A method and system for multi-optical path defect detection and adaptive process correction of cover adhesive

By constructing a defect-process coupling mapping through multi-optical path imaging and deep learning algorithms, and building a multi-process decoupling control model, the problems of incomplete defect identification and rough process control in the production of PVG conveyor belt cover rubber were solved. Real-time online identification and dynamic adaptive adjustment of cover rubber defects were realized, improving production stability and quality.

CN123104024APending Publication Date: 2026-11-17SHANDONG XINBAOLONG IND TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611252649.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-18
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the current PVG conveyor belt cover rubber production process, incomplete defect identification, delayed detection, lack of quantitative coupling between defects and processes, rough process control, and lack of layered adaptive correction lead to difficulty in controlling extrusion process drift and unstable production.

Method used

By employing multi-path imaging combined with deep learning algorithms, spatial scale parameters of defects and surface and subsurface defect segmentation masks are extracted. A defect-process coupling mapping coefficient matrix is ​​constructed, and a multi-process decoupling control model is built. By iteratively optimizing the process variable correction coefficients, online defect identification and dynamic adaptive adjustment of process parameters are achieved.

Benefits of technology

It enables real-time online identification of cover adhesive defects and dynamic adaptive adjustment of production processes, effectively suppressing process drift and improving the processing quality of cover adhesive and the stability of production lines.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104024A_ABST
    Figure CN123104024A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for multi-optical path defect detection and adaptive process correction of PVG conveyor belt cover rubber, belonging to the field of image detection technology. The method includes: acquiring multi-optical path imaging image sequences of the cover rubber and production line process time-series parameters; extracting defect spatial scale parameters and surface and subsurface defect segmentation masks; constructing a defect-process coupling mapping coefficient matrix; building a multi-process decoupling control model; iteratively optimizing process variable correction coefficients with a good cover rubber benchmark as the target; acquiring PVG conveyor belt cover rubber defect data; calculating the defect month-on-month growth rate, comprehensive process drift rate, and coupling ratio to determine the production line process status; if the process drift exceeds the limit, solving for local fine-tuning correction and global process compensation correction respectively, and synchronously updating the process variable correction coefficients and segmented correction strategy. This invention achieves online identification of PVG conveyor belt cover rubber defects and dynamic adaptive adjustment of the production process, effectively suppressing process drift and improving the processing quality of the cover rubber and the production line stability.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of image detection technology, specifically a method and system for detecting multi-optical-path defects in cover adhesive and adaptive process correction. Background Technology

[0002] PVG conveyor belts are widely used in heavy-duty material conveying scenarios such as coal mines, ports, and power plants. The cover rubber is a key layer that determines the core performance of the conveyor belt in terms of wear resistance, tear resistance, and delamination resistance. Its extrusion molding process directly determines the finished product qualification rate. During the continuous extrusion production of PVG cover rubber, it is easily affected by multiple parameters such as rubber compound mixing uniformity, screw speed, die clearance, cooling water flow rate, and barrel temperature control, which can easily produce defects such as surface bubbles, surface scratches, rubber color spots, sub-interface delamination, internal micropores, and uneven thickness. Moreover, long-term operation of extrusion equipment can lead to problems such as screw wear, die accumulation, temperature control drift, and cooling system flow rate reduction, causing continuous process drift, producing batches of defective cover rubber, and significantly increasing raw material loss and production costs. At present, the quality control of PVG cover rubber production lines has several technical shortcomings: First, there is no quantitative coupling relationship between defects and extrusion process, and process adjustment relies on manual experience. Existing control schemes only monitor single process parameters such as screw speed, die clearance, and cooling water flow rate, without establishing a quantitative mapping relationship between defect scale, defect area, and multi-process parameters. When the production line experiences an increase in defects, maintenance personnel can only rely on experience to roughly adjust single process parameters, unable to distinguish whether the increase in defects is caused by screw speed fluctuations, die clearance deviations, or cooling drift. They lack multi-process decoupling control capabilities, resulting in blind parameter adjustments and repeated parameter adjustments and defect rebounds. Secondly, they cannot distinguish between local parameter deviations and global process drift, lacking a layered adaptive correction mechanism. Extrusion production line process drift is divided into short-term local fluctuations, such as instantaneous screw speed oscillations and local cooling flow fluctuations, and long-term global drifts, such as die wear and continuous temperature control deviations. The correction magnitude and correction process segments for the two are completely different. Existing control systems only use a uniform fixed correction coefficient to adjust the entire process synchronously, without segmented differentiated compensation according to different stages. They cannot match the defect generation mechanisms of different extrusion processes. Small local fluctuations are prone to triggering over-correction, while the correction strength is insufficient during global drift, making it difficult to suppress the continuous deterioration of defects.

