Processing system for high frequency response test data of radio frequency probe

Abstract

The application relates to the technical field of radio frequency test data processing, in particular to a processing system for high-frequency response test data of a radio frequency detector, which comprises a ripple period extraction module, a phase mutation point calibration module, an adaptive filtering module, a phase verification module and a bandwidth extraction module. The ripple period extraction module is used for acquiring amplitude and phase data sequences and extracting ripple periods. The phase mutation point calibration module is used for calibrating phase mutation points. The adaptive filtering module is used for dividing sections, filtering each section, constructing a suppression criterion driven window iteration convergence. The phase verification module is used for extracting actual phases of single-frequency components and performing phase consistency verification to identify ripples and intrinsic characteristics of devices. The bandwidth extraction module is used for splicing section results, extracting a 3dB bandwidth and outputting a confidence degree. Through the technical scheme, adaptive consideration of ripple suppression and roll-off characteristic reservation is realized, and the consistency and reliability of production line bandwidth grading are improved.

Description

Technical Field

[0001] This application relates to the field of radio frequency test data processing technology, and in particular to a system for processing high-frequency response test data of radio frequency detectors. Background Technology

[0002] As core components in optoelectronic communication and microwave sensing, radio frequency (RF) detectors' -3dB bandwidth is a key indicator for evaluating high-frequency response performance. In large-scale production environments, production lines must perform wide-band sweep frequency measurements on each device and classify its performance based on bandwidth values. The stability and accuracy of the test results directly determine the product's quality upon leaving the factory. As the operating frequency band of devices expands towards millimeter waves, the test frequency range increases, and the requirements for the accuracy of frequency response data processing continue to rise.

[0003] Currently, Chinese patent application CN117077043A discloses an evaluation method for CdZnTe photon counting detectors based on leakage current response. This method achieves batch evaluation of detector performance by processing detector test data through a fusion decision tree classification and feature ranking algorithm. This method represents a technical approach in the field of detector test data processing that utilizes algorithms to automatically analyze raw test data and output performance evaluation results, providing a reference direction for the automated processing of detector production line testing.

[0004] However, the core interference source in high-frequency response testing of RF detectors is the standing wave ripple caused by impedance mismatch in the test link. This ripple is superimposed on the amplitude-frequency response curve, and its amplitude and period fluctuate continuously with probe contact quality, cable connection status, and ambient temperature. The aforementioned evaluation methods focus on the classification and ranking decisions of test data, without addressing the directional suppression of standing wave ripple in the frequency response curve. They cannot preserve the complete shape of the device's true roll-off characteristics while suppressing ripple. When ripple suppression is insufficient or the roll-off region is excessively smoothed, the bandwidth readings of repeated tests on the same device fluctuate beyond the allowable tolerance, and the consistency of the classification results cannot meet production line requirements. Summary of the Invention

[0005] To address the technical issues of unstable bandwidth extraction and insufficient consistency of classification results caused by standing wave ripple interference in the high-frequency response test of the radio frequency detector production line, this application provides a data processing system for high-frequency response test data of radio frequency detectors.

[0006] This application provides a processing system for high-frequency response test data of a radio frequency detector. The processing system includes: a ripple period extraction module, used to sweep the frequency of the radio frequency detector under test to obtain amplitude and phase data sequences, and calculate the autocorrelation function after detrending the amplitude data sequence to extract the ripple period; a phase change point calibration module, used to dewind the phase data sequence and perform second-order difference, and calibrate the frequency points where the slope change rate exceeds a first preset threshold as phase change points; and an adaptive filtering module, used to divide the window into segments with the distance from each frequency point to the nearest phase change point as the upper bound, filter each segment with the ripple period as the initial window to obtain residuals, and use the ratio of the autocorrelation value of the residual at the hysteresis corresponding to the ripple period to the noise background as the suppression criterion, and increase the window size when the suppression criterion exceeds the limit. The process iterates until convergence or the upper bound of the window is reached. A phase verification module extracts the actual phase of the single-frequency component from the residual of segments that have reached the upper bound of the window. It then extrapolates the expected phase by combining this phase with the ripple phase of adjacent converged segments. If the difference between the actual and expected phases of a single-frequency component exceeds a second preset threshold, the component is retained; otherwise, it is subtracted from the residual. A bandwidth extraction module splices the processing results of each segment to obtain a smooth curve and extracts the -3dB bandwidth from it. The processing result of converged segments is the filtered value at the time of iterative convergence. For segments that have reached the upper bound of the window and whose single-frequency components are not retained, the processing result is the corrected value of the residual minus the single-frequency component and then adding back the filtered value. For segments that have reached the upper bound of the window and whose single-frequency components are retained, the processing result is the filtered value at the upper bound of the window.

[0007] Preferably, the method for detrending the amplitude data sequence is as follows: a trend curve is obtained by least-squares fitting of the amplitude data sequence with a third-order polynomial, and the trend curve is subtracted from the amplitude data sequence to obtain the detrended amplitude sequence.

[0008] Preferably, the method for calculating the autocorrelation function to extract the ripple period is as follows: the lower bound of the search for the autocorrelation function is set to a preset multiple of the frequency resolution, and the upper bound of the search is a first preset proportion of the total frequency span; the condition for determining a local maximum is that the autocorrelation value is simultaneously greater than the adjacent previous value and the adjacent subsequent value and is greater than zero; if there is no local maximum within the search range that satisfies the determination condition, it is determined that the amplitude data sequence does not contain ripple, and the second preset proportion of the total frequency span is used as the substitute value for the ripple period.

[0009] Preferably, the first preset threshold is determined by: calculating the mean of the absolute value sequence of the slope change rate, and multiplying the mean by a preset multiple to obtain the first preset threshold.

[0010] Preferably, the method of dividing the segments by using the distance from each frequency point to the nearest phase change point as the upper bound of the window is as follows: the frequency axis is divided into several segments with the phase change point as the boundary, and the upper bound of the window for any frequency point in each segment is equal to the distance from that frequency point to the nearest phase change point; when the frequency interval between two adjacent phase change points is less than a preset multiple of the ripple period, the upper bound of the window for all frequency points in that segment is limited to half of that frequency interval.

[0011] The rule of dividing the segment by the phase change point and limiting the upper limit of the window to half the interval for the narrow segment forces the filter window in the segment with small frequency intervals near the roll-off region to be constrained within a safe range, eliminating the possibility of the window crossing the roll-off boundary and mixing and fitting the passband data with the roll-off data, thus protecting the integrity of the frequency response shape of the roll-off region.

[0012] Preferably, when filtering each segment using the ripple period as the initial window, a Savitzky-Gore filter is used, with a polynomial order of not less than third order, an odd window length, and an initial value that is not less than the minimum odd number of the ripple period.

