Blind spot-aware joint calibration method for time-interleaved adc clock skew error

Abstract

The application claims a time-interleaved ADC clock skew blind spot sensing joint calibration method and an implementation system. The method folds a target input frequency to a single-channel sampling bandwidth, generates a blind spot ratio or blind spot weight according to the distance of the folded frequency from 0, f ch / 4, f ch / 2, and other blind spot reference points; when the target frequency observation degenerates, a joint observation set containing at least an auxiliary frequency observation is constructed, and each frequency observation shares the same set of channel clock skew variables; and the channel clock skew is jointly estimated in combination with phase preliminary estimation, linearized correlation preliminary estimation, frequency weight, and regularization constraint. The scheme can stably suppress interleaved spur in special frequency points and Nyquist frequency band without increasing additional ADC channels, and is suitable for on-chip digital calibration implementation.

Description

Technical Field

[0001] This invention belongs to the fields of mixed-signal integrated circuit design, high-speed data converters and digital back-end calibration technology. Specifically, it relates to a method and system for estimating inter-channel clock skew error, identifying blind spot frequencies, solving multiple frequencies together and outputting skew estimates for time-interleaved analog-to-digital converters. Background Technology

[0002] Time-interleaved ADCs sample sequentially through multiple sub-ADC channels, increasing the system's equivalent sampling rate to M times that of a single channel. This architecture is commonly used in broadband communication receivers, high-speed oscilloscope measurements, radar front-ends, and high-speed link testing. In actual chips, channel mismatch is unavoidable. Offset mismatch and gain mismatch can be corrected relatively directly using the mean, amplitude, and multiplication coefficients. Clock skew between channels is a sampling timing error; the higher the input frequency, the easier it is for the same amount of skew to translate into significant phase error and image spurious signals.

[0003] For example, Chinese patent application CN119420355A, entitled "Digital Clock Skew Calibration Method Based on Similar Triangles," discloses a digital clock skew calibration approach that uses frequency domain phase relationships to estimate clock skew. While this type of frequency domain phase method is easy to implement digitally, the phase slope and phase difference statistics tend to degrade when the input frequency is folded to near DC or the single-channel Nyquist frequency, and the estimated values ​​are more sensitive to noise, leakage, and spurious signals.

[0004] For example, Yuan Jun et al. discussed a calibration algorithm for inter-channel clock bias in clock-interleaved ADCs in their paper "Algorithm for Calibrating Inter-channel Clock Bias in Clock-Interleaved ADCs". This type of method can improve inter-channel clock bias at ordinary frequencies, but it usually does not explicitly construct a blind reference set for special frequency families such as 0, fch / 4 and fch / 2 in the frequency folding domain, nor does it construct an auxiliary frequency observation stack that shares the same physical bias solution with the target frequency for the neighborhood of special frequency points.

[0005] Existing clock skew calibration methods often use phase difference, correlation, or adjacent sample difference components at a single frequency as skew indicators. These methods exhibit good convergence speed and low implementation complexity at common frequencies. However, when the input frequency is folded close to DC, a single-channel quarter-sampling rate, or a single-channel Nyquist frequency, phenomena such as repetition, symmetry, overlapping positive and negative frequency aliases, or phase reversal occur in the channel sampling sequence. This reduces the sensitivity of the skew information to the observed quantities, and decreases the reliability of the denominator, equivalent slope, or phase residual. Continuing to use single-frequency updates at common frequencies will cause the estimated value to jump due to quantization noise, spectral leakage, and spurious components, and the compensation amount may be incorrectly written to the register.

[0006] The blind spot referred to in this article is the area in the frequency folding domain where the folding frequency is close to 0 or f. ch / 4 or f ch At / 2, the sensitivity of clock skew to observations of phase, correlation, or equivalent slope decreases significantly, leading to frequency regions where estimated statistics degrade.