[0003] In summary, existing PVG conveyor belt cover rubber extrusion production lines suffer from multiple pain points, including incomplete defect identification, delayed detection, lack of quantitative coupling between defects and processes, extensive process control, lack of stratified adaptive correction, and lack of closed-loop verification for parameter tuning. These issues make it difficult to suppress extrusion process drift in real time and to stably control the defect level of the cover rubber. The industry urgently needs a complete technical solution that can identify surface and sub-interface defects online and achieve dynamic stratified adaptive correction of the extrusion process. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes a method and system for multi-optical path defect detection and adaptive process correction of PVG conveyor belt cover rubber. The method involves acquiring multi-optical path imaging image sequences of the cover rubber and production line process time-series parameters, extracting defect spatial scale parameters and surface and subsurface defect segmentation masks, and constructing a defect-process coupling mapping coefficient matrix. A multi-process decoupling control model is built, iteratively optimizing process variable correction coefficients with a good cover rubber benchmark as the target. Defect data of PVG conveyor belt cover rubber is collected, and the defect growth rate, comprehensive process drift rate, and their coupling ratio are calculated to determine the production line process status. If the process drift exceeds the limit, the local fine-tuning correction amount and the global process compensation correction amount are calculated separately, and the process variable correction coefficients and segmented correction strategies are updated simultaneously. This invention achieves online identification of PVG conveyor belt cover rubber defects and dynamic adaptive adjustment of the production process, effectively suppressing process drift and improving the processing quality of the cover rubber and the stability of the production line.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for multi-optical path defect detection and adaptive process correction of cover adhesive includes:

[0007] S1: Collect multi-path synchronous imaging image sequences and real-time process timing parameters of the PVG conveyor belt cover rubber to be inspected. Extract defect spatial scale parameters and surface and subsurface defect segmentation masks based on the multi-path synchronous imaging image sequences. Perform correlation analysis between the extracted defect spatial scale parameters and surface and subsurface defect segmentation masks and the synchronously acquired real-time process timing parameters to construct a defect-process coupling mapping coefficient matrix.

[0008] S2: Construct a multi-process decoupling control model based on the defect-process coupling mapping coefficient matrix. Take the defect spatial scale parameter and the good cover rubber benchmark corresponding to the surface and subsurface defect segmentation mask as the optimization target. Iteratively optimize the process variable correction coefficient of the multi-process decoupling control model until the difference between the total proportion of defects in the PVG conveyor belt cover rubber collected in real time and the proportion of defects in the good cover rubber benchmark is less than the preset first threshold. Output the process pre-correction ready signal.

[0009] S3: After receiving the process pre-correction ready signal, fine-tune the multi-process decoupling control model, collect full-dimensional defect statistics of PVG conveyor belt cover rubber segment by segment, and after continuously statistically analyzing N segments of cover rubber, calculate the defect month-on-month growth rate, comprehensive process drift rate and its coupling ratio based on the full-dimensional defect statistics.

[0010] S4: If the coupling ratio is less than the preset second threshold, the process variable correction coefficient and segmented correction strategy are used to maintain the continuous extrusion production of the cover rubber; if the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and segmented correction strategy are updated respectively based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics.

[0011] Specifically, the steps of S1 include:

[0012] S1.1: Based on the multispectral synchronous trigger acquisition device, acquire the bright field image sequence, dark field image sequence and polarization image sequence of the PVG conveyor belt cover rubber under preset lighting conditions, and merge them into a multi-optical path synchronous imaging image sequence.

