A real-time compensation method for mounting error based on dynamic servo vibration suppression and depth vision closed loop

Abstract

A real-time mounting error compensation method based on dynamic servo vibration suppression and depth vision closed loop is proposed. This method acquires the three-dimensional spatial coordinate data of reference points on the PCB board and the components to be mounted using a multi-view depth camera array. A depth point cloud model including surface normals and relative poses is constructed. The depth point cloud model is compared with preset standard position data to calculate the positional deviation of the mounting head in the X, Y, and Z axes, as well as the rotational deviation angle around these three axes. The dynamic servo vibration suppression module includes a high-frequency response piezoelectric ceramic actuator and a low-inertia mechanical linkage structure. This invention achieves high-density acquisition of spatial coordinate data and optimized depth measurement by setting the baseline distance between the left and right cameras in the multi-view depth camera array to within the range of 150mm to 200mm, and maintaining an angle of 30° to 45° between the optical axis and the PCB board plane. This allows for the capture of the three-dimensional features of minute components, and the optimized viewing angle avoids reflection interference and occlusion problems during vertical shooting.

Description

Technical Field

[0001] This invention belongs to the field of mounting compensation technology, specifically a real-time mounting error compensation method based on dynamic servo vibration suppression and depth vision closed loop. Background Technology

[0002] Surface mount technology (SMT) is a process used in microelectronic packaging and PCB component assembly. High-speed, precision mounting is the primary process for mounting high-end automotive chips, miniature resistors and capacitors, and precision semiconductor leads. Mounting accuracy determines the yield and stability of electronic product packaging. However, current industry-standard mounting correction solutions suffer from the following technical problems: First, existing mount visual positioning imaging systems have a single imaging configuration, mostly adopting a vertical orthogonal monocular vision acquisition architecture without standardized baseline angle adaptation parameters. On the one hand, the imaging depth is insufficient, the three-dimensional point cloud acquisition density of micro-components is low, and it is easily affected by PCB board pad reflections and component corner occlusion interference, resulting in many noise points and missing effective features in the original imaging point cloud. On the other hand, conventional point cloud matching algorithms have coarse convergence thresholds and simple iterative constraints, resulting in low accuracy in calculating the six-dimensional deviations of three-axis position and three-axis attitude. They can only identify planar two-dimensional mount errors and cannot accurately quantify the spatial rotation deviations of micro-components. The error vector acquisition itself has inherent accuracy errors, and the subsequent compensation and traceability benchmarks are inaccurate.

[0003] Secondly, existing servo correction and vibration suppression structures mostly rely on passive damping of the motor spindle for vibration reduction, lacking a separate high-frequency active vibration suppression linkage structure. They can only weaken low-frequency jitter and cannot separate and filter out the high- and low-frequency coupled vibration components during the placement process. The structural elastic deformation and high-frequency vibration at the end caused by the motor starting and stopping cannot be offset. At the same time, existing placement compensation systems lack a closed-loop linkage thermal distortion correction mechanism, and the six-dimensional fine-tuning mechanism mostly adopts an open-loop split module design. Temperature gradient deformation error and accumulated error of the open-loop module during operation cannot be traced and corrected in real time. Under high-speed continuous placement conditions, systematic offset errors continue to accumulate, resulting in poor long-term placement accuracy stability. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the present invention provides a real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop, so as to at least partially solve the above-mentioned technical problems.

[0005] The technical solution adopted in this invention is as follows: This invention proposes a real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop, comprising the following steps: S1: Obtain the three-dimensional spatial coordinate data of the reference points on the PCB board and the components to be mounted by a multi-view depth camera array, and construct a depth point cloud model including surface normal vectors and relative poses. S2: Compare the depth point cloud model with the preset standard position data, calculate the position deviation of the mounting head in the X, Y, and Z axis directions and the rotation deviation angle around the three axes, and generate an initial error vector; S3: Drive the dynamic servo vibration damping module according to the initial error vector. The dynamic servo vibration damping module includes a high-frequency response piezoelectric ceramic actuator and a low-inertia mechanical linkage structure. It uses the inverse piezoelectric effect of the piezoelectric ceramic actuator to generate a reverse micro-displacement force to counteract the elastic deformation and high-frequency vibration components during the movement of the robotic arm. S4: While the dynamic servo vibration damping module is in operation, the high-frequency acceleration signal of the end effector of the mounting head is collected in real time. The low-frequency jitter caused by the motor start-stop and the high-frequency vibration caused by mechanical resonance are separated by an adaptive filtering algorithm, and the separated vibration spectrum data is fed back to the control loop. S5: The initial error vector generated in step S2 is fused with the real-time vibration spectrum data fed back in step S4, and input into the Kalman filter for state estimation, and the corrected final error compensation vector is output. S6: Based on the final error compensation vector, send pulse sequence commands to the six-degree-of-freedom motion platform of the placement head to drive the servo motors of each axis to perform fine-tuning motion, so that the residual error of the end of the placement head relative to the target position of the PCB board is controlled within the sub-pixel range. S7: At the moment the placement action is completed, the depth camera array is triggered again to perform a second imaging of the placed component, extract the deviation value between the component center coordinates and the preset coordinates. If the deviation value exceeds the set threshold, an anomaly is marked and the re-placement process is triggered.

[0006] In one embodiment of the present invention, the multi-view depth camera array in step S1 adopts a binocular stereo vision architecture, with a baseline distance of 150mm-200mm between the left and right cameras, a camera resolution of not less than 1920×1080, a frame rate of not less than 60fps, and an angle of 30°-45° between the camera optical axis and the PCB board plane to optimize depth measurement accuracy.