[0007] Simply expanding the observation window or reducing the update rate can only reduce random fluctuations, but cannot determine whether the current frequency point has lost observability. Manually avoiding special frequencies is not suitable for broadband frequency sweeping, background adaptive calibration, and on-chip automatic testing. A calibration scheme is needed that actively identifies blind spots within the algorithm, uses adjacent frequency redundancy information to supplement degraded observations, and maintains hardware feasibility. Summary of the Invention

[0008] The technical problem this invention aims to solve is that existing time-interleaved ADC clock skew calibration methods can achieve effective estimation at normal frequency points, but when the input frequency is folded down to 0, f... ch / 4 or f ch Near specific frequency groups such as / 2, single-frequency phase or correlation statistics degrade, easily leading to estimation jumps, erroneous updates, and performance collapse at specific frequencies. A joint calibration method for time-interleaved ADC clock skew blind spot sensing is proposed. The technical solution of this invention is as follows:

[0009] A blind spot-aware joint calibration method for time-interleaved ADC clock skew error includes the following steps:

[0010] The target input frequency is folded into a single-channel sampling bandwidth to obtain the folded frequency. A blind spot ratio or blind spot weight is generated based on the distance between the folded frequency and the blind spot reference set. A joint observation set is constructed based on the blind spot ratio or blind spot weight, wherein when the blind spot ratio does not exceed a preset threshold, the joint observation set includes the target input frequency observation; when the blind spot ratio exceeds the preset threshold, the joint observation set includes at least one folded auxiliary frequency observation farthest from the nearest blind spot reference point, and the target input frequency observation and / or the auxiliary frequency observation share the same set of channel clock skew variables. At least one of frequency weight, phase position confidence, or regularization constraint is determined based on the blind spot ratio or blind spot weight, and the joint observation set is used to jointly estimate the same set of channel clock skew variables to obtain the channel clock skew estimate.

[0011] Furthermore, the time-interleaved ADC includes M interleaved channels, with a total sampling rate of f. s The single-channel sampling rate is f ch =f s / M, the folding frequency f fold We obtain it using the following formula, where r f Represents the target input frequency fin For a single-channel sampling rate f ch Take the remainder, f fold Located in [0, f ch / 2]:

[0012]

[0013]

[0014] Furthermore, the blind spot reference set includes at least 0 and f. ch / 4 and f ch / 2; When the distance from the folded frequency to the nearest reference point in the blind spot reference set is less than the preset protection radius, the target input frequency is determined to be located in a special frequency protection zone; the blind spot ratio is determined by the minimum distance from the folded frequency to the blind spot reference set and the blind spot protection radius. The larger the blind spot ratio, the lower the phase observation confidence corresponding to the current target input frequency.

[0015]

[0016] Where d represents the distance from the folding frequency to the nearest reference point in the blind reference set, and f b F represents any reference frequency point in the blind spot reference set. b R represents the blind spot reference set. b ρ represents the blind spot protection radius, and ρ represents the blind spot ratio.

[0017] Furthermore, the auxiliary frequency observation is obtained by screening a preset candidate frequency pool or by constructing frequency offsets on both sides of the target input frequency; the auxiliary frequencies are sorted according to at least one of blind spot distance, spectral leakage degree, amplitude stability and phase consistency.

[0018] Furthermore, when the folded frequency is close to the 0-frequency family, the observation weight of the target input frequency is reduced, or the target input frequency is removed from the joint observation set, and the auxiliary frequency that is far from the 0-frequency point after folding is selected as the main observation; when the folded frequency is close to f ch When dealing with a frequency family of / 2, reduce the observation weight of the target input frequency, or remove the target input frequency from the joint observation set, and select the frequency that is far from f after folding. ch The auxiliary frequency at / 2 frequency point is used as the main observation; when the folding frequency is close to f ch When the frequency family is / 4, the enhanced constraint mode is activated. The enhanced constraint mode includes at least one of the following: increasing the initial linearization correlation weight, increasing the smoothing constraint weight, increasing the prior constraint weight, and limiting the update step size.

[0019] Furthermore, the joint estimation includes: extracting the complex spectrum phase of each channel for each frequency observation in the joint observation set, and removing the common phase and ideal interleaved phase trend to obtain an initial phase estimate; constructing residual and Jacobian matrices based on the first-order approximation of the channel signal perturbation at the sampling time to obtain a linearized correlation initial estimate; generating fusion coefficients based on the blind spot ratio and phase position confidence, wherein the larger the blind spot ratio or the lower the phase position confidence, the greater the fusion weight of the linearized correlation initial estimate; and performing at least one of zero-mean constraint, smoothing, or amplitude limitation on the fused initial skew vector.

[0020] Furthermore, the joint estimation constructs an objective function using a shared channel clock skew vector as the independent variable. The objective function includes at least two of the following: a weighted multi-frequency observation residual term for each frequency observation in the joint observation set; a second-order difference smoothing term for adjacent channels; and a prior term relative to the initial estimate or historical effective estimate. Specifically, the larger the proportion of blind spots corresponding to a frequency observation, the smaller the weight of that frequency observation, and the larger the weight of the smoothing term or the prior term.

