3D grating mounting self-learning error compensation method and system

Abstract

The present application relates to a kind of 3D grating mounting self-learning error compensation method and system, the method includes: using the single industrial camera of being arranged on fixed reference point, Mark point image of COG display panel and 3D grating glass is respectively collected, initial alignment deviation is obtained by calculation;Three-level layered error decoupling and collaborative control architecture is constructed, the full-link dynamic inhibition of multiple source coupling error and steady precision maintenance are realized;Based on correction parameter driving manipulator and motion platform, complete fitting, and detect actual fitting deviation, and update compensation model parameter, form closed loop control.The present application is innovated by the coordination of same coordinate single camera hardware architecture and three-level layered self-learning compensation algorithm, eliminates system error from source and realizes submicron multiple source error inhibition, while greatly reducing equipment cost, improving yield and beat efficiency, break through the precision bottleneck of flip mounting, significantly improve 3D display performance and have the ability to CPO light engine and other light communication packaging scene extension.

Description

Technical Field

[0001] This invention relates to the field of 3D display manufacturing and precision mounting technology, and more specifically, to a 3D grating mounting self-learning error compensation method and system. Background Technology

[0002] The core bottleneck in the industrialization of parallax barrier-type naked-eye 3D display technology lies in the micron-level high-precision mounting of the 3D grating glass and the display panel. The industry requires the alignment deviation between the two to be controlled within ±5μm; otherwise, it will directly cause problems such as image crosstalk, ghosting, and dizziness, seriously affecting the 3D display effect. However, existing technologies still have shortcomings, such as: the dual-camera coordinate positioning has inherent system errors; there is a fundamental contradiction between the linear compensation model and the nonlinear multi-source coupling error; the offline calibration scheme cannot adapt to the dynamic errors of mass production, and the compensation effect continues to decay; the collaborative solution capability for multi-source coupling errors is seriously insufficient; and the algorithm is seriously out of touch with the flip-chip mounting process.

[0003] Therefore, it is necessary to design a 3D grating mounting self-learning error compensation method and system that eliminates the inherent system error of dual-camera coordinate positioning, adopts a nonlinear model to accurately adapt to multi-source coupling errors, realizes online dynamic calibration and self-learning update to avoid compensation attenuation, has the ability to jointly decouple and collaboratively solve multi-source errors, and deeply fits the actual scenario of flip-chip mounting process. Summary of the Invention

[0004] The purpose of this invention is to provide a 3D grating mounting self-learning error compensation method and system to solve the problems in the prior art, such as the inherent system error in dual-camera coordinate positioning, the essential contradiction between linear compensation models and nonlinear multi-source coupling errors, the inability of offline calibration schemes to adapt to mass production dynamic errors, the continuous decay of compensation effect, the serious insufficiency of collaborative solution capability for multi-source coupling errors, and the serious disconnect between the algorithm and flip-chip mounting process scenarios.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: A 3D grating mounting self-learning error compensation method includes: Using a single industrial camera set on a fixed reference, Mark point images of the COG display panel and 3D grating glass are acquired in the same visual coordinate system, and the initial alignment deviation is calculated. A three-level hierarchical error decoupling and collaborative control architecture is constructed. Through mechanical error pre-compensation method, glass warpage real-time compensation method and self-learning dynamic closed-loop compensation method, the full-link dynamic suppression and steady-state accuracy maintenance of multi-source coupling errors are achieved. The robot arm and motion platform are driven to complete the bonding based on the compensated correction parameters, and the actual bonding deviation is detected. The compensation model parameters are updated using the actual bonding deviation to form a closed-loop control; wherein, the mechanical error pre-compensation method includes: An improved DH homogeneous transformation matrix method is used to establish a kinematic error model between the manipulator and the XYR platform; The precise values ​​of each error term of the mechanical system are obtained through error calibration. The predicted value of the end pose error is calculated based on the error Jacobian matrix. The reverse compensation amount is added to the motion control command to offset the inherent error of the mechanical system. Based on the nine-point calibration and the measured data of the laser interferometer, the least squares method is used to identify the error Jacobian matrix and generate a pre-compensation parameter table to achieve efficient feedforward correction. The real-time glass warpage compensation method includes: By utilizing the telecentric characteristics of the bilateral telecentric lenses of the same coordinate vision system, online detection of warpage without any additional hardware is achieved through binocular stereo vision. By integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, the dual elimination of defocus positioning error and secondary offset of bonding is achieved. The self-learning dynamic closed-loop compensation method includes: A batch learning and real-time correction closed-loop architecture is adopted. By introducing a forgetting factor, recent deviations are given high weight to adapt to the slow-changing characteristics of errors. On the basis of regular batch statistical updates, a single-item large deviation instant triggering mechanism is superimposed to achieve dynamic adaptive optimization of compensation parameters and anomaly interception. The compensation parameter correction is calculated based on the incremental PID discretization expression. The optimal control coefficient is tuned by the critical proportional method to achieve rapid convergence of the fitting deviation. In addition, integral limiting and dead zone mechanisms are used to effectively suppress integral saturation and frequent fine-tuning of the actuator.

