A curved surface film covering flatness self-adaptive control method based on machine vision

Abstract

This application discloses a machine vision-based adaptive control method for the flatness of curved surface coating, relating to the field of robotic arm control technology. It is applied to a curved surface coating control system, which includes a machine vision perception module, a robot end effector, a composite controller, and a distributed gripper array. The method includes: acquiring a stripe projection image sequence and dynamic trajectory data of the end effector; performing phase calculation and 3D reconstruction on the image sequence to construct a digital elevation model and calculate flatness error data; inputting the dynamic trajectory data into a feedforward controller to generate compensation instructions, and inputting the flatness error data into a feedback controller to generate correction instructions, then superimposing them to obtain a comprehensive correction instruction; acquiring the strain sensitivity matrix, and using pseudo-inverse operations to allocate the comprehensive correction instruction to the displacement execution parameters of each gripper and issuing it for execution. This invention achieves real-time perception and active intervention of dynamic deformation, significantly improving the yield and batch consistency of curved surface coating.

Description

Technical Field

[0001] This invention relates to the field of robotic arm control technology, and more specifically to a machine vision-based adaptive control method for the flatness of curved surface coating. Background Technology

[0002] In recent years, the cross-integration of flexible electronics and embodied robots has continued to deepen, driving the evolution of robot body forms towards biomimetic flexibility and giving rise to a number of emerging applications such as electronic skin, adhesive biosensor arrays, and seamless human-machine interfaces. The core requirement of these applications is that the electronic functional layer must dynamically conform to the robot's curved surface and maintain stable electromechanical response during continuous movement. Therefore, the precision and reliability of the curved surface coating process directly determine the performance and lifespan of the final product.

[0003] However, when a robot dynamically performs a surface coating task, the film is subjected to the superposition of multiple complex stresses: the inertial and acceleration effects from the end effector motion cause wide fluctuations in the overall tension distribution of the film; the continuous change in substrate curvature causes local bending stress to evolve in real time; and the adhesive force between the film and the substrate generates transient interfacial shear during dynamic application. These factors are coupled with each other, forming a multimodal transient deformation field inside the film that encompasses tension, compression, and shear, and this deformation process has the triple characteristics of strong nonlinearity, fast time-varying, and spatial non-uniformity.

[0004] To address the aforementioned challenges, existing technologies primarily employ a static or quasi-static control strategy of "pause-measurement-adjustment." These methods typically rely on pausing the process after the robot reaches a preset pause point, using a discrete-point laser rangefinder or offline vision system to measure the height or shape of the film's fixed points, calculating the deviation, adjusting the fixture once, and then continuing operation. This model treats visual detection and execution control as two independent decision domains, resulting in control commands lagging significantly behind the evolution of dynamic deformation. Its control effectiveness largely depends on the selection of the pause position and random factors such as the measurement of transient states, making it difficult to achieve real-time tracking and proactive intervention of full-field dynamic deformation under high-speed continuous motion, and even more difficult to guarantee stable reproducibility between batches. Therefore, it is difficult to meet the integrated control requirements of high precision and high reliability for complex curved flexible coatings. Summary of the Invention

[0005] The purpose of this invention is to provide a machine vision-based adaptive control method for the flatness of curved surface coating. This method deeply integrates visual perception and motion information by simultaneously acquiring stripe projection image sequences and dynamic trajectory data of the end effector. A feedforward controller is used to preemptively offset predictable inertial and curvature variation disturbances, while a feedback controller corrects unmodeled residual errors in real time. The synergy of these two methods significantly improves dynamic response speed and control accuracy. Furthermore, pseudo-inverse operations are used to optimally distribute comprehensive correction commands to a distributed gripper array, achieving refined stress compensation through multi-degree-of-freedom collaboration. Therefore, this invention significantly improves the yield and batch consistency of curved surface coating.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a machine vision-based adaptive control method for the flatness of curved surface coating, applied to a curved surface coating control system. The system includes a machine vision perception module, a robot end effector, a composite controller, and a distributed gripper array. The method includes:

[0008] The stripe projection image sequence of the film-coating operation area collected by the machine vision perception module is acquired, and the dynamic trajectory data of the robot end effector is acquired simultaneously.

[0009] Phase calculation and 3D reconstruction are performed on the stripe projection image sequence to construct a digital elevation model characterizing the surface morphology of the thin film.

[0010] Calculation of flatness error data based on digital elevation model;

[0011] The dynamic trajectory data is input into the feedforward controller in the composite controller to generate compensation commands to counteract predictable disturbances.

[0012] The flatness error data is input into the feedback controller in the composite controller to generate correction instructions for correcting unmodeled residual errors;

[0013] By superimposing compensation and correction commands, a comprehensive correction command is generated.

[0014] Obtain the strain sensitivity matrix of the distributed gripper array, and use pseudo-inverse operation to allocate the comprehensive correction command as the displacement execution parameters of each gripper;

[0015] The displacement execution parameters are converted into coordinated control commands and sent to the servo controllers of each gripper for execution.