[0021]

[0022] Where L(τ) represents the objective function value, τ represents the shared channel clock skew vector, Q represents the number of frequency observations in the joint observation set, q represents the frequency observation sequence number, and w q Let x represent the weight of the q-th frequency observation. q This represents the channel observation corresponding to the q-th frequency. The model output is represented by the channel clock skew vector τ, where D represents the second-order difference matrix between adjacent channels, and λ represents the output of the model. s λ represents the smoothing constraint weight. p τ represents the prior constraint weight, and τ0 represents the initial estimate or the historical effective estimate.

[0023] Furthermore, the joint estimation employs a damped iterative solution, and the damping factor is adaptively adjusted or the step size is updated based on changes in the objective function, the step size, or the decrease in the residual.

[0024] Furthermore, under severe blind spot or low phase position confidence conditions, multiple candidate channel clock skew estimates are generated, and the candidate channel clock skew estimates are scored based on at least two of the following: multi-frequency observation residual, smoothness, relative initial estimate deviation, and amplitude limiting state. The estimated value that meets the preset rules is then output.

[0025] The estimated channel clock skew is subjected to hierarchical quantization; when the residual decrease after the previous quantization is less than the preset threshold, the number of iterations reaches the preset value, or the blind spot ratio is lower than the preset value, the estimation is continued by switching to a smaller quantization step size.

[0026] A time-interleaved ADC clock skew error calibration system for implementing the above method includes a frequency folding module, a blind spot ratio generation module, an auxiliary frequency observation construction module, a multi-frequency joint estimation module, a candidate solution scoring module, and an estimate output module. The frequency folding module folds the target input frequency to a single-channel sampling bandwidth and outputs the folded frequency. The blind spot ratio generation module generates a blind spot ratio or blind spot weight based on the distance between the folded frequency and the blind spot reference set. The auxiliary frequency observation construction module filters or constructs auxiliary frequency observations that are folded and farthest from the nearest blind spot reference point when the blind spot ratio exceeds a preset threshold. The multi-frequency joint estimation module performs joint estimation based on the channel clock skew variables shared by the target input frequency observation and / or the auxiliary frequency observations. The candidate solution scoring module scores multiple candidate channel clock skew estimates under severe blind spot or low phase position confidence conditions. The estimate output module outputs or writes channel clock skew estimates that meet preset scoring rules through a clock skew control interface.

[0027] Furthermore, the calibration system can be configured in the digital calibration circuit of the time-interleaved ADC chip. The digital calibration circuit connects M sub-ADC channels and a clock skew control interface, and outputs or writes channel clock skew estimates through the clock skew control interface.

[0028] The advantages and beneficial effects of this invention are as follows:

[0029] This invention can stably suppress interleaved spurious signals at specific frequency families, ensuring continuous calibration performance within the 0 to the first Nyquist band of the system; it is compatible with front-end calibration, back-end calibration, and on-chip automatic testing; the new processing is concentrated on the digital end, eliminating the need for additional ADC channels, and is easy to implement in on-chip digital calibration circuits.

[0030] Distinguishing features from related technologies

[0031] Compared to the clock skew digital calibration method based on similar triangles disclosed in CN119420355A, this invention does not simply rely on the frequency domain phase relationship at the target input frequency to obtain the clock skew. Instead, after the input frequency is folded to the single-channel sampling bandwidth, it first determines whether the folded frequency is close to 0. ch / 4 or f ch / 2 blind reference points; when the reliability of phase observations at the target frequency decreases, this invention introduces auxiliary frequency observations that are folded away from the blind reference points, and jointly solves the target frequency observations and auxiliary frequency observations on the same set of channel clock skew variables. Therefore, the distinguishing features of this invention are the blind reference set, the blind point ratio or blind point weight, the auxiliary frequency observation stack, and the multi-frequency joint estimation with shared skew variables.

[0032] Compared to the time mismatch calibration approach for time-interleaved ADCs based on correlation statistics described in CN117749181A, this invention does not solely rely on correlation coefficients or correlation quantities to determine inter-channel time deviations. Instead, it uses the initial correlation estimate as one of the stable seeds under blind spot conditions, and integrates it with the initial phase estimate in the frequency domain, frequency weights, smoothing constraints, and prior constraints. The key difference lies in the fact that the initial correlation estimate in this invention is controlled by the blind spot ratio and phase position confidence, preserving phase estimation accuracy at ordinary frequencies and enhancing the effects of linearized initial correlation estimates and regularization constraints in the blind spot neighborhood. This avoids erroneous updates caused by the degradation of a single correlation statistic or a single phase statistic at specific frequencies.

