A protective film slitting positioning system based on visual correction

Abstract

The present application relates to the technical field of film slitting and deviation correction, and discloses a protective film slitting positioning system based on visual deviation correction, which comprises an industrial camera and a dark field light source cooperatively installed at the upstream position of a slitting cutter group; a controller controls the industrial camera to acquire microscopic texture images, establishes a visual zero coordinate system and extracts a sparse feature point set; a two-dimensional affine transformation matrix is fitted through optical flow tracking operation, and the kinematic parameters and total lateral visual displacement of the protective film are solved; a control module calculates the time delay of the cutter gap to reach in combination with the set spatial physical distance and system time consumption, selectively outputs the servo drive instruction superimposed by the feedforward prediction instruction and the closed-loop feedback instruction according to the positive and negative state of the time delay, or only outputs the closed-loop feedback instruction, and drives the actuator to generate a lateral mechanical correction displacement. The present application extracts the surface texture to solve the displacement, introduces the feedforward prediction to offset the mechanical execution lag, and realizes accurate slitting positioning.

Description

Technical Field

[0001] This invention relates to the field of film slitting and correction technology, specifically to a protective film slitting and positioning system based on visual correction. Background Technology

[0002] In the slitting process of thin film materials such as protective films, the accuracy of the web alignment and positioning system directly determines the quality of the finished roll surface. Existing web alignment systems mostly rely on sensors to detect the physical edges of the film. However, in actual production, film edges are often damaged or have irregular shapes, which can easily interfere with the detection process, leading to inaccurate lateral position tracking. Simultaneously, under longitudinal traction tension, the film material undergoes normal lateral shrinkage deformation. Conventional control logic cannot separate this material's own stress deformation from the overall mechanical deviation of the film. The system may misinterpret the normal width reduction of the material as a deviation error, thus triggering incorrect web alignment actions.

[0003] On the other hand, the correction mechanism has an inherent mechanical response time after receiving the command. Traditional closed-loop feedback control usually only passively executes corrections after detecting deviations, lacking the ability to predict the film's operating state. When the protective film travels at high speed to the downstream slitting edge, the mechanical mechanism struggles to complete displacement compensation before the deviation position reaches the edge. If the system blindly outputs adjustment commands under these circumstances, it is prone to lag corrections due to insufficient response time, leading to reverse overshoot of the mechanical system and causing the dynamic correction process to become unstable. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a protective film cutting and positioning system based on visual correction. This system solves the problems of lateral shrinkage deformation of film materials under traction tension, difficulty in distinguishing between material deformation and roll offset caused by conventional edge detection methods, and limitations in dynamic correction accuracy due to mechanical lag in traditional feedback control.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solution: a protective film slitting and positioning system based on visual correction, comprising: an industrial camera, which is installed upstream of the slitting blade assembly in conjunction with a dark field light source;

[0006] The controller uses a vision acquisition module to control an industrial camera to synchronously and continuously acquire micro-texture images with a dark field light source. After establishing a vision zero-position coordinate system, it extracts a set of sparse feature points from the micro-texture images.

[0007] The controller uses the vector analysis module to obtain a sparse feature point set, performs optical flow tracking operation on it in consecutive image frames to fit a two-dimensional affine transformation matrix, and calculates the kinematic parameters of the protective film and the total lateral visual displacement relative to the visual null coordinate system based on the two-dimensional affine transformation matrix.

[0008] The controller uses a composite control module to obtain kinematic parameters and total lateral visual displacement. It calculates the blade arrival delay by combining the spatial physical distance set between the industrial camera and the slitting blade group with the system time consumption. Based on the positive or negative state of the blade arrival delay, it selectively outputs a servo drive command that combines feedforward prediction command and closed-loop feedback command to the correction actuator, or outputs only the closed-loop feedback command.

[0009] The correction actuator drives the protective film's running axis system to generate a lateral mechanical correction displacement according to servo drive commands, thereby achieving precise cutting and positioning of the protective film relative to the slitting blade assembly.

[0010] This invention captures the microscopic texture features of a protective film material surface through the combination of a dark-field light source and an industrial camera. An optical flow algorithm is used to track the random surface texture, resolving the spatial motion vector of the thin film material in a two-dimensional plane, eliminating dependence on the material's physical edges. The system calculates the time difference between a specific location on the material surface and its movement from the camera's field of view to the downstream slitting edge, introducing a feedforward prediction loop into the control circuit to compensate for the execution lag of the mechanical correction mechanism, performing mechanical correction before the deviation position reaches the cutting edge.

[0011] Furthermore, the system also includes an absolute position sensor; during the system initialization and calibration phase, when the cutting equipment establishes tension and is in a non-traction static state, the controller also uses the vision acquisition module to control the industrial camera to continuously acquire micro-texture images, calibrating the upper limit of visual strain noise corresponding to the inherent high-frequency mechanical vibration of the system.

[0012] Once the protective film enters a stable traction and centered operating state, the controller uses the vision acquisition module to capture a microscopic texture image to extract current image features as the initial zero-position reference frame. The absolute position calibration data from the absolute position sensor is then bound to the initial zero-position reference frame to establish a visual zero-position coordinate system. Subsequently, within the pixel matrix of the microscopic texture image, the gradient covariance matrix of the image pixels is calculated by scanning the region of interest, extracting pixels with strong bidirectional gradients as a sparse feature point set. Setting an upper limit for visual strain noise filters out high-frequency mechanical vibration interference generated during normal operation of the cutting equipment; combined with the zero-position reference established by the absolute position sensor, an absolute physical reference is provided for visual displacement calculation.

[0013] In a preferred embodiment of the present invention, the process of calibrating the upper limit of visual strain noise in the visual acquisition module includes: extracting displacement gradient data between adjacent frames of microtexture images, performing time-series statistical analysis, and calculating the peak value or root mean square value of the displacement gradient within a set confidence interval as the upper limit of visual strain noise. Using time-series statistical analysis to extract the peak value or root mean square value quantifies the inherent background noise baseline of the system, avoiding the problem of poor environmental adaptability caused by fixed threshold calibration.

[0014] Furthermore, the vector analysis module performs optical flow tracking on the sparse feature point set in consecutive image frames to fit a two-dimensional affine transformation matrix, and calculates the kinematic parameters and total lateral visual displacement of the protective film. This process includes: performing sparse optical flow tracking calculations on the sparse feature point sets of two adjacent micro-texture images in consecutive image frames, and running a random sampling consensus algorithm to remove mismatched points; after mismatch removal, fitting the two-dimensional affine transformation matrix to calculate the kinematic parameters, including longitudinal motion velocity, lateral drift velocity, and yaw angle parameters; and then performing a cumulative calculation based on the lateral drift velocity using time discrete integrals to obtain the total lateral visual displacement. Running the random sampling consensus algorithm to remove mismatched points avoids interference from local image noise on the affine transformation matrix fitting, improving the reliability of the kinematic parameter calculation.

[0015] Furthermore, during the optical flow tracking operation, the vector analysis module calculates the proportion of interior points in each fitting process by dividing the number of interior points corresponding to the two-dimensional affine transformation matrix by the total number of feature points participating in optical flow matching. When the proportion of interior points is less than the safety confidence threshold, a confidence downgrade condition is triggered. When the confidence downgrade condition is triggered or the cumulative period of the time discrete integral reaches the preset maximum full threshold, the vector analysis module calls the absolute position calibration data to overwrite the currently saved total lateral visual displacement to complete the forced reset of the coordinate system origin, and triggers the visual acquisition module to synchronously capture the microscopic texture image of the current scene to refresh the initial zero-position reference frame.

[0016] By introducing confidence degradation conditions and full integration thresholds, the absolute position data is forcibly retrieved to refresh the baseline when the visual tracking matching rate decreases, thus eliminating the cumulative drift error caused by time discrete integration.

[0017] Preferably, the system also includes a tension sensor; the controller is further equipped with a parameter identification module and an error decoupling module. The controller uses the parameter identification module to extract visual strain from the two-dimensional affine transformation matrix, combines this with the longitudinal traction tension data from the tension sensor to calculate longitudinal mechanical stress, and extracts dynamic mechanical parameters. The controller uses the error decoupling module to obtain the dynamic mechanical parameters to calculate the lateral shrinkage of the protective film, eliminates non-positional offsets caused by material deformation, and separates the deviation error corresponding to the overall roll offset from the total lateral visual displacement. The composite control module uses the deviation error as input, runs a proportional-integral-derivative control algorithm, and calculates the closed-loop feedback command. By using dynamic mechanical parameters to separate material deformation and deviation error, physical decoupling of the displacement components is achieved, preventing the system from misjudging the material shrinkage caused by tension as mechanical deviation.

[0018] Furthermore, visual strain includes visual longitudinal strain and visual transverse strain, and dynamic mechanical parameters include dynamic elastic modulus and dynamic Poisson's ratio. The parameter identification module extracts visual strain and calculates longitudinal mechanical stress by combining it with longitudinal traction tension data. The process of extracting dynamic mechanical parameters includes: stripping translational motion components from the two-dimensional affine transformation matrix to extract the displacement gradient matrix, and using partial derivatives to calculate visual longitudinal strain and visual transverse strain; combining the initial width and thickness of the protective film to obtain the initial cross-sectional area, and dividing the longitudinal traction tension data by the initial cross-sectional area to obtain the longitudinal mechanical stress.