[0016] In some embodiments, phase calculation and three-dimensional reconstruction are performed on the stripe projection image sequence to construct a digital elevation model characterizing the surface morphology of the thin film, including:

[0017] Filtering and background correction are performed on multiple images in a fringe projection image sequence; the fringe projection image sequence includes multiple consecutive fringe projection images with sequentially shifted phases;

[0018] A four-step phase-shifting algorithm is used to extract the folded phase distribution at each pixel;

[0019] The phase unfolding algorithm is used to restore the folded phase distribution to the absolute phase distribution;

[0020] Based on pre-calibrated system geometric parameters, the absolute phase distribution is mapped to the height variation field of the thin film relative to the reference surface;

[0021] A digital elevation model is generated based on the height variation field.

[0022] In some embodiments, the formula for extracting the folded phase distribution at each pixel using a four-step phase-shifting algorithm is as follows:

[0023] ;

[0024] in, I1 represents the folded phase at coordinates (x, y), and I4 represents the light intensity values ​​of the four consecutive fringe projection images at the corresponding pixel points.

[0025] In some embodiments, calculating flatness error data based on a digital elevation model includes:

[0026] Based on the digital elevation model, the global root mean square error of the relative mean of the height deviation of the thin film surface is calculated.

[0027] Calculate the maximum surface fluctuation amplitude in the digital elevation model and generate local peak-valley errors.

[0028] In some embodiments, the feedforward controller is constructed based on a first-order inertial plus pure delay model of the thin-film planarization process, and the transfer function of the first-order inertial plus pure delay model is:

[0029] ;

[0030] in, For process gain, Let θ be the time constant, θ be the pure delay time, and s be the Laplace operator;

[0031] The feedforward controller uses the dynamic trajectory data and the look-ahead steps determined by the pure delay time to generate compensation commands.

[0032] In some embodiments, the method further includes:

[0033] A recursive least squares method with a forgetting factor is used to update the perturbation matrix in the first-order inertia plus pure delay model online.

[0034] The update law formula for the perturbation matrix is:

[0035] ;

[0036] ;

[0037] ;

[0038] Where λ is the forgetting factor, and d(k) is a measurable perturbation vector composed of dynamic trajectory data. This is the estimated value of the perturbation matrix at time k.

[0039] In some embodiments, the strain sensitivity matrix of the distributed gripper array is obtained, and the comprehensive correction command is allocated as displacement execution parameters for each gripper using pseudo-inverse operation, including:

[0040] The recursive least squares method is used to identify and update the strain sensitivity matrix that represents the mapping relationship between the small displacement of the gripper and the global flatness online;

[0041] The Mohr-Penrose generalized inverse matrix is ​​obtained from the strain sensitivity matrix;

[0042] The generalized inverse matrix is ​​multiplied by the comprehensive correction command to obtain the optimal adjustment displacement vector that minimizes the sum of the squares of the displacements of each gripper, and this vector is used as the displacement execution parameter.

[0043] In some embodiments, the feedback controller employs an incremental digital PID control law, and its gain scheduling rules include:

[0044] The proportional gain increases non-linearly with the increase of the flatness error data.

[0045] The integral gain is automatically reduced when the compensation command has a significant effect.

[0046] The differential gain is adjusted in real time based on the sign of the product of the flatness error data and its rate of change.

[0047] Secondly, the present invention also provides a machine vision-based adaptive control device for the flatness of curved surface coating, the device comprising:

[0048] The data acquisition module is used to acquire the stripe projection image sequence of the film-coating operation area collected by the machine vision perception module, and simultaneously acquire the dynamic trajectory data of the robot end effector;

[0049] The model building module is used to perform phase calculation and three-dimensional reconstruction on the stripe projection image sequence to build a digital elevation model that characterizes the surface morphology of the thin film.

[0050] The error calculation module is used to calculate flatness error data based on the digital elevation model.

[0051] The first instruction module is used to input dynamic trajectory data into the feedforward controller in the composite controller and generate compensation instructions to counteract predictable disturbances.

[0052] The second instruction module is used to input the flatness error data into the feedback controller in the composite controller and generate correction instructions for correcting unmodeled residual errors.

[0053] The instruction overlay module is used to overlay compensation instructions and correction instructions to generate a comprehensive correction instruction;

[0054] The instruction allocation module is used to obtain the strain sensitivity matrix of the distributed gripper array and use pseudo-inverse operation to allocate the comprehensive correction instruction as the displacement execution parameter of each gripper.

[0055] The instruction execution module is used to convert displacement execution parameters into coordinated control instructions and send them to the servo controllers of each gripper for execution.

[0056] Thirdly, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the machine vision-based adaptive control method for the flatness of curved surface coating provided in the first aspect.

[0057] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the machine vision-based adaptive control method for the flatness of curved surface coating provided in the first aspect.

