Wafer test carrier table micro-shaking real-time compensation method and system based on machine vision

Abstract

The application discloses a wafer test bearing table micro-shaking real-time compensation method and system based on machine vision, relates to the technical field of micro-vibration compensation, and comprises the following steps: S1, collecting the vibration acceleration signal of the bearing table, and driving the inner ring compensation executor to continuously apply active vibration compensation force to the bearing table; S2, triggering the visual check according to the preset trigger condition, wherein the trigger condition is at least one of working condition gap triggering, periodic triggering and abnormal triggering. The wafer test bearing table micro-shaking real-time compensation method and system have a double-ring structure combining an inner ring compensation system and an outer ring calibration architecture, the inner ring compensation system undertakes the rapid suppression of high-frequency vibration, and the outer ring calibration architecture can check and correct the control parameters of the inner ring compensation system under the premise that the inner ring compensation system continuously operates, thereby helping to alleviate the long-term influence of slow drift of the control parameters on the compensation performance.

Description

Technical Field

[0001] This invention relates to the field of micro-vibration compensation technology, specifically to a real-time micro-shaking compensation method and system for a wafer testing platform based on machine vision. Background Technology

[0002] Wafer testing is a critical process in semiconductor manufacturing to ensure chip yield. During testing, the stage needs to perform high-precision stepping positioning to provide a stable alignment between the probe card and the wafer pads. As wafer size continues to increase and testing cycle time continues to decrease, the stage's ability to suppress residual micro-vibrations after stepping motion and its ability to maintain long-term positioning accuracy have become key factors restricting the improvement of testing efficiency and yield.

[0003] Existing vibration suppression solutions mainly employ a combination of passive isolation and active compensation. Passive isolation attenuates external foundation vibrations through air-bearing or spring elements, but its effectiveness in suppressing transient responses caused by structural resonance and stepping motion of the bearing platform itself is limited. Active compensation solutions typically use accelerometers to collect vibration signals and drive actuators via feedback controllers to apply reverse compensation forces, thereby achieving rapid suppression of disturbances in the inner loop.

[0004] However, in the long-term operation of actual wafer testing, the control parameters of the inner loop compensation system will slowly drift due to changes in ambient temperature, mechanical structure creep and electronic device aging, resulting in degradation of compensation performance over time. There is no cross-system collaborative calibration mechanism between the inner loop compensation system and the external vision alignment system, and the long-term accuracy of the inner loop cannot be adaptively corrected through external observation. The transient vibrations caused by stepping motion and the continuous micro-vibrations under the test steady state have different spectral characteristics, and the compensation strategy with fixed parameters is difficult to take into account the suppression effect under multiple operating conditions.

[0005] The aforementioned issues collectively make it difficult for existing solutions to balance real-time compensation bandwidth with long-term accuracy and adaptive capability. To address this, we propose a machine vision-based real-time micro-wobbling compensation method and system for wafer testing platforms. Summary of the Invention

[0006] To address the problems raised in the background art, this invention provides a real-time compensation method for micro-wobbles on a wafer testing platform based on machine vision, specifically including the following steps:

[0007] S1. Collect the vibration acceleration signal of the bearing platform and drive the inner ring compensation actuator to continuously apply active vibration compensation force to the bearing platform;

[0008] S2. Trigger visual verification according to preset triggering conditions, wherein the triggering conditions are at least one of: working condition interval triggering, periodic triggering, and abnormal triggering;

[0009] S3. After the visual verification is triggered, the visual acquisition module continuously acquires multiple frames of images at a frame rate lower than the inner loop compensation control frequency. Each frame simultaneously contains the fixed reference mark on the vibration isolation base and the follow-up calibration mark on the bearing platform. The relative displacement of the two marks is extracted frame by frame to obtain the displacement time series.

[0010] S4. Calculate the steady-state zero-position deviation and residual vibration characteristic quantities based on the displacement time series; the steady-state zero-position deviation is the offset of the series mean relative to the calibrated zero position; the residual vibration characteristic quantities include at least one of the following: peak-to-peak value, standard deviation, and frequency band amplitude distribution.

[0011] S5. Based on the steady-state zero-position deviation and residual vibration characteristics, perform at least one layer of correction on the inner loop compensation parameters according to the parameter correction strategy. The parameter correction strategy includes four types of correction: zero-position bias, compensation gain, phase parameter and model parameter iteration. Each layer of correction sets an independent dead zone threshold. If the deviation is less than the corresponding threshold, the correction layer is not performed.

[0012] During the parameter correction process, the inner loop compensation continues to operate.

[0013] Preferably, the working condition gap triggering is specifically triggered within the steady-state waiting window before the probe contacts the test after the bearing platform completes the stepping motion, and the verification process overlaps with the test preparation time;

[0014] The periodic triggering specifically involves determining a fixed time interval for triggering based on the drift rate of the inner loop compensation parameters.

[0015] The specific abnormal triggering is triggered when the accelerometer detects that the vibration amplitude exceeds a preset threshold, or when it detects a change in operating conditions. The change in operating conditions includes: before and after probe insertion and the start and stop of the platform.

[0016] If the periodic triggering time conflicts with the working condition interval triggering time, it will be postponed to the next working condition interval.

[0017] If multiple triggering conditions are met simultaneously, the triggers will be executed according to their priority, from abnormal triggering, intermittent triggering, to periodic triggering.

[0018] Preferably, the method for determining the inner loop compensation control frequency is as follows:

[0019] The lower limit is determined by using the highest effective frequency of the micro-vibration of the bearing platform as the basic lower limit, combined with the timing alignment constraint of an integer multiple of the visual sampling frame rate and the upper limit of the actuator response bandwidth, satisfying the following conditions:

[0020]

[0021] In the formula, For inner loop compensation control frequency; To control the multiplication factor of the frequency relative to the dominant vibration frequency; This is the highest effective frequency of the micro-vibration of the support platform; It is an integer multiple coefficient of the inner loop control frequency relative to the visual frame rate; This refers to the sampling frame rate of the visual acquisition module. This is a ratio factor between the control frequency and the actuator bandwidth; This is the inherent response bandwidth of the inner loop compensation actuator.

[0022] Preferably, the single exposure time for visual acquisition is determined based on the sampling frame rate and the target subpixel positioning accuracy. Each time a verification is triggered, multiple frames of images are continuously acquired and a displacement time sequence within the sampling window is formed. The subpixel positioning includes at least one of gradient-based subpixel interpolation, phase-correlation-based subpixel registration, or moment-based subpixel positioning.

[0023] Extract the subpixel-level center positions of the fixed reference marker and the follow-up calibration marker from a single frame image, and calculate their relative displacement.

[0024] The relative displacement of each frame is continuously calculated within the sampling window to form a displacement time series.

