Anti-disturbance mobile manipulator vibration measurement pre-alignment visual servoing method

Abstract

The application discloses a kind of anti-disturbance mobile manipulator vibration measurement prenormal alignment visual servo method, belong to robot visual servo and mechanical equipment state monitoring field.For mobile manipulator executes contact type vibration measurement prenormal alignment task, chassis residual microvibration causes depth observation fluctuation, eye in hand configuration causes attitude-image coupling, fixed hand-eye calibration residual error causes ideal zero-error convergence cannot be realized coupling problem, and perception stabilization hybrid visual servo method (PS-HVS) is proposed.The method generates robust depth map and normal estimation through multi-frame space-time fusion enhancement and distance weighted principal component analysis;Through observation gate inhibition, single-frame abnormal observation is suppressed;Through uncertainty coupling gain adjustment, action intensity is adaptively controlled;Through two-threshold hysteresis switching, phased execution is realized;And control target is reconstructed as allowed domain convergence rather than absolute zero-error convergence.The application significantly improves the normal alignment stability under non-stationary observation conditions, and provides a reliable pre-contact attitude reference for unmanned contact vibration measurement.

Description

Technical Field

[0001] This invention belongs to the field of robot vision servoing and mechanical equipment condition monitoring technology, specifically relating to a visual servoing method for vibration-adjusting before vibration measurement of an anti-disturbance mobile manipulator. Background Technology

[0002] With the rapid development of industrial intelligence and unmanned equipment operation and maintenance technologies, industrial equipment inspection and condition monitoring based on mobile operators has become an important development direction in the field of intelligent manufacturing. Contact vibration measurement, as a core detection method for fault diagnosis of rotating machinery and power equipment, relies heavily on the alignment accuracy of the vibration sensor's sensitive axis with the normal of the measured target surface. If there is a deviation in the normal angle, the vibration signal will exhibit projection distortion, directly leading to amplitude measurement errors, and consequently affecting the identification and judgment of equipment faults. Therefore, high-precision normal alignment before contact is a crucial prerequisite for ensuring the reliability of unmanned vibration measurement.

[0003] Visual servoing is the mainstream technology for automatic pose adjustment at the end effector of robotic arms. Currently, it is mainly divided into three categories: image-based visual servoing (IBVS), pose-based visual servoing (PBVS), and hybrid visual servoing (HVS). It has been widely used in mid-to-long-distance workpiece grasping and pose adjustment scenarios in structured industrial environments. However, in close-range normal alignment scenarios where the moving manipulator performs contact vibration measurement, existing visual servoing methods face the coupling effects of multiple disturbances due to limitations in the working environment and system configuration, making it difficult to meet the requirements for stable and high-precision alignment.

[0004] Specifically, existing technologies have the following three core limitations: First, observation disturbances caused by residual vibrations of the mobile platform. After the mobile manipulator stops at the target workstation, the chassis drive system still exhibits idling micro-vibrations. These vibrations are significantly amplified in close-range depth sensing scenarios, causing time-varying fluctuations and random noise in the depth data acquired by the RGB-D camera. The surface normal estimation results based on single-frame depth data exhibit severe jitter, directly causing frequent jumps in servo control quantities, affecting alignment accuracy and easily triggering system oscillations. Second, motion coupling disturbances caused by the eye-on-hand configuration. To achieve close-range local observation, depth cameras often adopt an eye-on-hand mounting method, with a fixed offset between the camera's optical center and the rotation center of the robotic arm's end effector. Parasitic translations inevitably occur during attitude fine-tuning, causing the target point position on the image plane to shift. This results in the coupling of normal attitude correction and image target centering, causing mutual interference between control quantities and significantly reducing servo convergence speed and steady-state stability. Third, system disturbances caused by hand-eye calibration residuals. Due to limitations imposed by installation and processing errors, sensor noise, and calibration algorithm accuracy, the hand-eye transformation matrix always has a fixed residual that cannot be completely eliminated. If the ideal zero error is still used as the convergence target, the closed-loop control system will continuously fight against this inherent deviation, causing steady-state chatter of the end effector. This will not only fail to achieve true high-precision alignment but also exacerbate the wear of the robotic arm joints.