[0058] Fifthly, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the machine vision-based adaptive control method for the flatness of curved surface coating provided in the first aspect.

[0059] The beneficial effects of this invention are as follows: By simultaneously acquiring stripe projection image sequences and dynamic trajectory data of the end effector, this invention deeply integrates visual perception and motion information; it utilizes a feedforward controller to preemptively offset predictable inertial and curvature change disturbances, while a feedback controller corrects unmodeled residual errors in real time; the synergy of these two significantly improves dynamic response speed and control accuracy; furthermore, through pseudo-inverse operations, it optimally distributes comprehensive correction commands to a distributed gripper array, achieving refined stress compensation through multi-degree-of-freedom collaboration. Therefore, this invention significantly improves the yield and batch consistency of curved surface coating.

[0060] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, the preferred embodiments of the present invention are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0061] Figure 1 This is a flowchart illustrating an adaptive control method for the flatness of curved surface coating based on machine vision, according to an embodiment of the present invention.

[0062] Figure 2 This is a schematic diagram of the structure of a machine vision-based adaptive control device for the flatness of curved surface coating, according to an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram of an electronic device structure provided in an embodiment of this application. Detailed Implementation

[0064] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] It should be noted that references to "an embodiment," "embodiment," "example embodiment," etc., in this specification refer to the described embodiment including specific features, structures, or characteristics; however, not every embodiment must include these specific features, structures, or characteristics. Furthermore, such expressions do not refer to the same embodiment. Moreover, when describing specific features, structures, or characteristics in conjunction with embodiments, whether or not explicitly described, it is indicated that incorporating such features, structures, or characteristics into other embodiments is within the knowledge of those skilled in the art.

[0066] Furthermore, the technical features involved in the different embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0067] In some embodiments, such as Figure 1As shown, a machine vision-based adaptive control method for the flatness of curved surface coating is provided. This method is applied to a curved surface coating control system, which includes a machine vision perception module, a robot end effector, a composite controller, and a distributed gripper array. The machine vision perception module is used to acquire image data of the coating operation area in real time. Its core hardware includes a digital light processing projector and a high-speed monocular industrial camera. The projector projects coded structured light stripes onto the film surface, and the camera captures the stripe images distorted by the film's undulations at a high frame rate. The robot end effector is the end device of the robotic arm that carries the film and performs the bonding action. Its dynamic trajectory data is synchronously acquired through a hard real-time industrial bus. The composite controller is a decision unit that integrates feedforward prediction and feedback correction functions to generate comprehensive correction commands to suppress film deformation. The distributed gripper array consists of multiple sets of three-degree-of-freedom precision electric grippers, used to apply controllable displacement compensation to the film edge. Specific methods include:

[0068] S101, acquire the stripe projection image sequence of the film-coating operation area collected by the machine vision perception module, and simultaneously acquire the dynamic trajectory data of the robot end effector.

[0069] Specifically, the system acquires a sequence of stripe projection images of the coating area from the machine vision perception module, and simultaneously acquires dynamic trajectory data of the robot's end effector. The stripe projection image sequence refers to a series of images containing structured light stripe patterns continuously acquired by a high-speed industrial camera. These images record the light stripe distortion information caused by deformation of the thin film surface. The dynamic trajectory data refers to the state information of the robot's end effector during movement, such as pose, velocity, and acceleration. This data is acquired via an industrial real-time bus and driven by a unified precision clock, ensuring that the end-to-end closed-loop delay is strictly controlled within 100 milliseconds. By synchronously acquiring these two types of data, a spatiotemporally aligned information foundation is provided for subsequent deformation perception and disturbance prediction, thus solving the lag problem caused by the asynchrony between perception and control data in existing technologies.

[0070] Optionally, before acquiring the stripe projection image sequence of the coating operation area collected by the machine vision perception module, the method of this application further includes: acquiring the historical calibration residuals of the machine vision perception module, generating perception confidence data based on the statistical distribution of the historical calibration residuals; automatically adjusting the sampling frame rate and phase stepping strategy of the stripe projection image sequence based on the perception confidence data, generating enhanced perception reference data containing redundant perception reserves, so that when acquiring the stripe projection image sequence of the coating operation area collected by the machine vision perception module, the acquisition can be performed based on the enhanced perception reference data.

[0071] Specifically, after initialization, the system uses its built-in data mining engine to extract historical calibration residuals of the sensing module over a preset time period (e.g., the past 30 work shifts). These historical calibration residuals refer to the root mean square deviation between the reconstructed point cloud and the actual geometry in the calibration target scene. Preferably, in one embodiment, the system uses a normal distribution or Gaussian mixture model to statistically fit these residual samples, thereby calculating the execution confidence data of the current sensing module. This execution confidence data characterizes the sensing system's probabilistic ability to accurately capture thin-film deformation when faced with uncertainties such as ambient light interference, lens contamination, or projection fringe degradation.