[0025] Preferably, the mean value of the displacement time series within the sampling window is calculated to obtain the steady-state zero position;

[0026] Calculate the offset of the steady-state zero position relative to the zero position reference during the initial calibration of the system to obtain the steady-state zero position deviation;

[0027] Time-domain statistical calculations were performed on the displacement time series to obtain the peak-to-peak value and standard deviation, which were used as time-domain characteristic quantities of residual vibration.

[0028] The displacement time series is subjected to frequency domain transformation to extract the amplitude components of each frequency band, forming the residual vibration spectrum distribution;

[0029] The residual vibration spectrum distribution is compared with the calibration reference spectrum to identify the target frequency band with higher amplitude as the frequency domain subdivision feature of the residual vibration.

[0030] By combining time-domain features with frequency-domain subdivision features, complete residual vibration features are obtained.

[0031] Preferably, the zero-position offset correction specifically includes:

[0032] When the steady-state zero deviation exceeds the corresponding dead zone threshold, a reverse DC bias is added to the output of the inner loop controller.

[0033] The magnitude of the DC bias corresponds to the value of the steady-state zero-point deviation;

[0034] Based on the stability margin of the inner loop compensation system and the rated stroke of the inner loop compensation actuator, the upper limit of the single correction step size of the DC bias is set.

[0035] Preferably, the compensation gain correction specifically includes:

[0036] When the residual vibration amplitude of the target frequency band is higher than the calibration reference value and exceeds the corresponding dead zone threshold in multiple consecutive verifications, the compensation gain of the inner loop controller in that frequency band is increased.

[0037] When the residual vibration amplitude of the target frequency band is lower than the calibration reference value and exceeds the corresponding dead zone threshold in multiple consecutive verifications, the compensation gain of the inner loop controller in that frequency band is reduced.

[0038] The upper and lower limits of the compensation gain adjustment are determined based on the stability boundary of the inner loop compensation system; the upper limit of the single adjustment step size of the compensation gain is determined based on the gain margin of the inner loop compensation system.

[0039] Preferably, the phase parameter correction specifically includes:

[0040] The displacement phase of the target frequency band is extracted from the displacement time series, and the phase of the inner loop feedforward channel excitation signal in the target frequency band is obtained synchronously. The phase difference between the two is calculated.

[0041] Adjust the delay parameters of the inner loop feedforward channel according to the value and direction of the phase difference;

[0042] The upper limit of the single adjustment step size of the delay parameter is determined based on the phase margin of the inner loop compensation system;

[0043] The effective frequency band for phase correction does not exceed the Nyquist frequency corresponding to the visual sampling frame rate.

[0044] Preferably, the iterative correction of the model parameters specifically involves:

[0045] The displacement time series acquired visually is used as the true value reference and compared with the output of the inner ring vibration prediction model to calculate the prediction error. When the mean square value of the prediction error after multiple consecutive verifications exceeds a preset threshold, the parameters of the inner ring vibration prediction model are updated through an online parameter identification algorithm. The preset threshold is determined according to the model accuracy requirements of the inner ring compensation system.

[0046] The corrections for each layer are performed sequentially in the order of zero-bias correction, compensation gain correction, phase parameter correction, and model parameter iterative correction.

[0047] The upper limit of the single adjustment step size for each level of correction is determined independently based on the system stability margin requirements of the corresponding level;

[0048] All outer loop correction actions are frozen when any of the following conditions are met: the inner loop compensation actuator is in saturated output state; or the visual acquisition detects image blurring or abnormal marker occlusion.

[0049] The independent dead zone thresholds for each layer are determined based on the visual measurement noise level of the corresponding layer and the inherent drift rate of the inner loop parameters.

[0050] It also provides a machine vision-based real-time micro-wobbling compensation system for wafer test stage, including:

[0051] An accelerometer sensor is mounted on the support platform to collect the vibration acceleration signal of the support platform.

[0052] An inner-loop compensation actuator, connected to the acceleration sensor, is used to continuously apply an active vibration compensation force to the support platform based on the vibration acceleration signal.

[0053] The reference mark assembly includes a fixed reference mark fixed on the vibration isolation base and a follow-up calibration mark fixed on the support platform;

[0054] The visual acquisition module is fixedly installed on the vibration isolation base and is used to continuously acquire multiple frames of images at a frame rate lower than the inner loop compensation control frequency after visual verification is triggered. Each frame contains both the fixed reference mark and the follow-up calibration mark.

[0055] The calibration control module is used to extract the relative displacement of two markers frame by frame from continuous images to obtain the displacement time series, calculate the steady-state zero-position deviation and residual vibration characteristics, and perform at least one layer of correction on the inner loop compensation parameters according to the parameter correction strategy. The parameter correction strategy includes four types of correction: zero-position bias, compensation gain, phase parameter and model parameter iteration, and each layer of correction sets an independent dead zone threshold.

[0056] During the parameter correction process performed by the calibration control module, the inner loop compensation actuator continues to operate continuously.

[0057] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0058] The wafer test stage micro-wobbling real-time compensation method and system proposed in this invention adopts a dual-loop structure combining an inner-loop compensation system and an outer-loop calibration architecture. The inner-loop compensation system is responsible for the rapid suppression of high-frequency vibrations; the outer-loop calibration architecture can verify and correct the control parameters of the inner-loop compensation system while it is running continuously, which helps to mitigate the long-term impact of slow drift of control parameters caused by changes in ambient temperature, mechanical creep, or aging of electronic components on compensation performance. Zero-point offset correction is used to eliminate steady-state DC offset, and compensation gain correction is verified through multiple consecutive checks to confirm the adjustment trend of frequency band amplitude to reduce random fluctuations. To mitigate the possibility of erroneous adjustments, phase parameter correction corrects timing deviations by extracting the phase relationship between the displacement phase and the excitation signal phase of the inner loop feedforward channel. Iterative correction of model parameters addresses systematic deviations between the inner loop vibration prediction model and the actual response. Each correction layer has an independently set dead zone threshold, with progressively increasing levels, which helps maintain parameter stability under small fluctuations and trigger timely updates under significant deviations. When the inner loop compensation actuator is in a saturated output state or the visual acquisition conditions are not met, all outer loop correction actions are frozen and restored after the anomaly is eliminated, thus preventing the injection of erroneous correction parameters when visual verification conditions are abnormal. Attached Figure Description

[0059] Figure 1 This is a flowchart of the real-time compensation method for micro-shaking of the wafer testing stage according to the present invention;

[0060] Figure 2 This is a schematic diagram of the real-time micro-shake compensation system module for the wafer testing stage of the present invention. Detailed Implementation

[0061] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0062] Reference Figure 1 As shown, the real-time compensation method for micro-shake of a wafer testing platform based on machine vision includes the following steps:

[0063] S1. Acquire the vibration acceleration signal of the bearing platform and drive the inner ring compensation actuator to continuously apply active vibration compensation force to the bearing platform; the inner ring compensation actuator is a linear drive device based on piezoelectric ceramics, voice coil motors or magnetostrictive materials. Its input end receives the compensation control signal after power amplification, and its output end is mechanically connected to the bearing platform. It is used to apply active vibration compensation force to the bearing platform according to the compensation control signal.

