Oca defect linkage cutting method based on computer vision

Abstract

The application discloses a transparent optical adhesive defect linkage slitting method based on computer vision and belongs to the technical field of optical adhesive online detection and coiled material slitting control. In order to solve the problem that after the online identification of the transparent optical adhesive defect, it is difficult to accurately bind the length position in the roll, thereby causing the slitting to fail to automatically isolate the defect section, the application collects an image sequence, an encoder pulse sequence and a tension data sequence, aligns time stamps, obtains visual displacement by using interframe registration, constructs an amplification state estimation model containing length coordinates, slip parameters, elongation parameters and time delay parameters, outputs corrected length coordinates, generates a defect map, combines delay compensation, tension rebound compensation and safety margin to determine a target cutting point, and according to actual cutting point errors, updates compensation parameters online, realizes millimeter-level linkage slitting of defect section jump cutting, segmentation or marking, reduces missed detection and mis-cutting, and reduces the inflow of defects into downstream processes.

Description

Technical Field

[0001] This invention relates to the field of roll material slitting control, and more particularly to a computer vision-based OCA defect linkage slitting method. Background Technology

[0002] Transparent optical adhesives are widely used in the manufacturing processes of display module bonding, touch panels, and cover plates. Their production typically employs a continuous conveyor belt process for coating, curing, inspection, slitting, and rewinding. To reduce the rate of appearance defects, current technologies generally incorporate linear or area array cameras during the conveyor belt process, combined with backlighting, coaxial lighting, or dark-field light sources to achieve online visual inspection. Some production lines also introduce conveyor belt encoders, tension sensors, and speed closed-loop control to record the location of defects, trigger alarms, or perform manual marking. In the slitting stage, a common practice is to set cutting points based on the encoder's cumulative length or the roll length calculation results, and then manually pick out defective sections or rework them afterward.

[0003] However, the aforementioned existing technologies still have shortcomings:

[0004] 1. Transparent optical adhesives have many types of defects and low contrast. They are affected by light fluctuations, weak material texture, and contamination interference, resulting in insufficient stability in defect identification, high false detection and false negative rates, and most systems remain at the alarm level.

[0005] 2. The binding accuracy between the defect location and the length coordinates within the roll is insufficient. During the tape feeding process, there are slippage, elongation caused by tension fluctuations, and asynchronous and delayed sensor sampling. As a result, it is difficult to obtain a stable defect map by relying solely on encoder length measurement or simple correction, and it is difficult to automatically isolate defect segments after winding.

[0006] 3. The slitting and rewinding processes are subject to dynamic effects such as response delays and material springback. Existing cutting points often lack comprehensive compensation and error feedback correction mechanisms, making it difficult to implement high-precision skip cutting, segmentation, or marking based on defect maps. This can easily lead to defects flowing into downstream bonding processes, causing batch appearance defects and increased rework costs. Therefore, a computer vision-based OCA defect linkage slitting method is needed to address the shortcomings of existing technologies. Summary of the Invention

[0007] One objective of this invention is to propose a computer vision-based method for coordinated cutting of transparent OCA defects. Addressing the problems in existing technologies where online identification of transparent OCA defects is difficult to stably and accurately bind to the in-roll length position, and where the cutting and rewinding processes cannot achieve skip cutting, segmentation, or marking based on a defect map, leading to difficulties in automatically isolating defect segments, the following technical solution is proposed: OCA image sequences, encoder pulse sequences, and tension data sequences are acquired and timestamps are generated, aligning each timestamp to a unified time axis; the encoder calculates the uncorrected length coordinate sequence, and inter-frame registration is performed on adjacent frames to obtain a visual displacement sequence; based on the encoder length increment, tension data, and visual displacement, an amplified state estimation model is constructed, including conveyor length coordinates, slippage parameters, elongation parameters, and time delay parameters; delay compensation is performed on the visual displacement, and high-frequency prediction and low-frequency correction are implemented, outputting a corrected length coordinate sequence; defect identification is performed on the image sequence, and a defect map is generated; based on the defect map, a base cutting point is determined, and delay compensation, tension rebound compensation, and safety margin are superimposed to obtain the target cutting point; cutting control commands are output to coordinate the cutting and rewinding mechanisms, and the cutting point compensation amount is updated online according to the actual cutting point error. This invention has the technical effects of improving the accuracy of defect location length binding, realizing millimeter-level cutting point control and automatic isolation of defect segments, reducing erroneous cutting and missed cutting, and reducing the flow of defects into downstream processes.

[0008] This invention provides a computer vision-based method for linked segmentation of OCA defects, including:

[0009] S1. Acquire the OCA image sequence, encoder pulse sequence, and tension data sequence during the transparent OCA conveyor belt transport process, and generate timestamps for each. S2. Align each timestamp to a unified time axis and calculate the uncorrected length coordinate sequence from the encoder pulse sequence. S3. Perform inter-frame registration on adjacent OCA images to obtain a visual displacement sequence. S4. Construct an augmented state estimation model based on encoder length increment, tension data, and visual displacement, including conveyor belt length coordinates, slippage parameters, elongation parameters, and time delay parameters. Perform delay compensation on the visual displacement and output a corrected length coordinate sequence after prediction and correction. S5. Perform defect identification on the OCA image sequence to obtain defect information and map it to the corrected length coordinate sequence to generate a defect map. S6. Determine the basic cutting point based on the defect map and calculate the cutting point compensation amount, including performing delay compensation, tension rebound compensation, and safety margin. Superimpose these to obtain the target cutting point and output a slitting control command to link the slitting mechanism and the winding mechanism. S7. Acquire the actual cutting point and calculate the cutting point error, updating the cutting point compensation amount for subsequent slitting control command generation.

[0010] Optionally, S1 includes:

[0011] During the transparent OCA transport process, the camera continuously acquires and forms an OCA image sequence. Each time the camera acquires an image, the camera trigger timestamp corresponding to that image is recorded, resulting in an OCA image sequence with camera trigger timestamps.

[0012] The encoder pulse sequence output by the belt encoder is acquired synchronously, and the corresponding encoder timestamp is recorded for the pulse count value in the encoder pulse sequence to obtain the encoder pulse sequence with encoder timestamp;

[0013] The tension data sequence output by the tension sensor is acquired synchronously, and the corresponding tension timestamp is recorded for the tension sample value in the tension data sequence to obtain the tension data sequence with tension timestamp;

[0014] The camera trigger timestamp, encoder timestamp, and tension timestamp are generated based on the same clock source, or they are generated based on different clock sources respectively. The timestamps are converted to a unified time axis through time synchronization or through pre-calibrated clock bias parameters and clock drift parameters, and the timestamped OCA image sequence, encoder pulse sequence, and tension data sequence are output.