[0013] S1.2: Input the multi-optical path synchronous imaging image sequence into a pre-trained multi-branch feature fusion network to extract the surface rubber texture gradient feature map, edge bubble diffraction feature map and subsurface stress distribution feature map respectively, and cascade them in the channel dimension to generate a multi-optical path fusion feature tensor;

[0014] S1.3: Input the multi-optical-path fusion feature tensor into a defect detection decoder based on depthwise separable convolution, and output defect spatial scale parameters and surface and subsurface defect segmentation masks; the surface and subsurface defect segmentation masks are binary masks, where a pixel with a mask value of 1 indicates that the corresponding pixel belongs to a defect region, and a pixel with a mask value of 0 indicates that the corresponding pixel belongs to a non-defect region; based on the surface and subsurface defect segmentation masks, count the total number of pixels contained in the defect region, and record it as the pixel area of ​​the defect region; the defect spatial scale parameters include at least the defect equivalent diameter, which is defined as the diameter of an equivalent circle with the same area as the pixel area of ​​the defect region.

[0015] Specifically, the specific steps of S1 further include:

[0016] S1.4: Synchronously collect real-time process timing parameters of the PVG cover rubber extrusion production line, perform sliding window mean filtering and timestamp alignment interpolation processing to obtain a standardized process parameter vector that corresponds one-to-one with each frame of multi-optical path synchronous imaging image sequence in the time dimension; the real-time process timing parameters include extruder screw speed, rubber extrusion feed speed, die extrusion gap and temperature-controlled cooling water flow rate.

[0017] S1.5: Calculate the Pearson correlation coefficient between the equivalent diameter of the defect and the extruder screw speed and rubber extrusion feed speed in the standardized process parameter vector to obtain the first correlation feature value;

[0018] S1.6: Calculate the mutual information entropy of the pixel area of ​​the defect region with the die extrusion gap and temperature-controlled cooling water flow rate in the standardized process parameter vector to obtain the second associated feature value;

[0019] S1.7: The first associated feature value and the second associated feature value are weighted and fused according to the preset defect type weight to construct the defect-process coupling mapping coefficient matrix.

[0020] Specifically, the steps of S2 include:

[0021] S2.1: Based on the defect-process coupling mapping coefficient matrix, a multi-process decoupling control model is constructed; the multi-process decoupling control model adopts a non-negative matrix decomposition structure based on sparse constraints, which decomposes the defect-process coupling mapping coefficient matrix into the product of the defect basis matrix and the process contribution coefficient matrix;

[0022] S2.2: Load the pre-stored PVG cover rubber good product benchmark database, extract the benchmark defect spatial scale parameter vector and benchmark surface and subsurface defect segmentation mask corresponding to the good product cover rubber with the same formula and extrusion batch as the PVG conveyor belt cover rubber to be inspected, and merge them as the cover rubber good product benchmark.

[0023] S2.3: Set the optimization objective function; the optimization objective function is composed of a weighted sum of the defect scale error term, the defect mask coincidence error term, and the process variable smoothing constraint term. The defect scale error term is obtained based on the defect spatial scale parameter and the reference defect spatial scale parameter vector, and the defect mask coincidence error term is obtained based on the surface and subsurface defect segmentation mask and the reference surface and subsurface defect segmentation mask.

[0024] Specifically, the steps of S2 further include:

[0025] S2.4: Initialize the process variable correction coefficients as unit vectors, and use a gradient descent-based optimization algorithm to iteratively update the initial process variable correction coefficients. After each iteration, calculate the updated process variable correction coefficients, and multiply the updated process variable correction coefficients element by element with the standardized process parameter vector to generate a virtual corrected extrusion process parameter sequence.

[0026] S2.5: Input the virtual modified extrusion process parameter sequence into the digital twin simulation platform to obtain the simulation defect spatial scale parameters and the simulation surface and subsurface defect segmentation mask, and substitute them into the optimization objective function to calculate the current objective function value;

[0027] S2.6: When the rate of change of the current objective function value between two adjacent iterations is less than the preset convergence threshold, stop the iteration and take the updated process variable correction coefficient corresponding to the last iteration as the current optimal process variable correction coefficient.

[0028] S2.7: Apply the current optimal process variable correction coefficient to the real-time PVG cover rubber extrusion production line, continuously produce multiple sections of cover rubber and collect actual defect detection data, calculate the total percentage of actual defects for each section of cover rubber, and calculate the average of the total percentage of actual defects for multiple sections of cover rubber, which is recorded as the real-time collected total percentage of defects in cover rubber; calculate the difference between the real-time collected total percentage of defects in cover rubber and the percentage of defects in the good cover rubber corresponding to the good cover rubber benchmark, and when the difference is less than the preset first threshold, output the process pre-correction ready signal.