[0072] To quantify the impact of this uncertainty on the sensing baseline, the system introduces a dynamic adjustment mechanism for the sampling frame rate and phase stepping strategy. In one embodiment, the logic for generating enhanced sensing baseline data follows the calculation formula:

[0073] ;

[0074] Where Fe is the adaptively adjusted sampling frame rate; Fbase is the system nominal frame rate; Csense is the perceived confidence data, which ranges from [0,1]; and κ is the preset sensitivity adjustment coefficient, preferably ranging from [0.1,0.5].

[0075] Specifically, when the perceived confidence data Csense is low (e.g., below 0.70), the frame rate gain term in the formula will automatically increase. The system compensates for single-frame uncertainty by increasing the sampling frequency and the number of encryption phase steps, using redundant perception. When Csense ≥ 1, the gain term is set to 0, and the system maintains the nominal frame rate.

[0076] In a specific implementation scenario, if the reflectivity distribution of the target-fitted curved surface exhibits obvious local high-saturation characteristics, the system will perform logical reasoning using a pre-trained confidence evaluation model. This model is built upon a convolutional neural network (CNN), with training conditions set as follows: a learning rate preferably of 0.0005, a batch size of 32, and 300 iterations. The training samples cover structured light acquisition scenarios with different materials, curvatures, and ambient light intensities. The model training employs the Adam optimizer, using the mean squared error between the predicted residual and the actual calibrated residual as the loss function, and a Dropout strategy is introduced during training to prevent overfitting. In another embodiment, a lightweight MobileNet can be used instead of CNN to reduce the edge computing load.

[0077] By transforming the original static sampling strategy into enhanced sensing reference data containing redundant sensing reserves, physical redundancy is provided for subsequent phase calculations, which greatly improves the execution stability of sensing commands in complex optical environments and reduces downstream control drift caused by sensing quality degradation from the source.

[0078] S102, perform phase calculation and three-dimensional reconstruction on the stripe projection image sequence to construct a digital elevation model characterizing the surface morphology of the thin film.

[0079] Optionally, phase calculation and 3D reconstruction are performed on the fringe projection image sequence to construct a digital elevation model characterizing the thin film surface morphology. This includes: filtering and background correction of multiple images in the fringe projection image sequence; the fringe projection image sequence includes multiple consecutive fringe projection images with sequentially shifted phases; extracting the folded phase distribution at each pixel using a four-step phase-shifting algorithm; restoring the absolute phase distribution to an absolute phase distribution using a phase unfolding algorithm; mapping the absolute phase distribution to a height variation field of the thin film relative to a reference surface based on pre-calibrated system geometric parameters; and generating a digital elevation model based on the height variation field. The formula for extracting the folded phase distribution at each pixel using the four-step phase-shifting algorithm is as follows: ;in, I1 represents the folded phase at coordinates (x, y), and I4 represents the light intensity values ​​of the four consecutive fringe projection images at the corresponding pixel points.

[0080] Specifically, the digital elevation model (DEM) is a two-dimensional matrix data describing the height of each point on the thin film surface relative to a reference plane, acquired through structured light 3D measurement technology. In practice, the acquired fringe projection image is first filtered and background corrected to eliminate ambient light interference and image noise. This embodiment employs a four-step phase-shifting method, where the projector sequentially projects four fringe patterns with phases shifted by π / 2, and the camera simultaneously acquires the corresponding four images. Based on the four-step phase-shifting algorithm, the folded phase distribution at each pixel is extracted. The folded phase refers to the principal phase value enclosed in the interval (-π, π], and its calculation formula is... ,in Let I be the folded phase at coordinates (x, y), and I1 to I4 be the light intensity values ​​of the four consecutive fringe projection images at the corresponding pixels. Since the folded phase exhibits a 2π periodic jump, a phase unfolding algorithm is needed to restore it to a continuous absolute phase distribution. This embodiment employs a quality-guided path-based phase unfolding strategy, prioritizing unfolding from high signal-to-noise ratio (SNR) regions to low SNR regions. Finally, based on pre-calibrated system geometric parameters, the absolute phase distribution is linearly mapped to the height variation field of the thin film relative to the reference surface, as shown in the formula: ,in This is the proportionality coefficient. The absolute phase distribution is used to generate a digital elevation model. Through full-field, non-contact 3D reconstruction, this step can capture the complete morphological information of the film surface in real time, providing a high-precision data foundation for subsequent quantitative evaluation of flatness. Its beneficial effect is that it overcomes the shortcomings of traditional discrete point measurements in characterizing full-field dynamic deformation.

[0081] S103, calculates flatness error data based on digital elevation model.

[0082] Optionally, the flatness error data is calculated based on the digital elevation model, including: calculating the global root mean square error of the relative mean of the height deviation of the film surface based on the digital elevation model; calculating the maximum surface fluctuation amplitude in the digital elevation model and generating local peak-valley errors.