[0015] Optionally, S2 includes:

[0016] The camera trigger timestamp, encoder timestamp, and tension timestamp are mapped to a unified time axis to obtain an aligned encoder pulse sequence and an aligned tension data sequence. This mapping includes establishing a time correspondence using at least one of the following methods: precise time protocol synchronization, network time protocol synchronization, and hardware trigger synchronization. On the unified time axis, the difference between pulse counts at adjacent times in the aligned encoder pulse sequence is used to obtain an encoder length increment sequence. This pulse count difference is then converted into a length increment according to the length coefficient corresponding to the encoder pulse. The encoder length increment sequence is then integrated along the unified time axis to obtain an encoder length coordinate sequence, which serves as an uncorrected length coordinate sequence. This uncorrected length coordinate sequence represents the cumulative length position of the transparent OCA along the belt travel direction. Value rules are established on the unified time axis. At the prediction or correction time of the amplification state estimation, interpolation calculations are performed on the aligned encoder pulse sequence and the aligned tension data sequence to obtain the encoder length increment value and tension value corresponding to the prediction or correction time.

[0017] Optionally, S3 includes:

[0018] Adjacent OCA images in an OCA image sequence are selected as image pairs on a unified timeline.

[0019] The image pairs are preprocessed to reduce the impact of illumination fluctuations on inter-frame registration;

[0020] The inter-frame displacement of the image pair is calculated using at least one of the following methods: feature point matching, dense optical flow, phase correlation, and template matching. The component of the inter-frame displacement in the transport direction is converted into the corresponding inter-frame length displacement to obtain a visual displacement sequence corresponding to a unified time axis.

[0021] Based on the results of the inter-frame registration, the observation quality parameters are calculated, including one or more of the following: feature matching in-point ratio, correlation peak ratio, and inter-frame displacement consistency error.

[0022] Optionally, S4 includes:

[0023] Using encoder length increment sequence, tension data sequence, visual displacement sequence, and observation quality parameters as inputs, an amplified state estimation model is established, with the amplified state vector representing the true tape length coordinates, slippage rate, material elongation coefficient, and time delay. On a unified time axis, the encoder length increment sequence and tension data sequence are used as high-frequency prediction inputs. The encoder length increment sequence is corrected for slippage based on the slippage rate, and the length increment after slippage correction is compensated for elongation based on the material elongation coefficient and the tension data sequence, resulting in a predicted amplified state vector. On the unified time axis, the time delay is estimated, and the visual displacement sequence is compensated for delay to obtain a delayed-compensated visual displacement sequence. The delayed-compensated visual displacement sequence is used as a low-frequency observation input, and the observation weights are adjusted according to the observation quality parameters to correct the predicted amplified state vector, outputting a corrected length coordinate sequence. The prediction is executed at the time the encoder length increment sequence is updated, and the correction is triggered when the corresponding time of the delayed-compensated visual displacement sequence arrives.

[0024] Optionally, S5 includes:

[0025] Using an OCA image sequence as input, defect identification processing is performed frame by frame on the OCA image sequence to obtain a defect list. For each defect, the defect list records at least the pixel coordinates of the defect region in the corresponding frame image and the camera trigger timestamp of that frame image.

[0026] Based on the pre-acquired camera calibration parameters, the pixel coordinates of the defect region are converted into defect width coordinates;

[0027] The camera trigger timestamps in the defect list are mapped to the correction length coordinate sequence to determine the length coordinates of each defect in the correction length coordinate sequence, and the length coordinates are determined as the defect length coordinate range based on the continuous occurrence of the same defect in adjacent frame images.

[0028] A defect map is generated based on the defect width coordinates and the defect length coordinate interval. The defect map includes the defect width coordinates and the defect length coordinate interval.

[0029] Furthermore, it also includes a defect verification step: triggering the camera to acquire verification images at a preset lead time before the arrival of the unified time axis corresponding to the target cutting point, and performing secondary verification on the defects that will trigger the splitting linkage strategy based on the verification images. When the secondary verification result does not meet the preset confidence threshold, cancel the splitting linkage strategy or recalculate the target cutting point.

[0030] Optionally, S6 includes:

[0031] Using a defect map as input, a segmentation and linkage strategy is generated based on the defect length coordinate range in the defect map. The segmentation and linkage strategy includes at least one of skipping, segmented segmentation, and defect marking.

[0032] Extract the interval boundary from the defect length coordinate interval and determine the interval boundary as the basic tangent point;

[0033] The length compensation corresponding to the response delay of the actuator is calculated based on the response delay of the actuator. The length compensation corresponding to the response delay of the actuator is equal to the product of the conveyor linear speed and the response delay time of the actuator. The conveyor linear speed is obtained by the difference of the encoder length increment sequence on a unified time axis, or by conversion from the traction roller speed.

[0034] Tension rebound compensation is calculated based on tension data sequence, tension change is determined based on tension data sequence, and length change caused by tension change is determined by combining material elastic parameters and rebound propagation time. The length change is used as tension rebound compensation.

[0035] A safety margin is set, which is a preset length value or is adaptively determined according to the defect type and defect level. The higher the defect level, the larger the safety margin.

[0036] The length compensation corresponding to the response delay of the actuator, the tension rebound compensation, and the safety margin are combined into the tangent point compensation amount;

[0037] The target cut point is obtained by superimposing the cut point compensation amount onto the base cut point.

[0038] Based on the slitting linkage strategy and the target slitting point, a slitting control command is generated and sent to the slitting mechanism and the winding mechanism to achieve linkage execution.

[0039] Optionally, the S7 includes:

[0040] After the slitting mechanism and the winding mechanism execute the slitting control command, the execution results of the slitting mechanism and the winding mechanism are obtained, and the execution results include the length coordinates corresponding to the actual cutting point;

[0041] The length coordinates corresponding to the actual tangent point are compared with the target tangent point to obtain the tangent point error;

[0042] The tangent error is used as a feedback quantity to update the parameters of the length compensation and tension rebound compensation corresponding to the response delay of the actuator in order to obtain the updated tangent compensation quantity.

[0043] When generating subsequent slitting control instructions, the updated slitting point compensation amount is called to generate a new target slitting point, thereby achieving online correction of the slitting point compensation amount.

[0044] Furthermore, using the tangent error as feedback for parameter updates also includes: synchronously updating at least one of the time delay parameter, slippage parameter, or elongation parameter in S4, so as to reduce the prediction error between the subsequent corrected length coordinate sequence and the target tangent point.

[0045] The beneficial effects of this invention are:

[0046] 1. By unifying the time axis alignment of camera images, encoder pulses and tension data, and jointly estimating the tape length coordinates, slippage parameters, elongation parameters and time delay parameters in the augmented state estimation model, combined with high-frequency prediction and low-frequency visual correction, high-precision binding between the defect location and the in-roll length coordinates is achieved, reducing positioning drift caused by slippage, elongation and time delay.