[0019] An identification computational model is constructed using visual longitudinal strain, visual transverse strain, and longitudinal mechanical stress. The upper limit of visual strain noise is introduced into the model as a threshold stabilization term, forming a stable denominator term incorporating the noise upper limit. Amplitude limiting is applied using a limiting function to output the dynamic elastic modulus and dynamic Poisson's ratio. Introducing the upper limit of visual strain noise as a threshold stabilization term into the model prevents computational divergence caused by division approaching zero under small strains, ensuring stable identification of dynamic mechanical parameters.

[0020] In a preferred embodiment of the present invention, the process of separating and obtaining the deviation error by the error decoupling module includes: constructing a lateral deformation estimation model by combining longitudinal mechanical stress, dynamic elastic modulus, dynamic Poisson's ratio, and the initial width of the protective film to estimate the lateral shrinkage; obtaining the shrinkage distribution coefficient and directional compensation polarity of the equipment process calibration, and combining the lateral shrinkage to obtain the unilateral lateral shrinkage compensation displacement component with directional polarity; subtracting the unilateral lateral shrinkage compensation displacement component from the total lateral visual displacement to separate and obtain the deviation error unaffected by tension shrinkage. Subtracting the lateral shrinkage compensation displacement from the total visual displacement removes the interference of material stress deformation and extracts the true mechanical deviation displacement component.

[0021] Furthermore, the composite control module calculates the blade arrival delay by combining the set spatial physical distance and system time consumption. This process includes: comparing the longitudinal movement speed with a preset minimum safe speed threshold and taking the maximum value as the processed longitudinal movement speed; dividing the set spatial physical distance by the processed longitudinal movement speed and subtracting the system time consumption to obtain the blade arrival delay. Using the minimum safe speed threshold limits the movement speed to prevent the speed denominator from being too small and causing overflow in the delay calculation during low-speed or stopped states, thus ensuring stable system operation.

[0022] Preferably, the process of the composite control module outputting servo drive commands to the correction actuator includes: when the blade arrival delay is positive, generating a feedforward prediction command based on the yaw angle parameter and lateral drift velocity, superimposing it with the closed-loop feedback command, and outputting the servo drive command to the correction actuator; when the blade arrival delay is non-positive, actively cutting off the feedforward prediction path, and controlling the correction actuator to generate lateral mechanical correction displacement solely based on the closed-loop feedback command. The control loop is switched according to the positive or negative state of the blade arrival delay. When the response time is sufficient, feedforward superposition is introduced to improve the dynamic response speed of the correction; when the response time is insufficient, the feedforward path is cut off to prevent reverse overshoot caused by lag correction and maintain system convergence stability.

[0023] This invention provides a protective film cutting and positioning system based on visual correction. It has the following beneficial effects:

[0024] 1. This invention acquires microscopic texture images of the protective film surface using a dark-field light source and an industrial camera, and then uses an optical flow algorithm to track the texture features to calculate the film's kinematic parameters and visual displacement. This method does not rely on the material's physical edges as a reference, avoiding detection interference caused by material edge damage or irregularities. It directly extracts the true motion vector from the material itself, improving the accuracy of position detection.

[0025] 2. This invention calculates dynamic mechanical parameters by combining visual strain and longitudinal traction tension data through a parameter identification module, and estimates the lateral shrinkage of the film using an error decoupling module, separating it from the total lateral visual displacement. This technical feature achieves physical decoupling between material stress deformation and mechanical deviation error in the control logic, preventing the system from misinterpreting normal width reduction caused by traction tension as mechanical deviation, and avoiding erroneous correction actions.

[0026] 3. This invention calculates the blade arrival delay by combining the spatial physical distance between the camera and the slitting blade assembly, system time consumption, and longitudinal movement speed. Based on the delay status, it selectively outputs either a superimposed feedforward and feedback command or a single feedback command. This control strategy utilizes feedforward prediction to offset the mechanical response lag of the correction actuator, achieving early correction of the deviation position before it reaches the blade. Simultaneously, it cuts off the feedforward path when the response time is insufficient, preventing reverse overshoot caused by lag correction and ensuring the stability of the dynamic correction process. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the protective film cutting and positioning system architecture based on visual correction according to the present invention;

[0028] Figure 2 This is a flowchart of the protective film cutting and positioning method of the present invention;

[0029] Figure 3This is a flowchart of the motion vector analysis and anti-drift correction logic of the present invention;

[0030] Figure 4 This is a flowchart illustrating the visual strain extraction and dynamic parameter identification logic of the present invention.

[0031] Figure 5 This is a schematic diagram illustrating the principle of tension reduction decoupling and deviation error extraction of the present invention.

[0032] Figure 6 This is a flowchart of the time-delay composite correction control logic of the present invention;

[0033] Figure 7 This is a graph showing the tension fluctuation of the present invention;

[0034] Figure 8 The mechanical parameters of this invention are identified using convergence curves.

[0035] Figure 9 This is a comparison chart of error stripping and extraction in this invention;

[0036] Figure 10 This is a comparison diagram of the feedforward composite correction control response of the present invention. Detailed Implementation

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0038] See attached document Figure 1 The present invention provides a protective film cutting and positioning system based on visual correction, comprising: an industrial camera, a dark field light source, a tension sensor, an absolute position sensor, a controller, and a correction actuator.

[0039] An industrial camera and a dark-field light source are mounted upstream of the slitting blade assembly, maintaining a predetermined spatial physical distance between them. The dark-field light source illuminates the surface of the protective film at a low angle to highlight surface texture features, allowing the industrial camera to continuously capture microscopic texture images.

[0040] Tension sensors are arranged in the roller system along the conveying path of the protective film to simultaneously measure longitudinal traction tension data. Absolute position sensors are arranged near the edge area of ​​the protective film to acquire absolute position calibration data. The images and numerical data collected by each sensor are integrated into the controller for processing. The controller generates servo drive commands based on the processing results and sends them to the correction actuator to drive the running axis of the protective film to generate lateral mechanical correction displacement.

[0041] The controller is equipped with multiple data processing and control logic modules that work together. The controller first imports image and sensor data into the vision acquisition module, which then controls the industrial camera and the dark field light source to synchronize exposure and continuously acquire micro-texture images to calibrate the upper limit of visual strain noise. When the equipment is initialized or malfunctions, a visual zero-position coordinate system based on physical coordinate alignment is established, and after establishing the visual zero-position coordinate system, the sparse feature point set of the micro-texture is extracted.

[0042] The coordinates and feature data extracted by the visual acquisition module then flow into the vector analysis module. The vector analysis module performs optical flow tracking on the sparse feature point set in consecutive image frames, fits a two-dimensional affine transformation matrix, and solves for the kinematic parameters of the protective membrane. The two-dimensional affine transformation matrix output by the vector analysis module is passed to the parameter identification module. The parameter identification module extracts the displacement gradient matrix from the two-dimensional affine transformation matrix, extracts the visual strain, and calculates the longitudinal mechanical stress by combining it with the longitudinal traction tension data from the tension sensor, thereby identifying the constrained dynamic mechanical parameters online.

[0043] Based on the dynamic mechanical parameters generated by the parameter identification module, the error decoupling module calculates the lateral shrinkage of the protective film, eliminates non-positional offsets caused by the material's own deformation, and extracts the deviation error corresponding to the overall roll offset. The finally extracted deviation error is input to the composite control module. The composite control module combines the spatial physical distance and system time consumption to determine the blade arrival delay. Depending on the positive or negative state of the blade arrival delay, it selectively outputs a servo drive command that superimposes the feedforward prediction command and the closed-loop feedback command to the correction actuator, or outputs only the closed-loop feedback command, to drive the protective film to achieve slitting and positioning relative to the slitting blade group.

[0044] See attached document Figure 2 This invention provides a protective film cutting and positioning method based on visual correction, comprising the following steps:

[0045] S10. System Initialization Calibration and Zero-Position Reference Establishment. The vision acquisition module controls the industrial camera to continuously acquire micro-texture images. When the slitting device is under tension and in a non-traction static state, the upper limit of the visual strain noise corresponding to the inherent high-frequency mechanical vibration of the system is calibrated. After the upper limit of the visual strain noise is calibrated, the slitting device is started for traction. When the protective film enters a stable traction and centered operating state, the vision acquisition module captures the micro-texture image and extracts the current image features as the initial zero-position reference frame. At the same time, the absolute position calibration data of the absolute position sensor is read and bound to the initial zero-position reference frame to establish the visual zero-position coordinate system.

[0046] After binding is completed, the visual acquisition module runs a corner detection algorithm in the set region of interest to extract a sparse set of feature points of dark field micro-texture for subsequent optical flow tracing calculations.