[0033] Compared to the technique in CN111064469A that uses the autocorrelation function of adjacent channels for time error calibration, this invention is not limited to the single time-domain statistical relationship of adjacent channels. Instead, it actively identifies special frequency protection zones in the folded frequency domain and constructs a joint observation set based on the blind spot distance, spectral leakage, amplitude stability, or phase consistency of auxiliary frequencies. In other words, this invention uses "whether the current frequency is suitable for estimation" as an explicit judgment condition in the calibration process, and supplements observability with auxiliary frequencies when it is not suitable for estimation. This differs from conventional adjacent-channel correlation calibration schemes that directly update the skew estimate under target input conditions.

[0034] In summary, the essential difference between this invention and the aforementioned related technologies lies not in using phase methods, correlation methods, or damped iterative solutions individually, but in combining folded domain blind spot identification, auxiliary frequency selection, multi-frequency shared skew variable modeling, blind spot weight scheduling, and candidate solution scoring into a closed-loop calibration process oriented towards stability at specific frequency points. This process can maintain estimation continuity near DC families, quarter-channel single-channel sampling rate families, and single-channel Nyquist families, and reduces the risk of erroneous compensation due to blind spot observation degradation. Attached Figure Description

[0035] Figure 1 This is a flowchart of a preferred embodiment of the blind spot perception joint calibration method provided by the present invention;

[0036] Figure 2 This is a schematic diagram of the special frequency protection zone and auxiliary frequency structure;

[0037] Figure 3 It is a 2.5 GHz input folded to f ch / 4 FFT calibration effect diagram for special frequency families;

[0038] Figure 4 This is an FFT calibration result diagram when the 28 GHz input is folded to the 0 frequency family;

[0039] Figure 5This is a diagram showing the FFT calibration results under a special family of high-frequency inputs at 55.5 GHz;

[0040] Figure 6 It is a stability graph of frequency sweep calibration across the entire Nyquist band from 0 to 56 GHz;

[0041] Figure 7 This is a comparison chart of the performance of the method of the present invention with the comparison of special frequency points;

[0042] Figure 8 This is a histogram of significant digits before and after Monte Carlo calibration. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0044] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0045] This invention proposes a joint calibration method for blind spot sensing. The method folds the input frequency into a single-channel sampling bandwidth and establishes a blind spot reference set based on DC, a quarter-channel sampling rate, and a single-channel Nyquist frequency. The distance from the folded frequency to the reference set generates the blind spot proportion or blind spot weight. Based on the blind spot proportion, a joint observation set including the target frequency and / or auxiliary frequencies is constructed, and joint estimation is performed on the same channel skew variable.

[0046] At ordinary frequencies, this invention can utilize channel FFT phase and cross-channel linear trend residuals to form a high-precision initial phase estimate. In the blind spot neighborhood, this invention constructs an auxiliary frequency observation stack near the target frequency, enabling adjacent frequency observations to share the same physical skew solution with the target frequency. Furthermore, it improves estimation stability through linearized correlation initial estimates, blind spot weights, smoothing constraints, prior constraints, and candidate solution scoring. Damped iterative optimization, specific fusion formulas, and parameter adjustment methods can be implemented as examples, but do not constitute the sole limitation on the core method.

[0047] The symbol M represents the number of interleaving channels, f s f represents the total sampling rate. ch T represents the single-channel sampling rate. s Indicates the total sampling period. This indicates the clock skew of the m-th channel relative to the ideal sampling time.

[0048] Implementation Step 1: System Model

[0049] In an M-channel time-interleaved ADC, the actual sampled value of the m-th channel at the n-th sampling point of its own channel can be written as:

[0050]

[0051] in It includes quantization noise, thermal noise, and residual non-ideal terms. The calibration objective is to estimate the vector under zero-mean constraints. The estimation results are then converted into digital time-series compensation quantities.

[0052]

[0053] Implementation Step Two: Blind Spot Identification and Auxiliary Frequency Stack

[0054] For any input frequency f in Fold it to [0, f ch Within [ / 2], the folding frequency and blind spot ratio are defined as follows:

[0055]

[0056]

[0057]

[0058]

[0059]

[0060] A blind spot ratio close to 0 indicates a normal frequency, while a ratio close to 1 indicates the current frequency falls within or is close to a blind spot protection zone. Protection radius R b It can be determined by the FFT frequency resolution, the single-channel sampling rate ratio, and the noise margin.