[0064] S2. Trigger visual verification according to preset trigger conditions. The trigger conditions are: at least one of the following: working condition interval trigger, periodic trigger, and abnormal trigger.

[0065] The specific triggering of the working condition gap is as follows: after the bearing platform completes the stepping motion, the probe is triggered within the steady-state waiting window before the test, and the verification process overlaps with the test preparation time.

[0066] The periodic triggering process is as follows: obtain the maximum allowable drift amount of the inner loop compensation parameter, divide the maximum drift amount by the drift rate of the inner loop compensation parameter per unit time to obtain a fixed time interval, and determine the fixed time interval for triggering based on the drift rate of the inner loop compensation parameter.

[0067] The maximum drift is determined based on the control accuracy requirements of the inner loop compensation system. The drift rate per unit time is obtained by collecting the parameter offset after the inner loop compensation system has been running continuously for a preset observation period offline and dividing it by the observation period. The preset observation period needs to cover the complete transition process of the inner loop compensation system from cold start to thermal steady state, and on this basis, it should be extended for a sufficient period of time to obtain the parameter offset trend under steady state. In a specific example, the observation period is 24 hours.

[0068] The specific abnormal triggering is as follows: it is triggered when a change in working condition is detected, or when the acceleration sensor detects that the vibration amplitude exceeds the preset threshold, the time exceeding the threshold is recorded, and visual verification is triggered after the vibration amplitude falls back below the preset threshold; when the inner loop compensation system is running normally and the bearing platform is in steady-state test conditions, the acceleration signal is continuously collected for a preset duration and its root mean square value is calculated. This root mean square value is used as the steady-state vibration reference value. The preset threshold is a preset multiple of the steady-state vibration reference value. The value of this multiple must be higher than the normal fluctuation range of the vibration amplitude under steady-state conditions. In a specific example, this multiple is 3. The working condition switching includes: before and after probe insertion and the start and stop of the platform;

[0069] If the cycle trigger and the working condition gap trigger conflict, the working condition gap trigger will be executed first, and the cycle trigger will be postponed to the next working condition gap. The working condition gap trigger corresponds to the steady-state waiting window in the test process. This window is limited in time and directly related to the production cycle. If it is missed, it must wait for the next complete test cycle before it can be used again. The essence of the cycle trigger is to complete the verification before the parameter drift accumulates to an unacceptable level. Its trigger time interval is determined by the drift rate. A moderate delay will not cause the parameter drift to exceed the allowable range.

[0070] If multiple triggering conditions are met simultaneously, the triggers will be executed according to their priority, from abnormal triggering, intermittent triggering, to periodic triggering.

[0071] The priority ranking is based on the following: abnormal triggers correspond to sudden changes in operating conditions, such as probe puncture or platform start-up and shutdown. At this time, the dynamic characteristics of the bearing platform change significantly, and the inner loop compensation parameters need to be updated as soon as possible to adapt to the new operating conditions. Therefore, the priority is the highest.

[0072] The verification is triggered during the break between working conditions, taking advantage of the idle window of the test preparation time. It overlaps with the test process time and does not affect the production cycle, thus having a time efficiency advantage.

[0073] Periodic triggering is a timed verification based on parameter drift rate. Its time constraint is the most lenient, and it can be executed in subsequent idle windows without affecting the compensation performance.

[0074] S3. After the visual verification is triggered, the visual acquisition module continuously acquires multiple frames of images at a frame rate lower than the inner loop compensation control frequency. Each frame simultaneously contains the fixed reference mark on the vibration isolation base and the follow-up calibration mark on the bearing platform. The relative displacement of the two marks is extracted frame by frame to obtain the displacement time series.

[0075] The specific method for determining the inner loop compensation control frequency is as follows:

[0076] The control frequency is determined based on the highest effective frequency of the platform's micro-vibration as the lower limit, combined with the timing alignment constraint (an integer multiple of the visual sampling frame rate) and the upper limit of the actuator response bandwidth. The following conditions must be met: the inner-loop compensation control frequency must simultaneously satisfy the lower limit constraint, the timing alignment constraint, and the upper limit constraint. The timing alignment constraint determines the candidate values ​​for the control frequency, while the lower limit constraint and the upper limit constraint are used to verify whether the candidate values ​​are within the feasible region. Specifically:

[0077]

[0078] In the formula, For inner loop compensation control frequency; To control the multiplication factor of the frequency relative to the dominant vibration frequency; This is the highest effective frequency of the micro-vibration of the support platform; It is an integer multiple coefficient of the inner loop control frequency relative to the visual frame rate. Its value is determined by the timing alignment requirements of the visual frame rate and the inner loop control frequency, so that a single frame visual acquisition cycle contains exactly k complete inner loop control cycles. This refers to the sampling frame rate of the visual acquisition module. This is a ratio factor between the control frequency and the actuator bandwidth; This is the inherent response bandwidth of the inner loop compensation actuator.

[0079] Multiplier The value of this parameter must satisfy two constraints: the lower limit must ensure that the control frequency is sufficiently higher than the dominant vibration frequency to provide adequate phase compensation margin; the upper limit must prevent the control frequency from being too high, which would cause the computational load on the digital controller to exceed its real-time processing capacity. A sufficient number of inner-loop control cycles must be included within a single-frame visual acquisition cycle to ensure the phase statistical accuracy of the inner-loop behavior during outer-loop verification and the timing alignment accuracy between the visual and inner-loop sequences. (Proportional coefficient) The value of is determined based on the flat range of the actuator's amplitude-frequency response. It is necessary to ensure that the control frequency is within the flat range of the actuator's amplitude-frequency characteristics so that the control signal has no significant amplitude attenuation or phase lag.

[0080] Based on the above constraints, in a specific example, the highest effective frequency of the micro-vibration of the support platform is... The sampling frame rate of the visual acquisition module is 100Hz. The inherent response bandwidth of the inner loop compensated actuator is 50fps. For 10kHz, take =10、 =100、 =0.5, calculated as follows =5kHz, simultaneously satisfying the above three constraints. The visual acquisition module and the inner loop compensation control module share the same hardware clock source. The trigger time of the visual frame is generated by the timer interrupt of this hardware clock source. The timing start of the inner loop control cycle is aligned with the trigger time of each visual frame, thereby ensuring that a single visual acquisition cycle contains k complete inner loop control cycles and avoiding timing mismatch caused by independent clock drift.

[0081] The single exposure time for visual acquisition is determined based on the sampling frame rate and the target subpixel localization accuracy. Each time a verification is triggered, multiple frames of images are continuously acquired and a displacement time series within the sampling window is formed. Subpixel localization includes at least one of gradient-based subpixel interpolation, phase-correlation-based subpixel registration, or moment-based subpixel localization.