[0047] S20. Motion Vector Analysis and Anti-Drift Correction. Based on the sparse feature point set extracted by the visual acquisition module, the vector analysis module performs sparse optical flow tracking calculations on the sparse feature point set of two consecutive frames and runs a random sampling consensus algorithm to remove mismatched point sets. After mismatch removal, the vector analysis module fits a two-dimensional affine transformation matrix and calculates the kinematic parameters, including longitudinal motion velocity, lateral drift velocity, and yaw angle parameters. Based on the lateral drift velocity, the vector analysis module performs cumulative calculations based on time discrete integration to obtain the total lateral visual displacement relative to the visual null coordinate system. Since long-term continuous integration will produce cumulative errors, when the confidence degradation condition is triggered or the cumulative period is reached, the absolute position calibration data of the absolute position sensor is called to force a coordinate reset and refresh the initial null reference frame as a new reference for subsequent optical flow tracking calculations.

[0048] S30. Visual Strain Extraction and Dynamic Parameter Identification. Upon receiving the two-dimensional affine transformation matrix, the parameter identification module extracts the translational motion components to obtain the displacement gradient matrix and calculates the visual strain using partial derivatives. The visual strain includes visual longitudinal strain and visual transverse strain. Simultaneously, the parameter identification module acquires the longitudinal traction tension data from the tension sensor to calculate the longitudinal mechanical stress. The longitudinal mechanical stress and visual longitudinal strain are then calculated to extract the dynamic elastic modulus. Furthermore, the visual transverse strain, visual longitudinal strain, and longitudinal mechanical stress are fused to extract the dynamic Poisson's ratio. During the calculation process, an upper limit for visual strain noise is added as a threshold stabilization term for amplitude limiting. Finally, the dynamic mechanical parameters, including the dynamic elastic modulus and dynamic Poisson's ratio, are output.

[0049] S40. Tension Shrinkage Decoupling and Misalignment Error Extraction. After acquiring the dynamic mechanical parameters, the error decoupling module estimates the current lateral shrinkage of the protective film based on the longitudinal traction tension data and the dynamic mechanical parameters. To eliminate material deformation interference, the error decoupling module introduces the shrinkage distribution coefficient calibrated by the equipment process, subtracting the lateral shrinkage from the total lateral visual displacement according to its directional attribute, thereby separating and obtaining the misalignment error unaffected by tension shrinkage.

[0050] S50, Time-Delay Composite Correction Control. After acquiring the deviation error, the composite control module first acquires the longitudinal motion speed and compares it with the preset minimum safe speed threshold, taking the maximum value of the two as the processed longitudinal motion speed; then, the spatial physical distance is divided by the processed longitudinal motion speed, and the system time consumption is subtracted. The system time consumption includes camera exposure time, algorithm processing time, and servo response time, to obtain the blade arrival delay.

[0051] When the blade arrival delay is positive, the composite control module generates a feedforward prediction command based on the yaw angle parameter and lateral drift velocity. This command is then superimposed on the closed-loop feedback command generated based on the deviation error, and a servo drive command is output to the correction actuator to drive the protective film to generate a lateral mechanical correction displacement, thereby achieving the slitting positioning relative to the slitting blade group. When the blade arrival delay is non-positive, to avoid overshoot in the system response, the composite control module actively cuts off the feedforward prediction path and controls the correction actuator to generate a lateral mechanical correction displacement solely based on the closed-loop feedback command to complete the protective film slitting positioning.

[0052] When the visual acquisition module executes step S10, it mainly performs the coordinated synchronization of the underlying image acquisition hardware and the calibration of the initial reference coordinate system to obtain the surface texture information of the high-transmittance material. Specifically, it includes the following sub-steps:

[0053] S11. Dark-field optical imaging hardware configuration and micro-texture features are highlighted. To obtain stable surface texture information, a global shutter area array industrial camera is specifically used to eliminate the rolling shutter motion distortion caused by line-by-line exposure under high-speed traction of the protective film. The dark-field light source specifically uses a low-angle illumination industrial LED strip light source or ring light source.

[0054] When a dark field light source illuminates the protective film surface at a preset low incident angle of 5° to 20°, the reflected light from the flat area will deviate from the light path received by the lens of the industrial camera. The micro-texture, fine scratches, and local micro-undulations on the surface of the protective film will scatter the light, and some of the scattered light will enter the lens of the industrial camera perpendicularly.

[0055] The optical arrangement of low-angle dark-field illumination creates high-contrast bright-colored micro-texture features in the image, compensating for the lack of macroscopic printed patterns and clear physical boundaries inherent in transparent composite protective films. The exposure control principle of global shutter area array industrial cameras and the optical path arrangement rules of low-angle dark-field illumination are both well-known technologies in this field.

[0056] S12. System Static Calibration and Low-Frequency Mechanical Vibration Compensation Vector Extraction. When the slitting equipment is under tension and in a non-traction static state, the vision acquisition module controls the industrial camera to continuously acquire a preset number of frames of micro-texture images. The preset number of frames is calculated by multiplying the acquisition frame rate of the industrial camera by a set calibration time window. The set calibration time window is set based on the principle of fully covering the lowest frequency vibration cycle of the production line's mechanical environment, and is usually set to 1 to 5 seconds.

[0057] Due to motor vibrations and roller system micro-vibrations in the industrial production line environment, there is an objective high-frequency, minute relative displacement between the industrial camera and the frame, which in turn generates optical flow pseudo-displacement between two consecutive still images. The vision acquisition module calculates the displacement gradient data between adjacent frames of micro-texture images and performs time-series statistical analysis. It calculates the peak value or root mean square value of the displacement gradient within a set confidence interval (usually set to 95% or 99%), and sets the extracted numerical features as the upper limit of visual strain noise.

[0058] The visual strain noise upper limit characterizes the inherent vibration background noise level of the system under the current mechanical environment. As an important safety limiting parameter, it is used to set a threshold stability term in subsequent mechanical parameter identification to prevent the system from misinterpreting frame vibration as tensile deformation of the protective membrane. Furthermore, the visual acquisition module records the initial reference coordinates of the rigid mechanical reference target mounted at the edge of the industrial camera's field of view during the static calibration step. After entering the traction operation phase, the visual acquisition module identifies and tracks this rigid mechanical reference target in real time, continuously acquiring the overall low-frequency offset of the equipment frame during operation. This allows for the extraction of the dynamic compensation vector for the system's low-frequency mechanical vibration, which is temporarily stored in the controller's memory to eliminate spatial low-frequency displacement interference introduced by equipment frame swaying during subsequent operation.

[0059] S13. Establishment of a visual zero-position coordinate system based on absolute position feedback. After calibrating the upper limit of visual strain noise, the slitting equipment starts the traction shaft system, and the protective film begins continuous conveying. The visual acquisition module monitors the output data of the absolute position sensor in real time. The absolute position sensor specifically adopts an ultrasonic edge detector, an infrared edge-aligning photocell, or a high-precision laser rangefinder. When the absolute position sensor detects that the edge of the protective film is within the preset physical center tolerance zone, and the edge tolerance meets the set time period for maintaining the state, the visual acquisition module determines that the protective film has entered a stable traction and centered operating state. Among them, the physical center tolerance zone is calibrated according to the slitting accuracy requirements of the slitting blade group, and the set time period is adjusted according to the conveying linear speed of the protective film and the span between the two sets of guide rollers.

[0060] At this point, the vision acquisition module captures the micro-texture image, extracts the current image features as the initial zero-position reference frame, and simultaneously reads the absolute position calibration data from the absolute position sensor. This absolute position calibration data is then data-bound with the image space center point of the initial zero-position reference frame, thereby establishing a visual zero-position coordinate system that incorporates physical scale meaning. This data-bound mechanism provides the continuous surface micro-texture with an absolute reference origin in device space, enabling the visually extracted pure relative displacement increment to be converted into an absolute deviation position in device space.

[0061] S14. Extraction of sparse feature point sets based on region of interest planning and feature response threshold. After establishing the visual null coordinate system, in order to reduce the computational load and meet the real-time requirements of the control cycle, the visual acquisition module plans the region of interest in the pixel matrix of the micro-texture image. The region of interest avoids the reflective blind zone at the edge of the protective film that is easily interfered with by ambient light, and is located in a geometric region with uniform dark-field illumination and dense texture distribution.

[0062] The visual acquisition module runs a corner detection algorithm within the region of interest, scanning and calculating the gradient covariance matrix of image pixels to extract pixels with strong bidirectional gradients as a sparse feature point set for the micro-texture. By transforming the intensive computation of the entire image pixel matrix into processing operations on a small set of discrete feature points, the computational overhead of the vision system is reduced. Specifically, the Shi-Tomasi corner detection algorithm is used. The Shi-Tomasi algorithm determines corner attributes based on whether the minimum eigenvalue of the pixel's neighborhood gradient matrix is ​​greater than a set feature response threshold, thus obtaining stable and relatively uniform feature pixel coordinates. The set feature response threshold is dynamically evaluated and extracted based on the grayscale contrast of the micro-texture image and the ambient light background noise.