[0013] Preferably, the noise background is determined as follows: the autocorrelation value of the autocorrelation function of the residual at each lag point is calculated, and after excluding the zero lag point and the lag point corresponding to an integer multiple of the ripple period, the arithmetic mean of the absolute values ​​of the autocorrelation values ​​at the remaining lag points is taken as the noise background; when the noise background is less than a preset minimum value, the preset minimum value is used to replace the noise background.

[0014] The noise background is determined by excluding zero-lag points and lag points corresponding to integer multiples of the ripple period, and then taking the average of the absolute values. This exclusion strategy removes the contribution of the periodic ripple component from the background estimate, ensuring that the noise background only reflects the level of random fluctuations in the residuals that do not possess ripple periodicity. The suppression criterion calculated in this way can sensitively and accurately indicate the degree of ripple residual, avoiding situations where the background is artificially low due to ripple leakage components, or where the ripple is not actually sufficiently suppressed but is mistakenly judged as converged.

[0015] Preferably, the method of increasing the window and iterating until convergence or reaching the upper bound of the window when the suppression criterion exceeds the limit is as follows: the smallest odd number of the product of the ripple period and the preset ratio is used as the window increment, and the current window length is increased by the window increment in each iteration; the convergence condition is that the suppression criterion does not exceed the preset convergence threshold; a maximum number of iterations is set, and the iteration is terminated and the upper bound of the window is reached when the maximum number of iterations is reached.

[0016] Preferably, the method for extracting the actual phase of the single-frequency component in the residual is as follows: a bandpass filter is constructed with the reciprocal of the ripple period as the center frequency, and the residual is bandpass filtered to obtain the single-frequency component; the actual phase of the single-frequency component is obtained by taking the inner product of the single-frequency component with the sine basis function and the cosine basis function of the center frequency, and then taking the arctangent.

[0017] A bandpass filter is constructed with the reciprocal of the ripple period as the center frequency to extract single-frequency components. The actual phase is obtained by taking the arctangent of the inner product of sine and cosine. This method simultaneously and completely extracts the amplitude and phase information of the ripple components, providing accurate phase measurements for subsequent phase consistency verification. The narrowband characteristics of the bandpass filter directionally separate ripple energy from the broadband residual, avoiding the contamination of phase estimation by broadband noise and improving the reliability of phase identification.

[0018] Preferably, after extracting the -3dB bandwidth, the method is further used to perform a confidence assessment, including: calculating the ratio of the number of segments in all segments that touch the upper bound of the window and whose single-frequency component is retained to the total number of segments as the anomaly ratio; when the anomaly ratio is less than or equal to a first ratio threshold, the confidence level is high; when the anomaly ratio is greater than the first ratio threshold and less than a second ratio threshold, the confidence level is medium; and when the anomaly ratio is greater than or equal to the second ratio threshold, the confidence level is low.

[0019] The confidence assessment transforms the segment status information generated during the processing into the reliability level output of the final measurement result, enabling production line operators to identify abnormal measurement conditions without manual waveform interpretation. Low confidence results can be promptly retested, reducing the grading error rate caused by single probe contact failure or cable condition abnormalities at the process level.

[0020] The technical solution of this application has the following beneficial technical effects:

[0021] To address the issue of unstable bandwidth extraction caused by standing wave ripple interference in high-frequency response testing of RF detector production lines, an adaptive ripple suppression processing chain with amplitude-phase dual-channel collaboration was established. By extracting the ripple period from the amplitude data as a global parameter and calibrating the roll-off start position from the phase data as a filtering boundary constraint, combined with residual autocorrelation feedback to drive window iteration convergence and phase consistency verification, the ripple is distinguished from the intrinsic characteristics of the device. This narrows the bandwidth reading fluctuation of repeated tests on the same device to within the allowable tolerance range, improving the consistency and reliability of the classification results and meeting the accuracy requirements of automated data processing for high-throughput testing on the production line.

[0022] The amplitude channel provides the frequency domain periodic characteristics of the ripple, while the phase channel provides the spatial location characteristics of the roll-off. These two channels, acquired independently, form a constrained synergy in the filtering operation. Specifically, the ripple period determines the initial size and iteration step of the filtering window, and the phase abrupt change point determines the upper bound of the window. This dual-channel synergistic structure physically confines the ripple suppression operation to the passband, ensuring that the roll-off characteristics are not blurred or shifted by smoothing operations. The extraction accuracy of the -3dB cutoff frequency is unaffected by ripple amplitude and period fluctuations. Attached Figure Description

[0023] Figure 1 This is a structural block diagram of a system for processing high-frequency response test data of a radio frequency detector according to an embodiment of this application; Figure 2 These are detrended amplitude sequence waveforms and autocorrelation function curves according to embodiments of this application; Figure 3 This is a comparative schematic diagram of the original residual, extracted single-frequency component, and corrected residual of a certain arrival window upper boundary segment according to an embodiment of this application. Detailed Implementation

[0024] This application provides a system for processing high-frequency response test data of radio frequency detectors. Figure 1 This is a structural block diagram of a system for processing high-frequency response test data of a radio frequency detector according to an embodiment of this application. Figure 1 As shown, the processing system for the high-frequency response test data of the radio frequency detector includes a ripple period extraction module, a phase change point calibration module, an adaptive filtering module, a phase verification module, and a bandwidth extraction module, which are described in detail below.

[0025] The ripple period extraction module is used to sweep the frequency of the RF detector under test to obtain amplitude data sequence and phase data sequence; after detrending the amplitude data sequence, the autocorrelation function is calculated, and the frequency interval corresponding to the first local maximum value is taken as the ripple period.

[0026] A photoelectric company's RF detector production line performs high-frequency response tests on hundreds of devices daily, covering a frequency band from 0.1 GHz to 40 GHz. A vector network analyzer uses RF probes to contact the pads of the chip under test, completing a wide-band sweep measurement. During the sweep, the instrument uniformly collects several frequency points between the start and cutoff frequencies. At each frequency point, the amplitude and phase values ​​of the transmission coefficient are recorded simultaneously, forming amplitude data sequences and phase data sequences, respectively. The unit for each point in the amplitude data sequence is decibels, representing the transmission gain at that frequency; the unit for each point in the phase data sequence is degrees, representing the transmission phase shift at that frequency.

[0027] Due to the continuous fluctuations in probe contact quality, cable connection status, and ambient temperature within the production line environment, the original amplitude data sequence is superimposed with standing wave ripples of unstable amplitude and period. To extract the oscillation period of the ripple from the amplitude data sequence, it is necessary to first remove the overall variation trend in the amplitude data, so that the ripple component is fully exposed in the residual sequence.