[0083] Specifically, flatness error data is an indicator used to quantify the degree of thin film deformation, encompassing both macroscopic and microscopic scales. At the macroscopic scale, the global root mean square error (ERMS) of the relative mean of the film surface height deviation is calculated using the following formula: Where Ω is the total number of pixels in the effective measurement area, and h is the height of the current pixel. The average height value within the region serves as a global feedback variable in the main control loop, characterizing the dispersion of overall flatness. At the microscale, the maximum surface fluctuation amplitude in the digital elevation model is calculated, generating the local peak-valley error (EPV), calculated as EPV = max(h) - min(h). This index serves as a microscopic reference for defect diagnosis, sensitive to local wrinkles and folds. Furthermore, the maximum height gradient value on the film surface can be calculated simultaneously. , The partial derivative is used as an auxiliary reference index to identify the spatial distribution of local wrinkle edges and potential stratified regions. By extracting multi-scale smoothness error data, this step provides the controller with a complete state description from macro to micro, which is beneficial in that it enables the controller to simultaneously achieve the dual objectives of overall smoothness optimization and local defect suppression.

[0084] S104 inputs dynamic trajectory data to the feedforward controller in the composite controller to generate compensation commands to counteract predictable disturbances.

[0085] The feedforward controller is constructed based on a first-order inertial plus pure delay model of the thin-film planarization process. The transfer function of the first-order inertial plus pure delay model is: ;in, For process gain, θ is the time constant, θ is the pure delay time, and s is the Laplace operator; the feedforward controller uses the dynamic trajectory data and the look-forward steps determined by the pure delay time to generate compensation commands.

[0086] Specifically, dynamic trajectory data is input to the feedforward controller in the composite controller to generate compensation commands to counteract predictable disturbances. The feedforward controller is a model-predictive control unit that uses known disturbance information to calculate control actions in advance to offset the impact of disturbances on the system output. In this embodiment, the feedforward controller employs a first-order inertial plus pure delay model of the thin-film planarization process, whose transfer function is... Where Kp is the process gain, reflecting the influence of the control input on the steady-state smoothness; Let be the time constant, representing the system inertia; θ be the pure time delay, derived from the cumulative delay of image acquisition, communication, and phase calculation; and s be the Laplacian operator. Based on this model and dynamic trajectory data, the feedforward controller utilizes the robot's future Na-step trajectory, where... Ts is the look-forward step count and the sampling period. Feedforward commands are applied in advance to compensate for system delays, generating compensation commands. Feedforward predictive control can proactively counteract predictable transient shear and tension disturbances caused by robot motion. Its beneficial effect is that it significantly reduces the pressure on the feedback loop and improves the system's response speed to major disturbances.

[0087] Optionally, a recursive least squares method with a forgetting factor can be used to update the perturbation matrix in the first-order inertia plus pure delay model online; wherein the update law formula for the perturbation matrix is: ; ; Where λ is the forgetting factor, and d(k) is a measurable perturbation vector composed of dynamic trajectory data. This is the estimated value of the perturbation matrix at time k.

[0088] Specifically, this embodiment employs a recursive least squares method with a forgetting factor to update the perturbation matrix Bd online for each sampling period. The recursive least squares method is a recursive parameter estimation algorithm that updates parameter estimates recursively after each new observation, without needing to store all historical data. The forgetting factor λ gradually reduces the weight of historical data in parameter estimation, enabling the algorithm to track changes in time-varying parameters more quickly. In this embodiment, the forgetting factor λ is preferably 0.98. This value strikes a balance between rapid tracking capability and estimation stability, allowing for timely responses to changes in the viscoelastic properties of the film without introducing estimation noise due to oversensitivity.

[0089] The online update process for the perturbation matrix is ​​executed within each control cycle. First, the system obtains the state prediction error Δxp(k) for the current sampling cycle, which reflects the difference between the state prediction based on the current model and the actual observation. Then, based on the difference between the state prediction error and the actual observation, the system corrects the perturbation matrix cycle by cycle. The specific update law formula is as follows:

[0090] ;

[0091] in, The value of the perturbation matrix at time k is the estimated value. Let L(k) be the estimated value from the previous time step, L(k) be the gain vector, and d(k-1) be the measurable perturbation vector from the previous time step. The update law means that the estimated value of the perturbation matrix from the previous time step is corrected based on the prediction error at the current time step. The magnitude of the correction is determined by both the gain vector L(k) and the prediction error. The larger the prediction error, the larger the correction, thus allowing the estimated value to quickly approach the true value.

[0092] The formula for calculating the gain vector L(k) is:

[0093] ;

[0094] Where P(k-1) is the covariance matrix of the previous time step, λ is the forgetting factor, and d(k-1) is the measurable perturbation vector of the previous time step. The gain vector L(k) determines the strength of the correction to the parameter estimate by the current observation data. Its denominator includes the forgetting factor λ. When λ is small, the gain vector is more sensitive to the prediction error, and the parameters are updated faster.