[0029] Specifically, the steps of S3 include:

[0030] S3.1: Receive the process pre-correction ready signal and activate the fine-tuning mode controller;

[0031] S3.2: The fine-tuning mode controller loads the current optimal process variable correction coefficient as the initial value for fine-tuning, and loads the defect basis matrix and process contribution coefficient matrix as the objects to be fine-tuned. It uses a Bayesian optimization-based hyperparameter fine-tuning algorithm to update the defect basis matrix and process contribution coefficient matrix online, and obtains the updated defect basis matrix and process contribution coefficient matrix, which is denoted as the fine-tuned multi-process decoupling control model.

[0032] S3.3: Collect full-dimensional defect statistics of PVG conveyor belt cover rubber segment by segment; the full-dimensional defect statistics include defect type label sequence, defect spatial scale parameter vector, defect depth distribution histogram, defect location spatial coordinate set, and standardized process parameter vector output at the corresponding time for each segment of cover rubber.

[0033] S3.4: Store the full-dimensional defect statistics of each section of the cover adhesive into a circular cache queue in the order of collection time.

[0034] Specifically, the steps of S3 further include:

[0035] S3.5: When the number of cover adhesive segments stored in the circular cache queue reaches a preset number N, extract the full-dimensional defect statistics of the most recent N consecutive cover adhesive segments from the circular cache queue, and then divide the N cover adhesive segments into the first N / 2 groups and the last N / 2 groups according to the collection time order. Calculate the total number of defects in the first N / 2 groups and the last N / 2 groups respectively. Divide the total number of defects in the last N / 2 groups by the total number of defects in the first N / 2 groups and subtract 1 to obtain the month-on-month growth rate of defects.

[0036] S3.6: Based on the extracted full-dimensional defect statistics of N segments of continuous cover rubber, extract the actual values ​​of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate from the standardized process parameter vector corresponding to each segment of cover rubber. These constitute four actual value sequences. Linear fitting is performed on each actual value sequence to obtain the slope of the fitted line. The absolute value of the slope of the fitted line is then divided by the arithmetic mean of the actual value sequences to obtain the drift rate of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate, respectively. These drift rates are then weighted according to the preset process parameter weights to obtain the comprehensive process drift rate.

[0037] S3.7: Calculate the ratio of the year-on-year growth rate of defects to the overall process drift rate, denoted as the coupling ratio.

[0038] Specifically, if the coupling ratio is less than a preset second threshold, the current production line is determined to be in a stable process state. The current optimal process variable correction coefficient remains unchanged, and the current segmented correction strategy remains unchanged. The segmented correction strategy is divided into a plasticizing section, a die extrusion molding section, and a cooling and shaping section according to the extrusion process of the cover rubber. Each process stage corresponds to a set of sub-vectors of process variable correction coefficients. The segmented correction strategy is set to be a unit vector for each process stage sub-vector during initialization.

[0039] Specifically, if the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and the segmented correction strategy are updated based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics, including:

[0040] S4.1: If the coupling ratio is greater than or equal to a preset second threshold, the current production line is determined to be in a process drift exceeding the limit state, triggering adaptive correction and extracting the latest full-dimensional defect statistics of the M-segment cover adhesive from the circular buffer queue; where M is a preset positive integer less than or equal to N;

[0041] S4.2: Based on the extracted full-dimensional defect statistics of the M-segment cover adhesive, the local weighted regression algorithm is used to solve for the local fine-tuning correction amount; the local fine-tuning correction amount is an incremental vector with the same dimension as the correction coefficient of the current optimal process variable;

[0042] S4.3: Based on the full-dimensional defect statistics of all N segments of cover adhesive in the circular buffer queue, the time-series prediction model is used to predict the defect evolution trend of P segments of cover adhesive, and the model prediction control algorithm is used to perform rolling optimization to obtain the global process compensation correction amount sequence; the global process compensation correction amount sequence contains P correction amount vectors, and each correction amount vector corresponds to the global compensation value of a segment of cover adhesive.