[0007] In one embodiment of the present invention, the high-frequency response piezoelectric ceramic actuator in step S3 is made of PZT-5H material, with a maximum stroke range of ±10μm and a resonant frequency greater than 20kHz. The piezoelectric ceramic actuator is connected to the mechanical linkage through a flexible hinge mechanism to form a parallel adjustable active vibration damping unit.

[0008] In one embodiment of the present invention, the adaptive filtering algorithm in step S4 uses wavelet packet decomposition technology to decompose the acceleration signal into 8 frequency bands, wherein the high-frequency jitter component is located in the 6th-8th frequency band and the low-frequency jitter component is located in the 1st-3rd frequency band, and noise interference is automatically identified and eliminated by the energy threshold method.

[0009] In one embodiment of the present invention, the Kalman filter in step S5 adopts an extended Kalman filter architecture, and its state vector includes position error, velocity error, acceleration error and vibration phase information. The process noise covariance matrix Q and the observation noise covariance matrix R are dynamically adjusted according to the real-time signal-to-noise ratio.

[0010] In one embodiment of the present invention, the six-degree-of-freedom motion platform in step S6 includes three linear modules and three rotary modules. Each module is driven by a frameless torque motor, and the encoder resolution is not less than 20 bits. The modules are connected by an aluminum alloy frame to form a closed motion chain structure.

[0011] In one embodiment of the present invention, the secondary imaging process in step S7 adopts a global shutter mode for shooting, and the exposure time is set to 100μs-200μs to eliminate motion blur caused by high-speed movement. The image processing algorithm adopts sub-pixel edge detection technology, and the positioning accuracy reaches the 0.1 pixel level.

[0012] In one embodiment of the present invention, a real-time temperature compensation step is also included: during the mounting process, the ambient temperature change is monitored synchronously, and when the temperature change rate exceeds 0.5℃ / min, a pre-stored thermal deformation compensation curve is automatically loaded to perform thermal expansion correction on the coordinate data in the depth point cloud model.

[0013] In one embodiment of the present invention, the calibration calculation process in step S2 adopts the ICP iterative nearest point algorithm, the convergence threshold is set to 0.01mm, the maximum number of iterations is 100, and the point cloud registration error is calculated in each iteration until the error is less than the set threshold or the maximum number of iterations is reached.

[0014] In one embodiment of the present invention, the execution cycle of the entire compensation method does not exceed 5ms, wherein the time for depth image acquisition and preprocessing does not exceed 1.5ms, the time for error vector calculation does not exceed 1ms, the time for servo vibration suppression control does not exceed 1.5ms, and the time for motion platform fine-tuning does not exceed 1ms, ensuring that the real-time requirements of high-speed mounting processes are met.

[0015] The beneficial effects of the technical solution of this invention are as follows: This invention sets the baseline distance between the left and right cameras in a multi-view depth camera array to within the range of 150mm to 200mm, maintains an angle of 30° to 45° between the optical axis and the PCB board plane, and combines a binocular stereo vision architecture with image acquisition capabilities of at least 1920×1080 resolution and 60fps frame rate to achieve high-density acquisition of spatial coordinate data and optimize depth measurement. It can capture the three-dimensional features of minute components, while the optimized viewing angle avoids reflection interference and occlusion problems during vertical shooting, providing high-quality raw point cloud data for subsequent registration. After the depth point cloud model is constructed, the ICP iterative nearest point algorithm, with a convergence threshold of 0.01mm and a maximum iteration limit of 100, matches the real-time acquired point cloud with preset standard position data layer by layer. By repeatedly calculating the point cloud registration error and updating the transformation matrix, until the convergence condition is met or the iteration limit is reached, the positional deviation of the mounting head in the X, Y, and Z axes and the rotational deviation angle around the three axes are calculated, generating an initial error vector.

[0016] This invention uses an initial error vector input into a dynamic servo vibration damping module to drive a high-frequency response piezoelectric ceramic actuator. The actuator is connected to a low-inertia mechanical linkage structure via a flexible hinge mechanism, forming a parallel adjustable active vibration damping unit. This allows the micro-displacement force generated by the actuator to act on key nodes of the robotic arm, counteracting the elastic deformation and high-frequency vibration components caused by motor start-stop during movement. Simultaneously, an adaptive filtering algorithm applies wavelet packet decomposition technology to decompose the high-frequency acceleration signal collected by the end effector of the mounting head into eight frequency bands. It identifies the high-frequency jitter components in bands 6 to 8 and the low-frequency jitter components in bands 1 to 3, and automatically removes background noise interference using an energy threshold method.

[0017] This invention employs a closed kinematic chain structure in a six-degree-of-freedom motion platform, with three linear modules and three rotary modules working collaboratively to eliminate the error accumulation effect of open-loop structures and ensure attitude stability during high-speed fine-tuning. A real-time temperature compensation step is embedded in the main loop, monitoring the rate of change in ambient temperature. When the rate exceeds 0.5℃ / min, a pre-stored thermal deformation compensation curve is loaded to perform thermal expansion correction on the coordinate data in the depth point cloud model, eliminating systematic coordinate offsets caused by temperature gradients.

[0018] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0019] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1This is a method framework diagram of the real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop proposed in an embodiment of the present invention. Figure 2 This is a functional diagram of the real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop proposed in an embodiment of the present invention. Detailed Implementation

[0020] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0021] The following describes a real-time mounting error compensation method based on dynamic servo vibration suppression and depth vision closed loop according to an embodiment of the present invention, with reference to the accompanying drawings.

