masking individual candidate operator bases in the order suboperators, keeping the order base coefficients of the unmasked candidate operator bases unchanged, compensation parameters of the same nonlinear order are regenerated.

[0045] Compare the changes in compensation parameters before and after shielding, use the compensation parameters generated after shielding to perform nonlinear compensation on the digital sampling sequence, perform fast Fourier transform on the compensated digital sampling sequence, and extract the spectral components in the order support domain.

[0046] Calculate the amplitude difference between the spectral components in the order support domain and the standard spectral characteristics, normalize and sum the changes in the compensation parameters and the amplitude difference to obtain the contribution value of the candidate operator basis;

[0047] Candidate operator bases are arranged in descending order of contribution value. Candidate operator bases are selected according to a preset lightweighting ratio. The order base coefficients of the selected candidate operator bases are refitted and weighted to obtain lightweight sub-operators of different nonlinear orders.

[0048] Optionally, S6 specifically includes:

[0049] The calibration signal is input into the ADC to be calibrated, the online sampling sequence during the operation of the ADC is collected, and a fast Fourier transform is performed to obtain the online spectral characteristics;

[0050] By matching online spectral characteristics and standard spectral characteristics based on frequency location, the difference calculation is performed on the spectral amplitude at the same frequency location to obtain the online spectral residual;

[0051] Frequency positions are read according to the order support domain of different nonlinear orders. Based on the frequency resolution of the online spectrum characteristics, the frequency positions are converted into frequency indices in the online spectrum residuals. The amplitude residuals of the corresponding frequency positions and adjacent frequency positions are extracted according to the frequency indexes and arranged in ascending order of frequency to obtain the online order residuals of different nonlinear orders.

[0052] Online order residuals of different nonlinear orders are input into lightweight order sub-operators. The nonlinear coefficients of the same nonlinear order in the ADC nonlinear model are fitted according to the output of the lightweight order sub-operators, and online compensation parameters of different nonlinear orders are generated according to the nonlinear coefficients.

[0053] Calculate the absolute difference between online compensation parameters and compensation parameters of the same nonlinear order, arrange the nonlinear orders in descending order of absolute difference, and identify the nonlinear order with the largest absolute difference as the target drift order.

[0054] Optionally, S7 specifically includes:

[0055] The difference between the online order residual and the order spectrum residual of the target drift order is calculated to obtain the residual difference value.

[0056] The residual difference is projected onto the candidate operator basis preserved by the lightweight order sub-operator of the target drift order to obtain the coefficient correction amount of different candidate operator basis;

[0057] The order base coefficients of the target drift order are adjusted according to the coefficient correction amount, the candidate operator basis remains unchanged, and the candidate operator basis is weighted and combined according to the adjusted order base coefficients to obtain the adjusted lightweight order suboperator.

[0058] The online order residual of the target drift order is input into the adjusted lightweight order suboperator to refit the nonlinear coefficients of the target drift order, and the updated compensation parameters of the target drift order are generated based on the nonlinear coefficients.

[0059] The updated compensation parameters of the target drift order are combined with the compensation parameters of nonlinear orders other than the target drift order. The nonlinear components in the online sampling sequence are calculated based on the updated compensation parameters. The nonlinear components are then subtracted from the online sampling sequence to obtain the calibration sampling sequence.

[0060] The beneficial effects of this invention are:

[0061] (1) The present invention constructs an order support domain based on different nonlinear orders and introduces an order decomposition mechanism in the improved DeepONet model to project and subtract the frequency features at the shared frequency position step by step to form a candidate operator basis for different nonlinear orders, thereby reducing feature aliasing at the same frequency position for different nonlinear orders and improving the distinguishability of order sub-operators.

[0062] (2) Knowledge distillation is performed based on the contribution of the order sub-operators to the compensation parameters. Candidate operator bases are selected according to the contribution values ​​of the candidate operator bases and the order base coefficients are refitted to form lightweight order sub-operators. While retaining the main nonlinear compensation information, the number of candidate operator bases is reduced, thereby reducing the model computation load in the online calibration process.

[0063] (3) Generate online compensation parameters based on online order residuals and lightweight order sub-operators. Identify the target drift order by the difference between the online compensation parameters and the compensation parameters. Adjust the lightweight order sub-operators for the target drift order to avoid updating the overall nonlinear order and improve the targeting and updating efficiency of online calibration when the ADC operating state changes. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of a lightweight online calibration method for ADC based on knowledge distillation proposed in this invention;

[0066] Figure 2 This is a structural diagram of the improved DeepONet model proposed in this invention;

[0067] Figure 3 This is a flowchart of the knowledge distillation process proposed in this invention;

[0068] Figure 4 This is a flowchart of the online drift recognition and local update process proposed in this invention. Detailed Implementation

[0069] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0070] refer to Figures 1-4 A lightweight online calibration method for ADC based on knowledge distillation includes the following steps:

[0071] S1. Obtain the standard spectral characteristics of the calibration signal and the ADC spectral characteristics of the ADC to be calibrated to obtain the spectral residual;

[0072] S2. Use Taylor polynomials to construct the ADC nonlinear model, and divide the nonlinear terms in the ADC nonlinear model into multiple nonlinear orders.

[0073] S3. Calculate the harmonic frequencies and intermodulation frequencies of different nonlinear orders based on the input frequency of the calibration signal. Form the order support domain by the frequency positions corresponding to the harmonic frequencies and intermodulation frequencies, and extract the spectral components in the order support domain from the spectral residual to obtain the order spectral residual.