[0028] Detrending processing is performed on the amplitude data sequence. A third-order polynomial is used to perform least-squares fitting on the entire amplitude data sequence to obtain a trend curve. This trend curve is then subtracted point-by-point from the amplitude data sequence to obtain the detrended amplitude sequence. The third-order polynomial is chosen because the overall frequency response envelope of the RF detector consists of a flat passband segment plus a monotonically rolled-off segment. This shape can be well described by a polynomial of order three or less. A third-order polynomial is sufficient to capture the gradually changing shape of the overall trend without absorbing ripple components, as the spatial oscillation frequency of ripple is much higher than the gradually changing frequency that a third-order polynomial can characterize. Without detrending processing, the overall downward trend in the amplitude data will dominate the autocorrelation function shape, causing it to exhibit a monotonically slow decay, and the periodic peaks corresponding to the ripple will be completely submerged and cannot be extracted.

[0029] The autocorrelation function is calculated for the detrended amplitude sequence. The normalized correlation coefficient between the sequence and itself is calculated sequentially at different lags, resulting in a curve with the lag as the horizontal axis and the autocorrelation coefficient as the vertical axis. The lower bound of the autocorrelation function search is set to 5 times the frequency resolution, and the upper bound is set to 50% of the total frequency span. The lower bound eliminates spurious correlations caused by sampling discreteness at excessively short lags, while the upper bound covers the ripple period range corresponding to RF test cable lengths of 0.3 meters to 3 meters in engineering practice.

[0030] The search involves finding the first local maximum within the search range. The criteria for a local maximum are: the autocorrelation value at that point is simultaneously greater than both the adjacent preceding and following values, and this autocorrelation value is greater than zero. The frequency interval represented by the hysteresis value corresponding to the first point satisfying this condition is the ripple period. The additional condition of a value greater than zero excludes peaks within troughs in negatively correlated regions; that is, although such points constitute local maxima, their autocorrelation coefficients are negative and do not physically correspond to true periodic oscillations.

[0031] If no local maxima satisfying the above criteria are found within the search range, the current amplitude data sequence is determined to be free of significant ripple. In this case, 3% of the total frequency span is used as a substitute value for the ripple period. This substitute value is only used for initializing the filter window in subsequent processing, allowing the processing flow to continue with a unified data interface. The basis for choosing 3% of the total frequency span as the substitute value is that the corresponding filter window is sufficiently narrow, performing only slight smoothing to suppress measurement noise, and will not cause any perceptible change to the intrinsic response of the device.

[0032] If the autocorrelation function has multiple local maxima that satisfy the criteria within the search range, the frequency interval corresponding to the first local maximum is taken as the ripple period. The first maximum corresponds to the periodic component with the strongest energy in the residual sequence, i.e., the main ripple period.

[0033] Figure 2 The waveform of the detrended amplitude sequence and the autocorrelation function curve are shown. Figure 2 The sub-chart above is a waveform chart of the detrended amplitude sequence. The vertical axis is in decibels, and the horizontal axis is frequency. The sequence exhibits a regular oscillation pattern. Figure 2 The subplot below shows the autocorrelation function curve of the detrended amplitude sequence. The vertical axis represents the autocorrelation coefficient, and the horizontal axis represents the hysteresis. The left area of ​​the figure marks the lower bound of the search, and the right area marks the upper bound of the search. The first local maximum within the search range occurs at a hysteresis of 1.2 GHz, corresponding to an autocorrelation coefficient of approximately 0.72. Therefore, the ripple period is determined to be 1.2 GHz.

[0034] The extraction process of the ripple period is illustrated using a specific test scenario: the test frequency band is from 0.1 GHz to 40 GHz, with a total frequency span of 39.9 GHz, 801 sampling points, and a frequency resolution of approximately 0.05 GHz. The lower bound of the search is 0.05 × 5 = 0.25 GHz, and the upper bound is 39.9 × 50% = 19.95 GHz. After calculating the autocorrelation function of the detrended amplitude sequence, the first local maximum value satisfying the judgment condition appears at 1.2 GHz corresponding to the hysteresis, with an autocorrelation coefficient of 0.72, which is greater than 0 and greater than the adjacent values. Therefore, the ripple period is determined to be 1.2 GHz. The two-way electrical length between impedance mismatch points in the test link corresponding to this value is approximately the speed of light divided by 1.2 GHz and then divided by 2, which is approximately 0.125 meters, consistent with the typical physical distance between the probe and the cable connector.

[0035] The ripple period is a frequency domain mapping of the two-way electrical length between impedance mismatch points in the test link. It depends only on the physical structure of the test link and is independent of the intrinsic frequency response of the device under test. This parameter plays four roles in subsequent processing: the hysteresis position parameter for residual detection in iterative optimization, the step size parameter for window adjustment, the notch filter frequency parameter in directional separation, and the slope parameter for phase extrapolation, providing the frequency domain positioning basis for the entire processing flow.

[0036] The phase abrupt change point calibration module is used to obtain the slope change rate by second-order difference after unwinding the phase data sequence, and to calibrate the frequency points where the slope change rate exceeds a first preset threshold as phase abrupt change points.

[0037] Phase dewinding is performed on the phase data sequence acquired by the frequency sweep. The original phase values ​​output by the vector network analyzer are limited to the main value range of -180 degrees to 180 degrees. When the actual phase changes continuously beyond this range, the recorded value jumps by 360 degrees. The specific operation of the dewinding process is as follows: starting from the second frequency point, the phase difference between the current point and the previous point is calculated point by point. If the difference is greater than 180 degrees, 360 degrees is subtracted from the phase value of the current point and all subsequent points. If the difference is less than -180 degrees, 360 degrees is added to the phase value of the current point and all subsequent points. After this processing, the phase data sequence becomes a curve that changes continuously along the frequency axis, eliminating artificial jumps. If the dewinding process is not performed, the 360-degree periodic jumps in the original phase data will produce huge spikes in subsequent differential operations, which will be misidentified as the phase change corresponding to the actual roll-off, causing the upper bound of the subsequent window to be set at a completely incorrect position.

[0038] A second-order difference operation is performed on the unwound phase sequence. The first-order difference is obtained by taking the difference between the phase values ​​of every two adjacent frequency sampling points and dividing it by the frequency step interval, thus obtaining the phase slope value at each frequency point, forming a phase slope sequence. The phase slope is physically equivalent to the group delay, reflecting the change in time delay of the signal passing through the device under test with frequency. The second-order difference is obtained by taking the difference between two adjacent points in the phase slope sequence and dividing it by the frequency step interval, thus obtaining a slope change rate sequence. The slope change rate characterizes the degree of drastic change in group delay along the frequency axis.