[0043] Specifically, the step of updating the process variable correction coefficient and the segmented correction strategy based on the local fine-tuning correction amount and the global process compensation correction amount obtained by solving the full-dimensional defect statistics data if the coupling ratio is greater than or equal to the second threshold also includes:

[0044] S4.4: Add the obtained local fine-tuning correction amount to the current optimal process variable correction coefficient element by element to obtain the updated process variable correction coefficient;

[0045] S4.5: The obtained global process compensation correction sequence is averaged segment by segment according to the process stage division boundary of the defined segment correction strategy: the arithmetic mean of the correction vectors corresponding to the cover rubber covered by the rubber plasticizing section, the die extrusion molding section and the cooling and shaping section are calculated respectively, and the sub-vectors of the corresponding process stages in the current segment correction strategy are replaced with the three arithmetic means respectively to obtain the updated segment correction strategy.

[0046] Specifically, after calculating the local fine-tuning correction amount and the global process compensation correction amount based on defect statistics, and updating the process variable correction coefficients and the segmented correction strategy respectively, the process further includes:

[0047] The updated process variable correction coefficients and the updated segmented correction strategy are applied to the online extrusion production of the cover adhesive;

[0048] Continuously collect full-dimensional defect statistics of the L-segment cover adhesive, where L is less than N;

[0049] Extract full-dimensional defect statistics of the L-segment cover adhesive from the circular cache queue, calculate the actual total defect percentage of each segment of cover adhesive, calculate the average of the actual total defect percentage of the L-segment cover adhesive, and record it as the real-time total defect percentage. At the same time, calculate the difference between the real-time total defect percentage and the defect percentage of the cover adhesive good product benchmark corresponding to the cover adhesive good product benchmark, and record it as the first real-time difference.

[0050] Extract the full-dimensional defect statistics of the last N segments of cover adhesive before the update operation is performed from the circular cache queue. Calculate the actual total defect percentage of each segment of cover adhesive and the average of the actual total defect percentage of the N segments of cover adhesive, which is recorded as the historical total defect percentage. Then calculate the difference between the historical total defect percentage and the good product benchmark defect percentage of the cover adhesive, which is recorded as the second historical difference.

[0051] Specifically, after calculating the local fine-tuning correction amount and the global process compensation correction amount based on defect statistics, and updating the process variable correction coefficients and the segmented correction strategy respectively, the process further includes:

[0052] If the first real-time difference is less than the second historical difference, and the first real-time difference is less than the preset first threshold, then the update is determined to be valid, and the updated process variable correction coefficient and the updated segmented correction strategy are maintained.

[0053] If the first real-time difference is greater than or equal to the second historical difference, or if the first real-time difference is greater than or equal to a preset first threshold, the update is deemed invalid, and the process variable correction coefficient and the segmented correction strategy before the update are rolled back.

[0054] If the update is deemed invalid, an alarm signal is triggered and sent to the host computer system.

[0055] A multi-optical-path defect detection and adaptive process correction system for cover adhesive includes:

[0056] The data acquisition module is used to simultaneously acquire multi-path synchronous imaging image sequences and real-time process timing parameters of the PVG conveyor belt cover rubber to be inspected;

[0057] The coupling analysis module is used to input multi-optical path synchronous imaging image sequences into a multi-branch feature fusion network to generate multi-optical path fusion feature tensors, and then output defect spatial scale parameters, surface and subsurface defect segmentation masks through a defect detection decoder; the correlation calculation between defect parameters and process parameters is completed by Pearson correlation coefficient and mutual information entropy, and the defect-process coupling mapping coefficient matrix is ​​constructed after weighted fusion.

[0058] The process pre-correction module is used to build a multi-process decoupling control model based on the defect-process coupling mapping coefficient matrix, construct an optimization objective function with the good product benchmark of the cover adhesive as the optimization target, and determine and output the process pre-correction ready signal by comparing the actual total defect ratio of the cover adhesive with the defect ratio of the good product benchmark of the cover adhesive.

[0059] The status determination module receives the process pre-correction ready signal and performs Bayesian optimization fine-tuning on the multi-process decoupling control model. It calculates the defect month-on-month growth rate, the comprehensive process drift rate and its coupling ratio to determine whether the production line is in a stable process state or a process drift exceeding the limit state.

[0060] The process correction module is used to update the process variable correction coefficients, calculate the global process compensation correction sequence, and update the segment correction strategy according to the processing segment.