[0063] If the number of sparse feature points extracted within the current region of interest is lower than a preset minimum limit, the visual acquisition module will expand the pixel boundary of the region of interest and re-execute feature extraction until the required number of basic feature points for optical flow tracking is met. The preset minimum limit is set to an integer greater than or equal to 4, based on the minimum degrees of freedom required for subsequent 2D affine transformation matrix fitting and redundancy tolerance requirements; in practical engineering, it is typically set to 50 to 100 feature points. The eigenvalue calculation criteria and the specific mathematical solution process for the corner response function in the Shi-Tomasi corner detection algorithm are both well-known techniques in this field.

[0064] The visual acquisition module ultimately inputs a set of sparse feature points with coordinate information into the vector analysis module, providing initial data for the dynamic optical flow tracking of the protective film.

[0065] See attached document Figure 3 When the vector analysis module executes step S20, it mainly extracts the mechanical motion parameters in the physical coordinate system through relative micro-texture displacement to eliminate the cumulative integral drift error caused by surface reflection or wrinkle interference. Specifically, it includes the following sub-steps:

[0066] S21. Sparse Optical Flow Tracking and Mismatch Point Removal Calculation for Continuous Image Frames. Based on the sparse feature point set extracted by the visual acquisition module, the vector analysis module performs sparse optical flow tracking calculation on the sparse feature point set of two consecutive image frames and runs a random sampling consensus algorithm to remove mismatch point sets. In specific implementation, the vector analysis module acquires two adjacent frames of micro-texture images (i.e., the previous image frame and the current image frame) continuously acquired by the industrial camera, and applies the pyramid Lucas-Carnard optical flow algorithm to search for corresponding matching feature points in the pixel matrix of the current image frame, thereby obtaining the two-dimensional pixel displacement vector of each feature point.

[0067] In actual production line environments, localized reflective jumps or physical wrinkles on the protective film surface can cause some optical flow tracking results to fail, resulting in the inclusion of erroneous matching points in the matched pixel coordinate set. Therefore, after obtaining the two-dimensional pixel displacement vector, the vector analysis module calls the random sampling consensus algorithm to randomly select three feature point pairs as the minimum solution set, calculates a temporary affine transformation matrix, and then substitutes the remaining feature point pairs into the temporary affine transformation matrix to calculate the reprojection error.

[0068] The vector analysis module classifies feature points with reprojection errors less than a preset pixel distance threshold as inliers. This preset pixel distance threshold is determined based on image resolution and optical flow tracking accuracy requirements, typically ranging from 0.5 to 2.0 pixels. After multiple random sampling and iterative verifications, the vector analysis module selects the temporary affine transformation matrix containing the most inliers, while simultaneously discarding feature points that fail to be classified as inliers as mismatched points. The set number of random sampling iterations is derived based on the desired inlier probability and confidence level to ensure algorithm convergence.

[0069] S22. Fitting the 2D Affine Transformation Matrix and Calculating Kinematic Parameters. After mismatch removal, the vector analysis module fits the 2D affine transformation matrix and calculates the kinematic parameters, including longitudinal velocity, lateral drift velocity, and yaw angle. The 2D affine transformation matrix mathematically contains a translation parameter vector and a linear deformation parameter matrix. The vector analysis module extracts the longitudinal and lateral pixel displacements from the translation parameter vector and converts them into spatial physical displacements using the camera-calibrated physical pixel equivalents.

[0070] Subsequently, the vector analysis module subtracts the displacement increment of the system's low-frequency mechanical vibration dynamic compensation vector, which is synchronously extracted during the current control cycle, from the spatial physical displacement. Then, it removes the physical displacement after canceling vibration interference and calculates the longitudinal motion velocity by the time interval between two adjacent frames. and lateral drift speed Simultaneously, the vector analysis module decomposes the rotation component from the linear deformation parameter matrix and calculates the yaw angle parameter of the protective film relative to the conveying direction using inverse trigonometric functions. .

[0071] The longitudinal motion velocity was calculated. Lateral drift speed and yaw angle parameters Then, based on the lateral drift velocity, the vector analysis module performs a time-discrete integral accumulation operation from the moment the visual null coordinate system was established or the moment of the last forced coordinate reset to obtain the total lateral visual displacement relative to the visual null coordinate system. Total lateral visual displacement The mathematical model for cumulative operations satisfies the following formula:

[0072] ;

[0073] in, Indicates the total lateral visual displacement. This indicates the total number of control cycles since the last forced coordinate reset. Indicates the first Lateral drift velocity per control cycle Indicates the first The time interval of each control cycle. As the system runs, This will be continuously accumulated to fully record the absolute displacement trajectory of the protective film across the entire time domain as it deviates from the visual origin.

[0074] S23. Forced coordinate reset and reference frame refresh logic based on confidence level and accumulation period. Because long-term continuous integration generates accumulated errors, when the confidence level degradation condition is triggered or the accumulation period reaches its maximum, the absolute position calibration data from the absolute position sensor is used to force a coordinate reset and refresh the initial zero-position reference frame as a new reference for subsequent optical flow tracking calculations. During the specific execution state diagnosis to determine whether the confidence level degradation condition has been triggered, the vector analysis module calculates the proportion of interior points in each fitting process. Interior point ratio This is calculated by dividing the number of interior points corresponding to the two-dimensional affine transformation matrix by the total number of feature points participating in optical flow matching. The vector analysis module sets a security confidence threshold. Safety confidence threshold The value range is typically set between 0.6 and 0.85. When the calculated interior point ratio satisfies... When the condition is met, it is determined that the micro-texture features of the current operation cycle have failed over a large area, and the confidence downgrade condition is triggered.

[0075] When the confidence degradation condition is triggered, the vector resolution module pauses the optical flow parameter update for the current cycle, maintains and outputs the lateral drift velocity of the previous stable cycle. Alternatively, the feedback signal from the absolute position sensor can be switched for transitional control to avoid sending abnormal commands to downstream modules. Simultaneously, the vector analysis module uses an internal timer to accumulate degradation time. If the degradation time exceeds the set safety tolerance limit, a shutdown alarm signal is sent to the controller. The set safety tolerance limit is extracted offline based on the protective film's own tension tolerance limit and is typically set to 1 to 3 seconds to prevent conveyor belt failure caused by prolonged loss of visual feedback.

[0076] When the confidence degradation condition is met, or the cumulative period of the time discrete integral reaches the preset maximum threshold, the vector analysis module pauses the cumulative calculation, and uses absolute position calibration data to overwrite and refresh the currently saved total lateral visual displacement. This process forces a reset of the coordinate system origin. Simultaneously, the vector analysis module sends a synchronization command to the visual acquisition module to capture the microscopic texture image of the current scene, refresh the initial zero-position reference frame, and restart a new round of sparse feature point extraction and tracking. Through this periodic intervention with physical boundary data, the system can effectively suppress the long-period drift defect caused by the pure optical flow algorithm.

[0077] See attached document Figure 4 When the parameter identification module executes steps S30, visual strain extraction, and dynamic parameter identification, the elastic modulus and Poisson's ratio of the polymer composite protective film, being a viscoelastic material, will nonlinearly drift with changes in the longitudinal traction tension and environment during actual operation. Relying on fixed static empirical parameters will introduce calculation errors. To achieve online identification of constrained dynamic mechanical parameters, the parameter identification module specifically executes the following sub-steps:

[0078] S31. Displacement Gradient Matrix Extraction and Visual Strain Extraction Based on 2D Affine Transformation Matrix. Upon receiving the 2D affine transformation matrix, the parameter identification module extracts the translational motion components from the matrix to obtain the displacement gradient matrix. Since the overall rigid translation only changes the spatial position of the protective film without altering the relative spacing of the internal microtextures, the separately extracted displacement gradient matrix can independently characterize the relative deformation of the local microtextures. Based on the displacement gradient matrix, the parameter identification module uses partial derivatives to calculate and extract the visual strain, which includes visual longitudinal strain and visual transverse strain.

[0079] Specifically, assuming the two-dimensional continuous displacement field of sparse feature points in the micro-texture includes a longitudinal displacement field and a transverse displacement field, the visual longitudinal strain is the partial derivative of the longitudinal displacement field with respect to the longitudinal spatial coordinates, used to characterize the relative elongation of the protective film along the traction direction; the visual transverse strain is the partial derivative of the transverse displacement field with respect to the transverse spatial coordinates, used to characterize the relative shrinkage of the protective film perpendicular to the traction direction. Through partial derivative calculations, the discrete relative displacement of pixels is transformed into a continuous strain distribution feature.

[0080] In specific mathematical operations, a two-dimensional affine transformation matrix involving translation and deformation can be decomposed into a translation vector and a 2×2 linear deformation parameter matrix. In this case, the diagonal elements of the linear deformation parameter matrix after removing the translational components represent the scaling ratio in the spatial coordinate system. (Visual longitudinal strain) The value is equal to the element on the main diagonal corresponding to the longitudinal direction in the linear deformation parameter matrix minus 1; visual lateral strain The value is equal to the element on the main diagonal of the corresponding horizontal direction in the linear deformation parameter matrix minus 1. Through the above algebraic mapping, the system realizes the transformation and extraction of strain parameters from pure visual affine operations to continuous medium mechanics.