[0006] In one embodiment, the step of using a single industrial camera mounted on a fixed reference to acquire Mark point images of the COG display panel and the 3D grating glass in the same visual coordinate system, and calculating the initial alignment deviation, includes: The COG display panel to be bonded is fixed on the vacuum adsorption platform, which is located directly below the field of view of the vision camera; the vision system is activated, and the camera captures the Mark points at the two corners of the LCM screen. The coordinates of the COG screen are digitally positioned in a unique vision coordinate system, and the reference coordinate matrix MCOG of the COG screen Mark points is recorded. The robotic arm moves the adsorbed 3D grating glass to the same field of view between the vision camera and the COG screen, and adjusts the grating plane and the COG screen plane to be at the same object height; the vision camera captures the Mark points at the two corners of the grating, completes the grating coordinate digital positioning in the same vision coordinate system, and records the coordinate matrix MGR of the grating Mark points. Based on two sets of coordinate matrices under the same coordinate system, the alignment deviation between the grating and the COG screen is calculated directly using the least squares method alignment matching function. This includes the translational deviations Δx and Δy of the X and Y axes, as well as the rotation angle deviation Δθ. There is no coordinate transformation step, thus eliminating system transformation error at its source. The control system synchronously sends the deviation data to the robot and the XYR platform. After the robot completes the attitude correction of the grating, it moves precisely to the top of the COG screen to complete the pre-lamination and vacuum full lamination process.

[0007] In one embodiment, the method of utilizing the telecentric characteristics of the bilateral telecentric lenses of a co-coordinate vision system to achieve online detection of warpage without additional hardware through binocular stereo vision includes: The vision system focuses on the upper and lower surfaces of the glass respectively, records the height difference between the two focusing positions, and obtains the glass thickness reference value. The Mark point areas at the two corners of the glass are focused in layers. The optimal focusing height of each Mark point is calculated using the Tenengrad gradient function to obtain the actual height values ​​of the four corner points. Based on the height values ​​of two corner points, the warping deformation surface of the entire glass is restored by fitting a quadratic surface, thereby obtaining the warping distribution, detection time, and detection accuracy of the entire surface.

[0008] In one embodiment, the method of achieving dual elimination of defocus positioning error and secondary offset by fusing imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, includes: Based on the actual height deviation of the Mark point, combined with the visual system depth model and lens distortion model, the changes in imaging magnification and edge distortion caused by the height deviation are calculated, and the pixel coordinates of the Mark point are corrected to eliminate the positioning error caused by defocus. Based on the thin plate bending model, the in-plane slippage caused by glass warping and flattening during vacuum adsorption bonding is calculated, and the pre-alignment coordinates of the grating are compensated in advance.

[0009] In one embodiment, the batch learning and real-time correction closed-loop architecture introduces a forgetting factor to assign high weight to recent deviations to adapt to the slow-changing characteristics of errors, and superimposes a single-item large deviation instant triggering mechanism on the basis of conventional batch statistical updates, to achieve dynamic adaptive optimization of compensation parameters and anomaly interception, including: After every 8-12 pieces of product are bonded, the vision system automatically detects the actual bonding deviation and calculates the average deviation. The average deviation is used as feedback input, and the correction amount of the compensation parameter is calculated by the incremental PID controller to update the pre-compensation parameter table. For large deviations in a single product, a single compensation correction is immediately triggered to avoid batch defects. By introducing a forgetting factor, more weight is given to recent data, which adapts to the slow variation of error and avoids the interference of historical data on the current compensation effect.

[0010] In one embodiment, the robot arm and motion platform are driven to complete the bonding based on the compensated correction parameters, and the actual bonding deviation is detected. The compensation model parameters are then updated using the actual bonding deviation to form a closed-loop control, including: S301. Power on the device, construct vision system calibration and mechanical system error modeling, generate pre-compensation parameter table and write it into the controller, and complete hardware self-test and origin regression. S302. During mass production, the motion error of the mechanical system is corrected by the pre-compensation algorithm, and the corrected motion control parameters are output to the robot and the XYR platform. S303. By using a single camera to take pictures in the same position, the coordinates of the Mark points of the COG display screen and the 3D grating are collected in the same coordinate system, and the initial alignment deviation is calculated. S304. The vision system acquires glass warping data online, calculates the positioning deviation caused by warping through a real-time compensation algorithm, performs secondary correction on the Mark point coordinates, and obtains the final alignment correction parameters. S305. The control system drives the robotic arm to complete alignment and vacuum bonding with the XYR platform according to the final corrected parameters. S306. After completing a single bonding process, the vision system detects the actual bonding deviation and updates the compensation parameters through a self-learning algorithm to achieve dynamic closed-loop optimization. S307. Repeat steps S301-S306 to achieve continuous optimization of positioning accuracy during continuous production.

[0011] A 3D grating mounting self-learning error compensation system includes: The same-coordinate vision positioning module is used to acquire Mark point images of the COG display panel and the 3D grating glass respectively in the same vision coordinate system using a single industrial camera set on a fixed reference, and calculate the initial alignment deviation. The three-level hierarchical compensation algorithm module is used to construct a three-level hierarchical error decoupling and collaborative control architecture. Through mechanical error pre-compensation method, glass warping real-time compensation method and self-learning dynamic closed-loop compensation method, it realizes full-link dynamic suppression and steady-state accuracy maintenance of multi-source coupling error. The bonding execution and iteration module is used to drive the robot and motion platform to complete bonding based on the compensated correction parameters, detect the actual bonding deviation, and update the compensation model parameters using the actual bonding deviation to form closed-loop control. The mechanical error pre-compensation method includes: An improved DH homogeneous transformation matrix method is used to establish a kinematic error model between the manipulator and the XYR platform; The precise values ​​of each error term of the mechanical system are obtained through error calibration. The predicted value of the end pose error is calculated based on the error Jacobian matrix. The reverse compensation amount is added to the motion control command to offset the inherent error of the mechanical system. Based on the nine-point calibration and the measured data of the laser interferometer, the least squares method is used to identify the error Jacobian matrix and generate a pre-compensation parameter table to achieve efficient feedforward correction. The real-time glass warpage compensation method includes: By utilizing the telecentric characteristics of the bilateral telecentric lenses of the same coordinate vision system, online detection of warpage without any additional hardware is achieved through binocular stereo vision. By integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, the dual elimination of defocus positioning error and secondary offset of bonding is achieved. The self-learning dynamic closed-loop compensation method includes: A batch learning and real-time correction closed-loop architecture is adopted. By introducing a forgetting factor, recent deviations are given high weight to adapt to the slow-changing characteristics of errors. On the basis of regular batch statistical updates, a single-item large deviation instant triggering mechanism is superimposed to achieve dynamic adaptive optimization of compensation parameters and anomaly interception. The compensation parameter correction is calculated based on the incremental PID discretization expression. The optimal control coefficient is tuned by the critical proportional method to achieve rapid convergence of the fitting deviation. In addition, integral limiting and dead zone mechanisms are used to effectively suppress integral saturation and frequent fine-tuning of the actuator.