[0061] Compared with the prior art, the beneficial effects of the present invention are:

[0062] 1. This invention relies on multi-optical path imaging combined with deep learning algorithms to accurately extract the defect features and scale parameters of PVG conveyor belt cover rubber, such as bubbles, scratches, color spots, sub-interface delamination, and internal micropores. By using Pearson correlation coefficient and mutual information entropy quantitative analysis, a coupled mapping relationship between cover rubber defects and extrusion process parameters is established, enabling online real-time identification of defects in the extrusion process. This solves the industry pain points of traditional single-optical-path visual inspection missing sub-interface latent defects, offline manual sampling inspection being lagging, and incomplete defect feature extraction, providing full-dimensional and high-precision defect data support for adaptive control of the extrusion process.

[0063] 2. This invention constructs a multi-process decoupled control model adapted to the PVG cover rubber extrusion process. It iteratively optimizes the correction coefficients of extrusion process variables by combining gradient descent and Bayesian optimization algorithms. Based on the coupling ratio between the defect growth rate and the overall process drift rate, it distinguishes between local parameter fluctuations and global long-term drift. Local fine-tuning and global segmented compensation are used to achieve layered differentiated corrections for the plasticizing section, die forming section, and cooling and shaping section. At the same time, a closed-loop verification of process correction effects, automatic parameter rollback, and upper computer linkage audible and visual alarm mechanism are added. This can dynamically offset the process drift caused by extruder screw wear, die accumulation, and temperature control decay, continuously stabilize the continuous extrusion condition of cover rubber, reduce the incidence of defects such as bubbles, delamination, and uneven thickness, and effectively improve the yield of PVG conveyor belt cover rubber products and the continuous production stability of the entire extrusion line. Attached Figure Description

[0064] Figure 1 This is a schematic diagram of a method for detecting multi-optical-path defects and adaptively correcting processes in a cover adhesive according to the present invention;

[0065] Figure 2 This is a flowchart illustrating the principle of a multi-optical-path defect detection and adaptive process correction method for cover adhesive according to the present invention.

[0066] Figure 3 This is a system architecture diagram of a multi-optical-path defect detection and adaptive process correction system for cover adhesive according to the present invention. Detailed Implementation

[0067] Example 1:

[0068] Please see Figure 1 and Figure 2 The present invention provides an embodiment of a method for multi-optical path defect detection and adaptive process correction of cover adhesive, the method comprising:

[0069] S1: Collect multi-path synchronous imaging image sequences and real-time process timing parameters of the PVG conveyor belt cover rubber to be inspected. Extract defect spatial scale parameters and surface and subsurface defect segmentation masks based on the multi-path synchronous imaging image sequences. Perform correlation analysis between the extracted defect spatial scale parameters and surface and subsurface defect segmentation masks and the synchronously acquired real-time process timing parameters to construct a defect-process coupling mapping coefficient matrix.

[0070] In this embodiment, the test object is PVG conveyor belt cover rubber used for heavy-duty transportation in coal mines and ports. During the continuous extrusion molding process of this rubber compound, it is easily affected by the mixing uniformity, extrusion equipment conditions, temperature control, and cooling conditions, resulting in defects such as surface bubbles, surface scratches, rubber color spots, uneven thickness, sub-interface delamination, and internal micropores. The four core process parameters of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature control cooling water flow rate directly determine the probability of occurrence, defect size, and spatial distribution range of various defects. This method is fully adaptable to the online quality control scenario of continuous extrusion of PVG cover rubber.

[0071] The specific steps of S1 include:

[0072] S1.1: Based on the multispectral synchronous trigger acquisition device, acquire the bright field image sequence, dark field image sequence and polarization image sequence of the PVG conveyor belt cover rubber under preset lighting conditions, and merge them into a multi-optical path synchronous imaging image sequence.

[0073] In this embodiment, the multispectral synchronous trigger acquisition device is installed at the online detection station downstream of the extrusion die and the cooling and shaping section. The entire device spans the entire cover adhesive conveying platform. The device integrates three major optical paths: bright field imaging, dark field imaging, and polarization imaging. Each imaging unit is equipped with an independent industrial camera, supplementary light source, and optical lens. All imaging units share a set of hardware trigger signal modules to achieve synchronous shooting of the three images and eliminate data matching errors caused by timing misalignment.