[0095] The update formula for the covariance matrix P(k) is:

[0096] ;

[0097] The covariance matrix P(k) characterizes the uncertainty of the current parameter estimation. Its update follows a recursive form; as new data is continuously added, the covariance matrix gradually converges, and the accuracy of parameter estimation continuously improves. The forgetting factor λ exponentially decays the weight of historical data, enabling the algorithm to adapt to parameter changes.

[0098] Through the above recursive update, the perturbation matrix The system can be adjusted based on the latest observation data in each sampling period, thereby continuously approximating the actual dynamic properties of the thin film material. This online update mechanism is consistent with... Figure 2 The machine vision-based adaptive control device for the flatness of curved surface coating shown in the figure works in conjunction with the flatness error data provided by the perception and calculation module as feedback input. The feedforward controller in the composite decision module generates compensation instructions based on the updated disturbance matrix, forming a complete closed loop from perception to decision and then to model adaptation.

[0099] S105 inputs the flatness error data to the feedback controller in the composite controller to generate a correction command for correcting the unmodeled residual error.

[0100] Optionally, the feedback controller adopts an incremental digital PID control law, whose gain scheduling rules include: the proportional gain increases nonlinearly with the increase of the flatness error data; the integral gain automatically decreases when the compensation command has a significant effect; and the derivative gain is adjusted in real time according to the sign of the product of the flatness error data and its rate of change.

[0101] Specifically, the flatness error data is input to the feedback controller in the composite controller to generate correction instructions for correcting unmodeled residual errors. The feedback controller is an error-adjusting control unit that adjusts the control action in real time based on the deviation between the system output and the set target. In this embodiment, the feedback controller uses an incremental digital PID control law, and its control quantity is... Where e(k) is the smoothness error at the current moment. To enhance robustness, gain scheduling is introduced to adjust the PID parameters online: the proportional gain Kp increases nonlinearly with the increase of the smoothness error data to accelerate the response under large deviations; the integral gain Ki automatically decreases when the feedforward compensation is significant to avoid integral oversaturation; the derivative gain Kd is adjusted in real time according to the sign of the product of the smoothness error data and its rate of change to suppress overshoot. The feedback controller can handle unmodeled dynamics and random disturbances that cannot be accurately described by the feedforward model. Its beneficial effect is that it achieves a balance between response speed and robustness, ensuring that the system can operate stably under various operating conditions.

[0102] S106, superimpose compensation instructions and correction instructions to generate comprehensive correction instructions.

[0103] Specifically, compensation and correction commands are superimposed to generate a comprehensive correction command. The comprehensive correction command is the final control output that integrates feedforward prediction and feedback correction, calculated as u(k) = u_ff(k) + u_fb(k). This feedforward and feedback collaborative control architecture clearly defines the roles of each: the feedforward is responsible for offsetting predictable major disturbances, while the feedback handles model errors, unmodeled dynamics, and random disturbances. Its beneficial effect lies in achieving synergistic optimization of response speed and control accuracy.

[0104] S107: Obtain the strain sensitivity matrix of the distributed gripper array, and use pseudo-inverse operation to allocate the comprehensive correction command as the displacement execution parameters of each gripper.

[0105] Optionally, the strain sensitivity matrix of the distributed gripper array is obtained, and the comprehensive correction command is allocated to the displacement execution parameters of each gripper using pseudo-inverse operation. This includes: using recursive least squares method to identify and update the strain sensitivity matrix that characterizes the mapping relationship between the small displacement of the gripper and the global flatness online; obtaining the Moore-Penrose generalized inverse matrix of the strain sensitivity matrix; multiplying the generalized inverse matrix with the comprehensive correction command to obtain the optimal adjustment displacement vector that minimizes the sum of squares of the displacements of each gripper, and using this vector as the displacement execution parameter.

[0106] Specifically, the strain sensitivity matrix of the distributed gripper array is obtained, and the comprehensive correction command is allocated to the displacement execution parameters of each gripper using pseudo-inverse operations. The strain sensitivity matrix J(k) is a Jacobian matrix characterizing the mapping relationship between the small displacements of each gripper and the global flatness, where each element represents the contribution of a unit displacement of a gripper to the overall flatness index. In this embodiment, a recursive least squares method with a forgetting factor is used to identify and update this matrix online to adapt to the time-varying characteristics of the viscoelasticity of the thin film material. Subsequently, the Mohr-Penrose generalized inverse matrix is ​​obtained from this strain sensitivity matrix. The generalized inverse matrix is ​​then multiplied by the comprehensive correction command u(k) to obtain the optimal adjustment displacement vector that minimizes the sum of squares of the displacements of each gripper in the least squares sense. The vector is used as the displacement parameter. This centralized decision-making and step-by-step execution design has the advantage of maintaining global optimality at the strategy level while providing sufficient mechanical degrees of freedom to cope with non-uniform deformation.

[0107] S108 converts the displacement execution parameters into collaborative control commands and sends them to the servo controllers of each gripper for execution.