[0082] The single exposure time must simultaneously satisfy both frame rate constraints and motion blur constraints. The frame rate constraint is that the single exposure time does not exceed a preset proportion of the sampling frame rate corresponding to the period. The motion blur constraint is that the amount of motion marked on the image within the single exposure time does not exceed the maximum blur displacement allowed for subpixel positioning. The maximum blur displacement is determined based on the target subpixel positioning accuracy. Specifically, the micro-vibration of the support platform is an oscillating motion rather than a unidirectional displacement, within the single exposure time... The maximum displacement of the inner mark is:

[0083]

[0084] In the formula, This represents the maximum displacement marked within a single exposure time. The amplitude of the vibration. The vibration frequency is used; when the vibration amplitude A is on the order of micrometers and the exposure time is on the order of milliseconds, the displacement of the marker motion within a single exposure is much smaller than the maximum blur displacement allowed by subpixel positioning, and visual measurement is effective.

[0085] Taking the sub-pixel localization method based on gray-scale moments as an example, its specific implementation process is as follows:

[0086] The region of interest containing the marked area is extracted from the acquired image as input; the binarized region of the marked area is extracted through adaptive threshold segmentation, and the zero-order gray-level moment of the image within this region is calculated. First-order grayscale moment and The zero-order grayscale moment The sum of the gray values ​​of all pixels within the region, the first-order gray-scale moment. The first gray-scale moment is the sum of the products of the gray values ​​of each pixel and their x-coordinates. The sum of the products of each pixel's grayscale value and its ordinate; the sub-pixel coordinates of the marker center ( , Calculated using the following formula:

[0087]

[0088]

[0089] This method uses the weighted characteristics of the gray-level distribution of the marked region to determine the center position. When the signal-to-noise ratio of the marked region image meets the input requirements of the sub-pixel localization algorithm, the theoretical localization error of the gray-level moment method is... for:

[0090]

[0091] in, Where is the pixel size, and SNR is the signal-to-noise ratio of the labeled region image. The effective number of pixels within the marked area; the lower limit of the effective number of pixels must ensure that the statistical estimate variance of the gray-scale moments of the marked area is lower than the allowable positioning error variance for subpixel positioning; in one example, when the target positioning accuracy is one-tenth of the pixel size, the signal-to-noise ratio is not less than 30dB and the effective number of pixels is not less than 100.

[0092] Gradient-based subpixel interpolation methods are suitable for markers with high edge contrast, determining edge positions by fitting the subpixel peaks of the edge gradient distribution; phase-correlation-based subpixel registration methods are suitable for scenarios with overall translation, obtaining subpixel-level displacement through phase inverse transformation of the cross power spectrum.

[0093] Extract the subpixel-level center positions of the fixed reference marker and the follow-up calibration marker from a single frame image, and calculate their relative displacement.

[0094] The relative displacement of each frame is continuously calculated within the sampling window to form a displacement time series. If the marker recognition of a certain frame fails during continuous acquisition, the data of that frame is discarded. The displacement time series is composed of the frames that are actually successfully extracted and their corresponding acquisition times. The timestamp of each frame retains the actual acquisition time.

[0095] S4. Calculate the steady-state zero-position deviation and residual vibration characteristic quantities based on the displacement time series; the steady-state zero-position deviation is the offset of the series mean relative to the calibration zero-position reference; the residual vibration characteristic quantities include at least one of the following: peak-to-peak value, standard deviation, and frequency band amplitude distribution.

[0096] The mean value of the displacement-time series within the sampling window is used to obtain the steady-state zero position.

[0097] Calculate the offset of the steady-state zero position relative to the calibrated zero position reference to obtain the steady-state zero position deviation;

[0098] Time-domain statistical calculations were performed on the displacement time series to obtain the peak-to-peak value and standard deviation, which were used as time-domain characteristic quantities of residual vibration.

[0099] The displacement time series is transformed in the frequency domain to extract the amplitude components of each frequency band, forming the residual vibration spectrum distribution.

[0100] The frequency domain transformation is implemented using Fast Fourier Transform. Before the transformation, a Hanning window is applied to the shift time series to suppress spectral leakage. After windowing, energy recovery correction is performed on the spectral amplitude. The correction coefficients are:

[0101]

[0102] in, The amplitude values ​​at various points of the rectangular window. The amplitude at each point of the Hanning window is summed by iterating through all sampling points of the window function; for a Hanning window of length N, this correction coefficient is approximately 2. Frequency band division adopts an equal-interval method, uniformly dividing the analysis frequency band into several sub-bands according to a preset frequency resolution.

[0103] The residual vibration spectrum distribution is compared with the calibration reference spectrum to identify target frequency bands with excessively high amplitudes as frequency domain subdivision features of the residual vibrations. The criterion for determining excessively high amplitudes is: the amplitude of the target frequency band exceeds a preset proportion of the amplitude of the corresponding frequency band in the calibration reference spectrum, and this excess state is continuous. The segment check persists. The frequency confirmation count is determined according to the hypothesis testing framework of independent testing. Let the probability that a single validation will misclassify statistical fluctuation as a true offset be . Then continuous The average probability of a misclassification is , The value of needs to be made The signal level is lower than the preset level; for example, in this embodiment... Set the value to 3; this preset ratio must be set higher than the statistical fluctuation range of the spectrum analysis. In a specific example, this ratio is set to 20%.

[0104] By combining time-domain features with frequency-domain subdivision features, complete residual vibration features are obtained.

[0105] S5. Based on the steady-state zero-position deviation and residual vibration characteristics, perform at least one layer of correction on the inner loop compensation parameters according to the parameter correction strategy. The parameter correction strategy includes four types of correction: zero-position bias, compensation gain, phase parameter and model parameter iteration. Each layer of correction sets an independent dead zone threshold. If the deviation is less than the corresponding threshold, the correction layer is not executed.

[0106] During the parameter correction process, the inner loop compensation continues to operate.

[0107] The zero-point offset correction is specifically as follows:

[0108] When the steady-state zero-point deviation exceeds the corresponding dead zone threshold, a reverse DC bias is superimposed on the output of the inner loop controller. The inner loop controller is the digital control algorithm module in the inner loop compensation system. Its input is the vibration acceleration digital signal after analog-to-digital conversion. It contains a feedforward control law and a feedback control law. The feedforward control law directly generates the feedforward compensation amount based on the acceleration signal, and the feedback control law generates the feedback compensation amount based on the integral amount of the acceleration signal. The two are superimposed at the output to generate the compensation control amount. The feedforward control law and the feedback control law together constitute the inner loop controller. The output of the inner loop controller drives the inner loop compensation actuator after digital-to-analog conversion and power amplification.