[0005] Current research primarily focuses on independent optimizations of individual problems, such as filtering and denoising depth images or improving the decoupled control of visual servoing. It fails to consider the coupling mechanism between observation disturbances, motion coupling, and system residuals in near-range normal alignment scenarios, and lacks a complete servo control scheme adapted to the non-stationary base conditions of the moving manipulator. In summary, existing visual servoing methods struggle to guarantee the stability and robustness of pre-normal alignment in unmanned contact vibration measurement, failing to meet the practical application requirements of unmanned operation and maintenance in industrial settings. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a visual servo method for vibration pre-measurement alignment of a disturbance-resistant mobile operator. Under non-ideal conditions where residual micro-vibration of the chassis, eye-hand configuration coupling, and fixed hand-eye calibration residuals coexist, this method achieves stable approximation and precise alignment of the sensitive axis of the vibration sensor with the normal of the target surface, thereby improving the reliability and engineering practicality of unmanned contact vibration measurement.

[0007] To achieve the above objectives, this invention adopts the following technical solution: Based on the Perceptually Stable Hybrid Visual Servoing (PS-HVS) architecture, a perturbation-resistant moving manipulator pre-vibration normal alignment visual servoing method is proposed, comprising the following steps: Step S1: Obtain close-range RGB-D observation data of the target surface, extract the target area, and construct a local depth point cloud.

[0008] Specifically, a depth camera with an eye-on-hand configuration acquires color images and near-field depth information of the target surface in real time. A target detection algorithm is used to extract the visual label region attached to the target surface. The center of the detection box is used as the target image observation point, and the depth data of the label neighborhood is extracted to construct a local depth point cloud.

[0009] Step S2: Spatiotemporal fusion enhancement processing is performed on multiple consecutive frames of depth data to generate a local robust depth map; based on the robust depth map, distance-weighted principal component analysis combined with robust outlier removal is used to estimate the target surface normal and calculate the observation quality index.

[0010] Further, the specific steps of the spatiotemporal fusion enhancement process in step S2 are as follows: Set a sliding time window of length L, extract the core receptive field centered on the observation point of the target image as the depth processing range; count the frequency of effective depth values ​​of each pixel in the core receptive field within the time window; retain pixels with frequencies higher than the quantile threshold as a stable pixel set, and remove noise pixels and flying points with poor temporal consistency; perform temporal median fusion on the depth values ​​in the stable pixel set to obtain the fused depth value of each pixel and generate a local robust depth map.

[0011] Furthermore, the specific steps of distance-weighted principal component analysis in step S2 are as follows: Project the robust depth map into a 3D point cloud set, and assign each point a weight inversely proportional to the square of the depth: In the formula, z i Let be the depth value of the i-th point, and ε be a small constant to prevent numerical singularities.

[0012] Calculate the weighted centroid and weighted covariance matrix. Perform eigenvalue decomposition on the covariance matrix and take the eigenvector corresponding to the smallest eigenvalue as the initial normal estimate. Calculate the perpendicular distance from each point to the initial fitted plane. Use the median absolute difference (MAD) to identify and remove outliers, retaining inliers with a distance less than κ times the MAD. Recalculate the weighted covariance matrix and normal vector using the retained inliers to obtain the final target surface normal estimate. Calculate the standard deviation of the inlier depth samples as the observation quality index σ. z The larger the value, the stronger the observation dispersion and the worse the stability of the normal estimation.

[0013] Step S3: Construct a dynamic safety boundary based on historical observation statistics, and perform gating screening on the current normal and image plane observations to suppress abnormal observations in a single frame.

[0014] Further, the specific steps of observation gating in step S3 are as follows: Maintain the set of normals and the set of image plane observation errors of the most recent H frames that have been observed; calculate the historical mean of normals, the historical mean of image errors and the corresponding historical dispersion; construct a dynamic safety boundary based on the basic tolerance boundary and the historical dispersion; calculate the deviation between the current observation and the historical mean; if the deviation exceeds the dynamic safety boundary, use the historical mean to replace the current observation; otherwise, accept the current observation and update the historical queue.