[0108] Specifically, the collaborative control commands are encapsulated according to the underlying communication protocol. In this embodiment, the EtherCAT industrial real-time bus is used to encapsulate the displacement commands into standard messages and synchronously send them to each gripper servo controller. Each servo controller responds in position mode, executing three-axis displacement movements with micron-level precision. After issuing the commands, the system continuously monitors the position feedback messages from each servo controller and sends the actual positioning deviation back to the control and optimization module as input for updating the strain sensitivity matrix in the next control cycle. By accurately executing the assigned displacement commands, the collaborative movements of each gripper achieve refined compensation for thin film stress. Its beneficial effect lies in transforming macroscopic flatness control commands into microscopic multi-degree-of-freedom collaborative execution, ensuring the physical accuracy of the control effect.

[0109] Based on the same inventive concept, this application also provides a machine vision-based adaptive control device for curved surface coating flatness to implement the aforementioned machine vision-based adaptive control method for curved surface coating flatness. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations in one or more embodiments of the machine vision-based adaptive control device for curved surface coating flatness provided below can be found in the limitations of the machine vision-based adaptive control method for curved surface coating flatness described above, and will not be repeated here.

[0110] In one embodiment, such as Figure 2As shown, a machine vision-based adaptive control device for the flatness of curved surface coating is provided. The device includes:

[0111] The data acquisition module 30 is used to acquire the stripe projection image sequence of the film-coating operation area collected by the machine vision perception module, and simultaneously acquire the dynamic trajectory data of the robot end effector.

[0112] Model building module 31 is used to perform phase calculation and three-dimensional reconstruction on the stripe projection image sequence to build a digital elevation model that characterizes the surface morphology of the thin film.

[0113] Error calculation module 32 is used to calculate flatness error data based on digital elevation model;

[0114] The first instruction module 33 is used to input dynamic trajectory data into the feedforward controller in the composite controller and generate compensation instructions to counteract predictable disturbances.

[0115] The second instruction module 34 is used to input the flatness error data to the feedback controller in the composite controller and generate correction instructions for correcting unmodeled residual errors.

[0116] The instruction overlay module 35 is used to overlay compensation instructions and correction instructions to generate a comprehensive correction instruction;

[0117] The instruction allocation module 36 is used to obtain the strain sensitivity matrix of the distributed gripper array and use pseudo-inverse operation to allocate the comprehensive correction instruction as the displacement execution parameters of each gripper.

[0118] The instruction execution module 37 is used to convert displacement execution parameters into coordinated control instructions and send them to the servo controllers of each gripper for execution.

[0119] This application also provides an electronic device, in some embodiments, referring to... Figure 3 As shown, the electronic device 700 includes an input unit 710, a memory 720, a processor 730, and an output unit 740. The memory 720 stores program instructions that can be executed on the processor 730. The processor 730 can execute the machine vision-based adaptive control method and / or technical solution for curved surface coating flatness based on the aforementioned embodiments by calling the program instructions. This electronic device 700 can be a mobile terminal device such as a mobile phone or computer.

[0120] Furthermore, embodiments of this application also provide a computer-readable storage medium for storing a computer program that executes a machine vision-based adaptive control method for the flatness of curved surface coatings. For example, computer program instructions, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. The program instructions that invoke the methods of this application may be stored in a fixed or removable storage medium, and / or transmitted via data streams in broadcast or other signal carrying media, and / or stored in a storage medium that operates according to the program instructions.

[0121] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. Optionally, they can be implemented using computer-executable program code, thereby storing them in a storage device for execution by a computing device, or fabricating them separately as individual integrated circuit modules, or fabricating multiple modules or steps as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0122] The technical features of the above embodiments can be arbitrarily integrated. For the sake of brevity, not all possible integrations of the technical features in the above embodiments are described. However, as long as the integration of these technical features does not contradict each other, they should be considered to be within the scope of this specification.

[0123] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A machine vision-based curved surface film covering flatness self-adaptive control method, characterized in that, An application is made in a curved surface coating control system, the system comprising a machine vision perception module, a robot end effector, a composite controller, and a distributed gripper array, the method comprising: The stripe projection image sequence of the film-coating operation area collected by the machine vision perception module is acquired, and the dynamic trajectory data of the robot end effector is acquired simultaneously. Phase calculation and three-dimensional reconstruction are performed on the stripe projection image sequence to construct a digital elevation model characterizing the surface morphology of the thin film; Calculate the flatness error data based on the digital elevation model; The dynamic trajectory data is input to the feedforward controller in the composite controller to generate compensation commands to counteract predictable disturbances. The feedforward controller is constructed based on a first-order inertial plus pure delay model of the thin-film planarization process, and the transfer function of the first-order inertial plus pure delay model is: ； Where Kp is the process gain, τp is the time constant, θ is the pure delay time, and s is the Laplace operator; The feedforward controller uses the dynamic trajectory data and the pure delay time to determine the number of look-forward steps to generate the compensation command; The perturbation matrix in the first-order inertial plus pure delay model is updated online using a recursive least squares method with a forgetting factor. The update law formula for the perturbation matrix is ​​as follows: ； L(k) = P(k-1)d(k-1) / [λ + dᵀ(k-1)P(k-1)d(k-1)]; P(k) = λ -1 [P(k-1) - L(k)dᵀ(k-1)P(k-1)]; wherein λ is a forgetting factor, d(k) is a measurable disturbance vector composed of the dynamic trajectory data, is the disturbance matrix estimate at time k. The flatness error data is input to the feedback controller in the composite controller to generate a correction instruction for correcting the unmodeled residual error; The compensation instruction and the correction instruction are superimposed to generate a comprehensive correction instruction; Obtain the strain sensitivity matrix of the distributed gripper array, and use pseudo-inverse operation to allocate the comprehensive correction command as the displacement execution parameters of each gripper; The displacement execution parameters are converted into collaborative control commands and sent to the servo controllers of each gripper for execution.