[0039] Determine the first preset threshold. Calculate the mean of the absolute value sequence of slope change rates, and multiply this mean by a preset multiple to obtain the first preset threshold. The preset multiple is set to 3, based on the three-times mean criterion in statistics: under the condition that the slope fluctuations caused by ripples approximately follow a symmetrical distribution, three times the absolute mean corresponds to the probability of extreme events in the tail of the distribution of approximately 2% to 5%, which can effectively distinguish between background fluctuations and real mutations.

[0040] The frequency points where the absolute value of the rate of change of the slope exceeds the first preset threshold are marked as phase abrupt change points. Based on the Cramer-Kronig causality, the frequency range where the amplitude response experiences a true roll-off corresponds to a sharp change in group delay, i.e., a significant jump in the rate of change of the slope. While standing wave ripples cause periodic fluctuations in phase, their amplitude after second-order difference is much smaller than the abrupt change amplitude caused by the true roll-off. This is because the ripples exhibit superimposed small-amplitude sinusoidal oscillations in phase; their first-order difference is a cosine oscillation of the same frequency, and the second-order difference remains a sinusoidal oscillation of the same frequency with unchanged amplitude. In contrast, the phase corresponding to the true roll-off suddenly changes from a gradual change to a rapid change, with a step in the first-order difference and a pulse in the second-order difference. The pulse amplitude far exceeds the peak value of the sinusoidal oscillation of the same frequency; therefore, a three-times mean threshold can effectively separate the two.

[0041] The threshold can be calibrated in batches using a data-driven approach: take the test data of the first 10 devices in the same batch, calculate the slope change rate distribution of each device, take the mean of the 95th percentile of each distribution as the adaptive threshold reference value of the batch, and take the larger of the two values ​​as the final threshold to adapt to the differences in the roll-off steepness of devices in different batches.

[0042] Using the aforementioned test scenario for illustration: with a frequency resolution of 0.05 GHz, the unwound phase sequence decreases approximately linearly within the 0.1 to 28 GHz range, with a slope of approximately -12 degrees per GHz. Near 28.5 GHz, the phase slope abruptly changes from -12 degrees per GHz to -58 degrees per GHz across three frequency points, corresponding to an absolute value of 920 degrees per GHz squared for the slope change rate. The average absolute value of the slope change rate for the entire sequence is 87 degrees per GHz squared, and the first preset threshold is 87 × 3 = 261 degrees per GHz squared. The slope change rate of 920 at 28.5 GHz far exceeds 261, and this point is designated as the phase abrupt change point. The peak slope change rate caused by ripple within the passband is approximately 150 degrees per GHz squared, below the 261 threshold, and was not mislabeled.

[0043] The phase abrupt change point marks the starting frequency position of the true roll-off of the amplitude curve. This position information does not depend on the amplitude data itself, but rather originates from the phase channel, which has a physical causal constraint with the amplitude. This cross-channel calibration method is unaffected by standing wave ripple, providing a reliable hard constraint for subsequent processing to prevent the filtering window from exceeding the roll-off start position.

[0044] An adaptive filtering module is used to divide the window into segments with the distance from each frequency point to the nearest phase change point as the upper bound of the window, filter each segment with the ripple period as the initial window, use the difference between the original value and the filtered value as the residual, and use the ratio of the autocorrelation value of the residual at the lag corresponding to the ripple period to the estimated noise background as the suppression criterion. When the suppression criterion exceeds the limit, the window is increased and iterated until convergence or the upper bound of the window is reached.

[0045] Using the aforementioned phase abrupt change points as boundaries, the entire frequency axis is divided into several continuous segments. The start and end boundaries of each segment are determined by two adjacent phase abrupt change points. The first segment extends from the lowest point of the frequency axis to the first phase abrupt change point, and the last segment extends from the last phase abrupt change point to the highest point of the frequency axis. The upper bound of the window for any frequency point within each segment is equal to the distance from that frequency point to the nearest phase abrupt change point. Frequency points at the center of a segment are farther from the abrupt change points on either side, resulting in a larger upper bound window; frequency points closer to the segment boundaries are closer to the abrupt change points, resulting in a smaller upper bound window.

[0046] When the frequency interval between two adjacent phase abrupt change points is less than twice the ripple period, the maximum available window within this segment is insufficient to cover a complete ripple oscillation cycle, and conventional time-domain smoothing cannot effectively suppress ripple in this situation. For such narrow segments, the upper bound of the window for all frequency points within the segment is uniformly limited to half of this frequency interval. This limitation ensures that the filter window will not cross the segment boundary into adjacent segments at any position, avoiding the mixing of roll-off data into the fitting window of the passband data. The basis for setting the multiplier to 2 is as follows: when the upper bound of the window is half the frequency interval, taking the center point of the segment as an example, extending half a window to the left and right just touches the two side boundaries; and the segment width corresponding to twice the ripple period just allows the window at the center point to cover a complete ripple cycle, and if it is lower than this width, complete coverage cannot be achieved at the center point either.

[0047] An initial filtering window is set for each segment, and filtering is performed. A Savitzky-Gore filter is used, with a third-order polynomial and an odd-numbered window length. The initial window length is the smallest odd number not less than the number of sampling points corresponding to the ripple period. The third-order polynomial is selected based on the fact that the local shape of the frequency response curve in the flat passband region is close to a horizontal straight line or a shallow arc. The third-order polynomial can preserve the slow trend change and shallow arc characteristics without tracking ripple details due to excessively high order. The constraint of an odd-numbered window length is an inherent requirement of the Savitzky-Gore filter, that is, the filter takes an equal number of data points on both sides of the center point as the axis of symmetry for fitting, and the odd-numbered window ensures that the center point is located at the geometric center of the window.

[0048] Using the aforementioned test scenario, the process of determining the initial window is explained: the ripple period is 1.2 GHz, the frequency resolution is 0.05 GHz, and the number of sampling points corresponding to the ripple period is 1.2 ÷ 0.05 = 24 points. The smallest odd number not less than 24 is 25, therefore the initial window length is set to 25 points.

[0049] The filter operates as follows at each frequency point: Centered on that point, a data segment within the current window length is selected. A third-order polynomial is used to perform a least-squares fit on this data segment, and the calculated value of the fitted polynomial at the center point is taken as the filter output for that point. For frequency points near the segment boundary, if the window centered on that point exceeds the segment range, the window is truncated to the segment boundary, and only the available data points within the segment are used for fitting. The effective window length after truncation must not be less than the polynomial order plus one, i.e., not less than 4 points; otherwise, the polynomial fit will be underdetermined. For boundary positions where the effective window length is less than 4 points, the original value is directly taken as the filter output.