[0109] The feedforward control law is in the form of proportional feedforward, and the feedforward compensation is:

[0110]

[0111] In the formula, This is the feedforward compensation amount. This is the feedforward gain coefficient. The vibration acceleration value at the current sampling time is the result of analog-to-digital conversion; the feedback control law is in proportional and integral form, and the feedback compensation is:

[0112]

[0113] In the formula, To provide feedback on compensation amount, For proportional gain, For integral gain, To perform a discrete summation of the acceleration signal from the initial time to the current time k, This refers to the inner loop control cycle;

[0114] The total output of the inner loop controller is:

[0115]

[0116] In the formula, This is the total output of the inner loop controller.

[0117] The magnitude of the DC bias corresponds to the value of the steady-state zero-point deviation;

[0118] Specifically, DC bias The calculation method is as follows:

[0119]

[0120] In the formula For steady-state zero-position deviation, This is the zero-position correction proportional coefficient. Its lower limit should ensure that the single correction amount is not too small, resulting in slow convergence, while its upper limit should ensure that the correction amount is not too large, causing overshoot. Based on these constraints... The range of values ​​is 0.1 to 0.5 for example; the negative sign indicates that the bias direction is opposite to the deviation direction.

[0121] Upper limit of single correction step size Based on the rated stroke of the inner ring compensating actuator Confirm, take ,in The step size scaling factor is determined as follows:

[0122] Based on the minimum phase margin required by the inner loop compensation system, calculate the maximum allowable DC bias increment of the inner loop compensation actuator without causing closed-loop instability. Divide this maximum DC bias increment by the rated stroke of the inner loop compensation actuator to obtain the upper limit of β. The actual value shall not exceed this upper limit. In a specific example, β is taken as 0.05.

[0123] Based on the stability margin of the inner loop compensation system and the rated stroke of the inner loop compensation actuator, the upper limit of the single correction step size of the DC bias is set.

[0124] The specific compensation gain correction is as follows:

[0125] When the residual vibration amplitude of the target frequency band is higher than the calibration reference value and exceeds the corresponding dead zone threshold in multiple consecutive verifications, the compensation gain of the inner loop controller in that frequency band is increased.

[0126] When the residual vibration amplitude of the target frequency band is lower than the calibration reference value and exceeds the corresponding dead zone threshold in multiple consecutive verifications, the compensation gain of the inner loop controller in that frequency band is reduced.

[0127] The upper and lower limits of the compensation gain adjustment are determined based on the stability boundary of the inner loop compensation system; the upper limit of the single adjustment step size of the compensation gain is determined based on the gain margin of the inner loop compensation system.

[0128] Specifically, when continuous If the residual vibration amplitude of the target frequency band in the second verification is higher than the amplitude of the corresponding frequency band in the calibration reference spectrum and the deviation exceeds the corresponding dead zone threshold, the compensation gain of the frequency band is increased by one step; otherwise, it is decreased by one step.

[0129] The value of should be chosen so that the triggering condition has sufficient statistical confidence to distinguish between true parameter drift and random fluctuations. In a specific example... Take 3.

[0130] Single adjustment step size Set according to a fixed percentage of the current gain value, that is:

[0131]

[0132] in The gain adjustment proportional coefficient has a lower limit that ensures sufficient adjustment sensitivity to track slow parameter drift, and an upper limit that ensures a single adjustment does not cause the closed-loop gain to exceed the stability boundary. Based on these constraints, the value range of γ is, for example, 0.01 to 0.02. The upper and lower limits of the compensation gain adjustment are determined by the root locus stability boundary of the inner-loop compensation system to ensure that the gain adjustment always remains within the closed-loop stable region. For example, the upper limit of the adjustment... Take the nominal gain 1.5 times, adjust the lower limit Take the nominal gain 0.5 times.

[0133] The phase parameter correction is specifically as follows:

[0134] The displacement phase of the target frequency band is extracted from the displacement time series, and the phase of the inner loop feedforward channel excitation signal in the target frequency band is obtained synchronously. The phase difference between the two is calculated.

[0135] Adjust the delay parameters of the inner loop feedforward channel according to the value and direction of the phase difference;

[0136] Specifically, displacement phase The phase value is obtained by performing a fast Fourier transform on the displacement time series and extracting the phase value corresponding to the frequency component with the largest amplitude within the target frequency band.

[0137] feedforward channel excitation signal phase The excitation signal of the inner loop feedforward channel is obtained by synchronously acquiring it and then analyzing it in the same frequency domain.

[0138] The acquisition of both signals is triggered by the same clock source, ensuring strict time synchronization. Specifically, the excitation signal of the inner loop feedforward channel is synchronously stored in the data buffer by the calibration control module when generating the compensation control quantity. At the trigger time of each visual frame, the vision acquisition module simultaneously reads the feedforward excitation signal data for the corresponding time period in the buffer. The sampling time of both data streams is calibrated by the same hardware clock, eliminating time reference deviation across clock domains. Phase difference The calculation formula is:

[0139]

[0140] The adjustment amount Δτ of the feedforward channel delay parameter is calculated according to the following formula:

[0141]

[0142] in The center frequency of the target frequency band; The sign of the delay parameter determines the direction of its increase or decrease. A positive phase difference indicates that the feedforward channel phase lags behind the actual vibration, and the delay needs to be reduced to compensate in advance.

[0143] The upper limit of the single adjustment step size of the delay parameter is determined based on the phase margin of the inner loop compensation system;

[0144] The effective frequency band for phase correction does not exceed the Nyquist frequency corresponding to the visual sampling frame rate.

[0145] The iterative correction of model parameters is specifically as follows:

[0146] The displacement time series acquired visually is used as a true reference and compared with the output of the inner ring vibration prediction model to calculate the prediction error. Since the reference value contains visual measurement noise, the influence of measurement noise on the model update direction is suppressed by taking a moving average of the reference values ​​from multiple consecutive verifications. When the mean square value of the prediction error from multiple consecutive verifications exceeds a preset threshold, the parameters of the inner ring vibration prediction model are updated using an online parameter identification algorithm. The preset threshold is determined based on the model accuracy requirements of the inner ring compensation system.

[0147] Specifically, the inner ring vibration prediction model adopts a second-order autoregressive model. The model order is determined by the order test method in system identification. The test criterion is to stop increasing the order when the decrease in the prediction error variance after increasing the model order is less than a preset proportion.