[0012] As can be seen from the above technical solutions, the present invention has at least the following advantages and positive effects compared with the prior art: The 3D grating mounting self-learning error compensation method and system of the present invention, through the synergistic innovation of the same coordinate single camera hardware architecture and the three-level layered self-learning compensation algorithm, eliminates system errors from the source and achieves sub-micron level multi-source error suppression. While significantly reducing equipment costs, improving mass production yield and cycle efficiency, it breaks through the precision bottleneck of flip mounting, significantly improves 3D display performance, and has the ability to extend to advanced packaging scenarios of optical communication such as NPO / CPO optical engines. Attached Figure Description

[0013] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0014] Figure 1 This is a flowchart illustrating the 3D grating mounting self-learning error compensation method of the present invention. Figure 2 This is a schematic diagram of the process in step S1 of the present invention, which uses a single industrial camera set on a fixed reference to acquire Mark point images of the COG display panel and the 3D grating glass in the same visual coordinate system, and calculates the initial alignment deviation. Figure 3 This is a flowchart illustrating the mechanical error pre-compensation method in step S2 of the present invention. Figure 4 This is a flowchart illustrating the real-time glass warpage compensation method in step S2 of the present invention. Figure 5 This is a schematic diagram of the process in step S204 of the present invention, which utilizes the telecentric characteristics of the telecentric lenses on both sides of the same coordinate vision system to achieve online detection of warpage without any additional hardware through binocular stereo vision. Figure 6 This is a flowchart illustrating the process of step S205 in this invention, which involves integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, to achieve dual elimination of defocus positioning error and secondary offset in bonding. Figure 7 This is a flowchart illustrating the self-learning dynamic closed-loop compensation method in step S2 of the present invention. Figure 8This is a flowchart illustrating the process of adopting a batch learning and real-time correction closed-loop architecture in step S206 of the present invention. By introducing a forgetting factor, recent deviations are given high weight to adapt to the slow-changing characteristics of errors. On the basis of conventional batch statistical updates, a single-item large deviation instant triggering mechanism is superimposed to realize the dynamic adaptive optimization and anomaly interception of compensation parameters. Figure 9 This is a flowchart illustrating step S207 of the present invention, which calculates the compensation parameter correction amount based on the incremental PID discretization expression, tunes the optimal control coefficient through the critical proportional method to achieve rapid convergence of the fitting deviation, and uses integral limiting and dead zone mechanisms to effectively suppress integral saturation and frequent fine-tuning of the actuator. Figure 10 This is a flowchart illustrating the process of step S3 in this invention, which involves driving the robotic arm and motion platform to complete the bonding based on the compensated correction parameters, detecting the actual bonding deviation, and updating the compensation model parameters using the actual bonding deviation to form a closed-loop control. Figure 11 This is a schematic diagram of the structure of a 3D grating mounting self-learning error compensation system according to the present invention.

[0015] The annotations in the attached figures are explained as follows: 210. Coordinate-based visual positioning module; 220. Three-level hierarchical compensation algorithm module; 230. Fitting execution and iteration module. Detailed Implementation

[0016] To more clearly explain the purpose, technical solutions, and advantages of the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein. On the contrary, these embodiments are provided so that the present invention will be more comprehensive and complete, and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0017] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a full understanding of embodiments of the invention. However, those skilled in the art will recognize that the technical solutions of the invention can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of the invention.

[0018] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0019] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0020] The present invention will now be described in detail with reference to specific embodiments: Example 1

[0021] This invention provides a 3D grating mounting self-learning error compensation method, with reference to... Figure 1 , Figure 1 This is a flowchart illustrating the 3D grating mounting self-learning error compensation method of the present invention.

[0022] Depend on Figure 1 A known method for 3D grating mounting self-learning error compensation includes the following steps: S1. Using a single industrial camera set on a fixed reference, acquire Mark point images of the COG display panel and the 3D grating glass in the same visual coordinate system, and calculate the initial alignment deviation.

[0023] This device employs a single-camera, simultaneous imaging architecture. Through a single visual coordinate system, it digitally positions the grating and the workpiece to be bonded to the COG display within the same visual space, eliminating coordinate transformation and calibration errors inherent in traditional dual-camera solutions. The device consists of six core units: a frame support unit, a vision imaging unit, a workpiece support unit, a grating adsorption unit, and an electrical control unit. Their specific structures and functions are as follows: Rack support unit: Made of natural marble, with a flatness of ≤2μm / m and a coefficient of thermal expansion of ≤2×10. -6 / ℃ provides a high-rigidity and high-stability mounting reference for the vision system, isolating the effects of mechanical vibration and temperature deformation on positioning accuracy, and serves as the mounting reference platform for the entire device.

[0024] It should be noted that 3D mounting is extremely sensitive to vibration. Natural marble has extremely high internal damping characteristics, which can effectively absorb mechanical vibration and prevent visual image blurring. Factory ambient temperature may fluctuate, such as due to diurnal temperature variations. The thermal expansion coefficient data is calculated based on the temperature variation range of the working environment to ensure that the deformation of the frame itself will not exceed ±2μm during temperature changes, thus guaranteeing that the reference standard of the vision system mounted on it remains unchanged.