[0022] like Figures 1 to 2 As shown, this embodiment of the invention provides a real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop, including the following steps: S1: Obtain the three-dimensional spatial coordinate data of the reference points on the PCB board and the components to be mounted by a multi-view depth camera array, and construct a depth point cloud model including surface normal vectors and relative poses. S2: Compare the depth point cloud model with the preset standard position data, calculate the position deviation of the mounting head in the X, Y, and Z axis directions and the rotation deviation angle around the three axes, and generate the initial error vector; S3: Drive the dynamic servo vibration damping module according to the initial error vector. The dynamic servo vibration damping module includes a high-frequency response piezoelectric ceramic actuator and a low-inertia mechanical linkage structure. It uses the inverse piezoelectric effect of the piezoelectric ceramic actuator to generate a reverse micro-displacement force to counteract the elastic deformation and high-frequency vibration components during the movement of the robotic arm. S4: While the dynamic servo vibration damping module is in operation, the high-frequency acceleration signal of the end effector of the mounting head is collected in real time. The low-frequency jitter caused by the motor start-stop and the high-frequency vibration caused by mechanical resonance are separated by an adaptive filtering algorithm, and the separated vibration spectrum data is fed back to the control loop. S5: The initial error vector generated in step S2 is fused with the real-time vibration spectrum data fed back in step S4, and input into the Kalman filter for state estimation, and the corrected final error compensation vector is output. S6: Based on the final error compensation vector, send pulse sequence commands to the six-degree-of-freedom motion platform of the placement head to drive the servo motors of each axis to perform fine-tuning motion, so that the residual error of the end of the placement head relative to the target position of the PCB board is controlled within the sub-pixel range. S7: At the moment the placement action is completed, the depth camera array is triggered again to perform a second imaging of the placed component, extract the deviation value between the component center coordinates and the preset coordinates. If the deviation value exceeds the set threshold, an anomaly is marked and the re-placement process is triggered.

[0023] In a specific application of this invention, after the system is started, the multi-view depth camera array first performs a spatial data acquisition task. The array uses the principle of binocular stereo vision to obtain the three-dimensional coordinate information of the PCB board reference point and the component to be mounted. Based on the discrete point cloud data, a depth point cloud model containing surface normal vectors and relative poses is constructed, transforming the position information on the two-dimensional plane into geometric constraints in three-dimensional space. The system compares and calculates the position deviation of the mounting head in the X, Y, and Z axes and the rotation deviation angle around the three axes, thereby generating an initial error vector.

[0024] Furthermore, based on the obtained initial error vector, the system drives the dynamic servo vibration damping module to intervene. The module integrates a high-frequency response piezoelectric ceramic actuator and a low-inertia mechanical linkage structure, forming a linkage unit through a specific mechanical connection. When the system detects the initial error vector, the control unit applies a reverse voltage signal to the piezoelectric ceramic actuator based on the vector value, utilizing the inverse piezoelectric effect to generate a micro- or even nano-level reverse micro-displacement force. Simultaneously, while the dynamic servo vibration damping module is operating, the system acquires the high-frequency acceleration signal of the end effector of the mounting head in real time. The acquired raw signal contains multiple frequency components, including low-frequency jitter caused by motor start-stop and high-frequency vibration caused by mechanical resonance. The system processes the signal using an adaptive filtering algorithm, which can separate the vibration signals of different frequency components and extract key vibration spectrum data. The separated vibration spectrum data is fed back to the control loop in real time as a basis for correcting the control parameters.

[0025] Subsequently, the control unit fuses the initial error vector generated in step S2 with the real-time vibration spectrum data fed back in step S4, and inputs the fused data into the Kalman filter for state estimation. The Kalman filter, through prediction and update steps, combines the deterministic information of the initial error with the random characteristics of the vibration signal to filter out measurement noise and interference signals, and outputs the smoothed final error compensation vector. After obtaining the final error compensation vector, the system sends pulse sequence commands to the six-degree-of-freedom motion platform of the placement head according to the vector. After receiving the commands, the six-degree-of-freedom motion platform drives the servo motors of each axis to perform fine-tuning motion, and performs multi-dimensional spatial adjustment of the end effector of the placement head, so that the residual of the end of the placement head relative to the target position of the PCB board is controlled within the sub-pixel range.

[0026] At the moment the placement action is completed, the system triggers the depth camera array again to perform a secondary imaging of the placed component. The secondary imaging process uses global shutter mode to eliminate image blurring caused by high-speed movement. The image processing algorithm extracts the component center coordinates and compares them with preset coordinates to calculate the deviation value. If the deviation value exceeds the set threshold, the system will mark the abnormality and trigger the re-placement process to ensure that the final placement quality meets the process requirements.

[0027] In one specific implementation, the multi-view depth camera array in step S1 adopts a binocular stereo vision architecture, with the baseline distance between the left and right cameras being 150mm-200mm, the camera resolution being no less than 1920×1080, the frame rate being no less than 60fps, and the angle between the camera optical axis and the PCB board plane being 30°-45°, in order to optimize the depth measurement accuracy.

[0028] In specific applications of this invention, the multi-view depth camera array in step S1 adopts a binocular stereo vision architecture. The baseline distance between the left and right cameras is set in the range of 150mm to 200mm. The distance value is determined by comprehensively considering the field of view and ranging accuracy of the mounting operation space. When the baseline distance is less than 150mm, the triangulation error of the stereo matching algorithm at long distances will increase with the square of the distance, resulting in a decrease in the signal-to-noise ratio of the depth data. When the baseline distance exceeds 200mm, the overlapping area of ​​the field of view shrinks, and edge objects are prone to occlusion or mismatch. Moreover, the large baseline increases the size of the mounting structure, which is not conducive to equipment layout. The baseline range of 150mm to 200mm ensures sufficient triangulation base length while maintaining a wide field of view coverage, so that the reference points on the PCB board and the components to be mounted can be included in the imaging field of view. The camera resolution is not less than 1920×1080, ensuring that the feature points in the image can be fully sampled under the sub-millimeter mounting accuracy requirements. If the resolution is lower than this standard, the quantization error during feature point localization will account for a large proportion of the total error budget, leading to discrete jumps in the subsequently calculated 3D coordinates and failing to meet the requirements for high-precision compensation. In high-speed mounting scenarios, the dynamic response of mechanical structures often occurs within millisecond timescales. Low frame rates can cause phase lag, resulting in a time delay between visual feedback information and the actual physical position, thus introducing additional tracking errors. High frame rate data acquisition combined with high-resolution imaging provides sufficient data redundancy for subsequent depth calculations, effectively suppressing the impact of single-frame image noise on the overall stability of the depth model.