[0148] Online parameter identification employs recursive least squares with a forgetting factor; the inner ring vibration prediction model adopts a second-order autoregressive form, and its mathematical expression is:

[0149]

[0150] In the formula The model parameters to be identified. These are the model's predicted output values ​​for the first two time points; parameter vector. The model parameter estimates at the current time, observation vector Composed of the model outputs from the first two time points; Prediction error:

[0151]

[0152] In the formula The displacement reference value is obtained from visual acquisition; the parameter recursive update formula is:

[0153]

[0154] In the formula The gain vector represents the weight of the current prediction error on the parameter update. The formula for calculating the gain vector is:

[0155]

[0156] The covariance matrix is ​​used to estimate the parameters, representing the degree of uncertainty in the parameter estimation. The formula for updating the covariance matrix is ​​as follows:

[0157]

[0158] In the formula The identity matrix; initial values ​​of the covariance matrix. Take a diagonal matrix, and set the diagonal elements as follows: Large numbers indicate significant uncertainty in the initial parameter estimation; forgetting factor To control the decay rate of historical data, the lower limit of the value must ensure the model's ability to track parameter abrupt changes, while the upper limit must ensure the statistical accuracy of parameter estimation under steady-state conditions; considering the above constraints, An example range of values ​​is: .

[0159] The displacement time series acquired visually is used as the truth reference. , and model prediction output Compare and calculate prediction error :

[0160]

[0161] When continuous The mean squared prediction error of the second verification exceeded the preset threshold. At that time, a parameter update is performed once, that is, when Triggered at time; The value of should ensure that the parameter update has sufficient statistical confidence. In a specific example... Set the threshold to 5; Based on the model accuracy requirements of the inner ring compensation system, the mean square value of the prediction error under the nominal working condition has a statistical estimation variance. The trigger threshold must be higher than this estimation variance to avoid unnecessary parameter updates due to random fluctuations. For example, it is taken as twice the mean square value of the prediction error under the nominal operating conditions.

[0162] Each layer of correction is performed in the following order: zero-position offset correction, compensation gain correction, phase parameter correction, and model parameter iterative correction. After each layer of correction is completed, a new set of displacement time series is collected and the steady-state zero-position deviation and residual vibration characteristics are recalculated for use as the basis for the next layer of correction.

[0163] The above execution order is set based on:

[0164] The operating frequency bands and mechanisms of each correction layer are different. Zero-bias correction only affects the DC component of the displacement time series and does not change the frequency domain amplitude distribution and phase relationship. Therefore, it has no coupling effect on subsequent gain correction and phase correction. Compensation gain correction only adjusts the amplitude of a specific frequency band and does not change the phase relationship and DC component. Therefore, it has no coupling effect on subsequent phase correction and model correction. Phase parameter correction only adjusts the timing parameters of the feedforward channel and does not change the gain of the feedback channel and the model structure. Therefore, it has no coupling effect on subsequent model correction. After each correction layer is completed, data is reacquired so that the next correction layer is judged based on the latest corrected state, avoiding the propagation of deviations introduced by previous corrections to subsequent layers.

[0165] The first visual verification after the system is powered on only performs zero-position offset correction. Compensation gain correction, phase parameter correction and model parameter iteration correction are skipped due to the lack of prior verification data for trend judgment. Normal execution will resume after a sufficient number of verification data are accumulated.

[0166] The upper limit of the single adjustment step size for each level of correction is determined independently based on the system stability margin requirements of the corresponding level;

[0167] The outer loop correction relies on the reliability of vision measurements. When the actuator is saturated, its output is already near its physical limits. Adding an outer loop correction may cause the actuator to enter the nonlinear region, leading to control instability. When vision acquisition conditions are abnormal, sub-pixel positioning accuracy cannot be guaranteed. In this case, performing correction may inject incorrect parameters into the inner loop controller, resulting in deterioration rather than improvement in compensation performance. Therefore, all outer loop correction actions should be frozen when any of the following conditions are met:

[0168] The output of the inner loop compensating actuator is in continuous The rated stroke was reached within each control cycle. of If the output value is more than twice the normal value, it is considered to be in a saturated output state. The value of needs to ensure that the saturation determination has sufficient temporal continuity to distinguish between transient peaks and sustained saturation. In a specific example... Take 10; This is the saturation determination proportional coefficient, whose value should ensure that the saturation determination is triggered when the actuator is close to its physical limit but has not yet entered the nonlinear region. In a specific example... Take 0.9;

[0169] If the sharpness index of the visually acquired image is lower than the preset sharpness threshold, it is judged as an image blurry; if the confidence score of the centroid extraction of the marker is lower than the preset confidence score threshold, it is judged as an abnormal marker occlusion. The preset sharpness threshold is a preset proportion of the statistical mean of the sharpness index under normal working conditions. This proportion must be lower than the lowest value of normal fluctuation to distinguish between normal fluctuation and abnormal blur. In a specific example, this proportion is 0.5. The preset confidence score is determined based on the gray-scale moment signal-to-noise ratio of the marker area and is taken as the statistical lower limit of the centroid confidence score under normal recognition conditions. In a specific example, this lower limit is 0.6 times the average confidence score under normal working conditions.

[0170] Once the freeze condition is lifted, the outer loop correction action automatically resumes execution. The criterion for lifting the freeze condition is: for freezes caused by actuator saturation, the freeze condition is lifted when the actuator output continuously drops back to the rated stroke. The control will be released after a preset number of cycles;

[0171] For freezes caused by visual acquisition anomalies, the image quality and marker recognition status will be reassessed when the next visual verification is triggered. If the verification requirements are met, the freeze will be lifted. Unexecuted corrections accumulated during the freeze period will not be compensated for. After recovery, the latest status after the freeze is lifted will be used as the basis for correction judgment.

[0172] The independent dead zone thresholds for each layer are determined based on the visual measurement noise level and the inherent drift rate of the inner loop parameters at the corresponding layer. When visual verification is triggered, if the image clarity or marker recognition requirements cannot be met during a continuous preset number of frame acquisitions, the current round of visual verification is terminated, the acquired data is discarded, and visual verification is re-executed once the next trigger condition is met. For verifications initiated due to abnormal triggering, if the acceleration amplitude at the start of the verification is still higher than the preset threshold, the current verification is skipped, and the verification is restarted only after the acceleration amplitude falls below the preset threshold.

[0173] The waiting process is set with a timeout exit mechanism. If the vibration amplitude does not fall below the preset threshold within the preset maximum waiting time, the verification process for this abnormal trigger will be terminated and the process will be restarted after the next trigger condition is met.

[0174] The maximum waiting time must be greater than the typical decay time of transient vibration caused by a single operating condition switch, and not exceed the periodic trigger interval, so as to ensure that subsequent verification scheduling will not be blocked due to infinite waiting before the next periodic verification arrives.

[0175] The initial system calibration is completed after the system is powered on and before the test platform enters the formal testing process. The calibration process is triggered before all visual verifications to ensure that the calibration zero-point reference and calibration reference spectrum used in subsequent steps are valid.

[0176] Before performing the above steps, the system needs to complete the initial calibration to establish a benchmark reference for each calculation step.

[0177] The initial calibration of the system is carried out under the condition that the bearing platform is in a static state and the inner loop compensation system is running under no-load to thermal steady state, at which time no external excitation is applied to the bearing platform.