[0081] For the specific mathematical principles of calculating the strain of a spatial displacement field using partial derivatives, those skilled in the art can refer to textbooks on continuum mechanics. The principles are well-known in the field and will not be elaborated here.

[0082] S32. Extraction and Calculation of Longitudinal Mechanical Stress. Simultaneously with extracting visual strain, the parameter identification module acquires the longitudinal traction tension data from the tension sensor to calculate the longitudinal mechanical stress. Combining this with the known cross-sectional geometry of the protective film, the parameter identification module maps the longitudinal traction tension data into a force per unit cross-section. The specific calculation logic is as follows: Longitudinal Mechanical Stress... The value is equal to the longitudinal traction tension data. Divide by the initial cross-sectional area of ​​the protective film ,Right now initial cross-sectional area The initial width of the protective film With the thickness of the protective film The product of, i.e. Initial width With thickness All of these are pre-stored in the controller's memory as fixed process parameters.

[0083] S33. Dynamic Mechanical Parameter Fusion Extraction with Additional Thresholding and Limiting. After obtaining the longitudinal mechanical stress and visual strain, the parameter identification module calculates the longitudinal mechanical stress and visual longitudinal strain to extract the dynamic elastic modulus. Then, it fuses the visual transverse strain, visual longitudinal strain, and longitudinal mechanical stress to extract the dynamic Poisson's ratio. During the calculation process, an upper limit for visual strain noise is added as a threshold stabilization term for limiting. The final output is the dynamic mechanical parameters, including the dynamic elastic modulus and dynamic Poisson's ratio. To address the nonlinear drift problem of Poisson's ratio in polymer materials under different traction tensions and to prevent ratio divergence under industrial high-frequency micro-vibration conditions, an identification calculation model including stress coupling terms and threshold stabilization terms is configured. The identification calculation model satisfies the following formula:

[0084] ;

[0085] ;

[0086] in, Represents the dynamic elastic modulus; Indicates the dynamic Poisson's ratio; Indicates longitudinal mechanical stress; Indicates visual longitudinal strain; Indicates visual lateral strain; Indicates the safety factor; Indicates the upper limit of visual strain noise; Indicates the stress coupling coefficient; Indicates the lower bound of the elastic modulus; Indicates the upper bound of the elastic modulus; Indicates the lower bound of Poisson's ratio; Indicates the upper bound of Poisson's ratio; This represents the amplitude limiting function; This represents the threshold stability term; This indicates a stress coupling correction term; This represents the stable denominator term that incorporates the upper limit of noise. This represents the calculation item for the basic elastic modulus; This represents the basic Poisson's ratio calculation term.

[0087] In the specific implementation of the identification operation model, the amplitude limiting processing function is used to limit the basic Poisson's ratio calculation term and the basic elastic modulus calculation term within the set boundary parameter range. When the value of the basic calculation term is less than the lower bound, the lower bound value is forcibly output; when the value of the basic calculation term is greater than the upper bound, the upper bound value is forcibly output. The absolute value operator takes the absolute value of the visual longitudinal strain to prevent the reverse oscillation from causing the division into zero anomalies caused by the threshold stability term.

[0088] The safety factor is set based on the system's requirements for environmental vibration immunity, and is usually in the range of 1.2 to 1.5. The upper limit of visual strain noise is obtained during the system initialization calibration stage. The threshold stability term formed by the upper limit of visual strain noise and the safety factor is used to ensure the numerical stability of the stable denominator term of the combined noise upper limit.

[0089] The stress coupling coefficient is used to characterize the nonlinear correction gain of the longitudinal mechanical stress on the transverse deformation capability of the protective film under tension. The specific value of the stress coupling coefficient is obtained by conducting offline tensile tests on the same batch of materials and using the least squares method to perform polynomial curve fitting, thereby realizing the effective integration of longitudinal mechanical stress and bidirectional visual strain in the mathematical model.

[0090] The lower and upper bounds of the elastic modulus, Poisson's ratio, and other elastic modulus are set based on the median values ​​in the manufacturer's property data sheet for the protective film material. These are then broadened by a 10% to 20% fluctuation based on the offline calibration results of the tensile testing machine. These broadened lower and upper bounds serve as the execution boundaries of the amplitude limiting function, thereby filtering and blocking illegal mechanical values ​​generated by random disturbances such as external abnormal impacts. Finally, the parameter identification module continuously outputs the calculated and extracted dynamic mechanical parameters to the error decoupling module.

[0091] See attached document Figure 5 The error decoupling module, during the execution of steps S40, tension contraction decoupling, and deviation error extraction, primarily eliminates lateral contraction interference caused by the protective film under tension, thereby separating the mechanical deviation physical quantity from the visual data containing deformation. The error decoupling module specifically executes the following sub-steps:

[0092] S41. Real-time estimation of dynamic lateral shrinkage. In the traction slitting process, when the protective film is stretched by longitudinal traction tension, according to the Poisson effect, the protective film will shrink in physical width laterally. If the inward shrinkage of the protective film edge or texture is directly misjudged as the overall lateral deviation of the protective film, it will cause the correction actuator to generate inappropriate correction compensation. To eliminate deformation interference, the error decoupling module calculates the shrinkage change in real time based on mechanical laws. Since the protective film will undergo local physical property changes due to factors such as equipment friction heating and long-term tensile fatigue during continuous traction, using fixed static coefficients to estimate the deformation will introduce a fundamental calculation deviation.

[0093] To accurately estimate the deformation, the error decoupling module acquires the dynamic mechanical parameters continuously output by the parameter identification module, as well as the longitudinal mechanical stress calculated based on the longitudinal traction tension data measured by the tension sensor. The dynamic mechanical parameters include the dynamic elastic modulus and the dynamic Poisson's ratio. The error decoupling module obtains the pre-stored initial width of the protective film according to its production specifications and, combined with the longitudinal mechanical stress, dynamic elastic modulus, and dynamic Poisson's ratio, constructs a lateral deformation estimation model to estimate the lateral shrinkage of the protective film at the current moment. The lateral deformation estimation model satisfies the following formula:

[0094] ;

[0095] in, Indicates the amount of lateral contraction; Indicates the dynamic Poisson's ratio; Indicates longitudinal mechanical stress; Represents the dynamic elastic modulus; Indicates the initial width of the protective film; This is used to calculate the theoretical longitudinal strain value under the current stress state. For the fundamental theory of lateral dimensional deformation caused by the Poisson effect, those skilled in the art can refer to the materials mechanics handbook for verification; this is well-known technology in the field and will not be elaborated upon here.

[0096] S42. Decoupling of deviation error based on asymmetric deformation distribution. Obtaining lateral shrinkage. Then, the lateral shrinkage amount needs to be... Peeling away from the total lateral visual displacement. Ideally, when the protective film is stretched and narrowed on the guide rollers, it typically contracts symmetrically inwards with respect to the geometric center line. However, in actual production line environments, due to uneven distribution of friction on the guide roller system surfaces and deviations in the parallelism of the mechanical shaft system installation, the lateral contraction of the protective film exhibits asymmetrical characteristics.

[0097] The error decoupling module incorporates a shrinkage distribution coefficient derived from equipment process calibration. This coefficient characterizes the physical proportion of the actual shrinkage at the industrial camera monitoring side to the overall lateral shrinkage. The error decoupling module utilizes this shrinkage distribution coefficient in conjunction with the lateral shrinkage amount. The unilateral lateral contraction compensation displacement component with directional polarity is obtained, and this component is subtracted from the total lateral visual displacement to separate the deviation error unaffected by tension contraction. The mathematical model for decoupling the deviation error satisfies the following formula:

[0098] ;

[0099] in, This indicates the deviation error, which characterizes the lateral mechanical offset distance of the protective film relative to its ideal center trajectory. Indicates the total lateral visual displacement; Indicates the shrinkage distribution coefficient; Indicates the amount of lateral contraction; Indicates directional compensation polarity; This indicates the actual contraction displacement on the monitoring side of the industrial camera; This indicates a unilateral lateral contraction compensation displacement component with attached directional polarity.

[0100] Shrinkage distribution coefficient The frictional resistance distribution of the actual guide roller system is extracted through offline calibration. The specific calibration method is as follows: with the equipment stopped and no tension established, equidistant longitudinal reference baselines are drawn on both sides of the protective film surface; then the equipment is started and rated working tension is applied. By measuring the actual relative displacement of the two reference baselines towards the center, the ratio of the shrinkage displacement on the industrial camera monitoring side to the total shrinkage displacement on both sides is calculated, thereby determining... The specific value of the contraction distribution coefficient. The value ranges from 0 to 1. When the protective film exhibits ideal symmetrical shrinkage, the shrinkage distribution coefficient... The value is 0.5; when lateral sliding constraints cause uneven contraction, the contraction distribution coefficient is... Deviation of 0.5.