[0029] Furthermore, the angle between the camera's optical axis and the PCB board plane is controlled within the range of 30° to 45°. This angle configuration determines the geometric sensitivity and projection distortion of depth measurement. When the angle is too small, approaching horizontal, the light projection angle is too gentle, and the parallax displacement generated by minute height undulations on the surface is minimal, making depth calculation extremely sensitive to noise. Small image matching errors can be amplified into huge depth deviations. When the angle is too large, approaching vertical, although the vertical measurement sensitivity is improved, the projection of the camera's field of view in the horizontal direction is severely compressed, and it is easily affected by high-light reflections from metal pads or component pins on the PCB board surface, causing feature point extraction failures. A tilt angle of 30° to 45° finds the optimal balance point in triangulation geometry, ensuring sufficient parallax displacement to improve depth resolution while avoiding excessive projection distortion and strong reflection interference. At this angle, the parallax changes at corresponding points in the left and right images acquired by the binocular camera exhibit a good linear relationship with the actual depth changes, which helps simplify the subsequent depth calculation model and reduce computational complexity.

[0030] Furthermore, the left and right cameras are triggered to acquire data simultaneously. Using the parallax information generated by the baseline distance of 150mm to 200mm, combined with the feature matching of the high-resolution image of 1920×1080, the depth value corresponding to each pixel is calculated. The continuous frame stream of 60fps ensures that the depth point cloud model can be updated in real time during the movement of the robotic arm, avoiding the loss of feature points due to motion blur. The optical axis angle of 30° to 45° optimizes the estimation accuracy of the surface normal vector, so that the constructed depth point cloud model not only contains XYZ three-dimensional coordinates, but also accurately reflects the local curvature and tilt angle of the PCB board surface.

[0031] In one specific implementation, the calibration calculation process in step S2 adopts the ICP iterative nearest point algorithm, with a convergence threshold of 0.01 mm and a maximum number of iterations of 100. The point cloud registration error is calculated in each iteration until the error is less than the set threshold or the maximum number of iterations is reached.

[0032] In specific applications, the alignment process in step S2 of this invention employs an iterative nearest-point algorithm to handle the registration problem between the depth point cloud model and the preset standard position data. The algorithm gradually reduces the spatial difference between the two sets of point clouds by repeatedly performing nearest-point matching and rigid body transformation calculations. In the mounting operation scenario, the reference points on the PCB board surface and the actual point cloud data composed of the components to be mounted often have deviations caused by rotation, translation, and local deformation. Direct comparison cannot obtain accurate pose information. The ICP algorithm first selects a feature point in the current point cloud as a query point and searches for the corresponding point with the closest Euclidean distance in the preset standard point cloud set to establish a point pair matching relationship. Subsequently, based on these matching point pairs, the optimal rotation matrix and translation vector are calculated to transform the current point cloud to a new spatial position. The convergence threshold is set to 0.01mm. When the average registration error of the point cloud calculated in two consecutive iterations is less than this value, the point cloud is considered to be aligned. At this time, the output displacement is the position deviation of the mounting head in the X, Y, and Z axis directions and the rotation deviation angle around the three axes. If only a single iteration or a large convergence threshold is relied upon, the calculation results are prone to getting trapped in local minima, causing the initial error vector to contain false offset components, which in turn misleads the subsequent servo vibration damping module to issue incorrect compensation commands.

[0033] Furthermore, in the actual mounting process, due to PCB surface reflection, component obstruction, or camera noise, some areas cannot find effective nearest-point matching, leading to non-convergence of the error function. Without an iteration limit, the algorithm continues to run on invalid paths, consuming the real-time computing resources of the control system and causing delays in subsequent steps S3 and S4. Setting an upper limit of 100 iterations ensures that even under difficult matching conditions, the system can output a relatively reasonable transformation result within a limited time, guaranteeing the timing stability of the entire closed-loop control process. The point cloud registration error calculated in each iteration reflects the difference between the current estimated pose and the true pose. The error value is not only used to determine whether to stop the iteration but also recorded as a monitoring indicator. By observing the error decrease curve with the number of iterations, the matching quality of the point cloud data can be evaluated. If the error no longer decreases significantly after multiple iterations, it indicates that the current matching point set contains a large number of outliers or incorrect matches. In this case, the system can trigger a protection mechanism based on the maximum number of iterations to avoid outputting an unreliable initial error vector.

[0034] In one specific implementation, the high-frequency response piezoelectric ceramic actuator in step S3 is made of PZT-5H material, with a maximum stroke range of ±10μm and a resonant frequency greater than 20kHz. The piezoelectric ceramic actuator is connected to the mechanical linkage through a flexible hinge mechanism to form a parallel adjustable active vibration damping unit.