[0074] S4. Input the order spectral residual and the frequency position in the order support domain into the improved DeepONet model. Introduce an order decomposition mechanism into the backbone network of the improved DeepONet model and establish order sub-operators according to different nonlinear orders.

[0075] S5. Generate compensation parameters corresponding to different nonlinear orders based on the order sub-operators, and perform knowledge distillation based on the contribution of the order sub-operators to the compensation parameters to obtain lightweight order sub-operators.

[0076] S6. Obtain the online sampling sequence of the ADC to be calibrated and extract the online spectral features. Obtain the online spectral residual based on the online spectral features and the standard spectral features. Extract the online order residual according to the order support domain. Input the online order residual into the lightweight order sub-operator to obtain the online compensation parameters and identify the target drift order.

[0077] S7. Adjust the lightweight order sub-operator of the target drift order according to the online order residual of the target drift order to obtain the updated compensation parameter. Use the updated compensation parameter to perform nonlinear compensation on the online sampling sequence to obtain the calibration sampling sequence.

[0078] In this embodiment, S1 specifically refers to:

[0079] Select the input frequency of the single-tone calibration signal according to the effective input frequency band of the ADC to be calibrated, and select two different input frequencies within the effective input frequency band to form a two-tone calibration signal. Set the output amplitude of the signal source according to the input amplitude range of the ADC to be calibrated, and use a high-precision signal source to generate the single-tone calibration signal and the two-tone calibration signal respectively.

[0080] Keeping the input frequency and output amplitude of the high-precision signal source constant, the single-tone calibration signal and the dual-tone calibration signal are respectively connected to the high-precision spectrum analyzer. The frequency scanning range of the high-precision spectrum analyzer is set according to the effective sampling frequency band of the ADC to be calibrated, and the frequency position and frequency domain amplitude within the frequency scanning range are collected.

[0081] The frequency positions and frequency domain amplitudes collected by the high-precision spectrum analyzer are arranged in ascending order of frequency. The input frequency of the single-tone calibration signal and the two input frequencies of the two-tone calibration signal are recorded. The frequency positions and frequency domain amplitudes of the spectral peaks outside the input frequencies are also recorded to form a spurious distribution. The frequency, frequency domain amplitude, and spurious distribution are combined to form a standard spectral feature.

[0082] Keeping the input frequency and output amplitude of the single-tone calibration signal and the dual-tone calibration signal unchanged, the single-tone calibration signal and the dual-tone calibration signal are input into the ADC to be calibrated respectively. The number of sampling points of the digital sampling sequence is determined according to the sampling frequency and spectral resolution of the ADC to be calibrated. The digital signal output by the ADC to be calibrated is continuously sampled to obtain the digital sampling sequence.

[0083] Perform a Fast Fourier Transform on the digital sampling sequence with the same number of sampling points. Establish a mapping between the frequency position and the Fast Fourier Transform result based on the sampling frequency of the ADC to be calibrated. Extract the spectral amplitude at different frequency positions and arrange the frequency positions and spectral amplitudes in ascending order of frequency to form the ADC spectral characteristics.

[0084] Align the frequency positions in the standard spectral features and ADC spectral features according to the frequency resolution of the ADC spectral features, and match the standard spectral features and ADC spectral features at the same frequency position;

[0085] Subtract the spectral amplitude in the standard spectral characteristics from the spectral amplitude at the same frequency position in the ADC spectral characteristics, and arrange the frequency positions and spectral amplitude differences in ascending order of frequency to obtain the spectral residual.

[0086] In this embodiment, S2 specifically refers to:

[0087] The input analog signal of the ADC to be calibrated is taken as the independent variable, and the output digital signal of the ADC to be calibrated is taken as the dependent variable. The constant term, linear term and nonlinear term in the Taylor polynomial are constructed according to the zero power, first power and second power and above of the input analog signal. The polynomial terms corresponding to different powers are assigned independent coefficients and summed to form an ADC nonlinear model that represents the nonlinear mapping between the input analog signal and the output digital signal.

[0088] The coefficients of the constant term are configured as offset coefficients, the coefficients of the first term are configured as linear gain coefficients, and the coefficients of nonlinear terms of second degree or higher are configured as nonlinear coefficients. Each nonlinear term is formed by multiplying the corresponding power of the input analog signal with a nonlinear coefficient of the same power.

[0089] The constant term and the first-order term are retained as the linear part of the ADC nonlinear model, and the nonlinear terms of the second order and above are included as the nonlinear part of the ADC nonlinear model, and the nonlinear terms are arranged in ascending order of power.

[0090] Read the power of the input analog signal in each nonlinear term, take the power value as the nonlinear order, divide the quadratic term into the second nonlinear order, divide the cubic term into the third nonlinear order, and divide the remaining nonlinear terms above the quadratic order into the same nonlinear order as the power in the same way.

[0091] By recording different nonlinear orders and their corresponding nonlinear coefficients in ascending order, multiple nonlinear orders in the ADC nonlinear model can be obtained.

[0092] In this embodiment, harmonic frequencies and intermodulation frequencies of different nonlinear orders are calculated based on the input frequency of the calibration signal. The frequency positions corresponding to the harmonic frequencies and intermodulation frequencies are then used to form an order support domain. Specifically:

[0093] Extract the input frequency of the single-tone calibration signal and the two input frequencies of the two-tone calibration signal, and calculate the frequency in sequence according to the second-order nonlinear order, the third-order nonlinear order and higher nonlinear orders;

[0094] For each nonlinear order, the input frequency of the single-tone calibration signal is multiplied by the order value of the nonlinear order to obtain the harmonic frequencies generated by the single-tone calibration signal at different nonlinear orders; for the two-tone calibration signal, the two input frequencies are multiplied by the order value of the nonlinear order respectively to obtain the harmonic frequencies generated by the two input frequencies at different nonlinear orders.