[0178] The calibration process of the zero-position reference is as follows: the visual acquisition module continuously acquires a preset number of marked images, extracts the relative displacement between the fixed reference mark and the follow-up calibration mark frame by frame, and takes the arithmetic mean of the relative displacement of all frames as the calibration zero-position reference.

[0179] The preset number of frames should ensure that the statistical uncertainty of the calibration results is less than one-tenth of the subpixel positioning accuracy. In a specific example, this number of frames is set to 200.

[0180] The process of obtaining the calibration reference spectrum is as follows: under standard steady-state conditions, continuous sampling is performed for a duration of... The displacement time series is obtained, and a Hanning window is applied to the time series. Then, a fast Fourier transform is performed on the time series. The resulting spectrum is divided into segments according to a preset number of segments and averaged to obtain the calibration reference spectrum. The preset number of segments should be selected so that the variance of the spectrum estimate after segment averaging is lower than the preset proportion of the variance of the single segment spectrum. In a specific example, the number of segments is 8.

[0181] It should cover no less than ten complete cycles at the minimum analysis frequency, in a specific example. The time frame is 10 seconds. The initial reference for the dead zone threshold of each layer is determined through the following process: continuously acquire displacement time series under steady-state conditions and calculate their standard deviation, which is then recorded as the visual measurement noise amplitude. The initial value of the dead zone threshold for each layer is set to 3× This multiple relationship ensures that the dead zone threshold is significantly higher than the random fluctuation range of the measurement noise, avoiding noise-triggered meaningless correction actions.

[0182] Specifically, the method for obtaining the visual measurement noise level is as follows: continuously acquire displacement time series under steady-state system conditions, calculate its standard deviation, and use this standard deviation as the noise amplitude. .

[0183] The method for obtaining the intrinsic drift rate of the inner loop parameters is as follows:

[0184] Under offline conditions, the parameter offset of the inner loop compensation system after continuous operation for a preset observation period is recorded. The average drift per unit time is obtained by dividing the parameter offset by the observation period, and is denoted as the drift rate. Dead zone thresholds for each layer. Determine by the following formula:

[0185]

[0186] in, The minimum time interval between two adjacent visual verifications is used to maximize the dead zone threshold, ensuring that it covers both the random fluctuation range of measurement noise and the cumulative amount of parameter drift between the two verifications.

[0187] The dead zone threshold for the four-layer correction is set in ascending order according to the following hierarchical relationship:

[0188] The dead zone threshold of zero-offset correction is the smallest, followed by the dead zone threshold of compensation gain correction, then the dead zone threshold of phase parameter correction, and the dead zone threshold of model parameter iteration correction is the largest. Zero-offset correction directly adjusts the output DC component, and its target is the mean of the displacement time series. The mean is most sensitive to small constant offsets, so zero-offset correction has the smallest dead zone.

[0189] The compensation gain correction targets the amplitude of a specific frequency band and requires multiple consecutive verifications to confirm the trend before it can be triggered. It has a higher tolerance for random fluctuations than the zero-position correction. The phase parameter correction relies on frequency domain phase extraction. The statistical variance of the phase estimation is higher than that of the amplitude estimation, and a larger deviation is required to overcome the estimation uncertainty.

[0190] Iterative correction of model parameters involves an overall update of the structural parameters of the prediction model. It is triggered only when a systematic mismatch occurs between the model prediction and the actual response. Therefore, the dead zone is the largest. The hierarchical dead zone design avoids unnecessary triggering of high-level corrections caused by noise disturbances in low-level corrections.

[0191] Reference Figure 2 As shown, the machine vision-based wafer test platform micro-wobbling real-time compensation system includes:

[0192] An accelerometer is mounted on the support platform to collect vibration acceleration signals of the support platform.

[0193] The inner loop compensation actuator, connected to the acceleration sensor, is used to continuously apply active vibration compensation force to the bearing platform based on the vibration acceleration signal;

[0194] The reference marking assembly includes a fixed reference mark fixed on the vibration isolation base and a follow-up calibration mark fixed on the support platform;

[0195] The vision acquisition module is fixedly installed on the vibration isolation base. It is used to continuously acquire multiple frames of images at a frame rate lower than the inner loop compensation control frequency after the vision verification is triggered. Each frame contains both a fixed reference mark and a follow-up calibration mark.

[0196] The calibration control module is used to extract the relative displacement of two markers frame by frame from continuous images to obtain the displacement time series, calculate the steady-state zero-position deviation and residual vibration characteristics, and perform at least one layer of correction on the inner loop compensation parameters according to the parameter correction strategy. The parameter correction strategy includes four types of correction: zero-position bias, compensation gain, phase parameter and model parameter iteration, and each layer of correction sets an independent dead zone threshold.

[0197] During the parameter correction process performed by the calibration control module, the inner loop compensation actuator continues to operate continuously.

[0198] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A machine vision-based method for real-time compensation of micro-wobbles on a wafer testing platform, characterized in that, Includes the following steps: S1. Collect the vibration acceleration signal of the bearing platform and drive the inner ring compensation actuator to continuously apply active vibration compensation force to the bearing platform; S2. Trigger visual verification according to preset triggering conditions, wherein the triggering conditions are at least one of: working condition interval triggering, periodic triggering, and abnormal triggering; S3. After the visual verification is triggered, the visual acquisition module continuously acquires multiple frames of images at a frame rate lower than the inner loop compensation control frequency. Each frame simultaneously contains the fixed reference mark on the vibration isolation base and the follow-up calibration mark on the bearing platform. The relative displacement of the two marks is extracted frame by frame to obtain the displacement time series. S4. Calculate the steady-state zero-position deviation and residual vibration characteristic quantities based on the displacement time series; The steady-state zero-point deviation is the offset of the sequence mean relative to the calibrated zero point; the residual vibration characteristic includes at least one of the following: peak-to-peak value, standard deviation, and frequency band amplitude distribution. S5. Based on the steady-state zero-position deviation and residual vibration characteristics, perform at least one layer of correction on the inner loop compensation parameters according to the parameter correction strategy. The parameter correction strategy includes four types of correction: zero-position bias, compensation gain, phase parameter and model parameter iteration. Each layer of correction sets an independent dead zone threshold. If the deviation is less than the corresponding threshold, the correction layer is not performed. During the parameter correction process, the inner loop compensation continues to operate.

2. The real-time compensation method for micro-wobbling of a wafer testing platform based on machine vision according to claim 1, characterized in that, S2 includes: The specific triggering of the working condition gap is as follows: after the bearing platform completes the stepping motion, the probe is triggered within the steady-state waiting window before the test, and the verification process overlaps with the test preparation time. The periodic triggering specifically involves determining a fixed time interval for triggering based on the drift rate of the inner loop compensation parameters. The specific abnormal triggering is triggered when the accelerometer detects that the vibration amplitude exceeds a preset threshold, or when it detects a change in operating conditions. The change in operating conditions includes: before and after probe insertion and the start and stop of the platform. If the periodic triggering time conflicts with the working condition interval triggering time, it will be postponed to the next working condition interval. If multiple triggering conditions are met simultaneously, the triggers will be executed according to their priority, from abnormal triggering, intermittent triggering, to periodic triggering.