[0025] The vision imaging unit consists of a single 24-megapixel global shutter CMOS industrial camera and dual telecentric lenses, fixed to the top of the bracket. The lens optical axis is perpendicular to the plane of the workstation, forming the unique vision coordinate system for the entire system. The camera has a single pixel physical size of 3.45μm × 3.45μm and a frame rate of 35fps. The dual telecentric lenses have an optical magnification of 0.38×, a working distance of 120mm ± 10mm, optical distortion ≤ 0.015%, and an aperture range of F4-F16, which can eliminate perspective distortion and ensure constant magnification within the depth of field.

[0026] It should be noted that, to achieve sub-micron level positioning accuracy, based on the "4x pixel principle" (i.e., accuracy is typically 1 / 4 to 1 / 10 of the pixel size), combined with an optical magnification of 0.38×, a single pixel size of 3.45μm and a high resolution of 24 megapixels were derived. Because 3D gratings may be warped, ordinary lenses cause objects to appear larger in the foreground and smaller in the background. Using a dual-telecentric lens can eliminate perspective distortion and ensure a constant magnification within a certain depth of field. The Tenengrad gradient function is a classic algorithm used in image processing to evaluate sharpness; its data is determined through testing with a large number of image samples at different heights, and it is used to find the point where the image edge is sharpest.

[0027] Workpiece support unit: The core is a high-precision vacuum adsorption platform used to fix the COG display panel. It can achieve three-degree-of-freedom fine adjustment of XYR (X-axis translation / Y-axis translation / R-axis rotation). The platform flatness is ≤1μm / m. It is located in the field of view directly below the vision camera throughout the entire process, ensuring that the COG screen is always in a unified coordinate system.

[0028] It should be noted that if the COG display panel itself is not flat, it will cause the Mark point to go out of focus. The 1μm / m flatness is calculated based on the optical depth of field range to ensure that the entire panel is within the camera's clear imaging range. Since the robotic arm may have initial deviations when grasping the grating, the support unit needs to have fine-tuning capabilities. Its stroke and accuracy are determined based on the difference between the robotic arm's repeatability and placement accuracy.

[0029] Grating adsorption unit: It adopts a back-mounted non-contact vacuum suction head, with adsorption points only distributed on the edge of the non-functional area of ​​the grating glass. Adsorption and transportation can be completed from the back of the grating without contacting the functional surface of the grating cylindrical lens, thus avoiding scratching the microstructure. The flatness of the suction head is ≤1μm / m, and the adsorption deformation is ≤0.05mm, ensuring that the grating is not deformed during transportation, flipping, and shooting.

[0030] It should be noted that the 3D grating surface has a microlens structure, making it extremely fragile. Calculations based on the Young's modulus and maximum allowable stress of the grating glass ensure that the suction force is both stable and will not damage the microstructure or cause the central area to sag. A 1μm / m suction head flatness is matched to the support platform, ensuring that the grating and panel are parallel and aligned at the moment of contact, avoiding air bubbles or stress.

[0031] Electrical control unit: It adopts programmable logic controller (PLC) motion controller and integrated drive and control driver to realize high-speed synchronous control of camera, robot arm and adsorption platform. The modules communicate with each other through EtherCAT industrial real-time bus. The bus communication cycle is ≤1ms and the communication delay is ≤1ms, which ensures the synchronization of vision positioning and motion execution.

[0032] It's important to note that during high-speed placement, once the vision system detects the position, it must notify the robotic arm to stop or correct it within a very short time. The 1ms cycle is calculated based on the robotic arm's movement speed and allowable following error. If the communication is too slow, errors will occur in the robotic arm, leading to placement deviations.

[0033] refer to Figure 2 , Figure 2 This is a schematic diagram of the process in step S1 of the present invention, which involves using a single industrial camera set on a fixed reference to acquire Mark point images of the COG display panel and the 3D grating glass in the same visual coordinate system, and calculating the initial alignment deviation.

[0034] Using a single industrial camera set on a fixed reference, Mark point images of the COG display panel and the 3D grating glass are acquired in the same visual coordinate system, and the initial alignment deviation is calculated. The specific steps include: S101. Fix the COG display panel to be bonded onto the vacuum adsorption platform, with the platform located directly below the field of view of the vision camera; start the vision system, and the camera captures the Mark points at the two corners of the COG screen. Complete the digital positioning of the COG screen coordinates in the unique vision coordinate system and record the reference coordinate matrix MCOG of the COG screen Mark points.

[0035] S102. The robotic arm moves the adsorbed 3D grating glass to the same field of view between the vision camera and the COG screen, and adjusts the grating plane and the COG screen plane to be at the same object height; the vision camera captures the Mark points at the two corners of the grating, completes the digitization of the grating coordinates in the same vision coordinate system, and records the coordinate matrix MGR of the grating Mark points.

[0036] S103. Based on two sets of coordinate matrices under the same coordinate system, the alignment deviation between the grating and the COG screen is calculated directly using the least squares method alignment matching function. This includes the translation deviations Δx and Δy of the X and Y axes, as well as the rotation angle deviation Δθ. There is no coordinate transformation step, thus eliminating system transformation error at its source.

[0037] S104. The control system synchronously sends the deviation data to the robot and the XYR platform. After the robot completes the attitude correction of the grating, it moves precisely to the top of the COG screen to complete the pre-lamination and vacuum full lamination process.

[0038] S2. Construct a three-level hierarchical error decoupling and collaborative control architecture. Through mechanical error pre-compensation method, glass warpage real-time compensation method and self-learning dynamic closed-loop compensation method, realize full-link dynamic suppression and steady-state accuracy maintenance of multi-source coupling errors.