[0015] Step S4: Based on the attitude error, image plane observation error and observation quality index, the local rotation increment and lateral translation increment of the robotic arm end effector are generated by uncertainty coupling gain adjustment.

[0016] Furthermore, the specific formula for adjusting the uncertainty coupling gain in step S4 is as follows: Attitude adjustment gain: Image plane constraint gain: In the formula, These represent the maximum gain of the pose channel and the image plane channel, respectively. For error-driven intensity coefficient, To observe the uncertainty penalty intensity coefficient, e θ For attitude error, e μv For the image plane observation error, σ z For observation quality indicators.

[0017] The end-effector rotation increment is generated along the cross product of the sensor's sensitive axis and the target's normal, while the lateral translation increment is generated by a fixed linear mapping of the image plane observation error; all control increments are projected onto the motion constraints of the robotic arm.

[0018] Step S5 employs a dual-threshold hysteresis switching mechanism to manage the three execution stages of centripetal coarse adjustment, attitude optimization, and final approach, outputting an end-effector incremental control command that satisfies mechanical constraints, driving the robotic arm end-effector to converge to the normal alignment allowable domain.

[0019] Furthermore, the specific rules for the dual-threshold hysteresis switching in step S5 are as follows: Set the threshold for entering the translation-locked image plane. Image plane thresholding to unlock translation lock And satisfy When the image plane observation error e μ When switching from centripetal coarse adjustment to attitude optimization, lateral translation is locked, retaining only attitude adjustment; when the image plane observation error eμ When the attitude optimization state is switched back to the centripetal coarse adjustment state, the lateral translation is unlocked; when the image plane observation error is between the two, the current state remains unchanged.

[0020] Furthermore, step S5 also includes the determination and execution of the final approach phase: Set the number of continuous steady-state verification frames N and the attitude steady-state threshold. When the attitude error satisfies N consecutive frames When the attitude alignment is deemed satisfactory, the system switches to the final approach state. In the final approach state, lateral translation and attitude adjustment are paused, and only a small axial approach increment along the sensor's sensitive axis is output until contact is completed.

[0021] Furthermore, the control objective of the method is to drive the end effector of the robotic arm to simultaneously converge the posture error, image plane observation error, and observation quality index to a preset allowable range, rather than pursuing absolute zero error convergence. The allowed domain is defined as follows: in, These represent the maximum allowable steady-state attitude error, the maximum image plane offset, and the maximum observation dispersion, respectively.

[0022] Furthermore, the method also includes an abnormal reset mechanism: the number of consecutive abnormal observations is counted in real time, and when the number of consecutive abnormal observations reaches a preset reset threshold, the historical observation queue is cleared, the reference benchmark is reset and the system returns to the centripetal coarse adjustment state, and the alignment process is re-executed.

[0023] The present invention also provides a vibration-aligned visual servo system for an anti-disturbance moving manipulator before vibration measurement, comprising: The depth sensing module is used to acquire near-range RGB-D observation data of the target surface, extract the target area and construct a local depth point cloud, perform spatiotemporal fusion enhancement and distance-weighted principal component analysis on continuous multi-frame depth data, and output the target surface normal and observation quality index. The observation gating module, connected to the depth perception module, is used to construct a dynamic safety boundary based on historical observation statistics, and to perform gating and filtering on the current normal and image plane observations to suppress abnormal observations in a single frame. The hybrid servo control module, connected to the observation gating module, is used to generate the local rotation increment and lateral translation increment of the robotic arm end effector based on attitude error, image plane observation error and observation quality index, using uncertainty coupling gain adjustment. The hysteresis switching execution module, connected to the hybrid servo control module, is used to manage the three execution stages of centripetal coarse adjustment, attitude optimization and final approach using a dual-threshold hysteresis switching mechanism. It outputs end-effector incremental control commands that meet mechanical constraints, driving the robotic arm end-effector to converge to the normal alignment allowable domain.

[0024] The beneficial effects of this invention are: This invention effectively suppresses observation disturbances and improves the stability of normal estimation. Through a sensing scheme combining multi-frame spatiotemporal fusion enhancement and range-weighted principal component analysis, it performs temporal smoothing and robust estimation of depth fluctuations caused by residual micro-vibrations of the chassis. This significantly reduces noise and jitter in near-range depth observations, outputting stable and reliable target surface normals and observation quality indicators, thus mitigating the impact of vibration disturbances on servo control from the source of perception.