[0050] For segments where the upper bound of the window is less than the initial window length, the actual filtering window for that segment is set to the largest odd number corresponding to its upper bound. These segments still undergo Savitzky-Gorye filtering in this module to obtain the reference output, but because the window is insufficient to cover the complete ripple period, the ripple suppression effect is limited, and they will subsequently undergo directional separation processing.

[0051] After filtering is completed for each segment, the corresponding filtered output value is subtracted from each point value of the original amplitude data sequence to obtain the residual. The residual reflects the components that the filtering failed to remove, including ripple residue and measurement random noise. For segments where the initial window has covered the complete ripple period, the degree of ripple residue in the residual depends on the matching relationship between the ripple and the window. That is, if the ripple is strictly single-frequency sinusoidal and the window is exactly an integer multiple of its period, theoretically the ripple can be completely suppressed. In practice, the ripple often has frequency drift and non-sinusoidal distortion, and there is usually still a certain degree of ripple residue in the residual. Its sufficiency will be checked later through a feedback mechanism.

[0052] Setting the upper bound of the window transforms the phase analysis results into a physical constraint on the filtering operation: regardless of ripple suppression requirements, the filtering window will never cross the starting position of the true roll-off at any frequency point or in any segment. This constraint structurally eliminates the possibility of passband data and roll-off data being mixed into the same fitting window, preventing the filter from physically making the error of cross-region smoothing. Phase information does not participate in the direct calculation of amplitude values, but it determines the spatial boundary of amplitude processing, forming a constraint relationship across data channels.

[0053] The residuals of each section need to be checked for sufficiency of ripple suppression to determine whether the current filtering window has sufficiently suppressed the ripple.

[0054] Calculate the autocorrelation value of the residuals at the lag corresponding to the ripple period. Calculate the standard autocorrelation function for the residuals of each segment, and take the autocorrelation coefficient at the lag equal to the number of sampling points corresponding to the ripple period. Continuing with the previous scenario, the ripple period is 1.2 GHz, corresponding to 24 sampling points, so take the autocorrelation coefficient value at the lag of point 24. This value reflects whether there are still repeating patterns in the residuals at ripple period intervals. If the ripple has been sufficiently suppressed, the residuals should be approximately white noise, and the autocorrelation coefficient at this lag should be close to zero; if the ripple residual is significant, the autocorrelation coefficient at this point will deviate significantly from zero.

[0055] Estimating the noise background: The mean absolute value of the autocorrelation function of the residuals, excluding zero-lag points and lag points corresponding to integer multiples of the ripple period, is taken as the noise background. The autocorrelation coefficient at the zero-lag point is always 1 and does not carry frequency selection information, so it must be excluded. At lag points corresponding to integer multiples of the ripple period, even if the residual ripple is weak, the autocorrelation coefficient may be slightly higher due to the period superposition effect. Including it in the mean calculation would raise the noise background estimate and reduce the test sensitivity, so it is also excluded. After excluding the above specific locations, the autocorrelation coefficients at the remaining lag points reflect the statistical fluctuation level of the random component in the residuals that does not have ripple periodicity, and its mean absolute value is a reasonable estimate of the noise background.

[0056] When the noise background is less than a preset minimum value, the preset minimum value is used to replace the noise background. The preset minimum value is the inverse square root of the length of the residual sequence. This value is set based on the fact that the theoretical standard deviation of the autocorrelation coefficient of a finite-length white noise sequence at non-zero lag is approximately equal to the inverse square root of the sequence length; autocorrelation values ​​below this level are no longer statistically distinguishable. Taking the aforementioned scenario as an example, if a segment contains 200 sampling points, the preset minimum value is 1 divided by the square root of 200, which is approximately 0.071. When the calculated noise background is less than 0.071, it is replaced with 0.071 to avoid the ratio value diverging when the noise background approaches zero.

[0057] Suppression criterion: The autocorrelation value of the residual at the hysteresis corresponding to the ripple period is divided by the noise background.

[0058] The judgment rule is as follows: when the suppression criterion does not exceed the preset convergence threshold, the ripple is considered to have been sufficiently suppressed, and the filtered output of the current segment is the final smoothing result, at which point the iteration terminates. The preset convergence threshold is set to 2. When this ratio equals 1, the autocorrelation level at the ripple lag is equal to the background of random fluctuations, and the ripple residue is indistinguishable from noise; setting it to 2 provides a margin of 1 to address statistical fluctuations caused by finite sequence lengths and non-ideal noise distributions.

[0059] When the suppression criterion exceeds the preset convergence threshold, the ripple suppression is deemed insufficient, and the window needs to be increased. The window increment is determined by multiplying the ripple period by a preset ratio of 0.5 to obtain the increment baseline value. The smallest odd number not less than the number of sampling points corresponding to this baseline value is taken as the window increment. The preset ratio of 0.5 is based on the following: after each step increases by half a ripple period, the window coverage of the ripple increases by half an oscillation period, resulting in a substantial improvement in smoothing capability; if the step is too small, the iteration efficiency is low, and if the step is too large, it may exceed the optimal value or even reach the upper bound in one step. Taking the aforementioned scenario as an example, the ripple period corresponds to 24 points, the increment baseline value is 24 × 0.5 = 12 points, and the smallest odd number not less than 12 is 13. Therefore, the window increases by 13 points in each iteration. After the first iteration, the window increases from 25 points to 38 points, and the smallest odd number not less than 38 is taken as 39 points.

[0060] The execution flow for each iteration is as follows: The Savitzky-Gore filter is re-executed on the segment with the new window length; the residuals are recalculated; the autocorrelation value and noise floor of the residuals at the hysteresis corresponding to the ripple period are recalculated; and the suppression criterion is recalculated. The iteration terminates if the suppression criterion falls below the convergence threshold. If the new window length has reached or exceeded the upper bound of the segment, the window is truncated to the largest odd number corresponding to the upper bound, the iteration terminates after the final filtering, and the segment is marked as having reached the upper bound of the window.

[0061] The maximum number of iterations is set to 10. If the upper bound is not reached after 10 iterations and the suppression criterion still exceeds the convergence threshold, the iteration is terminated and the segment is treated as if the upper bound of the window has been reached. The basis for setting this upper limit is that the step size is half a ripple period, and 10 iterations correspond to a window size that increases from 1 ripple period to approximately 6 ripple periods, which is far beyond what is normally required. In engineering practice, for standard standing wave ripples, convergence is usually achieved in 2 to 3 iterations. If convergence is still not achieved after 10 iterations, it indicates that the ripple characteristics of the segment are abnormally complex or non-single-frequency, and the time-domain smoothing method is no longer applicable, requiring a shift to frequency-domain directional processing.