[0035] In specific applications of this invention, the high-frequency response piezoelectric ceramic actuator in step S3 uses PZT-5H material. This material has an electromechanical coupling coefficient and low dielectric loss characteristics, ensuring rapid deformation and sufficient thrust output when a reverse voltage signal is applied. The high Curie temperature of PZT-5H material ensures that its piezoelectric performance will not drift or decay under the heat generated by the long-term continuous operation of the mounting equipment, thus maintaining the stability of the control loop output. If the stroke is too small, it cannot completely offset large-amplitude transient vibrations; if the stroke is too large, it will lead to an increase in actuator size and a decrease in control linearity, introducing nonlinear errors. A stroke range of ±10μm allows the actuator to fully compensate for minor disturbances in the mechanical link without entering the saturation region. A resonant frequency greater than 20kHz is a key indicator for suppressing high-frequency vibrations. The structural resonance generated by the robotic arm during high-speed start-up, stopping, and commutation is often concentrated in the range of tens to thousands of hertz, but some higher-order modes and motor commutation noise can extend to above 20kHz. When the driving frequency approaches or exceeds the actuator's resonant frequency, its output displacement will undergo a drastic phase jump, leading to control failure or even system instability. A resonant frequency greater than 20kHz provides sufficient phase margin for the control system, enabling the piezoelectric ceramic actuator to serve as an ideal force source and respond quickly to command signals over an extremely wide frequency band.

[0036] Furthermore, after the initial error vector generated in step S2 is input into the control loop, the system calculates the required compensation displacement based on the magnitude and direction of the deviation, and converts it into a driving voltage applied to the PZT-5H piezoelectric ceramic actuator. Due to the high electromechanical conversion efficiency of the material, the actuator generates corresponding deformation within microseconds, transmitting the displacement to the main nodes of the mechanical link through a flexible hinge mechanism. At this time, the reverse micro-displacement force generated by the actuator and the high-frequency vibration components caused by inertia, friction torque, or structural elastic deformation during the movement of the robotic arm are spatially superimposed, achieving vector cancellation. Since the resonant frequency is greater than 20kHz, the actuator can cover the main high-frequency vibration band of the robotic arm, effectively suppressing the vibration caused by motor commutation, gear meshing, or airflow disturbance. For example, during light-load, high-speed movement, the system can reduce the effective stiffness to improve sensitivity; during heavy-load, low-speed movement, the effective stiffness can be increased to enhance anti-interference capability.

[0037] In one specific implementation, the adaptive filtering algorithm in step S4 uses wavelet packet decomposition technology to decompose the acceleration signal into 8 frequency bands, where the high-frequency jitter component is located in the 6th-8th frequency band and the low-frequency jitter component is located in the 1st-3rd frequency band. Noise interference is automatically identified and eliminated by the energy threshold method.

[0038] In specific applications of this invention, the adaptive filtering algorithm in step S4 uses wavelet packet decomposition technology to process the high-frequency acceleration signal collected by the end effector of the mounting head. Utilizing the multi-resolution analysis characteristics of wavelet packet transform, the non-stationary vibration signal is finely divided in the time-frequency domain. Wavelet packet decomposition simultaneously decomposes the low-frequency and high-frequency components of the signal, constructing a complete binary tree structure, thereby more uniformly covering the entire frequency spectrum. The system decomposes the original acceleration signal into eight frequency bands, each with a bandwidth approximately one-eighth of the sampling frequency. This allows vibration components with different physical causes to be mapped to different frequency bands. High-frequency jitter components originate from high-speed commutation of the robotic arm joint motor, gear meshing impact, and high-order modal resonance of the cantilever beam structure. The vibration energy is concentrated in the higher frequency region, thus they are located in frequency bands 6 to 8. Low-frequency jitter components are typically caused by the start-stop inertia of the servo motor, transmission chain clearance, and external airflow disturbances. Their frequency is low and changes slowly, distributed in frequency bands 1 to 3.

[0039] Based on frequency band division, the system introduces an energy threshold method to automatically identify and eliminate noise interference. Accelerometers inevitably encounter environmental electromagnetic interference, thermal noise, and crosstalk from transmission lines during data acquisition. Noise often manifests as randomly distributed broadband signals with energy densities far lower than actual mechanical vibration signals. When the energy in a frequency band is below a set threshold, the band is determined to be primarily noise and is directly set to zero. When the energy is above the threshold, the band data is retained as valid vibration information. This process is dynamic; the threshold is automatically adjusted based on the background noise level under current operating conditions to ensure effective differentiation between useful signals and noise under different loads and speeds. For high-frequency jitter components located in bands 6 to 8, the system extracts their amplitude and phase information to assess the instantaneous stability of the mechanical structure. For low-frequency jitter components located in bands 1 to 3, the system records their trend changes to determine the steady-state performance of the drive system. Bands 4 and 5 typically contain mixed signals in transition regions; the algorithm categorizes or weights these signals based on their energy distribution to ensure the integrity of the spectral data.

[0040] Furthermore, the separated vibration spectrum data is fed back to the control loop, forming the input state variables of the Kalman filter together with the initial error vector generated in step S2. The high-frequency jitter component data reflects the transient response of the system on a millisecond time scale, revealing residual vibrations that the piezoelectric ceramic actuator has not completely canceled out. This data is used to correct the process noise covariance matrix in the Kalman filter, improving the real-time performance of state estimation. The low-frequency jitter component data reveals the long-term drift trend of the system, helping the filter predict future pose deviations, thereby optimizing the update strategy of the observation noise covariance matrix. If the original signal without frequency band decomposition is used directly, the control loop will respond to both high-frequency noise and low-frequency trends simultaneously, causing frequent acceleration and deceleration of the servo motor, resulting in unnecessary mechanical wear and even resonance. Wavelet packet decomposition technology, through frequency band isolation, enables the control algorithm to adopt different suppression strategies for different types of vibration. For example, it can use fast-response active compensation for high-frequency jitter and integral adjustment or feedforward compensation for low-frequency jitter.

[0041] In one specific implementation, the Kalman filter in step S5 adopts an extended Kalman filter architecture, whose state vector includes position error, velocity error, acceleration error and vibration phase information, and the process noise covariance matrix Q and the observation noise covariance matrix R are dynamically adjusted according to the real-time signal-to-noise ratio.