[0095] For the two-tone calibration signal, two non-negative integers are selected at each nonlinear order. The sum of the two non-negative integers is equal to the order value of the nonlinear order. The two input frequencies are multiplied by the two non-negative integers respectively, and the summation and difference operations are performed on the product of the two frequencies. The absolute values ​​of the summation and difference results are used as the intermodulation frequencies of the corresponding nonlinear orders.

[0096] Iterate through the combinations of non-negative integer values ​​that satisfy the order value constraints, collect the intermodulation frequencies obtained from the same nonlinear order, and delete the intermodulation frequencies with duplicate values.

[0097] Based on the sampling frequency of the ADC to be calibrated, the effective sampling frequency band is defined from zero frequency to half of the sampling frequency. For harmonic frequencies and intermodulation frequencies that exceed the effective sampling frequency band, periodic foldback is performed according to the sampling frequency. The harmonic frequency or intermodulation frequency is divided by the sampling frequency to obtain the remainder. If the remainder is not greater than half of the sampling frequency, the remainder is retained. If the remainder is greater than half of the sampling frequency, the remainder is subtracted from the sampling frequency to obtain the frequency position that falls within the effective sampling frequency band.

[0098] Harmonic frequency positions and intermodulation frequency positions after folding are collected according to nonlinear order. For frequency positions with the same value within the same nonlinear order, one frequency position is retained and arranged in ascending order of frequency to form order support domains for different nonlinear orders.

[0099] For the same frequency position that appears in different nonlinear orders, the same frequency position is retained in the order support domain of different nonlinear orders to form the frequency basis for subsequent comparison of the order support domain between different nonlinear orders.

[0100] In this embodiment, the spectral components within the order support domain are extracted from the spectral residual to obtain the order spectral residual, specifically as follows:

[0101] The frequency positions in the order support domain are read sequentially according to the nonlinear order, and the frequency interval between adjacent frequency sampling points in the ADC spectrum features is read as the frequency resolution.

[0102] Divide each frequency position in the order support domain by the frequency resolution, and take the nearest integer of the result to obtain the frequency index of the frequency position in the spectral residual.

[0103] For each frequency index, read the amplitude residual at the position of the frequency index, and read the amplitude residual at the previous and next positions of the frequency index; for frequency indices located at the boundary of the effective sampling frequency band, only read the adjacent positions that fall within the effective sampling frequency band.

[0104] The extraction results are checked according to the frequency position in the order support domain. For frequency indices that appear repeatedly within the same nonlinear order, the amplitude residual is retained once to avoid repeatedly extracting the same spectral components when adjacent frequency positions overlap.

[0105] The extracted amplitude residuals are arranged in ascending order of frequency position in the order support domain, and within each frequency position, the amplitude residuals are arranged in the order of the previous frequency position, the position of the frequency index, and the next frequency position.

[0106] The amplitude residuals after arranging the same nonlinear order are combined to form the order spectrum residuals, and the order spectrum residuals of different nonlinear orders are obtained respectively.

[0107] In this embodiment, S4 specifically refers to:

[0108] The improved DeepONet model includes a branch network and a backbone network. The branch network receives the order spectral residual, and the backbone network receives the frequency position and nonlinear order in the order support domain. The branch network outputs the order basis coefficients, and the backbone network outputs the candidate operator basis.

[0109] Arrange the order spectral residuals of the same nonlinear order in ascending order of frequency, and use the arranged amplitude residuals as elements in the residual vector in sequence. The number of elements in the residual vector is consistent with the number of amplitude residuals contained in the order spectral residuals.

[0110] The residual vector is input into the branch network. The input layer of the branch network receives the amplitude residual in the order of the elements of the residual vector. The residual is then weighted and summed layer by layer through multiple fully connected layers. Nonlinear activation operations are performed between adjacent fully connected layers to convert the amplitude changes at different frequency positions in the order spectrum residual into residual features.

[0111] The number of neurons in the output layer of the branch network is set to the number of candidate operator bases. The residual features are input into the output layer of the branch network, and the residual features are weighted and summed to obtain multiple order base coefficients according to the arrangement order of the neurons in the output layer.

[0112] Each frequency position in the order support domain and the order value of the corresponding nonlinear order are used as two input dimensions of the backbone network input layer. The backbone network is calculated sequentially for the frequency positions in the order support domain. In the backbone network, the frequency characteristics formed by different nonlinear orders at exclusive frequency positions and shared frequency positions are distinguished through the order decomposition mechanism.

[0113] The frequency features processed by the order decomposition mechanism are passed to the output layer of the backbone network. The output layer of the backbone network is set with the same number of neurons as the output layer of the branch network. Each frequency position receives the same output value as the number of order basis coefficients.

[0114] For the same nonlinear order, the output values ​​generated by neurons in the same output layer of the backbone network at different frequency positions are arranged in ascending order of frequency to form a candidate operator basis, and multiple candidate operator bases are formed according to the order of neurons in the output layer of the backbone network.

[0115] The order basis coefficients generated by the branch network are matched one by one with the candidate operator basis generated by the backbone network according to the same arrangement number. Each order basis coefficient is multiplied by the value of the candidate operator basis with the same arrangement number at different frequency positions, and the products obtained at the same frequency position are summed to obtain the output of different nonlinear orders at each frequency position in the order support domain.