[0047] 2. Based on the defect map, a cutting linkage strategy is generated, and the actuator response delay compensation, tension rebound compensation and safety margin are combined into the cutting point compensation amount to obtain the target cutting point. This enables millimeter-level control of defect segment skip cutting, segment cutting or marking, improves the accuracy of cutting and isolating defect segments, and reduces the flow of defects into the downstream bonding process.

[0048] 3. By acquiring the actual cutting point and calculating the cutting point error, the cutting point compensation amount is updated online, and the time delay parameter, slippage parameter or elongation parameter can be updated synchronously to achieve closed-loop self-correction of the cutting control, improve the long-term stability and consistency under different speed and tension conditions, and reduce the cost of incorrect cutting and rework. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a flowchart of a computer vision-based OCA defect linkage segmentation method.

[0051] Figure 2 This is a flowchart of step S4 of the present invention, which involves amplification state estimation and correction length coordinate output. Detailed Implementation

[0052] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0053] refer to Figures 1-2 The computer vision-based OCA defect linkage segmentation method includes:

[0054] S1. Acquire the OCA image sequence, encoder pulse sequence, and tension data sequence during the transparent OCA conveyor belt transport process, and generate timestamps for each. S2. Align each timestamp to a unified time axis and calculate the uncorrected length coordinate sequence from the encoder pulse sequence. S3. Perform inter-frame registration on adjacent OCA images to obtain a visual displacement sequence. S4. Construct an augmented state estimation model based on encoder length increment, tension data, and visual displacement, including conveyor belt length coordinates, slippage parameters, elongation parameters, and time delay parameters. Perform delay compensation on the visual displacement and output a corrected length coordinate sequence after prediction and correction. S5. Perform defect identification on the OCA image sequence to obtain defect information and map it to the corrected length coordinate sequence to generate a defect map. S6. Determine the basic cutting point based on the defect map and calculate the cutting point compensation amount, including performing delay compensation, tension rebound compensation, and safety margin. Superimpose these to obtain the target cutting point and output a slitting control command to link the slitting mechanism and the winding mechanism. S7. Acquire the actual cutting point and calculate the cutting point error, updating the cutting point compensation amount for subsequent slitting control command generation.

[0055] In this specific embodiment, S1 includes:

[0056] During the transparent OCA tape transport process, the detection controller synchronously acquires data from the camera, encoder, and tension sensor using a master clock with a unified time axis and generates timestamps.

[0057] The camera operates using an external hardware triggering method, with the trigger signal delivered by the digital output of the detection controller at a fixed frame rate. The camera exposes and outputs a grayscale image frame each time it receives a trigger. Simultaneously, the detection controller reads the master clock at microsecond resolution at the moment the trigger edge arrives and writes it to the camera's trigger timestamp, forming a timestamped OCA image sequence. ,in Indicates the first Frame OCA image, Indicates the frame number. Indicates the relationship with the first The camera trigger timestamp corresponding to the frame image and its unit is ;

[0058] The encoder is an incremental rotary encoder and is coaxially connected to the traction roller. The phase signal enters the high-speed counting module of the detection controller, and the detection controller uses a fixed sampling period. The current pulse count is read and the master clock is simultaneously read to generate encoder timestamps, forming a timestamped encoder pulse sequence. ,in Indicates the first The cumulative pulse count value at each read time. Indicates reading the sequence number. Indicates and The corresponding encoder timestamp and its unit is ;

[0059] The tension sensor is a bridge-type load cell that outputs a digital signal via a data acquisition module. The detection controller operates at a fixed sampling period. The tension sample value is read and the master clock is read simultaneously to generate a tension timestamp, forming a tension data sequence with timestamps. ,in Indicates the first The tension value obtained from the second sampling is in N. Indicates the sampling sequence number. Indicates and The corresponding tension timestamp and its unit is ;

[0060] When the timestamps of the camera, encoder, and tension sensor are generated by their respective internal clocks, the detection controller performs a clock conversion calibration once after power-on and maintains the parameters unchanged throughout the entire roll production process. The clock conversion uses a linear clock model to convert the local timestamps of the equipment to a unified time axis.

[0061] ;

[0062] in Represents a timestamp on a unified timeline, with units of 1. Indicates equipment Local timestamp and unit is This indicates the device identifier and its value is the camera. encoder or tension sensor It represents the clock drift factor of the device clock relative to a unified time axis and is a dimensionless real number. This represents the clock offset parameter of the device clock relative to a uniform time axis, and the unit is 1. ;

[0063] The calibration process involves simultaneously sending synchronization events to three types of devices via a detection controller and collecting data. Paired timestamps for group synchronization events Completed, the synchronization events for the camera were aligned using hardware trigger edges, and the encoder and tension sensor were aligned using controller sampling interrupt edges. The detection controller solved the problem for each device using least-squares fitting. and And write it into the parameter area for subsequent use. Convert to a unified timeline;

[0064] During the acquisition process, the detection controller writes the three types of sequences into the circular buffer in the order of arrival and records the sequence integrity flag. When frame loss, count wrap-around, or communication timeout occurs, the corresponding data point is marked as invalid and its timestamp is retained for subsequent steps to perform interpolation and rejection processing, thereby outputting OCA image sequence, encoder pulse sequence, and tension data sequence with unified time axis timestamps.

[0065] In this specific embodiment, S2 includes:

[0066] The detection controller uses a unified time axis as the sole reference and defines the start time of the roll material as the zero point of the unified time axis. ,in Represents a unified timeline timestamp with units of Specifically, the camera trigger timestamp of the first frame image The corresponding unified timeline time is set to zero, and the timestamps of subsequent data are shifted to the same zero point to obtain the aligned OCA image sequence. Aligned encoder pulse sequence With the aligned tension data sequence ,in Indicates the first Frame OCA image, Indicates the first Camera trigger timestamp of frame image Indicates the first The encoder's cumulative pulse count value at each reading time Indicates and Corresponding encoder timestamp, Indicates the first The tension value obtained from the second sampling, and the unit is... Indicates and Corresponding tension timestamp;

[0067] On the unified time axis, the detection controller calculates the pulse increment by the cumulative pulse count difference between adjacent encoder sampling times and converts it into a length increment using a pulse length coefficient, denoted as [missing information]. And the unit is After the equipment is installed, the value is obtained through calibration and solidified into a constant. The calibration process is as follows: the conveyor belt is moved at a constant linear speed and the roll material passes through the measurement reference length. Record the change in the cumulative pulse count of the encoder during this process. ,Will Set as and The ratio is written into the detection controller parameter area;

[0068] Based on the above The detection controller accumulates and integrates the encoder length increments in a uniform time axis sequence to obtain the uncorrected length coordinate sequence. The uncorrected length coordinates are indexed by the encoder sampling time and represent the cumulative length position of the roll material along the belt travel direction. Their calculation relationship is as follows:

[0069] ;

[0070] in Represents encoder timestamp The corresponding uncorrected length coordinates and the unit is Indicates the pulse length coefficient. Indicates the first The encoder's cumulative pulse count value at each reading moment This indicates that the encoder reads the sequence number and sets the initial conditions. and This is the cumulative pulse count value at time zero.