[0039] Based on the hardware of the same coordinate vision positioning device, this invention constructs a three-level hierarchical compensation framework for three core error sources: repeatable errors of mechanical systems, random errors of glass warping, and slow-varying errors in mass production. Through the hierarchical control logic of feedforward correction, online correction, and dynamic optimization, the positioning error is suppressed throughout the entire process.

[0040] refer to Figure 3 , Figure 3 This is a flowchart illustrating the mechanical error pre-compensation method in step S2 of the present invention.

[0041] The mechanical error pre-compensation method specifically includes the following steps: S201. Using the improved DH homogeneous transformation matrix method, a kinematic error model of the robot and the XYR platform is established to quantify the impact of various error terms such as joint rotation angle error, link length error, guide rail straightness error, and suction head perpendicularity error on the end-mounting accuracy, and to clarify the contribution and transmission law of each error source.

[0042] S202. Obtain the precise values ​​of each error term of the mechanical system through error calibration, calculate the predicted value of the end pose error based on the error Jacobian matrix, and add the reverse compensation amount to the motion control command to offset the inherent error of the mechanical system.

[0043] Let the pose of the end target be X.d Due to mechanical errors, the actual output pose is X. a =X d +ΔX, where ΔX is the mechanical pose error; the pre-compensation algorithm solves for the corrected joint control quantity θ. c =θ d +Δθ, so that the corrected actual pose X a1 =X d ,Right now: ,

[0044] The formula for calculating joint compensation is derived as follows: ,

[0045] Among them, J + Let be the Moore-Penrose generalized inverse of the error Jacobian matrix, and Δx, Δy, and Δθ be the measured planar pose errors.

[0046] S203. Based on the nine-point calibration and measured data from the laser interferometer, the error Jacobian matrix is ​​identified using the least squares method, and a pre-compensation parameter table is generated to achieve efficient feedforward correction. Using the nine-point calibration method, nine evenly distributed calibration points are selected within the robot's workspace. The theoretical and actual coordinates of each point are collected using a laser interferometer. The error Jacobian matrix and the values ​​of each error term are identified using the least squares method, generating a pre-compensation parameter table for the entire workspace, which is then written into the PLC controller's real-time database. During actual operation, the controller retrieves the pre-compensation parameters from the table in real time based on the target position.

[0047] In constructing the kinematic error model, this system adopts a two-stage modeling strategy of constructing an ideal model and introducing error elements. First, based on the improved DH parameter method, an ideal forward kinematic model between each joint of the device is established using the homogeneous transformation matrix (SE3). Second, considering the physical deviations in actual manufacturing and assembly, the standard homogeneous coordinate transformation method and helical quantity theory are introduced to transform geometric errors and non-rigid body errors caused by thermal / mechanical coupling into mathematical language. On this basis, by constructing the error Jacobian matrix, the small error elements of each independent motion axis at the bottom layer are linearly mapped, and finally the spatial comprehensive pose error of the end effector is accurately solved, providing a solid mathematical foundation for high-precision reverse compensation.

[0048] refer to Figure 4 , Figure 4 This is a flowchart illustrating the real-time glass warpage compensation method in step S2 of the present invention.

[0049] The real-time compensation method for glass warpage includes the following steps: S204. Utilizing the telecentric characteristics of the dual telecentric lenses of the same coordinate vision system, online detection of warpage without additional hardware is achieved through binocular stereo vision.

[0050] refer to Figure 5 , Figure 5 This is a schematic diagram of the process in step S204 of the present invention, which utilizes the telecentric characteristics of the telecentric lenses on both sides of the same coordinate vision system to achieve online detection of warpage without any additional hardware through binocular stereo vision.

[0051] By utilizing the telecentric characteristics of bilateral telecentric lenses in a co-coordinate vision system, online detection of warpage without additional hardware is achieved through binocular stereo vision. The specific steps include: S2041. The vision system focuses on the upper and lower surfaces of the glass respectively, records the height difference between the two focusing positions, and obtains the glass thickness reference value.

[0052] S2042. Perform layered focusing on the Mark point areas at the two corners of the glass, calculate the optimal focusing height for each Mark point using the Tenengrad gradient function, and obtain the actual height values ​​of the four corner points.

[0053] S2043. Based on the height values ​​of two corner points, the warping deformation surface of the entire glass is restored by fitting a quadratic surface to obtain the warping distribution, detection time, and detection accuracy of the entire surface.

[0054] S205 achieves dual elimination of defocus positioning error and secondary offset by integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation of in-plane slip based on thin plate bending theory.

[0055] refer to Figure 6 , Figure 6 This is a schematic diagram of the process in step S205 of the present invention, which achieves dual elimination of defocus positioning error and secondary offset by integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation of in-plane slip based on thin plate bending theory.

[0056] By integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, dual elimination of defocus positioning error and secondary offset in bonding is achieved. Specifically, the following steps are included: S2051. Based on the actual height deviation of the Mark point, and combining the depth-of-field model and lens distortion model of the visual system, calculate the changes in imaging magnification and edge distortion caused by the height deviation, correct the pixel coordinates of the Mark point, and eliminate the positioning error caused by defocus.

[0057] S2052. Based on the thin plate bending model, calculate the in-plane slippage caused by glass warping and flattening during vacuum adsorption bonding, and compensate the pre-alignment coordinates of the grating in advance to avoid secondary offset after bonding.

[0058] refer to Figure 7 , Figure 7 This is a flowchart illustrating the self-learning dynamic closed-loop compensation method in step S2 of this invention.

[0059] The self-learning dynamic closed-loop compensation method includes the following steps: S206. It adopts a batch learning and real-time correction closed-loop architecture. By introducing a forgetting factor, it assigns high weight to recent deviations to adapt to the slow-changing characteristics of errors. On the basis of regular batch statistical updates, it superimposes a single-item large deviation instant triggering mechanism to realize dynamic adaptive optimization of compensation parameters and anomaly interception.