[0062] Let's take a specific segment from the aforementioned scenario as an example to illustrate the iterative process: A segment lies between 5 GHz and 15 GHz, with an upper window bound of 4.8 GHz, corresponding to 96 points, and a maximum odd number of 95 points. After filtering with an initial window of 25 points, the autocorrelation value of the residual at lag 24 points is 0.43, the noise floor is 0.09, and the suppression criterion is 0.43 ÷ 0.09 ≈ 4.8, exceeding the convergence threshold of 2. In the first iteration, the window is increased to 39 points, and after re-filtering, the suppression criterion decreases to 2.6, still exceeding the limit. In the second iteration, the window is increased to 53 points, and after re-filtering, the suppression criterion decreases to 1.4, below the convergence threshold of 2, and the segment converges, with 53 points becoming the final window. Throughout the process of increasing the window from 25 points to 53 points, the upper window bound of 95 points is never exceeded, and the filtering operation is safely limited within the passband range.

[0063] This module establishes an automatic convergence closed-loop control mechanism: residual autocorrelation is used to measure the degree of ripple residue detected by the sensor, the window width is used to adjust the actuator response, and the ripple period serves as both the sensor's detection parameter and the actuator's stepping parameter. This feedback loop ensures that the filtering window is increased precisely to the minimum value where the ripple is just sufficiently suppressed, preventing ripple residue from being left due to an excessively small window, and avoiding obscuring the detailed characteristics of the device's intrinsic response due to an excessively large window. For sections that have not converged even after reaching the upper limit of the window, the ripple component in the residual will be processed through frequency-domain directional separation.

[0064] The phase verification module is used to extract the single-frequency component corresponding to the ripple period in the residual of the segment that reaches the upper boundary of the window, extract the ripple phase corresponding to the ripple period in the residual of the adjacent converged segment, and obtain the expected phase by linear extrapolation of the ripple phase along the frequency. If the difference between the actual phase of the single-frequency component and the expected phase exceeds a second preset threshold, the single-frequency component is retained; otherwise, the single-frequency component is subtracted from the residual.

[0065] For each segment that reaches the upper bound of the window during the above iteration process, a directional ripple separation operation is performed. Due to the window limitation, the ripple in such segments cannot be fully suppressed by time-domain smoothing. Therefore, the ripple components need to be extracted and separated in the frequency domain using the precise frequency information of the ripple period.

[0066] The single-frequency component corresponding to the ripple period in the residual is extracted. A bandpass filter is constructed with the reciprocal of the ripple period as the center frequency. This bandpass filter is applied to the residual in this section to obtain the single-frequency component. The passband width of the bandpass filter is set to 20% of the center frequency, i.e., 10% to the left and right of the center frequency. This passband width is set based on the fact that cable temperature drift and slight connector displacement in the production line environment can cause a slight shift in the ripple period during testing. The 20% passband width provides sufficient frequency tolerance to fully capture the ripple energy without being too wide to mistakenly include other frequency components. The bandpass filter adopts a zero-phase-shift finite impulse response structure, using forward filtering followed by reverse filtering to eliminate phase distortion and ensure that the extracted single-frequency component retains its true phase information.

[0067] Determine the actual phase of each single-frequency component. Calculate the inner product of the extracted single-frequency component with the sine and cosine basis functions of the center frequency: multiply the single-frequency component sequence point-by-point with the sine basis function and sum to obtain the sine projection coefficients; multiply the sequence point-by-point with the cosine basis function and sum to obtain the cosine projection coefficients. Take the arctangent of the sine and cosine projection coefficients to obtain the actual phase of the single-frequency component at the center frequency of the current segment. The frequencies of both the sine and cosine basis functions are equal to the reciprocal of the ripple period, and the independent variables are the offsets of each frequency point within the segment relative to the starting frequency of the segment.

[0068] Extract the ripple phase corresponding to the ripple period from the residuals of adjacent converged segments. Select a segment directly adjacent to the current upper bound segment of the window that has converged normally in the above iterations, and perform the same bandpass filtering and inner product operation on the final residual of this converged segment to obtain the ripple phase value at the converged segment. The suppression criterion of the converged segment has dropped below the convergence threshold, and the ripple residual in its residual is extremely weak, but the weak ripple phase information can still be extracted through the narrowband gain of the bandpass filter. If there are converged segments on both sides of the current segment, extract the ripple phase values ​​of both sides respectively, and take the weighted average of the two as the reference ripple phase. The weight is inversely proportional to the frequency distance of each from the center of the current segment. If there is only a converged segment on one side, use the ripple phase value of that side as the reference ripple phase.

[0069] The expected phase is obtained by linearly extrapolating the reference ripple phase along the frequency. The physical basis for extrapolation is that the standing wave ripple generated by the same pair of impedance mismatch points has a spatially continuous phase change over the entire frequency range, the ripple phase increases linearly with frequency, and the slope is uniquely determined by the electrical length between the mismatch points. The extrapolation slope is equal to 360 degrees divided by the ripple period, with the unit being degrees per gigahertz. Adding the reference ripple phase value to the extrapolation slope multiplied by the frequency difference between the center frequency of the reference segment and the center frequency of the current segment yields the expected phase at the current segment.

[0070] Taking the aforementioned scenario as an example, the extrapolation process is illustrated as follows: the ripple period is 1.2 GHz, and the extrapolation slope is 360 ÷ 1.2 = 300 degrees per GHz. The center frequency of a segment reaching the upper limit of the window is 27 GHz, and the center frequency of its adjacent converged segment to the left is 24 GHz. The ripple phase extracted from this converged segment is 45 degrees. The frequency difference is 27 - 24 = 3 GHz, and the expected phase is 45 + 300 × 3 = 945 degrees. Normalized to the 0 to 360 degree interval, this is 945 - 2 × 360 = 225 degrees.

[0071] The actual phase of a single-frequency component is compared with its expected phase, and the absolute value of the difference between the two is calculated. The determination rule is as follows: If the absolute value of the difference between the actual phase and the expected phase is greater than the second preset threshold, it is determined that the single-frequency component extracted in the current segment does not originate from the standing wave of the test link, but is a true feature in the intrinsic response of the device. This single-frequency component is retained without subtraction, and the final processing result for this segment is the output value of the Savitzky-Gore filter at the upper boundary of the window, thus retaining the filtering result containing intrinsic features. Simultaneously, this segment is marked as containing intrinsic features.