[0116] By taking the order spectrum residuals of the same nonlinear order as input and the numerical values ​​obtained by weighted combination of the order basis coefficients and the candidate operator basis as output, order sub-operators of different nonlinear orders are formed.

[0117] In this embodiment, the order decomposition mechanism is specifically as follows:

[0118] Compare the frequency positions in the order support domains of different nonlinear orders, match the frequency positions with the same frequency value, classify the frequency positions that appear only in one order support domain as exclusive frequency positions, and classify the frequency positions that appear in two or more order support domains as shared frequency positions.

[0119] The exclusive frequency position is divided by the highest frequency of the effective sampling frequency band of the ADC to be calibrated to obtain the normalized frequency value. The normalized frequency value and the order value of the nonlinear order are respectively input into the input layer of the backbone network. After being transformed layer by layer by the fully connected operation and nonlinear activation operation in the backbone network, the exclusive frequency characteristics of the exclusive frequency position under the corresponding nonlinear order are obtained.

[0120] For the shared frequency location, the nonlinear order involved in the shared frequency location is read respectively. The normalized frequency value of the same shared frequency location and the order value of different nonlinear orders are input into the backbone network. After the same number of fully connected operations and nonlinear activation operations, the shared frequency characteristics of the shared frequency location under different nonlinear orders are obtained.

[0121] The shared frequency features formed by arranging the same shared frequency position in ascending order of nonlinear order are retained as residual frequency features.

[0122] For two adjacent nonlinear orders, calculate the sum of the products of the same dimension elements in the residual frequency features of the lower-order nonlinear order and the shared frequency features of the higher-order nonlinear order, and calculate the sum of the squared values ​​of the different dimension elements in the residual frequency features of the lower-order nonlinear order. Divide the two sums to obtain the projection ratio.

[0123] Multiply each dimension element of the low-order nonlinear residual frequency feature by the projection ratio to obtain the projection component formed by the high-order nonlinear shared frequency feature along the direction of the low-order nonlinear residual frequency feature.

[0124] Subtracting the elements of the same dimension in the projected component from each dimension element in the shared frequency feature of the higher-order nonlinear order yields the residual frequency feature of the higher-order nonlinear order.

[0125] When there are three or more nonlinear orders sharing the same frequency position, the residual frequency features of the lower nonlinear orders that have been obtained are projected onto the shared frequency features of the higher nonlinear orders in ascending order of nonlinear order. The projection components formed by different lower nonlinear orders are then subtracted from the shared frequency features of the higher nonlinear orders to obtain the residual frequency features of the higher nonlinear orders.

[0126] According to the frequency position in the order support domain, the exclusive frequency position retains the exclusive frequency characteristics, and the shared frequency position retains the residual frequency characteristics of the corresponding nonlinear order. The exclusive frequency characteristics and residual frequency characteristics are arranged in ascending order of frequency.

[0127] The sorted exclusive frequency features and residual frequency features are passed to the output layer of the backbone network. Multiple output values ​​are generated for each frequency position. The output values ​​generated by neurons in the same output layer at all frequency positions of the same nonlinear order are arranged in ascending order of frequency to obtain candidate operator bases for different nonlinear orders.

[0128] In this embodiment, the improved DeepONet model inherits the basic structure of the DeepONet model, which uses a branch network and a backbone network to collaboratively represent nonlinear operators. The branch network receives the order spectral residuals and extracts residual features to generate order basis coefficients reflecting different spectral residual states. The backbone network receives the frequency positions in the order support domain and generates candidate operator bases. The order basis coefficients and candidate operator bases are weighted and combined according to their permutation numbers to form order sub-operators. Building upon the original DeepONet model's backbone network generating a unified operator base based on frequency positions, an order decomposition mechanism is introduced into the backbone network. This mechanism involves the nonlinear order and frequency positions jointly participating in the candidate operator base generation process, comparing the order support domains of different nonlinear orders and distinguishing between exclusive frequency positions and... For shared frequency locations, the shared frequency features formed at different nonlinear orders are projected and subtracted in ascending order of nonlinear order. The feature components formed by lower-order nonlinear orders at shared frequency locations are separated from the shared frequency features of higher-order nonlinear orders. The resulting residual frequency features and exclusive frequency features are arranged according to the order support domain to generate candidate operator bases. The improved DeepONet model can reduce the aliasing of order features between candidate operator bases when the same harmonic frequency or intermodulation frequency location exists at different nonlinear orders. This allows different order sub-operators to represent the spectral residual changes of the corresponding nonlinear orders, providing mutually distinguishable order feature representations for subsequent generation of compensation parameters based on order sub-operators, knowledge distillation, and identification of target drift orders.

[0129] In this embodiment, S5 specifically refers to:

[0130] The order spectral residuals of different nonlinear orders are input into the order sub-operators of the same nonlinear order. The output values ​​of the order sub-operators are read according to the frequency positions in the order support domain. The output values ​​of the order sub-operators are multiplied by the amplitude residuals at the same frequency positions in the order spectral residuals and summed. The summation result is divided by the summation result after the square of the output value of the order sub-operator to obtain the nonlinear coefficients of the same nonlinear order in the ADC nonlinear model.

[0131] Record the nonlinear coefficients according to the nonlinear order, and form compensation parameters by combining the nonlinear order with the nonlinear coefficients of the same nonlinear order. The nonlinear coefficients of the second nonlinear order are configured as second-order compensation parameters, the nonlinear coefficients of the third nonlinear order are configured as third-order compensation parameters, and the compensation parameters of the other nonlinear orders are formed in the same way.