[0060] refer to Figure 8 , Figure 8 This is a flowchart illustrating the process of adopting a batch learning and real-time correction closed-loop architecture in step S206 of the present invention. By introducing a forgetting factor to give high weight to recent deviations to adapt to the slow-changing characteristics of errors, and by superimposing a single-item large deviation instant triggering mechanism on the basis of conventional batch statistical updates, the dynamic adaptive optimization and anomaly interception of compensation parameters are realized.

[0061] A closed-loop architecture of batch learning and real-time correction is adopted. By introducing a forgetting factor to give high weight to recent deviations to adapt to the slow-changing characteristics of errors, and on the basis of regular batch statistical updates, an instant triggering mechanism for large single-item deviations is superimposed to achieve dynamic adaptive optimization of compensation parameters and anomaly interception. Specifically, the following steps are included: S2061. After every 8-12 pieces of product are bonded, the vision system automatically detects the actual bonding deviation and calculates the average deviation.

[0062] S2062. Using the average deviation as feedback input, the incremental PID controller calculates the correction amount of the compensation parameters and updates the pre-compensation parameter table.

[0063] S2063. For large deviations in a single product, immediately trigger a single-time compensation correction to avoid batch defects.

[0064] S2064. Introduce a forgetting factor to give higher weight to recent data, adapt to the slow-changing characteristics of error, and avoid the interference of historical data on the current compensation effect.

[0065] It should be noted that in this embodiment, the forgetting factor is set to 0.95. From the perspective of algorithm mechanism, the forgetting factor determines the length of the model's memory of historical data. If the forgetting factor is too large, the algorithm will rely too much on historical data, causing the system to respond slowly to real changes such as temperature drift or mechanical wear, resulting in data saturation. Conversely, if the forgetting factor is too small, the system will be overly sensitive to random disturbances or measurement noise of individual products, causing severe fluctuations in compensation parameters. Setting the forgetting factor to 0.95 means that the algorithm assigns a 5% correction weight to recent information data in each iteration update, while retaining 95% of the effective historical memory. On the one hand, it can effectively filter out occasional measurement noise, prevent the actuator from frequently fine-tuning due to small deviations, and ensure the stability of the mass production process. On the other hand, it gives the model sufficient sensitivity, enabling it to continuously track and offset the cumulative errors caused by environmental temperature rise or equipment aging. Therefore, a forgetting factor of 0.95 avoids overfitting the model to the initial offline calibration parameters while ensuring dynamic adaptive optimization of the compensation parameters in long-term operation, making it the best balance point for achieving high precision and high stability in mass production.

[0066] S207. The compensation parameter correction amount is calculated based on the incremental PID discretization expression. The optimal control coefficient is tuned by the critical proportional method to achieve rapid convergence of the fitting deviation. The integral limiting and dead zone mechanism is used to effectively suppress integral saturation and frequent fine-tuning of the actuator.

[0067] refer to Figure 9 , Figure 9 This is a flowchart illustrating step S207 of the present invention, which involves calculating the compensation parameter correction amount based on the incremental PID discretization expression, tuning the optimal control coefficient through the critical proportional gain method to achieve rapid convergence of the fitting deviation, and supplementing it with integral limiting and dead zone mechanisms to effectively suppress integral saturation and frequent fine-tuning of the actuator.

[0068] The compensation parameter correction is calculated based on the incremental PID discretization expression. The optimal control coefficient is tuned using the critical proportional gain method to achieve rapid convergence of the fit deviation. Integral limiting and dead zone mechanisms are used to effectively suppress integral saturation and frequent fine-tuning of the actuator. The specific steps include: S2071. The incremental PID control algorithm is used to calculate the correction amount of the compensation parameters.

[0069] S2072. The critical proportional method is used to tune the controller parameters and determine the optimal parameter combination so that the system can converge the fitting deviation to the target range within 3 learning cycles. At the same time, an integral limiting and dead zone mechanism is introduced to effectively avoid frequent small adjustments of the actuator due to integral depth saturation.

[0070] S3. Based on the compensated correction parameters, drive the robot arm and motion platform to complete the bonding, detect the actual bonding deviation, and use the actual bonding deviation to update the compensation model parameters to form a closed-loop control.

[0071] refer to Figure 10 , Figure 10 This is a flowchart illustrating the process of step S3 in this invention, which involves driving the robotic arm and motion platform to complete the bonding based on the compensated correction parameters, detecting the actual bonding deviation, and updating the compensation model parameters using the actual bonding deviation to form a closed-loop control.

[0072] The robot arm and motion platform are driven to complete the bonding based on the compensated correction parameters, and the actual bonding deviation is detected. The compensation model parameters are updated using the actual bonding deviation to form a closed-loop control, which specifically includes the following steps: S301. When the device is powered on, the vision system calibration and mechanical system error modeling are constructed, a pre-compensation parameter table is generated and written to the controller, and hardware self-test and origin regression are completed.

[0073] S302. During mass production, the motion error of the mechanical system is corrected by a pre-compensation algorithm, and the corrected motion control parameters are output to the robot and the XYR platform.

[0074] S303. By using a single camera to capture images in the same position, the coordinates of the Mark points of the COG display screen and the 3D grating are acquired in the same coordinate system, and the initial alignment deviation is calculated.

[0075] S304. The vision system acquires glass warping data online, calculates the positioning deviation caused by warping through a real-time compensation algorithm, performs secondary correction on the Mark point coordinates, and obtains the final alignment correction parameters.

[0076] S305. Based on the final corrected parameters, the control system drives the robotic arm to complete the alignment and vacuum bonding with the XYR platform.

[0077] S306. After completing a single bonding process, the vision system detects the actual bonding deviation and updates the compensation parameters through a self-learning algorithm to achieve dynamic closed-loop optimization.