[0074] Furthermore, considering the material characteristics of PVG rubber compounds—semi-transparent, easily reflective, and internally layered defects difficult to identify with the naked eye—the angles, intensity, and polarization angles of the three types of optical paths were precisely adjusted. For the bright-field imaging optical path, a ring-shaped diffused white light source was used to illuminate the surface of the cover rubber vertically downwards. Uniform lighting can fully reveal macroscopic surface defects such as scratches, spots, unevenness, and thickness variations. During adjustment, the light source power was gradually adjusted to avoid overexposure spots caused by specular reflection on the rubber surface, ensuring uniform brightness and complete boundaries for the entire roll of cover rubber. For the dark-field imaging optical path, a side-tilted narrow beam light source was used, with the light not directly entering the camera but only illuminating the rubber. Micro-bubbles, fine scratches, and missing material at the edges create highly scattering areas, which can magnify micron-level surface defects. During the debugging phase, multiple test shots are taken along the width of the cover rubber, and the tilt angle and illumination distance of the light source are adjusted to ensure that all micro-defects across the wide cover rubber can be clearly imaged. For the polarized imaging optical path, a polarized light source and a rotatable polarizing filter are used to penetrate the rubber surface with polarized light, weakening surface reflection interference and directly revealing hidden defects such as internal stress accumulation, sub-interface delamination, and internal closed micropores. During debugging, multiple PVG cover rubber samples with known delamination and micropore defects are replaced, and the polarizing filter is repeatedly rotated to select the optimal imaging angle and lock the fixed optical path operating parameters.

[0075] Furthermore, after completing the calibration of all optical path parameters, the synchronous triggering mechanism is activated, and the multispectral synchronous triggering acquisition device is connected to the infrared position sensor on the cover conveyor line. When the continuously extruded PVG cover adhesive is delivered to the fixed area of ​​the detection station, the sensor outputs a synchronous trigger pulse, which is simultaneously sent to three sets of industrial cameras. Exposure and shooting are completed at the same time, and each set of triggers synchronously outputs one frame of bright field, one frame of dark field, and one frame of polarization image. The device continuously acquires and generates three types of image sequences with time alignment, and the images with the same timestamp correspond one-to-one with the same piece of cover adhesive. The device's built-in high-speed cache module merges the three types of images according to the acquisition time sequence to generate a complete multi-optical path synchronous imaging image sequence.

[0076] S1.2: Input the multi-optical path synchronous imaging image sequence into a pre-trained multi-branch feature fusion network to extract the surface rubber texture gradient feature map, edge bubble diffraction feature map and subsurface stress distribution feature map respectively, and cascade them in the channel dimension to generate a multi-optical path fusion feature tensor;

[0077] In this embodiment, the multi-branch feature fusion network adopts a three-branch parallel structure. The three branches correspond to three types of data sources: bright field image, dark field image, and polarization image. The weights of the three branches are trained independently, and the back-end shares a unified channel fusion layer to complete feature stitching. The input is uniformly scaled to a 512×512 pixel RGB three-channel image.

[0078] Furthermore, the first branch processes the bright-field image sequence, with the core task of extracting the gradient feature map of the surface rubber texture. The bright-field image fully preserves information such as the overall rubber texture, planar undulations, large-area scratches, and color spot depressions of the PVG covering rubber surface. The first branch adopts a lightweight ResNet18 residual convolutional structure with a total of 18 layers, including one input convolutional layer, four sets of residual blocks, and one gradient extraction convolutional transition layer. Each residual block consists of two 3×3 standard convolutional layers, a batch normalized (BN) layer, a ReLU nonlinear activation function, and a shortened residual connection. The convolutional stride is uniformly set to 1, and the padding value is 1 to ensure that the feature map space size remains unchanged throughout. The image region is analyzed pixel by pixel through multi-layer convolutional operations. The network analyzes the grayscale variation trend, texture direction, and pixel gradient differences within the domain. In areas with intact rubber, pixel grayscale changes are gradual, and gradient features are uniform. However, in defective areas with surface scratches or unevenness, adjacent pixel grayscale changes drastically, and gradient features form obvious abrupt feature changes. The network incorporates a 0.2-scale Dropout layer to filter out irrelevant weak light and shadow noise and dust, ultimately outputting a surface rubber texture gradient feature map with 64 channels and a size of 512×512, which accurately represents the surface rubber texture state and surface defect distribution of the PVG cover rubber. Each brightfield image generates an independent surface rubber texture gradient feature map, forming a continuous feature map sequence.