[0061] When the blind spot ratio exceeds a threshold, the system does not rely solely on the target frequency for estimation; instead, it selects an auxiliary frequency near the target frequency. The auxiliary frequency is preferably approximately symmetrical about the target frequency and, after folding, far from the same blind spot reference point. Candidate frequencies can be generated from on-chip pilots or selected from a background candidate frequency pool. The set of auxiliary frequencies is denoted as:

[0062]

[0063] When the target frequency is located in a severe DC or Nyquist family blind spot, the target frequency can be removed from the joint estimation set; when the target frequency is located in a quarter-channel sampling rate blind spot, the target frequency can be retained but given blind spot weight, and the enhanced regularization parameter can be enabled simultaneously. Figure 2 The construction methods for protected areas and auxiliary frequencies in the folded frequency domain are given.

[0064] Step 3: Initial Phase Estimation and Initial Linearization Correlation Estimation

[0065] For each frequency f qFor channel m, extract the complex spectral phase of the channel near the corresponding FFT bin, and perform interpolation and robust unwrapping. The channel phase can be expressed as:

[0066]

[0067] in For common phase, For the ideal intertwined phase trend, This includes noise and interpolation errors.

[0068] Phase residuals after removing cross-channel linear trends Used to form an initial phase estimate:

[0069]

[0070] The initial phase estimate is accurate at normal frequencies, but its confidence decreases in the neighborhood of the blind spot due to reduced phase sensitivity. To avoid the initial solution being dominated by phase noise at the blind spot, this invention introduces a linearized correlation initial estimate. A first-order expansion of the perturbation at the sampling time is performed, and a residual r and Jacobian J are constructed near the current skew seed. The stable initial estimate direction is given by the derivative correlation term of the samples within the channel.

[0071]

[0072] Where D is the second-order difference matrix between adjacent channels. This initial estimate does not depend on a single phase residual and is suitable as a dominant seed when the proportion of blind spots is high or the phase position confidence is low.

[0073] Implementation Step Four: Adaptive Fusion and Joint Optimization

[0074] The initial phase estimate and linearized correlation estimate are unified to the zero-mean, limited channel skew space. Fusion coefficients. Based on the global blind spot ratio and phase position reliability A joint decision.

[0075]

[0076] For low-confidence channels, linearized correlation initial estimation is directly used to replace phase initial estimation; for scenarios with a high global blind spot ratio, the initial skew vector is smoothed and its maximum amplitude is limited.

[0077] Solving the shared channel skew vector during the joint refinement phase The objective function can be written as:

[0078]

[0079] in For the channel observation corresponding to the q-th frequency, For the model output given the skew vector, For frequency weights, To smooth out the weights, These are prior weights. The closer the frequency is to the blind spot, the higher the priority. The smaller the size, the higher the blind spot ratio. and The larger:

[0080]

[0081]

[0082] The joint estimation can be solved using a damped iterative approach. As an example, in the t-th iteration, let J be the model pair... The Jacobian matrix, where r is the observation residual. Let be the damping factor, and let the iteration step size satisfy:

[0083]

[0084]

[0085] If the trial step size decreases the objective function, then accept the step size and decrease the damping factor; if the objective function increases or changes beyond the divergence threshold, then reject the step size and increase the damping factor. Iteration stopping conditions can include relative cost change, maximum step size, and maximum number of iterations.

[0086] In a set of specific embodiments, the initial damping factor Pick Multiply by the diagonal mean of JTJ; when the trial step size is accepted, multiply the damping factor by 0.1; when the trial step size is rejected, multiply the damping factor by 10; the relative cost change threshold is set to... The maximum skew step size threshold is set to 0.005 ps, and the maximum number of iterations is set to 15. These parameters can be adjusted proportionally to the number of channels, noise level, and minimum register step size.

[0087] Implementation Step 5: Candidate Solution Scoring, Quantization, and Output Interface

[0088] Near severe blind spots and quarter-sampling-rate blind spots, this invention retains candidate solutions such as conventional joint estimation solutions, quarter-frequency enhanced constraint solutions, and conservative estimation solutions. Each candidate solution is scored based on multi-frequency model residuals, second-order difference smoothness, deviation from the initial solution, and amplitude limiting state, and the one with the lowest score is output.

[0089]

[0090] In hardware implementation, hierarchical quantization can be used for fine-tuning. Each level quantizes the skew control quantity to the current step size, and then performs a finite number of iterations. Typical step sizes include coarse adjustment 0.5 ps, regular adjustment 0.125 ps, and fine adjustment 0.03125 ps.