[0132] For the same nonlinear order, individual candidate operator bases are selected sequentially according to the order of candidate operator bases. The order base coefficients corresponding to the selected candidate operator bases are set to zero, while the order base coefficients of the remaining candidate operator bases remain unchanged. The remaining candidate operator bases and order base coefficients are weighted and combined to obtain the order sub-operator after masking individual candidate operator bases.

[0133] The order spectral residual is input into the order sub-operator after shielding a single candidate operator basis. The nonlinear coefficients are refitted according to the same nonlinear order and compensation parameters are generated. The absolute difference between the compensation parameters before and after shielding is calculated.

[0134] According to the nonlinear order in the ADC nonlinear model, the compensation parameters generated after shielding are multiplied by the same nonlinear order power of the digital sampling sequence, and the nonlinear components obtained from different nonlinear orders are subtracted from the digital sampling sequence to obtain the compensation sampling sequence after shielding a single candidate operator basis.

[0135] The compensated sampling sequence is subjected to fast Fourier transform, and the spectral components are extracted according to the order support domain of the same nonlinear order. The difference between the spectral amplitude at each frequency position in the order support domain and the spectral amplitude at the same frequency position in the standard spectral features is calculated, and the absolute difference obtained in the order support domain is summed.

[0136] For all candidate operator bases of the same nonlinear order, the absolute difference of the compensation parameter corresponding to each candidate operator base is divided by the sum of the absolute differences of the compensation parameters of all candidate operator bases to obtain the normalized result of the change of the compensation parameter. The sum of the absolute difference of the spectral amplitude corresponding to each candidate operator base is divided by the sum of the absolute differences of the spectral amplitude of all candidate operator bases to obtain the normalized result of the spectral amplitude difference.

[0137] The two normalization results corresponding to the same candidate operator basis are added together to obtain the contribution value of the candidate operator basis, and the candidate operator basis of the same nonlinear order are arranged in descending order of contribution value.

[0138] Based on the available storage space of the hardware running the improved DeepONet model and the storage space occupied by a single candidate operator basis, the number of candidate operator bases that can be retained during the online calibration process is determined, and the ratio of the number of candidate operator bases that can be retained to the number of candidate operator bases contained in the order sub-operator is used as the lightweight ratio.

[0139] The number of candidate operator bases to be retained for different nonlinear orders is determined based on the weighting ratio, and the corresponding number of candidate operator bases are selected in descending order of contribution value. The order base coefficients of the unselected candidate operator bases are set to 0.

[0140] For the selected candidate operator basis, the candidate operator basis is multiplied by the order basis coefficients and then summed at the same frequency position. The summation result is compared with the output value of the order sub-operator before lightweighting at the same frequency position. The order basis coefficients of the selected candidate operator basis are adjusted until the sum of the squared differences of all frequency positions in the order support domain reaches the minimum.

[0141] The adjusted order basis coefficients are matched and weighted with the selected candidate operator basis according to the arrangement number to form lightweight order suboperators of different nonlinear orders.

[0142] In this embodiment, S6 specifically refers to:

[0143] To ensure that the input frequency and output amplitude of the calibration signal are consistent with those used when forming the standard spectral characteristics, the calibration signal is input into the ADC to be calibrated. Digital signals are acquired according to the sampling frequency and number of sampling points used when forming the ADC spectral characteristics. The sampled values ​​are arranged in the order of sampling time to obtain the online sampling sequence.

[0144] Perform a Fast Fourier Transform on the online sampling sequence with the same number of sampling points as the number of sampling points. Establish the correspondence between the frequency position and the Fast Fourier Transform result based on the sampling frequency of the ADC to be calibrated. Extract the spectral amplitude at different frequency positions and arrange the frequency positions and spectral amplitudes in ascending order of frequency to obtain the online spectral characteristics.

[0145] Matching online and standard spectral features based on frequency position, subtracting the spectral amplitude in the standard spectral feature from the spectral amplitude at the same frequency position in the online spectral feature, and arranging the difference between the frequency position and the spectral amplitude in ascending order of frequency, the online spectral residual is obtained.

[0146] Read the frequency positions in the order support domain according to different nonlinear orders, divide each frequency position by the frequency resolution of the online spectrum feature and take the nearest integer to obtain the frequency index in the online spectrum residual, and read the amplitude residual of the corresponding frequency position and the adjacent frequency position according to the frequency index.

[0147] The amplitude residuals extracted from the same nonlinear order are arranged in ascending order of frequency position in the order support domain, and the amplitude residuals are arranged in the order of the previous frequency position, the position of the frequency index, and the next frequency position within each frequency position to obtain the online order residuals of different nonlinear orders.

[0148] The online order residuals of different nonlinear orders are input into the lightweight order sub-operator of the same nonlinear order, and the output values ​​of the lightweight order sub-operator at different frequency positions in the order support domain are read. Based on the output values ​​and the amplitude residuals at the same frequency positions in the online order residuals, the nonlinear coefficients of the same nonlinear order in the ADC nonlinear model are fitted.

[0149] Based on the nonlinear coefficients obtained by online fitting according to the nonlinear order, the nonlinear order and the nonlinear coefficients of the same nonlinear order are used to form online compensation parameters, and these parameters are matched one by one with the compensation parameters of the same nonlinear order generated before knowledge distillation.