[0042] In specific applications of this invention, the Kalman filter in step S5 adopts an extended Kalman filter architecture to adapt to the nonlinear characteristics of the robotic arm motion model in the mounting system. Through linearization approximation, the nonlinear state equation is transformed into a linear form for recursive calculation. The extended Kalman filter calculates the Jacobian matrix at the current state estimate and performs a first-order Taylor expansion on the system state transition function and observation function, thereby dynamically updating the linearization coefficients in each iteration. This ensures that the state estimation process can track the nonlinear motion trajectory of the robotic arm. The position error vector reflects the three-dimensional deviation calculated by the vision system and is the target variable for control loop correction. The velocity error vector characterizes the rate of position change and is used to predict the position trend at the next moment and suppress overshoot. The acceleration error vector captures transient impacts caused by inertial forces, helping the system respond quickly to high-frequency disturbances. Vibration phase information serves as a crucial state supplement, recording the instantaneous phase angle of the current vibration waveform, enabling the filter to distinguish interference signals of the same frequency but different phases, thus achieving phase cancellation when the piezoelectric ceramic actuator applies a reverse compensation force.

[0043] Furthermore, during the mounting process, environmental factors such as airflow fluctuations, conveyor belt vibrations, and changes in the reflectivity of the PCB surface can cause drastic fluctuations in the sensor's observed noise level. If the process noise covariance matrix Q is set too large, the filter will over-rely on model predictions and ignore actual observation data, causing state estimation to lag behind actual physical changes. If it is set too small, the filter will over-rely on noisy observation data, causing frequent oscillations in the estimated values. Similarly, the observation noise covariance matrix R reflects the reliability of the measurement system. When the signal-to-noise ratio decreases, the reliability of the observation data decreases. Increasing the value of R can reduce the weight of the observation data in state updates and prevent noise from contaminating the state estimation results. The dynamic adjustment mechanism estimates the current signal-to-noise ratio by calculating the power spectral density or root mean square value of the input signal in real time and automatically scales the diagonal elements of the Q and R matrices accordingly. When an increase in the energy of a high-frequency jitter component is detected, the system determines it to be a strong interference environment and automatically increases the observation noise covariance matrix R while decreasing the process noise covariance matrix Q, making the filter more inclined to make smooth predictions based on the robotic arm's dynamics model and avoiding following noise fluctuations. When the system is in a stable operating state and the signal-to-noise ratio is high, the system decreases R and increases Q to make the filter converge quickly to the latest observation data and improve the tracking sensitivity.

[0044] Specifically, the state update process of the extended Kalman filter is closely integrated with the frequency band separation data output in step S4. The high-frequency jitter and low-frequency dither components extracted through wavelet packet decomposition are input into the filter as part of the observation vector. The filter uses these features to update the various components of the state vector. The position error is initialized by the initial error vector provided by the depth vision closed loop; the velocity error is predicted by combining the time difference of the position error with the model; the acceleration error is derived from the rate of change of the velocity error; and the vibration phase information is directly derived from the phase angle data extracted from the frequency band analysis. In the prediction step, the state at the current moment is calculated using the state estimate from the previous moment and the system dynamics model; in the update step, the predicted value is corrected using the observation data at the current moment. This process is completed within microseconds, ensuring the real-time requirements of the control loop.

[0045] During the robotic arm's acceleration and start-up phase, the acceleration error component increases due to the large inertial force. At this time, the system automatically increases the process noise covariance matrix Q, allowing the state estimation to respond more quickly to changes in model predictions and avoiding estimation bias caused by model uncertainty. During the robotic arm's deceleration and stopping phase, the velocity error undergoes abrupt changes due to friction and back electromotive force. The system dynamically adjusts the observation noise covariance matrix R based on the signal-to-noise ratio to balance the relationship between model predictions and observed data, preventing state jumps caused by sudden data changes. Furthermore, the introduction of vibration phase information enables the filter to identify periodic interference. When periodic vibration at a specific frequency is detected, the filter automatically adjusts its internal gain to enhance its ability to suppress frequency components.

[0046] In one specific implementation, the six-degree-of-freedom motion platform in step S6 includes three linear modules and three rotary modules. Each module is driven by a frameless torque motor, and the encoder resolution is not less than 20 bits. The modules are connected by an aluminum alloy frame to form a closed motion chain structure.

[0047] In a specific application of this invention, the six-degree-of-freedom motion platform in step S6 is constructed by combining three linear modules and three rotary modules. Each module independently undertakes the task of displacement or attitude adjustment in a specific direction, and together realizes the full-degree-of-freedom motion control of the mounting head in three-dimensional space. The linear modules are responsible for translation in the X, Y, and Z axes, and the rotary modules are responsible for rotation around the X, Y, and Z axes. The three are directly driven by a frameless torque motor. The frameless torque motor separates the stator and rotor. The rotor is installed at the load end, and the stator is fixed on the frame. This structure makes the power transmission path the shortest and the mechanical inertia the smallest, which can reduce the response delay of the system. The high-resolution encoder provides real-time feedback on the absolute angle of the rotor or the actual position of the linear modules. The data is uploaded to the control core at high frequency, providing high-precision state observation values ​​for the Kalman filter in step S5, and supporting the rapid convergence of closed-loop control.

[0048] Furthermore, aluminum alloys, while ensuring sufficient strength, have a low mass density, which helps reduce the inertial load on moving parts and improve acceleration performance. In an open-loop series structure, the positional error of the end effector accumulates at each stage, and the vibration of the bottom module is easily transmitted and amplified to the top. The closed kinematic chain, through multi-point and multi-directional mechanical support, distributes the force of each module to the entire frame, effectively suppressing bending deformation under unidirectional force. When the piezoelectric ceramic actuator generates a high-frequency reverse compensation force, the closed frame can quickly absorb and disperse these dynamic loads, preventing structural resonance caused by local stress concentration.