2. The machine vision-based flatness control method of claim 1, wherein, Phase calculation and three-dimensional reconstruction are performed on the stripe projection image sequence to construct a digital elevation model characterizing the surface morphology of the thin film, including: Filtering and background correction are performed on multiple images in the striped projection image sequence; the striped projection image sequence includes multiple consecutive striped projection images with sequentially shifted phases; A four-step phase-shifting algorithm is used to extract the folded phase distribution at each pixel; The folded phase distribution is restored to an absolute phase distribution using a phase unfolding algorithm. Based on pre-calibrated system geometric parameters, the absolute phase distribution is mapped to a height variation field of the thin film relative to the reference surface; The digital elevation model is generated based on the height change field. 3.The machine vision-based flatness control method of curved surface coating according to claim 2, wherein, The formula for extracting the folded phase distribution at each pixel using the four-step phase-shifting algorithm is as follows: φw(x,y)=arctan[(I4(x,y)-I2(x,y)) / (I1(x,y)-I3(x,y))]; Where φw(x,y) is the folded phase at coordinate (x,y), and I1 to I4 are the light intensity values ​​of the four consecutive fringe projection images at the corresponding pixel points.

4. The machine vision-based adaptive control method for the flatness of curved surface coating as described in claim 1, characterized in that, The calculation of flatness error data based on the digital elevation model includes: Based on the digital elevation model, the global root mean square error of the relative mean of the height deviation of the thin film surface is calculated; Calculate the maximum surface fluctuation amplitude in the digital elevation model to generate local peak-valley errors. 5.The machine vision-based flatness control method for curved surface coating according to claim 1, wherein, Obtain the strain sensitivity matrix of the distributed gripper array, and use pseudo-inverse operation to allocate the comprehensive correction command as displacement execution parameters for each gripper, including: The recursive least squares method is used to identify and update the strain sensitivity matrix that represents the mapping relationship between the small displacement of the gripper and the global flatness online; The Moore-Penrose generalized inverse matrix is ​​obtained from the strain sensitivity matrix; The generalized inverse matrix is ​​multiplied by the comprehensive correction command to obtain the optimal adjustment displacement vector that minimizes the sum of the squares of the displacements of each gripper, and this vector is used as the displacement execution parameter. 6.The machine vision based flatness control method of curved surface coating according to claim 1, wherein, The feedback controller employs an incremental digital PID control law, and its gain scheduling rules include: The proportional gain increases non-linearly with the increase of the flatness error data; The integral gain is automatically reduced when the compensation command has a significant effect; The differential gain is adjusted in real time based on the sign of the product of the flatness error data and its rate of change.

7. A machine vision-based adaptive control device for the flatness of curved surface coating, characterized in that, The device includes: The data acquisition module is used to acquire the stripe projection image sequence of the film-coating operation area collected by the machine vision perception module, and simultaneously acquire the dynamic trajectory data of the robot end effector; The model building module is used to perform phase calculation and three-dimensional reconstruction on the stripe projection image sequence to build a digital elevation model that characterizes the surface morphology of the thin film. The error calculation module is used to calculate the flatness error data based on the digital elevation model; The first instruction module is used to input the dynamic trajectory data into the feedforward controller in the composite controller and generate compensation instructions to counteract predictable disturbances. The second instruction module is used to input the flatness error data into the feedback controller in the composite controller and generate correction instructions for correcting unmodeled residual errors. The instruction overlay module is used to overlay the compensation instruction and the correction instruction to generate a comprehensive correction instruction; The instruction allocation module is used to obtain the strain sensitivity matrix of the distributed gripper array and use pseudo-inverse operation to allocate the comprehensive correction instruction as the displacement execution parameters of each gripper. The instruction execution module is used to convert the displacement execution parameters into cooperative control instructions and send them to the servo controllers of each gripper for execution. The device is used to implement the machine vision-based adaptive control method for the flatness of curved surface coating as described in any one of claims 1 to 6.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the machine vision-based adaptive control method for the flatness of curved surface coating as described in any one of claims 1 to 6.

Images