[0150] The absolute difference between the online compensation parameter and the compensation parameter of the same nonlinear order is calculated. The absolute difference represents the parameter change of the corresponding nonlinear order during the operation of the ADC to be calibrated relative to the generation of the compensation parameter. The nonlinear orders are arranged in descending order of absolute difference, and the nonlinear order with the largest absolute difference is identified as the target drift order.

[0151] In this embodiment, S7 specifically refers to:

[0152] According to the frequency position in the order support domain, the online order residual of the target drift order is matched with the order spectrum residual of the target drift order item by item. The amplitude residual in the online order residual is subtracted from the order spectrum residual at the same frequency position, and the residual difference is obtained by arranging them in ascending order of frequency.

[0153] For each candidate operator basis retained by the lightweight order sub-operator of the target drift order, the values ​​at different frequency positions in the residual difference are multiplied by the values ​​of the candidate operator basis at the same frequency position and summed. The summation result is divided by the sum of the squares of the values ​​of the candidate operator basis at different frequency positions to obtain the coefficient correction amount of the candidate operator basis.

[0154] According to the arrangement number of the candidate operator basis, the coefficient correction amount is added to the order basis coefficients with the same arrangement number to obtain the adjusted order basis coefficients, while keeping the candidate operator basis retained by the lightweight order sub-operator unchanged.

[0155] The adjusted order base coefficients are multiplied by the candidate operator bases with the same arrangement number, and the products generated at the same frequency position are summed to obtain the updated output of the target drift order at different frequency positions in the order support domain, forming the adjusted lightweight order suboperator.

[0156] Input the online order residual of the target drift order into the adjusted lightweight order sub-operator, read the update output of the adjusted lightweight order sub-operator at different frequency positions, multiply the update output with the amplitude residual at the same frequency position in the online order residual item by item and sum them, divide the sum by the sum of the squared update output to obtain the nonlinear coefficient of the target drift order after update.

[0157] The target drift order and the updated nonlinear coefficients are used to form the updated compensation parameters for the target drift order. The compensation parameters for nonlinear orders other than the target drift order are kept unchanged. The updated compensation parameters and the compensation parameters for the other nonlinear orders are arranged in ascending order of nonlinear order.

[0158] For each sampled value in the online sampling sequence, the corresponding power of the sampled value is calculated according to the nonlinear order. The results of different powers are multiplied by the compensation parameter of the same nonlinear order to obtain the nonlinear components of different nonlinear orders.

[0159] The nonlinear components of different nonlinear orders at the same sampling time are added together, and the summation result is subtracted from the sampled values ​​at the same sampling time in the online sampling sequence. The subtracted sampled values ​​are then arranged in the order of sampling time to obtain the calibration sampling sequence.

[0160] Example 1: To verify the feasibility of this invention in practice, it was applied to the online calibration of a 14-bit, 125MSPS ADC in a communication receiving device. During continuous operation, the ADC is affected by rising operating temperature and changes in device state, causing varying degrees of drift in second-order, third-order, and higher-order nonlinear coefficients. This results in an increase in harmonics and intermodulation components in the output spectrum. Fixed compensation parameters are insufficient to maintain stable calibration results, and readjusting the complete model leads to significant online computation time and parameter update requirements.

[0161] The test employed a high-precision signal source to generate a 10MHz single-tone calibration signal and 18MHz and 21MHz dual-tone calibration signals. A high-precision spectrum analyzer acquired standard spectral characteristics. The same calibration signal was input into the ADC to be calibrated, and 16384 sampling points were acquired at a sampling frequency of 125MSPS. The ADC spectral characteristics were obtained through Fast Fourier Transform. Based on the standard spectral characteristics and the ADC spectral characteristics, spectral residuals were obtained. A Taylor polynomial was used to construct the ADC nonlinear model, and the nonlinear orders were divided into second to fifth orders. Order support domains were established based on harmonic frequencies and intermodulation frequencies, and order spectral residuals were extracted from the spectral residuals. The order spectral residuals were input into an improved DeepONet model. Order basis coefficients were generated through a branch network, and candidate operator bases were generated through a backbone network introducing an order decomposition mechanism. Order sub-operators were then established. Knowledge distillation was performed based on the contribution of the candidate operator bases to the compensation parameters, reducing the number of candidate operator bases retained for each order sub-operator from 32 to 12, forming lightweight order sub-operators.

[0162] The equipment temperature was gradually increased from 25℃ to 70℃. Online sampling sequences were collected at each temperature. Online compensation parameters were generated based on the online order residuals. The target drift order was identified based on the absolute difference between the online compensation parameters and the compensation parameters. The test results are shown in Table 1.

[0163] Table 1. Target drift order identification and calibration results under different operating conditions.

[0164] 25 0.0031 Second order 56.8 73.6 16.8 40 0.0074 Third order 55.9 73.1 17.2 55 0.0128 Third order 54.7 72.5 17.8 65 0.0186 Fourth level 53.8 71.9 18.1 70 0.0237 Fourth level 52.9 71.2 18.3

[0165] As shown in Table 1, with increasing operating temperature, the absolute difference in compensation parameters increased from 0.0031 to 0.0237, and the target drift order gradually shifted from second to third and fourth order, indicating that different nonlinear orders do not change synchronously. This invention can locate the target drift order based on online compensation parameters and adjust the lightweight order sub-operator accordingly. At 70℃, the SFDR increased from 52.9 dBc to 71.2 dBc, an improvement of 18.3 dB, demonstrating that it can maintain stable nonlinear compensation capability even under local order drift conditions.