[0091]

[0092] The switching conditions for different step sizes can be determined by the residual decrease, the number of iterations, and the blind spot ratio. For example, if the residual decrease is less than 1% after two consecutive iterations in the coarse adjustment stage, or after four iterations, or the blind spot ratio is less than 0.3, switch to the normal step size; if the residual decrease is less than 0.3% in the normal adjustment stage, or after six iterations, switch to the fine adjustment step size.

[0093] The final skew estimate can drive a digital fractional delay filter, a Farrow structure, a Lagrange interpolation FIR, or a multiphase interpolation unit, and can also be converted into a sampling phase control word or an analog delay control word. Therefore, the output of this invention is a channel skew estimate for timing calibration, which can then be calibrated by a digital compensation circuit or an analog clock adjustment circuit.

[0094] Implementation Step Six: Simulation Verification

[0095] To illustrate the effectiveness of this embodiment, a 56-channel time-interleaved ADC was used for frequency domain simulation. The total system sampling rate was 112 GSPS, the single-channel sampling rate was 2 GSPS, the input bit width was 7 bits, and quantization noise, thermal noise, systematic clock skew, and random clock skew were introduced to simulate actual clock tree and layout differences.

[0096] The simulation selected 2.5 GHz, 28 GHz, and 55.5 GHz as representative frequencies, respectively covering f ch / 4 special frequency point families, 0 frequency point families, and high-frequency special frequency point families; simultaneously, frequency sweeping is performed in the range of 0 to 56 GHz to cover the first Nyquist band of the 112 GSPS system. The dynamic performance of the representative frequency points is directly labeled on Figures 3 to 5 In the FFT spectrum, the full-band stability is determined by Figure 6 The sweep frequency curve is given.

[0097] To highlight the necessity of blind spot perception processing at specific frequency points, this embodiment includes a comparative example. The comparative example only uses the single-frequency phase residual of the target frequency to estimate the channel clock skew, without constructing an auxiliary frequency observation stack, and without introducing blind spot proportional weights and smoothing prior enhancement constraints.

[0098] like Figure 7 As shown, at 2.5 GHz f chIn the / 4 family neighborhood, the conventional single-frequency phase method yields SNDR, SFDR, and ENOB values ​​of 34.10 dB, 46.20 dB, and 5.37 bits, respectively, while the method of this invention improves these values ​​to 41.20 dB, 67.30 dB, and 6.55 bits, respectively. In the 0-frequency family neighborhood at 28 GHz, the conventional single-frequency phase method yields 31.40 dB, 42.30 dB, and 4.92 bits, respectively, while the method of this invention improves these values ​​to 39.60 dB, 62.80 dB, and 6.28 bits, respectively. In the Nyquist end special frequency family at 55.5 GHz, the conventional single-frequency phase method yields 29.70 dB, 40.80 dB, and 4.64 bits, respectively, while the method of this invention improves these values ​​to 38.40 dB, 58.60 dB, and 6.09 bits, respectively.

[0099] The above comparison shows that the single-frequency phase method is prone to dynamic performance collapse in the neighborhood of special frequency families due to the decrease in phase position confidence. The present invention, through auxiliary frequency observation, blind spot weighting and enhanced constraints, enables SNDR, SFDR and ENOB to maintain significant improvement at three representative special frequency points.

[0100] Figure 3 Give the FFT spectrum for a 2.5 GHz input. The frequency at which this frequency falls after folding is f0. ch Near the / 4 special frequency family, visible interleaving spurious signals exist before calibration; after adopting the method of this invention, the SNDR marked in the FFT plot is improved from 35.60 dB to 41.20 dB, the SFDR is improved from 47.80 dB to 67.30 dB, and the ENOB is improved from 5.62 bit to 6.55 bit.

[0101] Figure 4 The FFT spectrum is given at an input of 28 GHz. This frequency folds into the 0-frequency family, a position where traditional single-frequency phase estimation is prone to degradation. Before calibration, the SNDR is 32.50 dB, SFDR is 43.90 dB, and ENOB is 5.10 bits; after calibration, the SNDR is 39.60 dB, SFDR is 62.80 dB, and ENOB is 6.28 bits.