[0049] During the mounting process, the robotic arm needs to frequently start and stop at high speeds and track complex trajectories. The direct drive characteristics of the frameless torque motor enable the system to output a large instantaneous torque to meet the requirements of rapid response. The position feedback provided by the high-resolution encoder enables the control system to correct deviations caused by friction, gravity components or external disturbances in real time. The closed kinematic chain structure ensures that the relative positional relationship between each module remains stable under high-speed movement, avoiding positioning drift caused by frame deformation.

[0050] After the Kalman filter calculates the state vector containing vibration phase information, the control command is sent to the frameless torque motor. The motor adjusts its output torque according to the command, driving the linear and rotary modules for fine-tuning. Since there is no intermediate transmission link, the motor's response is almost instantaneous, enabling immediate execution of compensation actions. The high-resolution encoder monitors the actual motion state of each axis in real time and feeds the data back to the filter and control loop, forming a complete "perception-decision-execution" closed loop. If a traditional servo motor with a gearbox is used, the transmission backlash will cause a delay in command execution, making it impossible to promptly compensate for high-frequency vibrations. If the encoder resolution is insufficient, minute positional deviations will not be detected, leading to control loop failure.

[0051] In one specific implementation, the secondary imaging process in step S7 uses a global shutter mode for shooting, with the exposure time set to 100μs-200μs to eliminate motion blur caused by high-speed movement, and the image processing algorithm uses sub-pixel edge detection technology to achieve a positioning accuracy of 0.1 pixels.

[0052] In specific applications of this invention, the secondary imaging process in step S7 employs a global shutter mode to ensure that all pixels on the image sensor begin and end exposure at the same time, eliminating motion distortion caused by the line scanning time difference in rolling shutter mode. During the high-speed movement of the mounting equipment, there is a relative displacement between the robotic arm's end effector and the PCB board. If a rolling shutter is used, the acquisition time of the top and bottom of the image differs by milliseconds, resulting in a jelly effect where straight edges appear tilted or curved, affecting the accuracy of feature point extraction. Global shutter mode controls the exposure time within an extremely short window of 100μs to 200μs, so that the displacement of the object within this time span is much smaller than the physical size of a single pixel. This freezes the high-speed motion state at the physical level and avoids the motion blur caused by excessively long integration time. The lower limit of 100μs needs to balance the light intensity and signal-to-noise ratio. Too short an exposure will result in insufficient image brightness, increasing the difficulty of subsequent algorithm processing. The upper limit of 200μs is limited by the instantaneous speed of the robotic arm under maximum acceleration. Exceeding this value will produce visible motion blur, causing the edge contours to spread.

[0053] Furthermore, in the sharp image captured by the global shutter, the edge contours are sharp and free of geometric distortion, providing high-quality input data for the sub-pixel algorithm. If the image has motion blur, the grayscale transition at the edges will become gradual, causing the sub-pixel fitting algorithm to fail to converge accurately, or even produce divergence errors. Therefore, a short exposure time of 100μs-200μs is a prerequisite for achieving a positioning accuracy of 0.1 pixels. The sub-pixel-level coordinate data acquired by the vision system is converted into a position vector in three-dimensional space, which serves as the observation input to the Kalman filter and is fused with the state estimate output from step S5.

[0054] In one specific implementation, a real-time temperature compensation step is also included: during the mounting process, the ambient temperature change is monitored synchronously, and when the temperature change rate exceeds 0.5℃ / min, a pre-stored thermal deformation compensation curve is automatically loaded to perform thermal expansion correction on the coordinate data in the depth point cloud model.

[0055] The entire compensation method has an execution cycle of no more than 5ms, of which depth image acquisition and preprocessing take no more than 1.5ms, error vector calculation takes no more than 1ms, servo vibration suppression control takes no more than 1.5ms, and motion platform fine-tuning takes no more than 1ms, ensuring that the real-time requirements of high-speed mounting processes are met.

[0056] In practical applications, the real-time temperature compensation step of this invention is embedded in the main control loop of the mounting operation. A temperature sensor array integrated into the main nodes of the equipment base, gantry, and motion module continuously collects environmental and structural surface temperature data. The system performs differential calculations on the continuously collected temperature values ​​to calculate the rate of change per unit time. When the temperature change rate exceeds a set threshold of 0.5℃ / min, it is determined to have entered the thermal deformation sensitive zone, triggering the loading program of the pre-stored thermal deformation compensation curve. The threshold setting is based on measured analysis of the thermal expansion coefficient of typical mounting equipment materials and the thermal inertia of the structure. Below this rate, the structural deformation is in a slow drift state, which can be automatically corrected by conventional servo integral terms; above this rate, the nonlinear deformation speed caused by the thermal gradient exceeds the self-correction capability of the mechanical transmission system, introducing an active compensation strategy. The pre-stored thermal deformation compensation curve is a three-dimensional spatial deviation mapping table obtained by calibrating a laser tracker at different ambient temperature points before the equipment leaves the factory or during periodic maintenance. It records the changes in displacement, angular deflection, and Z-axis height of each degree of freedom axis from the low-temperature baseline state to the high-temperature state.

[0057] Once the triggering condition is met, the system reads the current actual temperature value and performs linear interpolation or spline fitting on the compensation curve to generate a corresponding thermal deformation vector field. This vector field contains translational components in the X, Y, and Z directions, as well as rotational components around these three axes, which are then superimposed onto the point cloud model coordinate data generated by the depth vision system. The depth point cloud model reflects the real-time spatial distribution of the PCB board surface and component pads. Thermal deformation causes a shift in the relative position between the camera coordinate system and the workpiece coordinate system. Without correction, the error vector calculated by vision will contain spurious deviations caused by thermal expansion and contraction. The thermal expansion correction process restores the true geometric dimensions of the physical space to the standard state at the reference temperature, ensuring that the subsequent calculated mounting errors only reflect dynamic factors such as mechanical vibration and positioning deviations, and are not affected by ambient temperature fluctuations.