[0101] Directional compensation polarity The value is either 1 or -1, determined by the installation position of the industrial camera and the definition of the positive direction of the image coordinate system. When the relative displacement direction of the texture within the industrial camera's field of view due to the shrinkage of the protective film coincides with the positive direction set by the mechanical misalignment, the direction compensation polarity... Set to 1; when the relative displacement direction is opposite to the positive direction set by the mechanical deviation setting, the direction compensation polarity is... Take -1.

[0102] The error decoupling module filters out the viscoelastic dimensional changes of the protective membrane by algebraically subtracting the total lateral visual displacement from the deformation components with accompanying polarity and asymmetry. The extracted deviation error is input as displacement deviation data to the composite control module, providing a reference for subsequent servo closed-loop control.

[0103] See attached document Figure 6 The composite control module, when executing the S50 and time-delay composite deviation correction control steps, primarily addresses the dual physical and temporal lag caused by the sensor's front-end positioning. In actual production lines, the industrial camera's field of view is typically installed upstream of the slitting blade assembly, with a physical distance between them. The combined time required for image acquisition, algorithm computation, and servo mechanism delays means that by the time the control system is ready to execute deviation correction, the protective film carrying the deviation error has already been moved forward the corresponding distance. To achieve decoupling between prediction and feedback, and precise action coordination, the composite control module specifically executes the following sub-steps:

[0104] S51. System Time Consumption Assessment and Blade Arrival Delay Calculation. The composite control module acquires the set physical distance between the industrial camera and the slitting blade assembly, and receives the longitudinal motion speed continuously calculated and output by the vector analysis module. The physical distance is usually measured during equipment installation or maintenance using a laser rangefinder or a high-precision mechanical ruler along the protective film's transport path, and is used as a constant input to the controller. In industrial web guiding systems, there is an inherent time delay from image exposure to mechanical completion; therefore, the composite control module needs to dynamically assess system time consumption in real time. System time consumption specifically includes camera exposure time, algorithm processing time, and servo response time.

[0105] The camera exposure time is directly read from the hardware shutter time parameter configured in the industrial camera's underlying driver; the algorithm processing time is calculated by the internal system's beat timer during the current control cycle when the controller performs image feature tracking and deviation error calculation; the servo response time is extracted based on a comprehensive evaluation of the current loop setup time and position loop tuning time within the servo driver of the correction actuator. Furthermore, to prevent mechanical oscillations caused by insufficient action time in the servo mechanism, the system forcibly incorporates a time response dead zone threshold into the servo response time evaluation. The time response dead zone threshold characterizes the shortest physical response time required for the correction actuator to overcome static friction and generate effective displacement, typically ranging from 10 to 30 milliseconds.

[0106] Therefore, the extracted servo response time is stored in the system as an empirical calibration constant. To prevent calculation overflow due to excessively low speed during equipment start-up and shutdown, the composite control module compares the longitudinal motion speed with a preset minimum safe speed threshold and takes the larger value. This preset minimum safe speed threshold is determined based on the minimum stable continuous operating linear speed of the traction shaft servo motor and the lower limit of the overflow prevention accuracy for the controller's underlying floating-point division operation. The composite control module uses the larger value as the processed longitudinal motion speed in the denominator calculation. The composite control module divides the spatial physical distance by the processed longitudinal motion speed and subtracts the system time consumption to obtain the tool edge arrival delay. The mathematical model for calculating the tool edge arrival delay satisfies the following formula:

[0107] ;

[0108] in, This indicates the blade arrival delay, used to characterize the effective available action preparation time for the physical point of the protective film observed in the field of view of the industrial camera to reach the cutting position of the slitting blade assembly; This indicates the physical spatial distance between the industrial camera and the slitting blade assembly. This indicates the longitudinal velocity after processing; Indicates the camera's exposure time; Indicates the processing time of the algorithm; Indicates the servo response time; This represents the overall system time consumption.

[0109] S52. Independent generation of closed-loop feedback commands and feedforward prediction commands. The composite control module obtains the deviation error separated by the error decoupling module, uses the deviation error as input, runs the proportional-integral-derivative (PID) control algorithm, and calculates the closed-loop feedback command. The parameter tuning principles for the proportional gain, integral time, and derivative time in the PID control algorithm are well-known in the field and will not be elaborated further here. Those skilled in the art can refer to the automatic control system design manual for configuration and deployment; these are common techniques in the field and will not be elaborated further.

[0110] While generating the closed-loop feedback command, the composite control module acquires the yaw angle parameters and lateral drift velocity calculated by the vector analysis module. Since the protective film not only exhibits parallel offset but also tilts with yaw angle parameters, it will continue to generate predictable lateral displacement increments within the knife edge arrival delay. The composite control module utilizes the extracted lateral drift velocity and yaw angle parameters, combined with the knife edge arrival delay, to construct a kinematic feedforward prediction model. Because the lateral translation compensation of the protective film and the deflection compensation of the guide roller are mechanically decoupled, and the lateral drift velocity itself already comprehensively includes all actual lateral relative sliding increments at the physical level of optical flow analysis, to prevent kinematic double-count overshoot, the kinematic feedforward prediction model is decomposed into independent lateral position feedforward and yaw angle feedforward, satisfying the following formula:

[0111] ; ;

[0112] in, This indicates a lateral position feedforward prediction instruction. This indicates the yaw angle feedforward prediction command. Indicates the feedforward gain coefficient for lateral position; This represents the feedforward gain coefficient for the yaw angle; Indicates the lateral drift velocity; Indicates the yaw angle parameter; Indicates the time delay before the blade arrives; This represents the feedforward predicted lateral displacement deviation within the knife edge arrival time delay. The aforementioned feedforward gain coefficients are used to adjust the weight ratio of the predicted compensation amount in the overall control output to prevent mechanical overshoot caused by relying solely on mathematical model predictions. The values ​​of each feedforward gain coefficient are typically set between 0.3 and 0.7. The specific values ​​of each feedforward gain coefficient are empirically tuned during the equipment commissioning phase based on step response tests conducted under no-load conveyor belt operation.

[0113] S53. Composite control path switching and servo drive command output based on time delay status. After calculating the tool edge arrival delay, the composite control module performs logical judgment on the value of the tool edge arrival delay and executes the corresponding control strategy branch.

[0114] When the blade arrival delay is positive, it indicates that the protective film offset observed by the current vision system has not yet reached the physical position of the slitting blade group. Furthermore, since the time response dead zone has been eliminated in the servo response time, the correction actuator has sufficient time margin to perform motion compensation. At this point, the composite control module combines the yaw angle parameters and lateral drift velocity to generate a feedforward prediction command, which is then superimposed with the closed-loop feedback command generated based on the deviation error, and outputs a servo drive command to the correction actuator.

[0115] The specific superposition logic is as follows: the system sets the closed-loop feedback command output by the proportional-integral-derivative control algorithm as the basic displacement compensation amount. The final integrated lateral servo drive command output to the correction actuator is set as Then the system performs linear algebraic sum operations, that is... It also sends yaw angle feedforward prediction commands through parallel independent communication channels. The above-mentioned comprehensive lateral servo drive commands With yaw angle feedforward prediction command Together, they constitute the servo drive commands output to the correction actuator.

[0116] The correction mechanism can be a ball screw slide base driven by an AC servo motor or a thrust cylinder. The servo drive command includes both the basic displacement component required for translation and the yaw correction command corresponding to the yaw angle. The servo drive command drives the correction mechanism to translate or deflect the guide roller system, thereby driving the running axis of the protective film to generate a lateral mechanical correction displacement. This allows for pre-positioning relative to the slitting blade group before the physical part of the protective film with the deviation error reaches the slitting position of the slitting blade group.

[0117] When the blade arrival delay is non-positive, it indicates that the protective film conveyor speed is too fast, or system fluctuations are causing the algorithm processing time to increase, resulting in the currently observed offset about to pass or already passing the cutting blade group. In this state, continuing to inject feedforward prediction commands will cause the mechanical system to produce delayed and useless correction actions, leading to control overshoot and belt instability in the correction frame. To avoid system response overshoot, the composite control module actively cuts off the feedforward prediction path, forcing the feedforward gain coefficient to zero. At this point, the servo drive commands no longer include yaw correction commands, and the system independently controls the correction actuator to generate lateral mechanical correction displacement based on closed-loop feedback commands to complete the protective film cutting and positioning.

[0118] Furthermore, the system control logic and data flow processing described above, from motion vector analysis and parameter identification decoupling to delay compensation command generation, are all continuously iterated within a set millisecond sliding window period in the controller's underlying operating system. This set millisecond sliding window period is matched to the highest effective acquisition frame rate interval of the industrial camera and the bus communication synchronization cycle of the correction actuator's servo driver, typically ranging from 1 to 5 milliseconds. Through this high-frequency iterative execution mechanism, the system maintains continuous displacement detection and dynamic positioning compensation feedback, thereby constructing a complete automated control closed loop that adapts to fluctuations in production line operating conditions.