[0166] Further comparisons were made between the fixed compensation parameters, the overall model update, and the online calibration performance of the present invention. 100 sets of online sampling sequences were processed continuously, and the results are shown in Table 2.

[0167] Table 2 Performance Comparison of Different ADC Online Calibration Methods

[0168] Fixed compensation parameters 0 0 0 61.4 1.26 Global order suboperator update 128 4096 436 73.5 45.1 Lightweight order suboperator overall update 48 1536 171 73.2 18.6 This invention updates the target drift order. 12 384 46 72.9 6.1

[0169] As shown in Table 2, although there is no model update time with fixed compensation parameters, the average SFDR is only 61.4 dBc, which cannot adapt to the drift of ADC nonlinear parameters. The overall order sub-operator update can reach 73.5 dBc, but it requires updating 4096 parameters per cycle, with an average update time of 436 ms. This invention reduces the candidate operator basis through knowledge distillation and only adjusts the lightweight order sub-operator of the target drift order. The number of parameters updated per cycle is reduced to 384, the average update time is reduced to 46 ms, and the average SFDR still reaches 72.9 dBc, significantly reducing the online update computation while maintaining calibration performance.

[0170] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A lightweight online calibration method for an ADC based on knowledge distillation, characterized in that, Includes the following steps: S1. Obtain the standard spectral characteristics of the calibration signal and the ADC spectral characteristics of the ADC to be calibrated to obtain the spectral residual; S2. Use Taylor polynomials to construct the ADC nonlinear model, and divide the nonlinear terms in the ADC nonlinear model into multiple nonlinear orders. S3. Calculate the harmonic frequencies and intermodulation frequencies of different nonlinear orders based on the input frequency of the calibration signal. Form the order support domain by the frequency positions corresponding to the harmonic frequencies and intermodulation frequencies, and extract the spectral components in the order support domain from the spectral residual to obtain the order spectral residual. S4. Input the order spectral residual and the frequency position in the order support domain into the improved DeepONet model, introduce an order decomposition mechanism into the backbone network of the improved DeepONet model, and establish order sub-operators according to different nonlinear orders. S5. Generate compensation parameters corresponding to different nonlinear orders based on the order sub-operators, and perform knowledge distillation based on the contribution of the order sub-operators to the compensation parameters to obtain lightweight order sub-operators. S6. Obtain the online sampling sequence of the ADC to be calibrated and extract the online spectral features. Obtain the online spectral residual based on the online spectral features and the standard spectral features. Extract the online order residual according to the order support domain. Input the online order residual into the lightweight order sub-operator to obtain the online compensation parameters and identify the target drift order. S7. Adjust the lightweight order sub-operator of the target drift order according to the online order residual of the target drift order to obtain the updated compensation parameter. Use the updated compensation parameter to perform nonlinear compensation on the online sampling sequence to obtain the calibration sampling sequence.

2. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, Specifically, S1 is: A high-precision signal source is used to generate single-tone calibration signals and two-tone calibration signals. The single-tone calibration signals and two-tone calibration signals are then connected to a high-precision spectrum analyzer to collect frequency domain amplitude, frequency and spurious distribution to form standard spectrum characteristics. The same single-tone calibration signal and two-tone calibration signal are input into the ADC to be calibrated. After sampling and quantization, a digital sampling sequence is obtained. The digital sampling sequence is then subjected to a fast Fourier transform to form the ADC spectral characteristics. Based on the frequency position matching standard spectral characteristics and ADC spectral characteristics, the difference calculation is performed on the spectral amplitude at the same frequency position to obtain the spectral residual.

3. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, Specifically, S2 is: The analog input signal of the ADC to be calibrated is used as the independent variable, and the digital output signal of the ADC to be calibrated is used as the dependent variable. A Taylor polynomial is constructed according to constant terms, first-order terms, and second-order and above nonlinear terms to obtain the nonlinear model of the ADC. Configure the coefficients of the constant term as offset coefficients, configure the coefficients of the first-order term as linear gain coefficients, and configure the coefficients of nonlinear terms of second order and above as nonlinear coefficients. The nonlinear orders are divided according to the power of the nonlinear terms in the ADC nonlinear model. The quadratic terms are divided into second-order nonlinear orders, the cubic terms are divided into third-order nonlinear orders, and the remaining nonlinear terms are divided in ascending order of power, resulting in multiple nonlinear orders.

4. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, The step involves calculating harmonic frequencies and intermodulation frequencies of different nonlinear orders based on the input frequency of the calibration signal, and forming an order support domain by locating the frequency positions corresponding to the harmonic frequencies and intermodulation frequencies. Specifically: The input frequencies of single-tone calibration signals and two-tone calibration signals are extracted, and the input frequencies are multiplied according to the nonlinear order to obtain harmonic frequencies of different nonlinear orders. Two input frequencies are extracted from the two-tone calibration signal. The two input frequencies are combined by integer multiples according to the nonlinear order to obtain the intermodulation frequencies of different nonlinear orders. Based on the sampling frequency of the ADC to be calibrated, the harmonic frequencies and intermodulation frequencies that are outside the effective sampling frequency band are folded back to the effective sampling frequency band to obtain the frequency positions of the harmonic frequencies and intermodulation frequencies. Based on the frequency positions of the harmonic frequencies and intermodulation frequencies gathered according to the nonlinear order, order support domains of different nonlinear orders are formed.

5. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, The step of extracting spectral components within the order support domain from the spectral residual to obtain the order spectral residual is specifically as follows: The frequency position in the order support domain is read according to the nonlinear order, and the frequency position is converted into the frequency index in the spectral residual according to the frequency resolution of the ADC spectral characteristics. Based on the frequency index, extract the amplitude residuals at the corresponding frequency position and adjacent frequency positions from the spectral residuals; The amplitude residuals extracted from the same nonlinear order are arranged in ascending order of frequency to obtain the order spectrum residuals of different nonlinear orders.

6. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, Specifically, S4 is: The improved DeepONet model includes a branch network and a backbone network; The order spectral residuals of the same nonlinear order are arranged into residual vectors in ascending order of frequency. The residual vectors are input into the branch network, and residual features are extracted through multi-layer fully connected operations and nonlinear activation operations. Multiple order base coefficients are generated based on the residual features. The frequency position and nonlinear order in the order support domain are input into the backbone network, and an order decomposition mechanism is introduced into the backbone network to generate candidate operator bases based on different nonlinear orders. By weighting and combining candidate operator bases of the same nonlinear order based on the order basis coefficients, order suboperators of different nonlinear orders can be established.

7. The lightweight online calibration method for ADC based on knowledge distillation according to claim 6, characterized in that, The order decomposition mechanism is as follows: By comparing the order support domains of different nonlinear orders, frequency positions that exist only in one order support domain are classified as exclusive frequency positions, and frequency positions that exist simultaneously in multiple order support domains are classified as shared frequency positions. Nonlinear mapping is performed on the exclusive frequency positions to obtain exclusive frequency features of different nonlinear orders, and shared frequency features of shared frequency positions are extracted according to the nonlinear order. Following the increasing order of nonlinearity, the shared frequency characteristics of the previous nonlinear order are projected onto the shared frequency characteristics of the next nonlinear order to obtain the projected components. The projection component is subtracted from the shared frequency characteristics of the next nonlinear order to obtain the residual frequency characteristics, and the projection and subtraction are repeated in ascending order of nonlinear order. By arranging the exclusive frequency features and residual frequency features according to their frequency positions in the order support domain, candidate operator bases for different nonlinear orders are obtained.

8. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, Specifically, S5 is: The order spectral residual is input into the order sub-operators of different nonlinear orders. The nonlinear coefficients of the same nonlinear order in the ADC nonlinear model are fitted according to the output of the order sub-operators, and the compensation parameters are generated according to the nonlinear coefficients. By sequentially masking individual candidate operator bases in the order suboperators, keeping the order base coefficients of the unmasked candidate operator bases unchanged, compensation parameters of the same nonlinear order are regenerated. Compare the changes in compensation parameters before and after shielding, use the compensation parameters generated after shielding to perform nonlinear compensation on the digital sampling sequence, perform fast Fourier transform on the compensated digital sampling sequence, and extract the spectral components in the order support domain. Calculate the amplitude difference between the spectral components in the order support domain and the standard spectral characteristics, normalize and sum the changes in the compensation parameters and the amplitude difference to obtain the contribution value of the candidate operator basis; Candidate operator bases are arranged in descending order of contribution value. Candidate operator bases are selected according to a preset lightweighting ratio. The order base coefficients of the selected candidate operator bases are refitted and weighted to obtain lightweight sub-operators of different nonlinear orders.

9. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, Specifically, S6 is: The calibration signal is input into the ADC to be calibrated, the online sampling sequence during the operation of the ADC is collected, and a fast Fourier transform is performed to obtain the online spectral characteristics; By matching online spectral characteristics and standard spectral characteristics based on frequency location, the difference calculation is performed on the spectral amplitude at the same frequency location to obtain the online spectral residual; Frequency positions are read according to the order support domain of different nonlinear orders. Based on the frequency resolution of the online spectrum characteristics, the frequency positions are converted into frequency indices in the online spectrum residuals. The amplitude residuals of the corresponding frequency positions and adjacent frequency positions are extracted according to the frequency indexes and arranged in ascending order of frequency to obtain the online order residuals of different nonlinear orders. Online order residuals of different nonlinear orders are input into lightweight order sub-operators. The nonlinear coefficients of the same nonlinear order in the ADC nonlinear model are fitted according to the output of the lightweight order sub-operators, and online compensation parameters of different nonlinear orders are generated according to the nonlinear coefficients. Calculate the absolute difference between online compensation parameters and compensation parameters of the same nonlinear order, arrange the nonlinear orders in descending order of absolute difference, and identify the nonlinear order with the largest absolute difference as the target drift order.

10. The lightweight online calibration method for ADC based on knowledge distillation according to claim 1, characterized in that, Specifically, S7 is: The difference between the online order residual and the order spectrum residual of the target drift order is calculated to obtain the residual difference value. The residual difference is projected onto the candidate operator basis preserved by the lightweight order sub-operator of the target drift order to obtain the coefficient correction amount of different candidate operator basis; The order base coefficients of the target drift order are adjusted according to the coefficient correction amount, the candidate operator basis remains unchanged, and the candidate operator basis is weighted and combined according to the adjusted order base coefficients to obtain the adjusted lightweight order suboperator. The online order residual of the target drift order is input into the adjusted lightweight order suboperator to refit the nonlinear coefficients of the target drift order, and the updated compensation parameters of the target drift order are generated based on the nonlinear coefficients. The updated compensation parameters of the target drift order are combined with the compensation parameters of nonlinear orders other than the target drift order. The nonlinear components in the online sampling sequence are calculated based on the updated compensation parameters. The nonlinear components are then subtracted from the online sampling sequence to obtain the calibration sampling sequence.

Images