[0025] Decoupling attitude and translation control mitigates motion coupling disturbances. This invention employs a hybrid visual servoing framework, separating the centripetal coarse adjustment and attitude optimization stages through a dual-threshold hysteresis switching mechanism. Combined with uncertainty coupling gain adaptive adjustment of control strength, it effectively reduces the cross-axis coupling between attitude correction and image centering in the eye-on-hand configuration, improving servo convergence speed and process stability.

[0026] This invention employs an allowable domain convergence mechanism to adapt to inherent system residuals. It abandons the traditional ideal zero-error convergence objective and reconstructs a multi-index allowable domain convergence mode, which can adaptively handle inherent system deviations such as hand-eye calibration residuals. This avoids steady-state chatter caused by the closed-loop system continuously resisting unavoidable fixed errors, significantly improving the robustness and long-term operational reliability of the alignment process.

[0027] The solution is comprehensive and highly practical for engineering applications. This invention constructs a complete closed-loop control system from sensing enhancement, anomaly suppression, servo control to staged execution. It can be adapted to existing mobile manipulator platforms without additional hardware modifications, and can provide a stable and reliable pre-contact attitude reference for unmanned contact vibration measurement in industrial sites. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of a visual servoing system and scene for a vibration-adjustable mobile manipulator before vibration measurement according to the present invention; Figure 2 This is a general framework diagram of the PS-HVS method of the present invention; Figure 3 This is a comparison chart of the zero-mean depth fluctuation of the target plane center point in this invention; Figure 4 This is a comparison chart of the normal estimation performance of the present invention, where (a) is a comparison of absolute angle error and (b) is a comparison of dynamic angle jitter amplitude. Figure 5The following is a comparative diagram of the observation of gated ablation experiments in this invention, where (a) is the control response without gate and (b) is the control response with gate. Figure 6 The above is a comparison of the overall convergence results of the present invention, where (a) is the normal residual distribution and (b) is the convergence process curve. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of the present invention.

[0030] Example 1 This embodiment provides a visual servoing method for vibration measurement normal alignment of an anti-disturbance moving manipulator. Based on the PS-HVS architecture, it implements complete anti-disturbance normal alignment control. The overall system configuration and method framework are as follows: Figure 1 , Figure 2 As shown, the specific implementation steps are as follows: S1, System Initialization and Data Acquisition. A mobile manipulator hardware platform is constructed, including an inspection mobile platform, a six-DOF collaborative robotic arm, an RGB-D camera mounted on the hand, and a contact vibration sensor. After coarsely locating the target equipment area, the mobile platform remains stationary at idle speed, with only local visual servo adjustments performed by the robotic arm. The RGB-D camera acquires real-time color images and near-field depth data of the target surface. A target detection algorithm extracts the visual label region attached to the target surface, using the center of the detection frame as the target image observation point, and extracts depth data from the label's neighborhood to construct a local depth point cloud.

[0031] S2, Multi-frame Depth Robustness and Normal Estimation. Spatiotemporal fusion enhancement processing is performed on continuous multi-frame depth data to generate a locally robust depth map. Based on the robust depth map, distance-weighted principal component analysis combined with robust outlier removal is used to estimate the target surface normal vector, and the standard deviation of the interior depth samples is calculated as the observation quality index σ. z The detailed implementation process of this step can be found in Example 2.

[0032] S3, Observation Gating Processing. Maintain the set of normals and the set of image plane observation errors for the most recent H frames (H=3 in this embodiment); calculate the historical mean normal, the historical mean image error, and the corresponding historical dispersion; construct a dynamic safety boundary based on the preset basic tolerance boundary and the historical dispersion. Calculate the deviation between the current observation and the historical mean. If the deviation exceeds the dynamic safety boundary, the current observation is determined to be abnormal, and the historical mean is used to replace the current observation input to the control loop; otherwise, the current observation is accepted and the historical queue is updated. This mechanism can effectively suppress the impact of abnormal observations from single-frame jumps on servo control, and its suppression effect is shown in the attached figure. Figure 5 As shown.