[0058] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0059] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A real-time compensation method for placement error based on dynamic servo dithering and deep vision closed-loop, characterized in that, Includes the following steps: S1: Obtain the three-dimensional spatial coordinate data of the reference points on the PCB board and the components to be mounted by a multi-view depth camera array, and construct a depth point cloud model including surface normal vectors and relative poses. S2: Compare the depth point cloud model with the preset standard position data, calculate the position deviation of the mounting head in the X, Y, and Z axis directions and the rotation deviation angle around the three axes, and generate an initial error vector; S3: Drive the dynamic servo vibration damping module according to the initial error vector. The dynamic servo vibration damping module includes a high-frequency response piezoelectric ceramic actuator and a low-inertia mechanical linkage structure. It uses the inverse piezoelectric effect of the piezoelectric ceramic actuator to generate a reverse micro-displacement force to counteract the elastic deformation and high-frequency vibration components during the movement of the robotic arm. S4: While the dynamic servo vibration damping module is in operation, the high-frequency acceleration signal of the end effector of the mounting head is collected in real time. The low-frequency jitter caused by the motor start-stop and the high-frequency vibration caused by mechanical resonance are separated by an adaptive filtering algorithm, and the separated vibration spectrum data is fed back to the control loop. S5: The initial error vector generated in step S2 is fused with the real-time vibration spectrum data fed back in step S4, and input into the Kalman filter for state estimation, and the corrected final error compensation vector is output. S6: Based on the final error compensation vector, send pulse sequence commands to the six-degree-of-freedom motion platform of the placement head to drive the servo motors of each axis to perform fine-tuning motion, so that the residual error of the end of the placement head relative to the target position of the PCB board is controlled within the sub-pixel range. S7: At the moment the placement action is completed, the depth camera array is triggered again to perform a second imaging of the placed component, extract the deviation value between the component center coordinates and the preset coordinates. If the deviation value exceeds the set threshold, an anomaly is marked and the re-placement process is triggered.

2. The real-time compensation method for placement error based on dynamic servo dithering and depth vision closed-loop according to claim 1, characterized in that, The multi-view depth camera array described in step S1 adopts a binocular stereo vision architecture, with a baseline distance of 150mm-200mm between the left and right cameras, a camera resolution of no less than 1920×1080, a frame rate of no less than 60fps, and an angle of 30°-45° between the camera optical axis and the PCB board plane to optimize depth measurement accuracy.

3. The real-time compensation method for placement error based on dynamic servo dithering and depth vision closed-loop according to claim 1, characterized in that, The high-frequency response piezoelectric ceramic actuator mentioned in step S3 is made of PZT-5H material, with a maximum stroke range of ±10μm and a resonant frequency greater than 20kHz. The piezoelectric ceramic actuator is connected to the mechanical linkage through a flexible hinge mechanism to form a parallel adjustable active vibration damping unit.

4. The real-time compensation method for placement error based on dynamic servo dithering and depth vision closed-loop of claim 1, wherein, The adaptive filtering algorithm described in step S4 uses wavelet packet decomposition technology to decompose the acceleration signal into 8 frequency bands, where the high-frequency jitter component is located in the 6th-8th frequency band and the low-frequency jitter component is located in the 1st-3rd frequency band. It also automatically identifies and removes noise interference using the energy threshold method.

5. The real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop as described in claim 1, characterized in that, The Kalman filter described in step S5 adopts an extended Kalman filter architecture. Its state vector includes position error, velocity error, acceleration error and vibration phase information. The process noise covariance matrix Q and the observation noise covariance matrix R are dynamically adjusted according to the real-time signal-to-noise ratio.

6. The real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop as described in claim 1, characterized in that, The six-degree-of-freedom motion platform described in step S6 includes three linear modules and three rotary modules. Each module is driven by a frameless torque motor, and the encoder resolution is not less than 20 bits. The modules are connected by an aluminum alloy frame to form a closed motion chain structure.

7. The real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop as described in claim 1, characterized in that, The secondary imaging process described in step S7 uses a global shutter mode for shooting, with an exposure time set to 100μs-200μs to eliminate motion blur caused by high-speed movement. The image processing algorithm uses sub-pixel edge detection technology, achieving a positioning accuracy of 0.1 pixels.

8. The real-time compensation method for mounting errors based on dynamic servo vibration suppression and depth vision closed loop as described in claim 1, characterized in that, It also includes a real-time temperature compensation step: during the mounting process, the ambient temperature changes are monitored synchronously. When the temperature change rate exceeds 0.5℃ / min, the pre-stored thermal deformation compensation curve is automatically loaded to correct the coordinate data in the depth point cloud model for thermal expansion.

9. The real-time compensation method for placement error based on dynamic servo dithering and depth vision closed-loop of claim 1, wherein, The calibration calculation process described in step S2 adopts the ICP iterative nearest point algorithm, with a convergence threshold of 0.01 mm and a maximum number of iterations of 100. The point cloud registration error is calculated in each iteration until the error is less than the set threshold or the maximum number of iterations is reached.

10. The real-time compensation method for placement error based on dynamic servo dithering and depth vision closed-loop of claim 1, wherein, The entire compensation method has an execution cycle of no more than 5ms, of which depth image acquisition and preprocessing take no more than 1.5ms, error vector calculation takes no more than 1ms, servo vibration suppression control takes no more than 1.5ms, and motion platform fine-tuning takes no more than 1ms, ensuring that the real-time requirements of high-speed mounting processes are met.

Images