[0072] If the absolute value of the difference between the actual phase and the expected phase is less than or equal to the second preset threshold, then the single-frequency component is determined to be the VSWR ripple of the test link, and the phase consistency verification is passed. The single-frequency component is subtracted from the residual to obtain the corrected residual. The corrected residual is added back to the Savitzky-Gore filter output value for that section to obtain the final processing result for that section.

[0073] Figure 3 This diagram illustrates the comparison between the original residual, the extracted single-frequency component, and the corrected residual for a certain arrival window upper bound segment. The original residual exhibits regular oscillations with ripple periods as intervals. The single-frequency component, extracted using a bandpass filter with the reciprocal of the ripple period as the center frequency, has an oscillation envelope that closely matches the ripple component in the original residual. The corrected residual, obtained by subtracting the single-frequency component from the original residual, significantly suppresses ripple oscillations, leaving only residual fluctuations at the level of random noise.

[0074] The second preset threshold is set at 60 degrees. This threshold is based on the following: under ideal conditions, the phase deviation of co-originating ripples should be 0 degrees. In engineering practice, changes in probe contact state and ambient temperature cause slight drifts in the cable's electrical length, resulting in a certain deviation between the actual ripple phase and the theoretical linear extrapolated value. 60 degrees corresponds to 60 degrees of the ripple period. Spatial offset covers the phase drift range under normal production line conditions. The data-driven calibration method for this threshold is as follows: insert and remove the probe 10 times consecutively at the same test station and measure the same device, calculate the standard deviation of the ripple phase fluctuation in each measurement, and take 3 times the standard deviation as the actual threshold. If the value deviates significantly from 60 degrees, the measured value is used instead.

[0075] The core physical basis for phase consistency verification is that standing wave ripples from the same physical source satisfy phase spatial continuity in the frequency domain, which is determined by the causality of electromagnetic wave propagation; while the intrinsic resonance of the device may be at the same frequency as the ripple in the frequency domain, its phase is independently determined by the physical parameters of the resonant structure and is independent of the electrical length of the test link. There is no linear extrapolation relationship between the phase values ​​of the two.

[0076] To illustrate the identification effect of phase verification with a specific scenario: A detector has a real depression caused by encapsulation resonance at 15 GHz. After the segment reaches the upper limit of the window, the difference between the actual phase of the single-frequency component extracted and the expected phase extrapolated from the adjacent converged segment is 138 degrees, which is far beyond the 60-degree threshold. It is judged as an intrinsic feature and retained. This real depression is not mistakenly treated as ripple and eliminated.

[0077] When there are no converged segments on either side of the segment reaching the upper limit of the window, a reference ripple phase cannot be obtained for extrapolation. In this case, phase consistency verification is not performed on this segment; instead, single-frequency component subtraction is performed directly, and the segment is marked as an unverified segment, which is then classified as a low-confidence segment in subsequent confidence assessments.

[0078] This module utilizes the ripple phase information generated by the converged segment to provide a reference for phase verification of the segment reaching the upper boundary of the window, forming information mutual correction between segments. In this module, the ripple period simultaneously serves as both a notch frequency parameter and a phase extrapolation slope parameter, forming a quadruple reuse together with the detection hysteresis parameter and window stepping parameter in the aforementioned module. These four roles are closely related through the same physical essence.

[0079] The bandwidth extraction module is used to stitch together the results of each segment to obtain a smooth curve, and extract the -3dB bandwidth from the smooth curve.

[0080] The final processing results of each segment are concatenated in frequency order to form a smooth curve covering the entire test frequency band. The final processing result of the converged segment is the Savitzky-Gorye filter output value at the time of iterative convergence; for the segment that reaches the upper limit of the window and passes the phase consistency verification, the final processing result is the corrected value of the filter output after subtracting the single-frequency component from the residual; for the segment that reaches the upper limit of the window and the single-frequency component is retained, the final processing result is the Savitzky-Gorye filter output value at the upper limit of the window.

[0081] The splicing position between adjacent segments is at the frequency of the phase abrupt change. Since the filtering window of each segment is truncated to the segment's interior near the boundary, the output values ​​of the segments on both sides of the splicing point are independently determined by the data within their respective segments, theoretically potentially resulting in slight numerical discontinuities. A linear transition is performed at the splicing point: within a range of 3 frequency points on each side of the splicing point, linear interpolation is used to smoothly connect the output values ​​on both sides, eliminating jumps. The reason for selecting 3 frequency points in the transition range is that this range is much smaller than the number of sampling points corresponding to the ripple period, and the transition operation will not introduce perceptible signal distortion, only eliminating numerical steps at the splicing point.

[0082] Extract the -3dB bandwidth from the smoothed curve. Determine the maximum amplitude value of the smoothed curve across the entire frequency band, and subtract 3 dB from this maximum value to obtain the -3dB threshold line. Scan the smoothed curve point by point from the high-frequency end to the low-frequency end, and find the frequency point where the curve first drops below the -3dB threshold line. Perform linear interpolation between this point and the previous point to calculate the precise frequency position where the curve is exactly equal to the -3dB threshold; this frequency is the -3dB cutoff frequency. The -3dB bandwidth value is obtained by subtracting the test start frequency from the -3dB cutoff frequency. The reason for searching from the high-frequency end to the low-frequency end rather than from the low-frequency end to the high-frequency end is that the frequency response of the RF detector is usually flat at the low-frequency end and rolls off at the high-frequency end. The -3dB cutoff frequency is located in the high-frequency roll-off region. Starting the search from the high-frequency end allows for faster target location and avoids false triggering caused by residual local ripple within the passband.

[0083] After extracting the -3dB bandwidth from the smoothed curve, a confidence assessment is performed. The number of segments that reach the upper bound of the window and whose single-frequency components are retained is counted, and the ratio of this number to the total number of segments is calculated as the anomaly ratio. This ratio reflects the proportion of segments in the global measurement that require phase verification and whose identification results are intrinsic features. A higher proportion indicates more complex test link conditions and a greater dependence of the processing results on identification accuracy.

[0084] The rules for determining the confidence level are as follows: When the proportion of abnormal segments is below the first proportion threshold, the confidence level is high. High confidence indicates that the vast majority of segments can achieve ripple suppression through regular iterations, the test link is in good condition, and the bandwidth extraction results can be directly used for device classification. The first proportion threshold is set to 0.1, meaning the proportion of abnormal segments is less than 10%.

[0085] When the proportion of anomalies falls between the first and second proportion thresholds, the confidence level is medium. Medium confidence indicates that some segments possess intrinsic characteristics or complex ripple environments, but the proportion is still within a controllable range. The bandwidth results can be used for classification, but it is recommended to record this confidence level in the database for future reference. The second proportion threshold is set to 0.3, meaning the proportion of anomaly segments is less than 30%.