[0071] Value rules are established on a unified timeline to provide the input required for prediction and correction in subsequent augmented state estimation. These value rules define the "prediction time" as the timestamp of each encoder. The encoder length increment and tension value are directly taken at the corresponding time, where the encoder length increment is multiplied by the difference between two consecutive cumulative pulse counts. The tension value is obtained through the tension sequence. The result is obtained by performing a linear interpolation based on the timestamp.

[0072] The value selection rule defines the "correction time" as the unified timeline time reached by visual observation in subsequent steps and in the encoder sequence. With tension sequence Performing a linear interpolation based on the timestamp yields the uncorrected length coordinates and tension values ​​corresponding to the corrected time. The linear interpolation is performed around the target time. The nearest forward valid sampling point and the nearest backward valid sampling point are used to construct an interpolation interval, and the results are calculated proportionally to time within the interpolation interval. Furthermore, when a sampling point around the target time is... When marked as invalid, the detection controller replaces the invalid sampling point with the most recent valid sampling point in time and performs interpolation within the replaced interval, thereby outputting an uncorrected length coordinate sequence consistent with the unified time axis, as well as encoder length increment values ​​and tension values ​​oriented towards the prediction time and correction time.

[0073] In this specific embodiment, S3 includes:

[0074] The detection controller operates on a unified timeline according to the camera trigger timestamps. Selecting two adjacent OCA images in sequence and Construct an image pair and define the displacement observation time of the image pair as... ,in Indicates the first The frame is an OCA image and is an 8-bit grayscale matrix. Indicates the first Frame OCA image, and Represents the camera trigger timestamp for the corresponding frame, in units of ;

[0075] To ensure that the displacement components of inter-frame registration are consistent with the transport direction, the detection controller first performs lens distortion correction and perspective correction on each frame and defines the corrected image coordinate system as follows: pixel horizontal axis Consistent with the transport direction and with the pixel's vertical axis aligned with its width, where Represents the pixel coordinates along the transport direction in the corrected image;

[0076] Then on and The same preprocessing procedure is performed to suppress illumination fluctuations and background gradients. This preprocessing includes cropping a fixed region of interest, removing the grayscale mean from the cropped region, applying a two-dimensional Hanning window to the cropped region to reduce boundary effects, and performing bandpass filtering on the cropped region to suppress low-frequency illumination fluctuations and high-frequency noise. The region of interest is fixedly set after device installation. The pixels are positioned at the center of the imaging field of view to avoid vignetting at the edges;

[0077] After obtaining the preprocessed image pairs, the detection controller uses the phase correlation method to calculate the two-dimensional translation between the two frames and takes its component in the travel direction as the inter-frame pixel displacement. ,in Indicates from arrive The pixel displacement in the transport direction is expressed in pixels. The phase-correlated cross-power spectrum is constructed from the two-dimensional discrete Fourier transform of two frames and the correlation peak position is obtained through the two-dimensional inverse transform. The correlation peak position is refined to sub-pixel level by fitting a quadratic surface to the peak neighborhood, and the fitting window is fixed. Pixel;

[0078] The detection controller will shift pixels. Calculate the inter-frame length displacement based on the pixel size in the transport direction. And generate a visual displacement sequence corresponding to a unified time axis. The conversion relationship is as follows: ,in Represents the inter-frame length displacement and the unit is . Indicates the pixel size in the transport direction, and the unit is . Indicates inter-frame pixel displacement, the After the equipment is installed, the pixel spacing is calculated and written into the parameter area of ​​the detection controller after being calibrated and solidified into a constant by imaging in the belt direction using a calibration plate with 1mm equally spaced lines.

[0079] To provide observation quality parameters for subsequent amplified state estimation, the detection controller calculates the peak correlation ratio (PCR) for the phase correlation results of each pair of images. With consistency error and Synchronous output, where Defined as the ratio of the amplitude of the primary peak to the amplitude of the secondary peak, and used to characterize the uniqueness of related peaks. Define as according to After translation and The mean absolute difference within the region of interest, expressed in gray levels, is used to characterize the registration residual. The detection controller will... or The observations are marked as low confidence and their observation weights are reduced in subsequent fusion to obtain an output result that includes visual displacement and observation quality parameters.

[0080] In this specific embodiment, S4 includes:

[0081] The detection controller uses encoder length increment and tension value as high-frequency prediction inputs and visual displacement as low-frequency observation inputs to construct an augmented state estimation model on a unified time axis and uses the augmented state vector. The coordinates representing the true length of the conveyor belt, slippage parameters, elongation parameters, and time delay parameters are as follows: Represents a unified timeline timestamp with units of Indicates time The corrected length coordinates and the unit is Represents the slip ratio, which is dimensionless and has a range of values ​​of 1. Indicates the elongation coefficient of the material and the unit is . This represents the equivalent time delay of visual observation, in milliseconds, and its value range is [value range missing]. ;

[0082] Model initialization in Execute and set at all times Simultaneously set reference tension for The tension sample value at time t is in N, and will be... As random walk parameters to adapt to changes in operating conditions, the amplified state estimation is implemented using an extended Kalman filter, and the covariance matrix is ​​initialized as a diagonal matrix.