[0078] S307. Repeat the above process to continuously optimize positioning accuracy during continuous production.

[0079] In summary, this application achieves sub-micron level multi-source error suppression by synergistic innovation of co-coordinate single-camera hardware architecture and three-level hierarchical self-learning compensation algorithm, thereby eliminating system errors at the source. This significantly reduces equipment costs, improves mass production yield and cycle time efficiency, breaks through the precision bottleneck of flip-chip mounting, significantly improves 3D display performance, and has the ability to extend to high-precision scenarios such as CPO light engines.

[0080] Example 2

[0081] refer to Figure 11 This invention provides a 3D grating mounting self-learning error compensation system. Figure 11 This is a schematic diagram of the structure of a 3D grating mounting self-learning error compensation system according to the present invention.

[0082] Depend on Figure 11 A 3D grating mounting self-learning error compensation system is known to include: a coordinate visual positioning module 210, a three-level hierarchical compensation algorithm module 220, and a bonding execution and iteration module 230.

[0083] The co-coordinate vision positioning module 210 is used to acquire Mark point images of the COG display panel and the 3D grating glass respectively in the same vision coordinate system using a single industrial camera set on a fixed reference, and calculate the initial alignment deviation.

[0084] The three-level hierarchical compensation algorithm module 220 is used to construct a three-level hierarchical error decoupling and collaborative control architecture. Through mechanical error pre-compensation method, glass warping real-time compensation method and self-learning dynamic closed-loop compensation method, it realizes full-link dynamic suppression and steady-state accuracy maintenance of multi-source coupling errors.

[0085] The bonding execution and iteration module 230 is used to drive the robot and motion platform to complete bonding based on the compensated correction parameters, detect the actual bonding deviation, and update the compensation model parameters using the actual bonding deviation to form closed-loop control.

[0086] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the claims. It should be understood that the invention is not limited to the precise structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A method for self-learning error compensation in 3D grating mounting, characterized in that, include: Using a single industrial camera set on a fixed reference, Mark point images of the COG display panel and 3D grating glass are acquired in the same visual coordinate system, and the initial alignment deviation is calculated. A three-level hierarchical error decoupling and collaborative control architecture is constructed. Through mechanical error pre-compensation method, glass warpage real-time compensation method and self-learning dynamic closed-loop compensation method, the full-link dynamic suppression and steady-state accuracy maintenance of multi-source coupling errors are achieved. Based on the compensated correction parameters, the robot arm and motion platform are driven to complete the bonding, and the actual bonding deviation is detected. The actual bonding deviation is used to update the compensation model parameters to form a closed-loop control. The mechanical error pre-compensation method includes: An improved DH homogeneous transformation matrix method is used to establish a kinematic error model between the manipulator and the XYR platform; The precise values ​​of each error term of the mechanical system are obtained through error calibration. The predicted value of the end pose error is calculated based on the error Jacobian matrix. The reverse compensation amount is added to the motion control command to offset the inherent error of the mechanical system. Based on the nine-point calibration and the measured data of the laser interferometer, the least squares method is used to identify the error Jacobian matrix and generate a pre-compensation parameter table to achieve efficient feedforward correction. The real-time glass warpage compensation method includes: By utilizing the telecentric characteristics of the bilateral telecentric lenses of the same coordinate vision system, online detection of warpage without any additional hardware is achieved through binocular stereo vision. By integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, the dual elimination of defocus positioning error and secondary offset of bonding is achieved. The self-learning dynamic closed-loop compensation method includes: A batch learning and real-time correction closed-loop architecture is adopted. By introducing a forgetting factor, recent deviations are given high weight to adapt to the slow-changing characteristics of errors. On the basis of regular batch statistical updates, a single-item large deviation instant triggering mechanism is superimposed to achieve dynamic adaptive optimization of compensation parameters and anomaly interception. The compensation parameter correction is calculated based on the incremental PID discretization expression. The optimal control coefficient is tuned by the critical proportional method to achieve rapid convergence of the fitting deviation. In addition, integral limiting and dead zone mechanisms are used to effectively suppress integral saturation and frequent fine-tuning of the actuator.

2. The 3D grating mounting self-learning error compensation method according to claim 1, characterized in that, The method utilizes a single industrial camera mounted on a fixed reference to acquire Mark point images of the COG display panel and the 3D grating glass in the same visual coordinate system, and calculates the initial alignment deviation, including: The COG display panel to be bonded is fixed on the vacuum adsorption platform, which is located directly below the field of view of the vision camera; the vision system is activated, and the camera captures the Mark points at the two corners of the COG screen. The coordinates of the COG screen are digitally positioned in a unique vision coordinate system, and the reference coordinate matrix MCOG of the COG screen Mark points is recorded. The robotic arm moves the adsorbed 3D grating glass to the same field of view between the vision camera and the COG screen, and adjusts the grating plane and the COG screen plane to be at the same object height; the vision camera captures the Mark points at the two corners of the grating, completes the digitization of the grating coordinates in the same vision coordinate system, and records the coordinate matrix MGR of the grating Mark points. Based on two sets of coordinate matrices under the same coordinate system, the alignment deviation between the grating and the COG screen is calculated directly using the least squares method alignment matching function. This includes the translational deviations Δx and Δy of the X and Y axes, as well as the rotation angle deviation Δθ. There is no coordinate transformation step, thus eliminating system transformation error at its source. The control system synchronously sends the deviation data to the robot and the XYR platform. After the robot completes the attitude correction of the grating, it moves precisely to the top of the COG screen to complete the pre-lamination and vacuum full lamination process.