3. The real-time compensation method for micro-shake of a wafer testing platform based on machine vision according to claim 1, characterized in that, The method for determining the inner loop compensation control frequency is as follows: The lower limit is determined by using the highest effective frequency of the micro-vibration of the bearing platform as the basic lower limit, combined with the timing alignment constraint of an integer multiple of the visual sampling frame rate and the upper limit of the actuator response bandwidth, satisfying the following conditions: In the formula, For inner loop compensation control frequency; To control the multiplication factor of the frequency relative to the dominant vibration frequency; This is the highest effective frequency of the micro-vibration of the support platform; It is an integer multiple coefficient of the inner loop control frequency relative to the visual frame rate; This refers to the sampling frame rate of the visual acquisition module. This is a ratio factor between the control frequency and the actuator bandwidth; This is the inherent response bandwidth of the inner loop compensation actuator.

4. The real-time compensation method for micro-wobbling of a wafer testing platform based on machine vision according to claim 1, characterized in that, S3 includes: The single exposure time for visual acquisition is determined based on the sampling frame rate and the target subpixel positioning accuracy. Each time a verification is triggered, multiple frames of images are continuously acquired and a displacement time sequence within the sampling window is formed. The subpixel positioning includes at least one of gradient-based subpixel interpolation, phase-correlation-based subpixel registration, or moment-based subpixel positioning. Extract the subpixel-level center positions of the fixed reference marker and the follow-up calibration marker from a single frame image, and calculate their relative displacement. The relative displacement of each frame is continuously calculated within the sampling window to form a displacement time series.

5. The real-time compensation method for micro-shake of a wafer testing platform based on machine vision according to claim 1, characterized in that, S4 includes: The mean value of the displacement-time series within the sampling window is used to obtain the steady-state zero position. Calculate the offset of the steady-state zero position relative to the zero position reference during the initial calibration of the system to obtain the steady-state zero position deviation; Time-domain statistical calculations were performed on the displacement time series to obtain the peak-to-peak value and standard deviation, which were used as time-domain characteristic quantities of residual vibration. The displacement time series is subjected to frequency domain transformation to extract the amplitude components of each frequency band, forming the residual vibration spectrum distribution; The residual vibration spectrum distribution is compared with the calibration reference spectrum to identify the target frequency band with higher amplitude as the frequency domain subdivision feature of the residual vibration. By combining time-domain features with frequency-domain subdivision features, complete residual vibration features are obtained.

6. The real-time compensation method for micro-wobbling of a wafer testing platform based on machine vision according to claim 1, characterized in that, The zero-position offset correction specifically refers to: When the steady-state zero deviation exceeds the corresponding dead zone threshold, a reverse DC bias is added to the output of the inner loop controller. The magnitude of the DC bias corresponds to the value of the steady-state zero-point deviation; Based on the stability margin of the inner loop compensation system and the rated stroke of the inner loop compensation actuator, the upper limit of the single correction step size of the DC bias is set.

7. The real-time compensation method for micro-shake of a wafer testing platform based on machine vision according to claim 1, characterized in that, The compensation gain correction specifically refers to: When the residual vibration amplitude of the target frequency band is higher than the calibration reference value and exceeds the corresponding dead zone threshold in multiple consecutive verifications, the compensation gain of the inner loop controller in that frequency band is increased. When the residual vibration amplitude of the target frequency band is lower than the calibration reference value and exceeds the corresponding dead zone threshold in multiple consecutive verifications, the compensation gain of the inner loop controller in that frequency band is reduced. The upper and lower limits of the compensation gain adjustment are determined based on the stability boundary of the inner loop compensation system; the upper limit of the single adjustment step size of the compensation gain is determined based on the gain margin of the inner loop compensation system.

8. The real-time compensation method for micro-shake of a wafer testing platform based on machine vision according to claim 1, characterized in that, The phase parameter correction is specifically as follows: The displacement phase of the target frequency band is extracted from the displacement time series, and the phase of the inner loop feedforward channel excitation signal in the target frequency band is obtained synchronously. The phase difference between the two is calculated. Adjust the delay parameters of the inner loop feedforward channel according to the value and direction of the phase difference; The upper limit of the single adjustment step size of the delay parameter is determined based on the phase margin of the inner loop compensation system; The effective frequency band for phase correction does not exceed the Nyquist frequency corresponding to the visual sampling frame rate.

9. The real-time compensation method for micro-wobbling of a wafer testing platform based on machine vision according to claim 1, characterized in that, The iterative correction of the model parameters specifically involves: The displacement time series acquired visually is used as the true value reference and compared with the output of the inner ring vibration prediction model to calculate the prediction error. When the mean square value of the prediction error after multiple consecutive verifications exceeds a preset threshold, the parameters of the inner ring vibration prediction model are updated through an online parameter identification algorithm. The preset threshold is determined according to the model accuracy requirements of the inner ring compensation system. The corrections for each layer are performed sequentially in the order of zero-bias correction, compensation gain correction, phase parameter correction, and model parameter iterative correction. The upper limit of the single adjustment step size for each level of correction is determined independently based on the system stability margin requirements of the corresponding level; All outer loop correction actions are frozen when any of the following conditions are met: the inner loop compensation actuator is in saturated output state; or the visual acquisition detects image blurring or abnormal marker occlusion. The independent dead zone thresholds for each layer are determined based on the visual measurement noise level of the corresponding layer and the inherent drift rate of the inner loop parameters.

10. A machine vision-based real-time micro-wobbling compensation system for a wafer testing platform, characterized in that, include: An accelerometer sensor is mounted on the support platform to collect the vibration acceleration signal of the support platform. An inner-loop compensation actuator, connected to the acceleration sensor, is used to continuously apply an active vibration compensation force to the support platform based on the vibration acceleration signal. The reference mark assembly includes a fixed reference mark fixed on the vibration isolation base and a follow-up calibration mark fixed on the support platform; The visual acquisition module is fixedly installed on the vibration isolation base and is used to continuously acquire multiple frames of images at a frame rate lower than the inner loop compensation control frequency after visual verification is triggered. Each frame contains both the fixed reference mark and the follow-up calibration mark. The calibration control module is used to extract the relative displacement of two markers frame by frame from continuous images to obtain the displacement time series, calculate the steady-state zero-position deviation and residual vibration characteristics, and perform at least one layer of correction on the inner loop compensation parameters according to the parameter correction strategy. The parameter correction strategy includes four types of correction: zero-position bias, compensation gain, phase parameter and model parameter iteration, with each correction layer having an independent dead zone threshold. During the parameter correction process performed by the calibration control module, the inner loop compensation actuator continues to operate continuously.

Images