[0083] ;

[0084] in This represents the initial state covariance matrix, with the four elements within the parentheses corresponding to... The unit is dimensionless variance The unit is The unit is Predicting at each encoder timestamp Upon arrival, trigger and read the encoder length increment corresponding to that moment. With tension value ,in Indicates the first The unified timeline timestamp of the secondary encoder sampling time. This indicates an increase in encoder length, in mm. The value represents the tension in N, and the encoder length increment is converted into a length coordinate increment in the prediction based on slippage correction and elongation compensation. The process noise covariance is superimposed while maintaining continuity using random walk. To reflect model uncertainty, where This represents the process noise covariance matrix for each prediction step, with the units of the four diagonal elements being... Dimensionless ;

[0085] Correction is triggered when the visual displacement observation arrives and is based on the observation time. Visual displacement for indexing and observation quality parameters and ,in Indicates the first Camera trigger timestamp of the frame image Represents the inter-frame length displacement and the unit is . This indicates the correlation peak ratio, which is dimensionless. Indicates consistency error, with units of gray levels;

[0086] To perform delay compensation for visual displacement, the detection controller uses the currently estimated time delay parameters. Map the visual observations to the effective observation intervals on a unified time axis. And at both ends of this interval, the predicted length trajectory is obtained. A linear interpolation calculation is performed to obtain the predicted length difference between the endpoints of the interval as the observed predicted value. Simultaneously, the observation noise variance is adjusted according to the observation quality parameter. and The observation noise variance is set to . ,when or The observation noise variance is set to . ,in Indicates the first The noise variance of the sub-visual observation, in units of Based on this, the gain calculation and state correction of the extended Kalman filter are completed to output the correction length coordinates;

[0087] The relationship between the above predictions and observations is determined in this embodiment by the following formula:

[0088] ;

[0089] in This represents the estimated corrected length coordinate of the filtered output. This represents the estimated slippage rate of the filtered output. This represents the estimated elongation coefficient of the filtered output. This represents the estimated time delay of the filtered output, expressed in milliseconds (ms). Indicates the first The unified timeline timestamp of the secondary encoder sampling time. and They represent the first Frame and the Camera trigger timestamp of the frame image Indicates the encoder length increment. Indicates the tension value. Indicates reference tension. Indicates visual displacement observation;

[0090] The detection controller will perform each prediction and correction after completion. The corrected length coordinate sequence is written in chronological order as the length reference for the S5 defect mapping and S6 tangent point calculation, and then corrected. Perform boundary truncation to satisfy the value range, thereby ensuring that the length coordinate is monotonically continuous throughout the entire roll and can be used for subsequent linkage cutting control.

[0091] In this specific embodiment, S5 includes:

[0092] The detection controller uses OCA image sequences with uniform timeline timestamps. As input, defect identification is performed frame by frame to generate a list of defects, where Indicates the first The frame is an OCA image and is an 8-bit grayscale matrix. Indicates the first The camera trigger timestamp of the frame image and the unit is ;

[0093] Defect identification employs a pixel-level segmentation network with fixed parameters and outputs a defect probability map and a defect category map on each frame of image. The input to the segmentation network is... Linear normalization to Single-channel image and cropped to The pixel region of interest has a backbone consisting of a 4-level downsampling encoder and a 4-level upsampling decoder, with each level's convolutional kernel size being [missing information]. It employs downsampling with a step size of 2 and bilinear interpolation upsampling, and the network output is a defect probability map and a defect category map of the same size as the input;

[0094] The detection controller categorizes the defect probability map according to a threshold. Binarization is used to obtain a defect mask, and the defect mask is then processed. Opening operations are performed on structuring elements to remove isolated noise points. Subsequently, connected component labeling is performed on the defect mask, with each connected component treated as a defect candidate. For each defect candidate, its bounding rectangle is calculated, and its pixel coordinate range within the frame is recorded. and and defect categories With Defect Confidence ,in and This represents the minimum and maximum pixel coordinates of the bounding rectangle of the defect in the conveyor direction. and Represents the minimum and maximum pixel coordinates of the bounding rectangle of the defect in the width direction. Indicates the defect category label, This represents the defect confidence level, calculated from the mean of the defect probability map within the defect candidate region, and its value range is [value range missing]. ;

[0095] The detection controller converts pixel width coordinates into defect width coordinates based on pre-acquired camera calibration parameters and outputs the defect width coordinate range. The camera calibration parameters are fixed as constants after the device is installed and include the width pixel size. With width reference pixel row number ,in This represents the physical length corresponding to each pixel in the width direction. The reference pixel row number represents the zero point of the roll material width direction within the area of ​​interest. The defect width direction coordinates take the roll material centerline as the zero point and the positive direction to one side is taken as the positive direction.

[0096] To precisely link defects to their position within the roll length, the detection controller uses the camera trigger timestamp of the frame to which the defect candidate belongs. Mapping to the corrected length coordinate sequence And calculate the length coordinates corresponding to the defect candidate. ,in Representing a unified timeline moment The corrected length coordinates are in mm. The time of falling between two adjacent correction length coordinate records and The result is obtained by linear interpolation between the two values:

[0097] ;

[0098] in Representing a unified timeline moment The corresponding correction length coordinates and the unit is Represents the unified timeline time to be mapped, with units of 1. and They respectively represent satisfying The most recent forward and most recent backward correction length coordinates record the time and the unit is and Representing time respectively and The corresponding correction length coordinates, in mm;

[0099] The detection controller merges the same defect that appears repeatedly across frames into a single defect entry through trajectory association and determines the defect length coordinate range accordingly. The trajectory association matches adjacent frames based on the intersection-union ratio (IU / U) of their bounding rectangles, and the IU / U is set to a threshold value. When a defect candidate has an intersection-union ratio (IU) of the last bounding rectangle of an existing defect entry in the previous frame that is not less than 1, And defect category If the same defect is found, merge it into the defect entry and update the last occurrence timestamp and width coordinate range of the defect entry; otherwise, create a new defect entry.

[0100] Each defect entry must record at least the defect category. Defect confidence level Defect width coordinate range The first time the defect appeared was marked by a timestamp. timestamp of last occurrence And the length coordinate interval is determined as A defect map is generated, which is stored in the form of a set of defect entries, and each defect entry contains both the width coordinate range and the length coordinate range of the defect.

[0101] To achieve the aforementioned defect verification, the detection controller checks the defects in the defect map that meet the following criteria: And defect category Defect entries belonging to the preset set requiring linkage are marked with a review flag and their required location information is saved. This location information includes the predicted camera trigger timestamp value corresponding to the center of the defect length coordinate interval and the defect width coordinate interval. When S6 calculates the target cut point that will trigger the split-linkage strategy and gives its corresponding unified time axis moment, Afterwards, the detection controller... Preset lead time before arrival The camera is triggered to acquire a verification image, and the same segmentation network inference and connected component extraction as described above are performed on the predicted regions of the corresponding defect entries in the defect map on the verification image to obtain the verification confidence. ,in This represents the time point on the unified timeline corresponding to the target tangent point, and the unit is 1. Indicates the lead time for review and the unit is... Indicates the confidence level of the review and its range is [value range missing]. The detection controller will With the confidence threshold for review Compare and If the defect entry is marked as a failed review, the S6 outputs a control flag to cancel the corresponding splitting linkage strategy for the defect or outputs an updated defect length coordinate range to trigger the S6 to recalculate the target cutting point.