[0033] S4, Uncertainty-coupled visual servo control. Based on the current attitude error eθ, image plane observation error eμ, and observation quality index σz, uncertainty-coupled gain adjustment is used to generate the local rotation increment and lateral translation increment of the robotic arm end effector. The attitude adjustment gain and image plane constraint gain are calculated using the following formulas: In the formula, These represent the maximum gain of the pose channel and the image plane channel, respectively. For error-driven intensity coefficient, To observe the uncertainty penalty strength coefficient, e θ For attitude error, e μv For the image plane observation error, σ z This is an indicator of observation quality; the more stable the observation (σ... z The smaller the value, the greater the gain; the greater the error, the smaller the gain, thus achieving adaptive matching between the intensity of the control action and the observation quality and error level.

[0034] The end-effector rotation increment is generated along the cross product of the sensor's sensitive axis and the target's normal, while the lateral translation increment is generated by a fixed linear mapping from the image plane observation error. All control increments are projected onto the joint velocity and displacement constraints of the robotic arm to ensure execution safety.

[0035] S5, Hysteresis State Switching and Allowable Domain Convergence Control. A dual-threshold hysteresis switching mechanism is used to manage the three execution phases. The system initially enters a centripetal coarse adjustment state, while simultaneously performing lateral translation and attitude adjustment.

[0036] Set the threshold for entering the translation-locked image plane. =50px, image plane threshold to unlock translation lock. =100px, satisfies < When the image plane observation error e μ At this time, the system switches from the centripetal coarse adjustment state to the attitude optimization state, locking the lateral translation degree of freedom and retaining only the normal attitude adjustment; when the image plane observation error e μ When the system switches from attitude optimization state back to centripetal coarse adjustment state, it unlocks lateral translation again; when the error is between the two thresholds, it keeps the current state unchanged to avoid chatter caused by frequent switching at the boundary.

[0037] Set the number of continuous steady-state verification frames N=5 and the attitude steady-state threshold. =0.5°. When the attitude error is less than or equal to 0.5° for 5 consecutive frames, the attitude alignment is determined to be up to standard, and the system switches to the final approach state. In this state, lateral translation and attitude adjustment are paused, and only a small axial approach increment along the sensor's sensitive axis is output. In this embodiment, it is 1 mm / step, until the vibration sensor makes contact with the target surface.

[0038] The control objective of this method is to drive the robotic arm's end effector to converge within the allowable normal alignment region, rather than pursuing absolute zero error. The allowable region is defined as: in, These represent the maximum allowable steady-state attitude error, the maximum image plane offset, and the maximum observation dispersion, respectively.

[0039] In addition, this embodiment sets up an abnormal reset mechanism: the number of consecutive abnormal observations is counted in real time. When the number of consecutive abnormal observations reaches the preset reset threshold, the historical observation queue is cleared, the reference benchmark is reset and returned to the centripetal coarse adjustment state, and the alignment process is restarted to avoid the system from becoming unstable due to continuous abnormal observations.

[0040] The complete normal alignment workflow is as follows: First, after the system starts, the mobile platform automatically moves to the vicinity of the target device to complete coarse positioning; then, the robotic arm drives the RGB-D camera into a close-range hovering state; next, the system automatically executes the processes of depth acquisition, multi-frame fusion, normal estimation, observation gating, servo control, and state switching in sequence, without any manual intervention; finally, when the system determines that the alignment has met the standard, it automatically drives the end effector to approach the target surface along the normal direction to complete the pre-contact attitude preparation.

[0041] Experimental verification shows that, under the conditions of residual micro-vibration of the chassis, eye-on-hand configuration coupling, and fixed hand-eye calibration residuals, the method of this embodiment achieves a median servo step count of 11 steps for 30 sets of random initial poses, and the final end-effector normal residuals are consistently maintained within the allowable range of 0.5°, meeting the accuracy requirements of contact vibration measurement. The overall convergence results are shown in the attached figure. Figure 6 As shown.