[0102] Figure 5 The FFT spectrum is given for a specific family of high-frequency inputs at 55.5 GHz. This scenario is close to the Nyquist junction of the system, making calibration more difficult than for low-frequency and mid-frequency scenarios. Before calibration, the SNDR was 30.80 dB, SFDR was 41.50 dB, and ENOB was 4.82 bits; after applying the method of this invention, the SNDR was 38.40 dB, SFDR was 58.60 dB, and ENOB was 6.09 bits. All three FFT plots above use the paper-style spectral annotation method, directly showing the FFT values ​​within the plot. sInput frequency, SNDR, SFDR, ENOB, and NFFT are available to facilitate simultaneous observation of spurious suppression and dynamic performance improvement.

[0103] Figure 6 The frequency sweep results for the first Nyquist band from 0 to 56 GHz are presented. The sweep curves cover DC bands and f... ch / 4 family and f ch For special frequency families such as the / 2 family, after calibration, the SNDR remained between 37.6 dB and 41.0 dB, the SFDR remained between 55.8 dB and 64.2 dB, and the ENOB remained between 5.96 bit and 6.53 bit. The calibrated curves are basically consistent with the Oracle reference curves, indicating that the present invention has stability in the entire Nyquist band and no obvious performance collapse occurs at special frequency points.

[0104] The advantages and beneficial effects of this invention are as follows:

[0105] This invention explicitly identifies blind spots in the DC family, quarter-channel single-channel sampling rate family, and Nyquist family in the frequency folding domain, avoiding erroneous updates using ordinary single-frequency formulas at unobservable frequency points.

[0106] This invention places the initial phase estimation and the initial linearization correlation estimation within the same estimation framework. The accuracy of the phase method is preserved at ordinary frequencies, while the stability is maintained at special frequencies using the time-domain linearization seed.

[0107] This invention constrains the blind spot neighborhood solution through an auxiliary frequency stack, multi-frequency weights, smooth priors, and candidate solution scoring, resulting in continuous calibration output, suitable for frequency sweep testing and background adaptive operation.

[0108] The parameters of this invention can be scaled proportionally to the number of channels and the sampling rate per channel, and are applicable to different time-interleaving architectures such as 4, 8, 16, 32 and 56 channels.

[0109] The new modules in this invention are concentrated on the digital side, and can reuse existing FFT, accumulation, comparison, sparse solution, register and digital compensation units without adding analog phase detection circuits or additional ADC channels.

[0110] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0111] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0112] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A blind spot sensing joint calibration method for time-interleaved ADC clock skew error, characterized in that, Includes the following steps: The target input frequency is folded into a single-channel sampling bandwidth to obtain the folded frequency; the blind spot ratio or blind spot weight is generated based on the distance between the folded frequency and the blind spot reference set. A joint observation set is constructed based on the blind spot ratio or blind spot weight. When the blind spot ratio does not exceed a preset threshold, the joint observation set includes a target input frequency observation. When the blind spot ratio exceeds the preset threshold, the joint observation set includes at least one auxiliary frequency observation that is folded away from the nearest blind spot reference point. The target input frequency observation and / or the auxiliary frequency observation share the same set of channel clock skew variables. At least one of frequency weight, phase position confidence, or regularization constraint is determined based on the blind spot ratio or blind spot weight. The joint observation set is then used to jointly estimate the same set of channel clock skew variables to obtain the channel clock skew estimate.

2. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 1, characterized in that, The time-interleaved ADC includes M interleaved channels with a total sampling rate of f. s The single-channel sampling rate is f ch =f s / M, the folding frequency f fold We obtain it using the following formula, where r f Represents the target input frequency f in For a single-channel sampling rate f ch Take the remainder, f fold Located in [0, f ch / 2]: ； 。 3. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 2, characterized in that, The blind spot reference set includes at least 0 and f. ch / 4 and f ch / 2; When the distance from the folding frequency to the nearest reference point in the blind point reference set is less than the preset protection radius, it is determined that the target input frequency is located in a special frequency protection zone; The blind spot ratio is determined by the minimum distance from the folding frequency to the blind spot reference set and the blind spot protection radius. The larger the blind spot ratio, the lower the confidence level of the phase observation corresponding to the current target input frequency. ； Where d represents the distance from the folding frequency to the nearest reference point in the blind reference set, and f b F represents any reference frequency point in the blind spot reference set. b R represents the blind spot reference set. b ρ represents the blind spot protection radius, and ρ represents the blind spot ratio.

4. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 2, characterized in that, The auxiliary frequency observation is obtained by screening a preset candidate frequency pool or by constructing frequency offsets on both sides of the target input frequency; the auxiliary frequencies are sorted according to at least one of blind spot distance, spectral leakage degree, amplitude stability and phase consistency.

5. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 2, characterized in that, When the folded frequency is close to the 0 frequency family, reduce the observation weight of the target input frequency, or remove the target input frequency from the joint observation set and select the auxiliary frequency that is far away from the 0 frequency after folding as the main observation; When the folding frequency is close to f ch When dealing with a frequency family of / 2, reduce the observation weight of the target input frequency, or remove the target input frequency from the joint observation set, and select the frequency that is far from f after folding. ch The auxiliary frequency at frequency point 2 is used as the main observation; When the folding frequency is close to f ch When the frequency family is / 4, the enhanced constraint mode is activated. The enhanced constraint mode includes at least one of the following: increasing the initial linearization correlation weight, increasing the smoothing constraint weight, increasing the prior constraint weight, and limiting the update step size.

6. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 1, characterized in that, The joint estimation includes: extracting the complex spectrum phase of each channel for each frequency observation in the joint observation set, and removing the common phase and ideal interleaved phase trend to obtain an initial phase estimate; constructing residual and Jacobian matrices based on the first-order approximation of the channel signal perturbation at the sampling time to obtain a linearized correlation initial estimate; generating fusion coefficients based on the blind spot ratio and phase position confidence, wherein the larger the blind spot ratio or the lower the phase position confidence, the greater the fusion weight of the linearized correlation initial estimate; and performing at least one of zero-mean constraint, smoothing, or amplitude limitation on the fused initial skew vector.

7. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 6, characterized in that, The joint estimation constructs an objective function using a shared channel clock skew vector as the independent variable. The objective function includes at least two of the following: a weighted multi-frequency observation residual term for each frequency observation in the joint observation set; a second-order difference smoothing term for adjacent channels; and a prior term relative to the initial estimate or historical effective estimate. Specifically, the larger the proportion of blind spots corresponding to a frequency observation, the smaller the weight of that frequency observation, and the larger the weight of the smoothing term or the prior term. ； Where L(τ) represents the objective function value, τ represents the shared channel clock skew vector, Q represents the number of frequency observations in the joint observation set, q represents the frequency observation sequence number, and w q Let x represent the weight of the q-th frequency observation. q This represents the channel observation corresponding to the q-th frequency. The model output is represented by the channel clock skew vector τ, where D represents the second-order difference matrix between adjacent channels, and λ represents the output of the model. s λ represents the smoothing constraint weight. p τ represents the prior constraint weight, and τ0 represents the initial estimate or the historical effective estimate.

8. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 7, characterized in that, The joint estimation employs a damped iterative solution, and the damping factor is adaptively adjusted or the step size is updated based on changes in the objective function, the step size, or the decrease in residuals.

9. The blind spot sensing joint calibration method for time-interleaved ADC clock skew error according to claim 1, characterized in that, Under conditions of severe blind spots or low phase position confidence, multiple candidate channel clock skew estimates are generated. The candidate channel clock skew estimates are scored based on at least two of the following: multi-frequency observation residuals, smoothness, deviation from the initial estimate, and amplitude limiting status. The estimated value that meets the preset rules is selected for output. The estimated channel clock skew is subjected to hierarchical quantization. When the decrease in residual after the previous quantization is less than a preset threshold, the number of iterations reaches a preset value, or the proportion of blind spots is lower than a preset value, the estimation is switched to a smaller quantization step size.

10. A time-interleaved ADC clock skew error calibration system for implementing the method of any one of claims 1 to 9, characterized in that, The system includes a frequency folding module, a blind spot ratio generation module, an auxiliary frequency observation construction module, a multi-frequency joint estimation module, a candidate solution scoring module, and an estimate output module. The frequency folding module folds the target input frequency to a single-channel sampling bandwidth and outputs the folded frequency. The blind spot ratio generation module generates a blind spot ratio or blind spot weight based on the distance between the folded frequency and the blind spot reference set. The auxiliary frequency observation construction module filters or constructs auxiliary frequency observations that are folded and farthest from the nearest blind spot reference point when the blind spot ratio exceeds a preset threshold. The multi-frequency joint estimation module performs joint estimation based on the channel clock skew variable shared by the target input frequency observation and / or auxiliary frequency observations. The candidate solution scoring module scores multiple candidate channel clock skew estimates under severe blind spot or low phase position confidence conditions. The estimate output module outputs or writes channel clock skew estimates that meet preset scoring rules through a clock skew control interface.

Images