[0102] In this specific embodiment, S6 includes:

[0103] The detection controller takes a defect map as input and sorts it by defect length coordinate interval in ascending order on the calibration length coordinate axis. Each defect entry in the defect map contains a defect category. Defect confidence level Defect width coordinate range With the defect length coordinate range Indicates the defect category label, This represents the confidence level of the defect and its value range is [value range missing]. and These represent the minimum and maximum physical coordinates of the defect in the width direction of the roll material, respectively, in mm. and These represent the start and end points of the defect on the correction length coordinate, respectively, with the unit being mm;

[0104] The detection controller receives the review flag and removes defect entries that fail the review from the defect map to ensure that the linkage strategy is triggered only for defects that pass the second confirmation.

[0105] The detection controller is based on the defect category. A slitting linkage strategy is generated, and defect entries with defect categories of bubbles, foreign objects, and contamination are set as the skip-cut strategy. Defect entries with defect categories of scratches and indentations are set as the segmented slitting strategy. Defect entries with defect categories of tiny particles are set as the defect marking strategy. The skip-cut strategy is used to perform two cross-cuts at the start and end positions of the defect segment and link the winding mechanism to guide the defect segment into the waste roll. The segmented slitting strategy is used to perform two cross-cuts at the start and end positions of the defect segment and make the defect segment form an independent roll segment. The defect marking strategy is used to trigger the marking device to print a mark on the edge of the roll at the defect position without triggering the cross-cut.

[0106] To avoid overly close switching due to adjacent defects, the detection controller adjusts the spacing between two adjacent defect entries. Defects no larger than 200mm are merged into the same processing interval, with the boundary of the merged interval serving as the subsequent base cutting point. This indicates the endpoint of the coordinate range of the defect length of the previous defect entry. Indicates the starting point of the coordinate interval for the defect length of the next defect entry;

[0107] The detection controller extracts the interval boundaries as basic cut points from each processing interval and generates a list of basic cut points to be executed, wherein basic cut points are generated for both the skip-cut strategy and the segmented cutting strategy. With the base tangent point Generate basic pointcuts for the defect marking strategy. ,in and All represent the position of the basic tangent point on the correction length coordinate, and the unit is mm;

[0108] The detection controller calculates the cut point compensation for each basic cut point and sums them to obtain the target cut point. The cut point compensation consists of execution delay compensation, tension rebound compensation, and safety margin, and is defined by the same notation as a signed length correction of the basic cut point. The calculation relationship for the target cut point is as follows:

[0109] ;

[0110] in This indicates the position of the target tangent point on the corrected length coordinate system, with units in mm. This represents the base tangent point corresponding to the target tangent point, and the unit is mm. This indicates that delay compensation is being performed and the unit is mm. This indicates tension rebound compensation, and the unit is mm. This indicates a safety margin, and the unit is mm.

[0111] Among them, the execution delay compensation Based on the response delay time of the actuator and the conveyor linear speed, and taking the trigger advance direction as negative, the response delay time of the traverse actuator is set as follows: The response delay time of the winding switching actuator is set to The response delay time of the marker actuator is set to The linear speed of the conveyor belt is denoted as And the unit is And by encoder length increment timestamps of adjacent encoders The average value obtained over the most recent 10 sampling windows is used to suppress speed jitter. Indicates the encoder length increment and the unit is Represents encoder timestamps in units of The traverse cutting action will Set as The take-up switching action will Set as The marking action will Set as ;

[0112] Tension rebound compensation Based on tension data sequence calculation and implemented using a tension-length empirical model, the detection controller uses the sliding average tension value of the 100ms preceding the target cutting point as the pre-cut tension. And the unit is N, based on the winding tension setting. As a reference, the tension change was calculated. Simultaneously, the tension rebound coefficient, calibrated through a tension step test after equipment installation, is adopted. With rebound propagation time Solidified to a constant, of which The coefficient representing the equivalent length change caused by the change in tension. This indicates the time delay in the propagation of tension disturbance to the cutter; the detection controller will... Set as And its sign direction is defined as when The target tangent point is increased along the belt direction to offset the shift of the tangent point caused by the increase in pre-shear tension;

[0113] Safety margin Defect Category Determine and fix the value, and detect the controller's cut-off point for the pre-cut target of the jump-cut strategy. And take the tangent point of the target after cutting. The target cutting point of the segmentation and splitting strategy is taken as the cutting point. And take the tangent point of the target after cutting. Defect marking strategy ;

[0114] The detection controller writes one or two target cut points corresponding to each strategy into the execution queue and executes a minimum volume segment length constraint to ensure the executability of the mechanism. The minimum volume segment length is set to... And when the distance between the target cutting points of two adjacent transverse cuts is less than The two cross-cuts are merged into one and the corresponding defect entries are changed to a defect marking strategy.

[0115] The detection controller generates slitting control commands based on the target slitting point and sends them to the slitting mechanism and the winding mechanism. The slitting control commands include the strategy type and the target slitting point. The corresponding action actuator identifier and action parameters include the cross-cutting action parameters, which include the cutter down signal and the cutter reset signal; the rewinding switching action parameters include the waste roll selection signal and the finished roll selection signal; and the marking action parameters include the inkjet switch signal and the inkjet content code.

[0116] During the slicing process, the controller reads the correction length coordinates in real time. And in Reach the smallest target cutoff point in the queue of execution. Immediately issue the corresponding slitting control command, thereby linking the slitting mechanism and the winding mechanism to complete jump cutting, segmented slitting or defect marking according to the defect map and isolate the defective segment into the finished roll.

[0117] In this specific embodiment, S7 includes:

[0118] After the slitting and winding mechanisms execute slitting control commands, the detection controller acquires the execution results and forms a cutting point feedback event stream. This cutting point feedback event is output by a cut detection sensor installed at a fixed distance downstream of the cutter. The cut detection sensor is a through-beam photoelectric sensor whose detection beam crosses the tape path and uses the change in light transmission after the cut as the trigger condition. The fixed distance in the tape path from the cut detection sensor to the cutter edge is denoted as... ,in This represents the geometric distance between the feedback sensor and the cutting edge of the blade, expressed in mm.

[0119] When the When a cut passes through the cut detection sensor, the detection controller records the trigger moment on a unified time axis. And the unit is And read the estimated value of the corrected length coordinate at that moment from the corrected length coordinate sequence. And the unit is mm, the length coordinates corresponding to the actual tangent point are defined as:

[0120] ;

[0121] in Indicates the first The length coordinates of each actual tangent point are given in mm.

[0122] The detection controller retrieves the target cutoff point that matches the cutoff feedback event from the queue to be executed. The unit is mm. The matching is completed according to the timing rule of "first to send and first to receive" and in combination with the action type identifier. The action type identifier includes traverse cutting action and rewinding switching action and is consistent with the actuator identifier in the S6 slitting control command.