[0086] When the anomaly ratio exceeds the second ratio threshold, the confidence level is low. Low confidence indicates that the ripple state of a large number of sections is abnormally complex or that there are multiple intrinsic features. The test link conditions may be unsatisfactory. It is recommended that the operator readjust the probe contact state or check the cable connection before performing a retest.

[0087] The first proportional threshold of 0.1 and the second proportional threshold of 0.3 are set based on the following: Under normal production line conditions, when the probe contact is good, only 1 to 2 sections near the roll-off zone may touch the upper limit of the window. For a typical division of 5 to 10 sections, the abnormality rate is usually below 0.1. When the probe contact begins to deteriorate or the cable condition is abnormal, window-limited sections also begin to appear in the middle of the passband, and the abnormality rate rises to the range of 0.1 to 0.3. At this time, the results are still basically reliable, but attention is required. When the abnormality rate exceeds 0.3, it means that more than 30% of the sections are in a non-standard processing state, and the reliability of the results cannot be guaranteed.

[0088] The residual statistics and phase verification results generated during the iteration process are comprehensively reused in this module as diagnostic criteria for measurement quality, realizing a closed-loop data flow from processing intermediate data to final quality assessment. This assessment mechanism allows production line operators to obtain the reliability level of each measurement without manual waveform interpretation, and promptly performs retests for low-confidence results, reducing the grading error rate caused by single measurement anomalies.

Claims

1. A system for processing high-frequency response test data of a radio frequency detector, characterized in that, The processing system includes: a ripple period extraction module, used to sweep the frequency of the RF detector under test to obtain amplitude and phase data sequences, and calculate the autocorrelation function after detrending the amplitude data sequence to extract the ripple period; a phase change point calibration module, used to dewind the phase data sequence and perform second-order difference, and calibrate the frequency points where the slope change rate exceeds a first preset threshold as phase change points; and an adaptive filtering module, used to divide the window into segments with the distance from each frequency point to the nearest phase change point as the upper bound, filter each segment with the ripple period as the initial window to obtain residuals, and use the ratio of the autocorrelation value of the residual at the hysteresis corresponding to the ripple period to the noise background as the suppression criterion, and increase the window and iterate until convergence or the upper bound of the window is reached when the suppression criterion exceeds the limit. The phase verification module is used to extract the actual phase of the single-frequency component in the residual of the segment that reaches the upper limit of the window, and extrapolate the expected phase by combining it with the ripple phase of the adjacent converged segment. If the difference between the actual phase and the expected phase of the single-frequency component exceeds the second preset threshold, the component is retained; otherwise, it is subtracted from the residual. The bandwidth extraction module is used to stitch together the processing results of each segment to obtain a smooth curve and extract the -3dB bandwidth from it. The processing result of the converged segment is the filter value at the time of iterative convergence. For the segment that reaches the upper limit of the window and the single-frequency component is not retained, the processing result is the correction value of the residual after subtracting the single-frequency component and adding back the filter value. For the segment that reaches the upper limit of the window and the single-frequency component is retained, the processing result is the filter value at the upper limit of the window.

2. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The method for detrending the amplitude data sequence is as follows: a trend curve is obtained by least-squares fitting of the amplitude data sequence with a third-order polynomial, and the detrended amplitude sequence is obtained by subtracting the trend curve from the amplitude data sequence.

3. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The method for calculating the autocorrelation function to extract the ripple period is as follows: the lower bound of the search for the autocorrelation function is set to a preset multiple of the frequency resolution, and the upper bound of the search is a first preset proportion of the total frequency span; the condition for determining a local maximum is that the autocorrelation value is simultaneously greater than the adjacent previous value and the adjacent subsequent value and is greater than zero; if there is no local maximum within the search range that satisfies the determination condition, it is determined that the amplitude data sequence does not contain ripple, and the second preset proportion of the total frequency span is used as the substitute value for the ripple period.

4. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The first preset threshold is determined by calculating the mean of the absolute value sequence of the slope change rate, and multiplying the mean by a preset multiple to obtain the first preset threshold.

5. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The method of dividing the frequency axis into segments by using the distance from each frequency point to the nearest phase change point as the upper bound of the window is as follows: the frequency axis is divided into several segments with the phase change point as the boundary, and the upper bound of the window for any frequency point in each segment is equal to the distance from that frequency point to the nearest phase change point; when the frequency interval between two adjacent phase change points is less than a preset multiple of the ripple period, the upper bound of the window for all frequency points in that segment is limited to half of that frequency interval.

6. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, When filtering each segment using the ripple period as the initial window, a Savitzky-Gore filter is used, with a polynomial order of not less than third order, an odd window length, and an initial value that is not less than the smallest odd number of the ripple period.

7. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The noise background is determined as follows: the autocorrelation value of the residual autocorrelation function at each lag point is calculated, and after excluding the zero lag point and the lag point corresponding to an integer multiple of the ripple period, the arithmetic mean of the absolute values ​​of the autocorrelation values ​​at the remaining lag points is taken as the noise background; when the noise background is less than a preset minimum value, the preset minimum value is used to replace the noise background.

8. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The method of increasing the window and iterating until convergence or reaching the upper bound of the window when the suppression criterion exceeds the limit is as follows: the smallest odd number of the product of the ripple period and the preset ratio is used as the window increment, and the current window length is increased by the window increment in each iteration; the convergence condition is that the suppression criterion does not exceed the preset convergence threshold. Set a maximum number of iterations. If the iteration fails to converge after reaching the maximum number of iterations, terminate the iteration and process the case where the upper bound of the window has been reached.

9. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, The method for extracting the actual phase of the single-frequency component in the residual is as follows: a bandpass filter is constructed with the reciprocal of the ripple period as the center frequency, and the residual is bandpass filtered to obtain the single-frequency component; the actual phase of the single-frequency component is obtained by taking the inner product of the single-frequency component with the sine basis function and the cosine basis function of the center frequency, and then taking the arctangent.

10. The system for processing high-frequency response test data of a radio frequency detector according to claim 1, characterized in that, After extracting the -3dB bandwidth, a confidence assessment is performed, including: calculating the ratio of the number of segments that touch the upper bound of the window and whose single-frequency component is retained to the total number of segments as the anomaly ratio; when the anomaly ratio is less than or equal to a first ratio threshold, the confidence level is high; when the anomaly ratio is greater than the first ratio threshold and less than a second ratio threshold, the confidence level is medium; and when the anomaly ratio is greater than or equal to the second ratio threshold, the confidence level is low.

Images