[0119] See attached document Figure 7 To be continued Figure 10To aid in understanding the technical solution of this invention, a specific application example is provided below, taking into account a particular high-transmittance PET composite protective film slitting process.

[0120] In this application embodiment, the initial width of the protective film to be processed is preset with parameters. 1000 mm, thickness preset parameters The initial cross-sectional area is calculated to be 0.05 mm. The area is 50 square millimeters. The slitting machine is set to a linear speed of 120 meters per minute (i.e., a nominal longitudinal speed of 2000 millimeters per second), and the system's reference working tension is set to 150 Newtons. The physical spatial distance between the industrial camera and the slitting blade assembly is also specified. The measured value is 500 mm. The correction mechanism uses a ball screw slide base driven by an AC servo motor, and the slide base is equipped with a guide roller system with translational and yaw adjustment degrees of freedom. The specific implementation steps and logic are as follows:

[0121] During the system initialization and feature point acquisition phase, after the controller powers on, the vision acquisition module drives the industrial camera to perform exposure acquisition. Specifically, the industrial camera is a 5-megapixel global shutter industrial area array camera, and the dark field light source uses a low-angle ring LED array with an incident angle set to 15 degrees. Within a 3-second time window during which the tension of the slitting equipment reaches 150 Newtons but the traction motor is braked and stationary, the vision acquisition module continuously acquires micro-texture images, calculates the optical flow pseudo-displacement gradient between adjacent frames, extracts peak data within the 95% confidence interval, and sets the upper limit of visual strain noise. The calibration value is 0.00005.

[0122] As the traction motor starts, when the infrared edge sensor confirms that the edge of the protective film is within the ±2 mm physical center tolerance zone for 2 consecutive seconds, the vision acquisition module captures the micro-texture image and extracts the current image features as the initial zero-position reference frame. Simultaneously, it reads the absolute position calibration data from the absolute position sensor. The system binds the absolute position calibration data to the image space center point of the initial zero-position reference frame, thereby establishing a visual zero-position coordinate system with physical scale significance. Subsequently, the vision acquisition module plans an 800×800 pixel region of interest in the central area of ​​the micro-texture image, avoiding the edge highlight band, and runs the Shi-Tomasi corner detection algorithm to extract a set of 100 sparse feature points with strong bidirectional gradients, which are then input into the subsequent computation.

[0123] During micro-texture tracking and kinematic parameter calculation, the vector analysis module acquires the previous and current image frames and performs pixel displacement calculations on a set of 100 sparse feature points based on the pyramid Lucas-Cannard optical flow algorithm. To eliminate mismatches caused by local wrinkles, the system executes a random sampling consensus algorithm with a preset pixel distance threshold of 1.0 pixel, iteratively fitting the optimal two-dimensional affine transformation matrix. The vector analysis module extracts the translational displacement from the two-dimensional affine transformation matrix, subtracts the displacement increment of the temporarily stored system low-frequency mechanical vibration dynamic compensation vector, and divides it by the camera's set 2-millisecond sampling period (corresponding to a 500Hz frame rate) to calculate the current longitudinal motion velocity in real time. (Approximately 2000 mm / s), Lateral drift speed and yaw angle parameters .

[0124] Meanwhile, the vector analysis module measures the lateral drift velocity. Execution time discrete integration, cumulative calculation of total lateral visual displacement During this process, if the calculated proportion of interior points is lower than the safety confidence threshold of 0.65, the micro-texture features of the current computation cycle are considered to have largely failed. At this point, a confidence degradation condition will be triggered, and the system will maintain the confidence level from the previous stable cycle. Output and enable an internal timer to accumulate degradation time to prevent malfunctions caused by dirt or scratches on the membrane surface.

[0125] For mechanical feature transformation and dynamic identification of constrained parameters, the parameter identification module receives a two-dimensional affine transformation matrix, extracts a 2×2 linear deformation parameter matrix after stripping the translation vector, and then extracts the values ​​of the main diagonal elements minus 1 according to the definition of partial derivatives to obtain the visual longitudinal strain. Visual lateral strain The system synchronously reads data from the tension sensor. (Fluctuating between 140 and 160 Newtons due to mechanical resonance), divided by area Longitudinal mechanical stress is obtained .

[0126] Subsequently, the parameter identification module runs a dynamic parameter fusion model with additional thresholds. A safety factor is then set based on the properties of PET material. The threshold stability term is derived by combining the upper limit of visual strain noise. The lower and upper bounds of the elastic modulus are set to 3.0 × 10⁻⁶. 3 Megapascals and 4.5 × 10 3 The lower and upper bounds for Poisson's ratio are set to 0.35 and 0.45, respectively. After substituting the constraints into the formula, the dynamic elastic modulus (MPa) and dynamic Poisson's ratio are calculated and output online at the current millisecond level. (See attached...) Figure 7 With appendix Figure 8 As shown, this process constitutes a complete online identification process for dynamic mechanical parameters. (See attached diagram.) Figure 7 In the middle section, time (s) is plotted on the horizontal axis and traction tension (N) on the vertical axis, recording the longitudinal traction tension fluctuation curve during system operation; while in the attached section... Figure 8 Similarly, using time (s) as the horizontal axis, the convergence curve of dynamic mechanical parameter identification is shown when the dynamic parameters tend to stabilize.

[0127] Based on the independent decoupling stage of mechanical deviation based on asymmetric coefficients, the error decoupling module uses the extracted dynamic parameters and stress state to substitute into the lateral deformation estimation model. The overall lateral shrinkage of the protective film was calculated. Due to a slight difference in bearing damping on one side of the on-site guide roller, the shrinkage distribution coefficient on the industrial camera side was extracted after preliminary offline calibration. The value is 0.45. The system compensates for polarity in a set direction. Substituting into the error model, the total lateral visual displacement including deformation is calculated. The displacement component caused by unilateral lateral contraction due to tension is precisely isolated, and the deviation error unaffected by the tension contraction amplitude is finally obtained. Combined with the appendix Figure 9 According to the records regarding tension amplitude reduction decoupling and deviation error, the system successfully generated a comparison curve for tension amplitude reduction decoupling and deviation error extraction, clearly separating the total lateral visual displacement. Unilateral contraction compensation component and actual deviation error .

[0128] In the time consumption assessment and feedforward composite correction execution stage, the composite control module first assesses the system time consumption. The camera exposure time is read as 0.5 milliseconds, and the algorithm processing time is 1.5 milliseconds. Combining the servo controller current and position loop tuning time with the minimum physical response time required to overcome static friction (the time response dead zone threshold is set to 15 milliseconds), the servo response time is extracted to be 20 milliseconds. The total system time consumption is 22 milliseconds. Next, by dividing the spatial physical distance of 500 mm by the current longitudinal movement speed of 2000 mm / s, the basic spatial transmission time is calculated to be 250 milliseconds. Subtracting the total system time consumption of 22 milliseconds, the actual blade arrival delay is calculated. It takes 228 milliseconds.

[0129] Because the arrival delay is greater than 0, the composite prediction condition is met. The system operates using a proportional-integral-derivative (PID) control algorithm based on... The closed-loop feedback command is used. Simultaneously, the feedforward gain coefficients for lateral position and yaw angle are extracted, and the lateral drift velocity is... Multiplying by 228 milliseconds yields the predicted offset, generating a lateral position feedforward prediction command. The system performs a linear algebraic summation on the two sets of lateral commands and combines them with the independently generated yaw angle feedforward prediction command to form a comprehensive servo drive command for issuance. The servo motor drives the lead screw to move the guide roller system, generating a mechanical correction displacement before the offset part completes the distance traveled. (Combined with...) Figure 10 The system demonstrates the time-delay composite correction control response by plotting a comparison curve with time (s) on the horizontal axis and the normalized correction execution displacement on the vertical axis. This effectively addresses mechanical deviation step disturbances.

[0130] For the parameter tuning principle in the proportional-integral-derivative control algorithm, and the construction and iterative solution process of the image pyramid in the pyramid Lucas-Cannard optical flow algorithm, those skilled in the art can refer to modern control theory and digital image processing related technical manuals for configuration and deployment. Their principles and applications are well-known technologies in this field and will not be elaborated here.

[0131] The technical verification conclusions of the present invention will be elaborated in detail based on the above specific application embodiments and accompanying data.

[0132] From the appendix Figure 7 With appendix Figure 8 Data characteristics reveal that in industrial roll-to-roll conveying, traction tension undergoes a dynamic build-up process upon startup, along with unavoidable low-frequency pulsations and moderate resonance. Characterizing the stress deformation of polymer films using traditional static mechanical parameters results in severe distortion. This invention, through online coupling of optical flow visual strain and sensor mechanical stress, identifies the dynamic elastic modulus and dynamic Poisson's ratio in real time under the constraints of amplitude limiting rules and noise thresholds. During the initial tension build-up and fluctuation phases, the parameters exhibit certain damping characteristics and converge rapidly within a short time, smoothing out parameter drift caused by the nonlinear viscoelastic properties of the material, and providing rigorous physical quantity support for subsequent peeling of spurious deformations.