[0042] Example 2 This embodiment is a specific implementation of step S2 in embodiment 1, and it forms the perception foundation of the entire visual servoing system. Its depth smoothing and normal estimation performance are shown in the attached figure. Figure 3 Appendix Figure 4 As shown, the specific steps are as follows: S2.1, Initialize the sliding time window. Set a sliding time window of length L, which is L=5 in this embodiment; extract the core receptive field centered on the target image observation point as the depth processing range, which is a 100×100 pixel area in this embodiment.

[0043] S2.2, Temporal Consistency Filtering. Count the frequency of valid depth values ​​for each pixel within the core receptive field in the sliding time window; set a quantile threshold of 0.8, retain pixels with a frequency higher than this threshold as a stable pixel set, and remove noisy pixels with poor temporal consistency and edge flying spots.

[0044] S2.3, Temporal Median Fusion. Temporal median fusion is performed on the depth values ​​within the stable pixel set to obtain the fused depth value for each pixel, generating a local robust depth map.

[0045] S2.4, Distance-Weighted Principal Component Analysis. The robust depth map is projected as a 3D point cloud set, and each point is assigned a weight inversely proportional to the square of the depth. In the formula, z i Let be the depth value of the i-th point, and ε be a small constant to prevent numerical singularities. Let 10 be the depth value. -6 To prevent numerical oddities.

[0046] Calculate the weighted centroid and weighted covariance matrix, perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the smallest eigenvalue as the initial normal estimate.

[0047] S2.5 Robust Outlier Removal. Calculate the vertical distance of each point to the initial fitting plane and use the absolute median difference (AMD) to identify outliers: calculate the median distance to the sample and the AMD, retain inliers with a distance less than κ times the MAD (in this embodiment, κ=3), and remove outliers caused by local reflections or edge-blended pixels.

[0048] S2.6, Secondary Normal Estimation and Quality Assessment. The weighted covariance matrix and normal vector are recalculated using the retained interior points to obtain the final target surface normal estimate; the standard deviation of the interior point depth samples is calculated as the observation quality index σ. z The larger the value, the stronger the observation dispersion and the worse the stability of the normal estimation.

[0049] Experimental verification shows that the method in this embodiment can reduce the standard deviation of the depth measurement of the center point of the target plane from 0.413 mm in the original single frame to 0.062 mm, reducing the dispersion by about 85%; the absolute mean error of the normal estimation is reduced from 0.94° to 0.23°, and the standard deviation of dynamic angle jitter is reduced by about 50.4%, which significantly improves the stability of near-range normal estimation under the condition of residual micro-vibration of the chassis.

[0050] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator, characterized in that, Includes the following steps: S1, acquire close-range RGB-D observation data of the target surface, extract the target area and construct a local depth point cloud; S2, perform spatiotemporal fusion enhancement processing on continuous multi-frame depth data to generate a local robust depth map; based on the robust depth map, use distance-weighted principal component analysis combined with robust outlier removal to estimate the target surface normal and calculate the observation quality index; S3 constructs a dynamic safety boundary based on historical observation statistics, and performs gating and filtering on current normal and image plane observations to suppress abnormal observations in a single frame. S4. Based on the attitude error, image plane observation error and observation quality index, uncertainty coupling gain adjustment is used to generate the local rotation increment and lateral translation increment of the end effector of the robotic arm. S5 employs a dual-threshold hysteresis switching mechanism to manage the three execution stages of centripetal coarse adjustment, attitude optimization, and final approach, outputting end-effector incremental control commands that satisfy mechanical constraints to drive the robotic arm end-effector to converge to the normal alignment allowable domain.

2. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The specific steps of the spatiotemporal fusion enhancement process in step S2 are as follows: set a sliding time window of length L, count the frequency of each pixel in the core perceptual domain having an effective depth value within the time window; retain pixels with a frequency higher than the quantile threshold as a stable pixel set; Temporal median fusion is performed on the depth values ​​within the stable pixel set to obtain a locally robust depth map.

3. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The specific steps of distance-weighted principal component analysis in step S2 are as follows: project the robust depth map into a three-dimensional point cloud set, and assign each point a weight that is inversely proportional to the square of the depth. Calculate the weighted centroid and weighted covariance matrix, perform eigenvalue decomposition on the covariance matrix, and take the eigenvector corresponding to the smallest eigenvalue as the initial normal estimate; Outliers in the plane fitting are removed based on the absolute median, and the weighted covariance matrix and normal are recalculated using the retained inliers to obtain the final target surface normal estimate. The standard deviation of the interior point depth samples is calculated as an indicator of observation quality. The larger the value, the stronger the observation dispersion and the worse the stability of the normal estimation.

4. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The specific steps of observation gating in step S3 are as follows: maintain the set of normals and the set of image plane observation errors of the most recent H frames that have been observed; calculate the historical mean of normals, the historical mean of image errors and the corresponding historical dispersion; construct a dynamic safety boundary based on the basic tolerance boundary and the historical dispersion; calculate the deviation between the current observation and the historical mean; if the deviation exceeds the dynamic safety boundary, use the historical mean to replace the current observation; otherwise, accept the current observation and update the historical queue.

5. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The specific formula for adjusting the uncertainty coupling gain in step S4 is as follows: Attitude adjustment gain: Image plane constraint gain: in, For maximum gain, For error-driven intensity coefficient, To observe the intensity coefficient of the uncertainty penalty; The end-rotation increment is generated along the cross product of the sensor's sensitive axis and the target's normal, while the lateral translation increment is generated by the image plane observation error through a fixed linear mapping.

6. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The specific rules for the dual-threshold hysteresis switching in step S5 are as follows: Set an image plane threshold for entering translation lock and an image plane threshold for releasing translation lock, with the image plane threshold for entering translation lock being less than the image plane threshold for releasing translation lock; when the image plane observation error is less than or equal to the image plane threshold for entering translation lock, switch from the centripetal coarse adjustment state to the attitude optimization state, locking the lateral translation and retaining only attitude adjustment; when the image plane observation error is greater than or equal to the image plane threshold for releasing translation lock, switch back from the attitude optimization state to the centripetal coarse adjustment state, unlocking the lateral translation; when the image plane observation error is between the two, maintain the current state.

7. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 6, characterized in that, Step S5 further includes: setting the number of continuous steady-state verification frames and the attitude steady-state threshold. When the number of frames in which the attitude error continuously meets the steady-state verification frame count is less than or equal to the attitude steady-state threshold, the system switches to the final approach state. In the final approach state, the system pauses lateral translation and attitude adjustment, and only outputs a small axial approach increment along the sensor's sensitive axis.

8. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The control objective of the method is to drive the end effector of the robotic arm to simultaneously converge the attitude error, image plane observation error and observation quality index to a preset allowable range, rather than pursuing absolute zero error convergence. The allowed domain is defined as follows: in, These represent the maximum allowable steady-state attitude error, the maximum image plane offset, and the maximum observation dispersion, respectively.

9. The visual servoing method for vibration measurement pre-normal alignment of an anti-disturbance mobile manipulator according to claim 1, characterized in that, The method also includes an abnormal reset mechanism: when the number of consecutive abnormal observations reaches a preset reset threshold, the historical observation queue is cleared, the reference benchmark is reset, and the system returns to the centripetal coarse adjustment state.

10. A vibration-damping pre-vibration alignment visual servo system for an anti-disturbance moving manipulator, characterized in that, include: The depth sensing module is used to acquire near-range RGB-D observation data of the target surface, extract the target area and construct a local depth point cloud, perform spatiotemporal fusion enhancement and distance-weighted principal component analysis on continuous multi-frame depth data, and output the target surface normal and observation quality index. The observation gating module, connected to the depth perception module, is used to construct a dynamic safety boundary based on historical observation statistics, and to perform gating and filtering on the current normal and image plane observations to suppress abnormal observations in a single frame. The hybrid servo control module, connected to the observation gating module, is used to generate the local rotation increment and lateral translation increment of the robotic arm end effector based on attitude error, image plane observation error and observation quality index, using uncertainty coupling gain adjustment. The hysteresis switching execution module, connected to the hybrid servo control module, is used to manage the three execution stages of centripetal coarse adjustment, attitude optimization and final approach using a dual-threshold hysteresis switching mechanism. It outputs end-effector incremental control commands that meet mechanical constraints, driving the robotic arm end-effector to converge to the normal alignment allowable domain.

Images