3. The 3D grating mounting self-learning error compensation method according to claim 1, characterized in that, The method of utilizing the telecentric characteristics of bilateral telecentric lenses in a co-coordinate vision system to achieve online detection of warpage without additional hardware through binocular stereo vision includes: The vision system focuses on the upper and lower surfaces of the glass respectively, records the height difference between the two focusing positions, and obtains the glass thickness reference value. The Mark point areas at the two corners of the glass are focused in layers. The optimal focusing height of each Mark point is calculated using the Tenengrad gradient function to obtain the actual height values ​​of the four corner points. Based on the height values ​​of two corner points, the warping deformation surface of the entire glass is restored by fitting a quadratic surface, thereby obtaining the warping distribution, detection time, and detection accuracy of the entire surface.

4. The 3D grating mounting self-learning error compensation method according to claim 1, characterized in that, The method achieves dual elimination of defocus positioning error and secondary offset by integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory. This includes: Based on the actual height deviation of the Mark point, combined with the visual system depth model and lens distortion model, the changes in imaging magnification and edge distortion caused by the height deviation are calculated, and the pixel coordinates of the Mark point are corrected to eliminate the positioning error caused by defocus. Based on the thin plate bending model, the in-plane slippage caused by glass warping and flattening during vacuum adsorption bonding is calculated, and the pre-alignment coordinates of the grating are compensated in advance.

5. The 3D grating mounting self-learning error compensation method according to claim 1, characterized in that, The aforementioned closed-loop architecture employs batch learning and real-time correction. It introduces a forgetting factor to assign high weight to recent deviations to adapt to the slow-changing nature of errors. Furthermore, it superimposes a single-item large deviation instant triggering mechanism on top of conventional batch statistical updates, achieving dynamic adaptive optimization of compensation parameters and anomaly interception. This includes: After every 8-12 pieces of product are bonded, the vision system automatically detects the actual bonding deviation and calculates the average deviation. The average deviation is used as feedback input, and the correction amount of the compensation parameter is calculated by the incremental PID controller to update the pre-compensation parameter table. For large deviations in a single product, a single compensation correction is immediately triggered to avoid batch defects. By introducing a forgetting factor, more weight is given to recent data, which adapts to the slow variation of error and avoids the interference of historical data on the current compensation effect.

6. The 3D grating mounting self-learning error compensation method according to claim 1, characterized in that, The process involves driving the robotic arm and motion platform to complete the bonding based on the compensated correction parameters, detecting the actual bonding deviation, and updating the compensation model parameters using the actual bonding deviation to form a closed-loop control, including: S301. Power on the device, construct vision system calibration and mechanical system error modeling, generate pre-compensation parameter table and write it into the controller, and complete hardware self-test and origin regression. S302. During mass production, the motion error of the mechanical system is corrected by the pre-compensation algorithm, and the corrected motion control parameters are output to the robot and the XYR platform. S303. By using a single camera to take pictures in the same position, the coordinates of the Mark points of the COG display screen and the 3D grating are collected in the same coordinate system, and the initial alignment deviation is calculated. S304. The vision system acquires glass warping data online, calculates the positioning deviation caused by warping through a real-time compensation algorithm, performs secondary correction on the Mark point coordinates, and obtains the final alignment correction parameters. S305. The control system drives the robotic arm to complete alignment and vacuum bonding with the XYR platform according to the final corrected parameters. S306. After completing a single bonding process, the vision system detects the actual bonding deviation and updates the compensation parameters through a self-learning algorithm to achieve dynamic closed-loop optimization. S307. Repeat steps S301-S306 to achieve continuous optimization of positioning accuracy during continuous production.

7. A 3D grating mounting self-learning error compensation system, characterized in that, include: The same-coordinate vision positioning module is used to acquire Mark point images of the COG display panel and the 3D grating glass respectively in the same vision coordinate system using a single industrial camera set on a fixed reference, and calculate the initial alignment deviation. The three-level hierarchical compensation algorithm module is used to construct a three-level hierarchical error decoupling and collaborative control architecture. Through mechanical error pre-compensation method, glass warping real-time compensation method and self-learning dynamic closed-loop compensation method, it realizes full-link dynamic suppression and steady-state accuracy maintenance of multi-source coupling error. The bonding execution and iteration module is used to drive the robot and motion platform to complete bonding based on the compensated correction parameters, detect the actual bonding deviation, and update the compensation model parameters using the actual bonding deviation to form closed-loop control. The mechanical error pre-compensation method includes: An improved DH homogeneous transformation matrix method is used to establish a kinematic error model between the manipulator and the XYR platform; The precise values ​​of each error term of the mechanical system are obtained through error calibration. The predicted value of the end pose error is calculated based on the error Jacobian matrix. The reverse compensation amount is added to the motion control command to offset the inherent error of the mechanical system. Based on the nine-point calibration and the measured data of the laser interferometer, the least squares method is used to identify the error Jacobian matrix and generate a pre-compensation parameter table to achieve efficient feedforward correction. The real-time glass warpage compensation method includes: By utilizing the telecentric characteristics of the bilateral telecentric lenses of the same coordinate vision system, online detection of warpage without any additional hardware is achieved through binocular stereo vision. By integrating imaging magnification correction based on depth of field and lens distortion characteristics, and pre-compensation for in-plane slip based on thin plate bending theory, the dual elimination of defocus positioning error and secondary offset of bonding is achieved. The self-learning dynamic closed-loop compensation method includes: A batch learning and real-time correction closed-loop architecture is adopted. By introducing a forgetting factor, recent deviations are given high weight to adapt to the slow-changing characteristics of errors. On the basis of regular batch statistical updates, a single-item large deviation instant triggering mechanism is superimposed to achieve dynamic adaptive optimization of compensation parameters and anomaly interception. The compensation parameter correction is calculated based on the incremental PID discretization expression. The optimal control coefficient is tuned by the critical proportional method to achieve rapid convergence of the fitting deviation. In addition, integral limiting and dead zone mechanisms are used to effectively suppress integral saturation and frequent fine-tuning of the actuator.

Images