[0123] The detection controller defines the tangent error as... ,in Indicates the first Each tangent point error is expressed in mm, and the following is... Together with the action type corresponding to the tangent point and the linear velocity of the conveyor belt near the trigger time of the tangent point. and pre-shear tension Write to the feedback cache, where This represents the conveyor linear velocity calculated using the average of 10 encoder sampling windows identical to those in S6, with units of . This represents the pre-shear tension calculated using the average tension sliding over 100 ms before the target tangent point is triggered, and the unit is N;

[0124] Detection controller with As a feedback quantity, the corresponding parameters of execution delay compensation and tension rebound compensation in the tangent point compensation quantity are updated online, and the time delay parameter, slippage parameter and elongation parameter in S4 are updated simultaneously to reduce the prediction error of the subsequent correction length coordinate sequence and the target tangent point. Specifically, the parameters to be updated are composed into a parameter vector. And update as follows:

[0125] ;

[0126] in This represents the updated parameter vector. This represents the response delay time of the transverse actuator, expressed in milliseconds, and is used in S6 to calculate execution delay compensation. This represents the response delay time of the take-up switching actuator, expressed in milliseconds, and is used in S6 to calculate execution delay compensation. This indicates the response delay time of the marker actuator, expressed in milliseconds (ms), and is used in S6 to calculate execution delay compensation. Represents the tension rebound coefficient and its unit is . And it is used for S6 calculation of tension rebound compensation. Represents the estimated value of the delay parameter and the unit is . This represents an estimate of the slippage rate and is dimensionless. This represents the estimated elongation coefficient, and the unit is 1. Describes the gain vector and its dimension is 1 / 2. Consistent and determined according to action type and working condition:

[0127] ;

[0128] When the cutting point is a transverse cutting action, the setting is as follows: When the cutting point is a take-up switching action, set When the cut point is a marker action, set , ,in This indicates a time delay for updating the gain, and it is dimensionless.

[0129] The tension rebound coefficient update gain is adaptively set according to the tension change, and the detection controller defines the tension change as... And the unit is N, where Set the winding tension value when Time setting And when Time setting ,in This represents the springback coefficient update gain, which is dimensionless.

[0130] S4 parameter synchronous update gain setting is , ,in This indicates that the time delay parameter update gain is dimensionless. This indicates that the slippage rate update gain is dimensionless. This indicates the elongation coefficient update gain, which is dimensionless. Indicates the time since the last incision feedback. Feedback time for this incision The sum of encoder length increments accumulated by S2, in mm, and when When To ensure stable updates;

[0131] This represents an operator that performs element-wise boundary truncation on the parameter vector, where... Cut off at ,Will Cut off at ,Will Cut off at ,Will Cut off at ,Will Cut off at Will Cut off at ,Will Cut off at ;

[0132] The detection controller will be updated and The parameters are written to the S6 parameter area for subsequent calculations of delay compensation and tension rebound compensation, and the updated parameters are then used to perform these calculations. The state vector of the S4 amplified state estimate is written as the initial value for prediction and correction at the next moment, realizing online closed-loop self-correction of the tangent point compensation amount and length coordinate estimation parameters.

[0133] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0134] This invention addresses the technical problems in online detection of transparent OCA defects, namely, the difficulty in stably binding the identification results to the in-roll position, the difficulty in automatically isolating defect segments after rewinding, and the difficulty in implementing linkage control based on the defect map during the slitting process. It integrates encoder length increments, tension data, and visual displacement into a unified time axis: on the one hand, the encoder provides high-frequency length prediction to ensure continuity; on the other hand, visual displacement obtained through adjacent frame registration corrects the cumulative length drift. Simultaneously, the elongation effect caused by tension and the visual observation delay are incorporated into the estimation process, outputting a corrected length coordinate sequence, enabling the defect identification results to be reliably mapped to a defect map. Based on this, the basic cutting point is extracted from the defect map and superimposed with actuator response delay compensation, tension rebound compensation, and safety margin to obtain the target cutting point. This directly transforms "defect localization" into "linkage control of slitting and rewinding," enabling skip cutting, segmentation, or marking of defect segments, reducing the probability of defects flowing into the downstream bonding process and reducing rework costs.

[0135] In terms of algorithm structure, this invention makes targeted improvements to address the aforementioned technical problems: First, it adopts amplified state joint estimation, elongation coefficient, time delay parameter, and conveyor length coordinates together to estimate slip rate, elongation coefficient, and time delay parameter, avoiding in-roll position drift caused by relying solely on length calculation by a single encoder or empirical correction; Second, it adopts a multi-rate asynchronous fusion mechanism, combining encoder high-frequency prediction with visual low-frequency correction, and combining timestamp alignment and delay compensation to improve the stability and consistency of length binding under different working conditions; Third, it constructs a cutting point compensation combination model, unifying execution delay, tension rebound propagation effects, and safety margin into the cutting point compensation amount, and achieving online correction through actual cutting point error feedback, thereby maintaining millimeter-level cutting point accuracy under speed and tension fluctuation conditions, and better achieving the technical effect of automatic isolation of defective segments.

Claims

1. A computer vision-based OCA defect linkage segmentation method, characterized in that, include: S1 acquires the OCA image sequence, encoder pulse sequence, and tension data sequence during the transparent OCA transport process, and generates timestamps for each. S2 aligns each timestamp to a unified time axis and calculates the uncorrected length coordinate sequence from the encoder pulse sequence; S3 performs inter-frame registration on adjacent OCA images to obtain a visual displacement sequence; S4 constructs an augmented state estimation model based on encoder length increment, tension data, and visual displacement, including belt length coordinates, slippage parameters, elongation parameters, and time delay parameters, performs delay compensation on the visual displacement, and outputs a corrected length coordinate sequence after prediction and correction; S5 performs defect identification on the OCA image sequence to obtain defect information and maps it to the corrected length coordinate sequence to generate a defect map; S6 determines the basic cutting point based on the defect map and calculates the cutting point compensation amount, including performing delay compensation, tension rebound compensation, and safety margin, superimposes it to obtain the target cutting point, and outputs the slitting control command to link the slitting mechanism and the winding mechanism; S7 obtains the actual cutting point and calculates the cutting point error, updates the cutting point compensation amount for subsequent slitting control command generation.

2. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S1 includes: During the transparent OCA transport process, the camera continuously acquires and forms an OCA image sequence. Each time the camera acquires an image, the camera trigger timestamp corresponding to that image is recorded, resulting in an OCA image sequence with camera trigger timestamps. The encoder pulse sequence output by the belt encoder is acquired synchronously, and the corresponding encoder timestamp is recorded for the pulse count value in the encoder pulse sequence to obtain the encoder pulse sequence with encoder timestamp; The tension data sequence output by the tension sensor is acquired synchronously, and the corresponding tension timestamp is recorded for the tension sample value in the tension data sequence to obtain the tension data sequence with tension timestamp; The camera trigger timestamp, encoder timestamp, and tension timestamp are generated based on the same clock source, or they are generated based on different clock sources respectively. The timestamps are converted to a unified time axis through time synchronization or through pre-calibrated clock bias parameters and clock drift parameters, and the timestamped OCA image sequence, encoder pulse sequence, and tension data sequence are output.

3. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S2 include: The camera trigger timestamp, encoder timestamp, and tension timestamp are mapped to a unified time axis to obtain an aligned encoder pulse sequence and an aligned tension data sequence. This mapping includes establishing a time correspondence using at least one of the following methods: precise time protocol synchronization, network time protocol synchronization, and hardware trigger synchronization. On the unified time axis, the difference between pulse counts at adjacent times in the aligned encoder pulse sequence is used to obtain an encoder length increment sequence. This pulse count difference is then converted into a length increment according to the length coefficient corresponding to the encoder pulse. The encoder length increment sequence is then integrated along the unified time axis to obtain an encoder length coordinate sequence, which serves as an uncorrected length coordinate sequence. This uncorrected length coordinate sequence represents the cumulative length position of the transparent OCA along the belt travel direction. Value rules are established on the unified time axis. At the prediction or correction time of the amplification state estimation, interpolation calculations are performed on the aligned encoder pulse sequence and the aligned tension data sequence to obtain the encoder length increment value and tension value corresponding to the prediction or correction time.

4. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S3 include: Adjacent OCA images in an OCA image sequence are selected as image pairs on a unified timeline. The image pairs are preprocessed to reduce the impact of illumination fluctuations on inter-frame registration; The inter-frame displacement of the image pair is calculated using at least one of the following methods: feature point matching, dense optical flow, phase correlation, and template matching. The component of the inter-frame displacement in the transport direction is converted into the corresponding inter-frame length displacement to obtain a visual displacement sequence corresponding to a unified time axis. Based on the results of the inter-frame registration, the observation quality parameters are calculated, including one or more of the following: feature matching in-point ratio, correlation peak ratio, and inter-frame displacement consistency error.

5. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S4 includes: Using encoder length increment sequence, tension data sequence, visual displacement sequence, and observation quality parameters as inputs, an amplified state estimation model is established, and the amplified state vector represents the true tape length coordinates, slippage rate, material elongation coefficient, and time delay. On a unified time axis, the encoder length increment sequence and tension data sequence are used as high-frequency prediction inputs. The encoder length increment sequence is corrected for slippage based on the slippage rate, and the length increment after slippage correction is compensated for elongation based on the material elongation coefficient and the tension data sequence to obtain the predicted amplified state vector. On a unified time axis, the time delay is estimated, and the visual displacement sequence is compensated for delay to obtain the delayed-compensated visual displacement sequence. Using the delayed-compensated visual displacement sequence as a low-frequency observation input, the observation weights are adjusted according to the observation quality parameters to correct the predicted amplified state vector, and a corrected length coordinate sequence is output. The prediction is performed at the time when the encoder length increment sequence is updated, and the correction is triggered when the corresponding time of the delayed-compensated visual displacement sequence arrives.

6. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S5 include: Using an OCA image sequence as input, defect identification processing is performed frame by frame on the OCA image sequence to obtain a defect list. For each defect, the defect list records at least the pixel coordinates of the defect region in the corresponding frame image and the camera trigger timestamp of that frame image. Based on the pre-acquired camera calibration parameters, the pixel coordinates of the defect region are converted into defect width coordinates; The camera trigger timestamps in the defect list are mapped to the correction length coordinate sequence to determine the length coordinates of each defect in the correction length coordinate sequence, and the length coordinates are determined as the defect length coordinate range based on the continuous occurrence of the same defect in adjacent frame images. A defect map is generated based on the defect width coordinates and the defect length coordinate range. The defect map includes the defect width coordinates and the defect length coordinate range.

7. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S6 include: Using a defect map as input, a segmentation and linkage strategy is generated based on the defect length coordinate range in the defect map. The segmentation and linkage strategy includes at least one of skipping, segmented segmentation, and defect marking. Extract the interval boundary from the defect length coordinate interval and determine the interval boundary as the basic tangent point; The length compensation corresponding to the response delay of the actuator is calculated based on the response delay of the actuator. The length compensation corresponding to the response delay of the actuator is equal to the product of the conveyor linear speed and the response delay time of the actuator. The conveyor linear speed is obtained by the difference of the encoder length increment sequence on a unified time axis, or by conversion from the traction roller speed. Tension rebound compensation is calculated based on tension data sequence, tension change is determined based on tension data sequence, and length change caused by tension change is determined by combining material elastic parameters and rebound propagation time. The length change is used as tension rebound compensation. A safety margin is set, which is a preset length value or is adaptively determined according to the defect type and defect level. The higher the defect level, the larger the safety margin. The length compensation corresponding to the response delay of the actuator, the tension rebound compensation, and the safety margin are combined into the tangent point compensation amount; The target cut point is obtained by superimposing the cut point compensation amount onto the base cut point. Based on the slitting linkage strategy and the target slitting point, a slitting control command is generated and sent to the slitting mechanism and the winding mechanism to achieve linkage execution.

8. The computer vision-based OCA defect linkage segmentation method according to claim 1, characterized in that, S7 includes: After the slitting mechanism and the winding mechanism execute the slitting control command, the execution results of the slitting mechanism and the winding mechanism are obtained, and the execution results include the length coordinates corresponding to the actual cutting point; The length coordinates corresponding to the actual tangent point are compared with the target tangent point to obtain the tangent point error; The tangent error is used as a feedback quantity to update the parameters of the length compensation and tension rebound compensation corresponding to the response delay of the actuator in order to obtain the updated tangent compensation quantity. When generating subsequent slitting control instructions, the updated slitting point compensation amount is called to generate a new target slitting point, thereby achieving online correction of the slitting point compensation amount.

9. The computer vision-based OCA defect linkage segmentation method according to claim 6, characterized in that, It also includes a defect verification step: triggering the camera to acquire verification images with a preset lead time before the arrival of the unified time axis corresponding to the target cutting point, and performing secondary verification on the defects that will trigger the splitting linkage strategy based on the verification images. When the secondary verification result does not meet the preset confidence threshold, cancel the splitting linkage strategy or recalculate the target cutting point.

10. The computer vision-based OCA defect linkage segmentation method according to claim 8, characterized in that, Using the tangent error as feedback for parameter updates further includes: synchronously updating at least one of the time delay parameter, slippage parameter, or elongation parameter in S4, so as to reduce the prediction error between the subsequent corrected length coordinate sequence and the target tangent point.

Images