[0133] From the appendix Figure 9 Data characteristics reveal that the unprocessed visual displacement curve exhibits a complex pattern, blending with mechanical offset. When tension fluctuates, the material's Poisson's contraction effect manifests in the visual image as edge shrinkage or texture shifting that closely resembles actual mechanical misalignment. If the controller directly uses this as the correction input, the system will execute incorrect compensation actions. This system, based on extracted dynamic mechanical parameters and combined with an asymmetric contraction distribution coefficient that reflects on-site physical constraints, performs decoupling. The decoupled compensation displacement component shown in the figure accurately reflects the degree of tension narrowing. Subtracting this component eliminates the coupling burrs with tension fluctuations in the resulting true error curve, restoring the pure mechanical off-center loading trajectory.

[0134] From the appendix Figure 10Data characteristics show that this solution exhibits excellent dynamic control performance in response to the spatial physical distance delay effect caused by the front-mounted sensor. Faced with step disturbances, conventional hysteresis feedback control responses often exhibit significant phase lag, slow tracking, and subsequent position overshoot because the actual offset has already passed the cutting edge by the time the system detects the deviation and drives the servo. In contrast, the predictive composite control response of this invention calculates the actual cutting edge arrival delay by dynamically subtracting the dead zone time of the system's hardware and software actions and introducing feedforward prediction correction. This composite superposition strategy allows the actuator to enter the response trajectory earlier, significantly reducing response lag time and suppressing overshoot when approaching the target position, ensuring the smoothness and high precision of the protective film cutting and positioning boundary at high linear speeds.

Claims

1. A protective film cutting and positioning system based on visual correction, characterized in that, include: An industrial camera, which is installed upstream of the slitting blade assembly in conjunction with a dark field light source; The controller uses a vision acquisition module to control the industrial camera to synchronously and continuously acquire micro-texture images with the dark field light source. After establishing a visual zero-position coordinate system, it extracts a set of sparse feature points from the micro-texture images. The sparse feature point set is obtained using the vector analysis module, and a two-dimensional affine transformation matrix is ​​fitted to it by optical flow tracking operation in consecutive image frames to solve the kinematic parameters of the protective film and the total lateral visual displacement relative to the visual null coordinate system. The kinematic parameters and the total lateral visual displacement are obtained by using the composite control module. The blade arrival delay is calculated by combining the spatial physical distance between the industrial camera and the slitting blade group and the system time consumption. Based on the positive and negative state of the blade arrival delay, a servo drive command that combines the feedforward prediction command and the closed-loop feedback command is selectively output to the correction actuator, or only the closed-loop feedback command is output. The correction actuator drives the operating axis of the protective film to generate a lateral mechanical correction displacement according to the servo drive command.

2. The protective film cutting and positioning system based on visual correction according to claim 1, characterized in that, It also includes an absolute position sensor. During the system initialization and calibration phase, when the cutting equipment establishes tension and is in a non-traction static state, the controller also uses the vision acquisition module to control the industrial camera to continuously acquire the micro-texture image and calibrate the upper limit of the visual strain noise corresponding to the inherent high-frequency mechanical vibration of the system. Once the protective film enters a stable traction and centered operating state, the controller uses the visual acquisition module to capture the micro-texture image to extract the current image features as the initial zero-position reference frame. The absolute position calibration data of the absolute position sensor is data-bound with the initial zero-position reference frame to establish the visual zero-position coordinate system. Then, the gradient covariance matrix of the image pixels is scanned and calculated within the region of interest planned in the pixel matrix of the micro-texture image, and pixels with strong bidirectional gradients are extracted as the sparse feature point set.

3. The protective film cutting and positioning system based on visual correction according to claim 2, characterized in that, The process by which the visual acquisition module calibrates the upper limit of the visual strain noise includes: After extracting displacement gradient data between adjacent frames of the microtexture image and performing time series statistical analysis, the peak value or root mean square value of the displacement gradient within a set confidence interval is calculated as the upper limit of the visual strain noise.

4. The protective film cutting and positioning system based on visual correction according to claim 2, characterized in that, The process by which the vector analysis module performs optical flow tracking operations on the sparse feature point set in the continuous image frames to fit the two-dimensional affine transformation matrix, and calculates the kinematic parameters of the protective film and the total lateral visual displacement, includes: Sparse optical flow tracking calculation is performed on the sparse feature point set of two adjacent frames of microtexture images in the continuous image frames, and a random sampling consensus algorithm is run to remove the set of mismatched points. After mismatch elimination, the kinematic parameters, including longitudinal velocity, lateral drift velocity, and yaw angle parameters, are calculated by fitting the two-dimensional affine transformation matrix. Then, based on the lateral drift velocity, a cumulative calculation is performed to obtain the total lateral visual displacement.

5. The protective film cutting and positioning system based on visual correction according to claim 4, characterized in that, During the optical flow tracking operation, the vector analysis module calculates the proportion of inliers in each fitting process by dividing the number of inliers corresponding to the two-dimensional affine transformation matrix by the total number of feature points participating in optical flow matching. When the proportion of inliers is less than the safety confidence threshold, a confidence downgrade condition is triggered. When the confidence degradation condition is triggered or the cumulative period of the time discrete integral reaches the preset maximum threshold, the vector analysis module calls the absolute position calibration data to overwrite the currently saved total lateral visual displacement to complete the forced reset of the coordinate system origin, and triggers the visual acquisition module to synchronously capture the microscopic texture image of the current scene to refresh the initial zero-position reference frame.

6. The protective film cutting and positioning system based on visual correction according to claim 2, characterized in that, It also includes a tension sensor, and the controller is further equipped with a parameter identification module and an error decoupling module; The controller uses the parameter identification module to extract visual strain from the two-dimensional affine transformation matrix, combines it with the longitudinal traction tension data of the tension sensor to calculate longitudinal mechanical stress, and fuses and extracts dynamic mechanical parameters. The controller uses the error decoupling module to obtain the dynamic mechanical parameters to calculate the lateral shrinkage of the protective film, eliminates the non-positional offset caused by the material's own deformation, and separates the deviation error corresponding to the whole roll offset from the total lateral visual displacement. The composite control module takes the deviation error as input, runs the proportional-integral-derivative control algorithm, and calculates the closed-loop feedback command.

7. The protective film cutting and positioning system based on visual correction according to claim 6, characterized in that, The visual strain includes visual longitudinal strain and visual transverse strain. The dynamic mechanical parameters include dynamic elastic modulus and dynamic Poisson's ratio. The parameter identification module extracts the visual strain and calculates the longitudinal mechanical stress by combining it with the longitudinal traction tension data. The process of fusing and extracting the dynamic mechanical parameters includes: The translational motion components are stripped from the two-dimensional affine transformation matrix to extract the displacement gradient matrix, and the visual longitudinal strain and visual transverse strain are calculated using partial derivatives. The initial cross-sectional area is obtained by combining the initial width and thickness of the protective film, and the longitudinal traction tension data is divided by the initial cross-sectional area to obtain the longitudinal mechanical stress. A recognition calculation model is constructed using the visual longitudinal strain, the visual transverse strain, and the longitudinal mechanical stress. The upper limit of visual strain noise is introduced into the identification operation model as a threshold stabilization term, forming a stable denominator term that combines the upper limit of noise. The dynamic elastic modulus and the dynamic Poisson's ratio are output by performing amplitude limiting processing through the amplitude limiting processing function.

8. The protective film cutting and positioning system based on visual correction according to claim 7, characterized in that, The process by which the error decoupling module separates and obtains the deviation error includes: A lateral deformation estimation model is constructed by combining the longitudinal mechanical stress, the dynamic elastic modulus, the dynamic Poisson's ratio, and the initial width of the protective film, and the lateral shrinkage is estimated. Obtain the shrinkage distribution coefficient and directional compensation polarity of the equipment process calibration, and combine the lateral shrinkage amount to obtain the unilateral lateral shrinkage compensation displacement component with directional polarity. The unilateral lateral contraction compensation displacement component is subtracted from the total lateral visual displacement to separate and obtain the deviation error that is not affected by the tension contraction amplitude.

9. A protective film cutting and positioning system based on visual correction according to claim 4, characterized in that, The process by which the composite control module calculates the blade arrival delay by combining the set spatial physical distance and the system time consumption includes: The longitudinal motion speed is compared with a preset minimum safe speed threshold, and the maximum value of the two is taken as the processed longitudinal motion speed. Divide the set spatial physical distance by the processed longitudinal motion speed and subtract the system time consumption to obtain the blade arrival delay.

10. A protective film cutting and positioning system based on visual correction according to claim 9, characterized in that, The process by which the composite control module outputs the servo drive command to the correction actuator includes: When the time delay of the blade edge is positive, the feedforward prediction command is generated by combining the yaw angle parameter and the lateral drift speed, and then superimposed with the closed-loop feedback command to output the servo drive command to the correction actuator. When the time delay of the blade arrival is non-positive, the feedforward prediction path is actively cut off, and the correction actuator is controlled to generate the lateral mechanical correction displacement solely based on the closed-loop feedback command.

Images