[0079] Furthermore, the second branch processes the dark-field image sequence, with the core task of extracting diffraction feature maps of edge bubbles. The imaging principle of dark-field images dictates their high sensitivity to structurally abrupt regions such as tiny bubbles in the adhesive, minor scratches, and missing material on the sides. The second branch employs a lightweight branch of MobileNetV1's depthwise separable convolution, comprising 16 computational units: one input convolutional layer, seven sets of depthwise separable convolutional blocks, and one diffraction feature enhancement convolutional layer. Each set of depthwise separable convolutions is split into 3×3 depthwise convolutions and 1×1 pointwise convolutions, coupled with a LeakyReLU activation function with a negative half-axis slope of 0.1. This branch weakens low-light noise in dark areas and focuses on edge contours and light bubble diffraction areas in images. It captures the integrity of workpiece contour edges and the light scattering and bubble diffraction features of local areas through feature operations. When the cover adhesive is intact and without minor defects, the bubble diffraction features of the edge areas are continuous and regular. When there are bubbles, scratches, or defects in the cover adhesive, the light scattering path is disordered, and the diffraction features are broken and bright. After the network strips away invalid background noise, it outputs an edge bubble diffraction feature map with 64 channels and a size of 512×512, which intuitively marks the location and area of ​​minor surface defects and forms a continuous feature sequence over time.

[0080] Furthermore, the third branch processes the polarization image sequence, with the core task of extracting subsurface stress distribution feature maps. Polarization images mainly reflect the internal stress distribution, interlayer separation, and internal closed micropores of the rubber compound. These defects do not show obvious surface changes but will alter the propagation path of polarized light. The third branch uses an improved U-Net downsampling encoder structure with 12 layers, including 4 downsampling convolutional blocks and multiple attention gating modules. The activation function is GELU, which is adapted to smooth the polarization light feature distribution. The polarization features in the normal areas of rubber compound mixing and extrusion processes are uniform and stable, while the polarization features in the delamination and micropore areas caused by insufficient plasticization, uneven cooling, and die pressure fluctuations are clustered and disordered. The network removes optical clutter, accurately locates the sub-interface hidden defect areas, and outputs a sub-interface stress distribution feature map with 64 channels and a size of 512×512.

[0081] Furthermore, after the three branches have completed feature extraction of three sets of images in a single frame and generated surface adhesive texture gradient feature map, edge bubble diffraction feature map, and subsurface stress distribution feature map, the surface adhesive texture gradient feature map, edge bubble diffraction feature map, and subsurface stress distribution feature map are spliced ​​and integrated at the data channel level without changing the spatial size and pixel position relationship of each feature map. Simply put, it is to combine three different types of surface adhesive texture gradient feature map, edge bubble diffraction feature map, and subsurface stress distribution feature map corresponding to the same cover adhesive at the same time into a whole data structure according to the channel order. This integrated multidimensional data structure is the multi-optical path fusion feature tensor.

[0082] S1.3: Input the multi-optical-path fusion feature tensor into a defect detection decoder based on depthwise separable convolution, and output defect spatial scale parameters and surface and subsurface defect segmentation masks; the surface and subsurface defect segmentation masks are binary masks, where a pixel with a mask value of 1 indicates that the pixel belongs to a defect region, and a pixel with a mask value of 0 indicates that the pixel belongs to a non-defect region; based on the surface and subsurface defect segmentation masks, count the total number of pixels contained in the defect region, and record it as the defect region pixel area; the defect spatial scale parameters include at least the defect equivalent diameter, which is defined as the diameter of an equivalent circle with the same area as the defect region pixel area;

[0083] Furthermore, the parameter configuration process before the defect detection decoder runs is as follows:

[0084] (1) Complete the network basic structure and parameter settings of the depthwise separable convolutional units. Specifically, the decoder body has a total of 18 computational network layers, including 15 depthwise separable convolutional layers, 2 ordinary convolutional transition layers, and 1 output mapping layer. Each depthwise separable convolutional unit is split into two independent operation parts: depthwise convolution and pointwise convolution. The depthwise convolution part is uniformly set with a kernel size of 3×3, a stride of 1, and a padding pixel count of 1. This parameter combination can ensure that the spatial size of the feature map remains unchanged after the depthwise convolution operation, thus avoiding the deviation of the defect location coordinates. The pointwise convolution part is uniformly set with a kernel size of 1×1, a stride of 1, and a padding pixel count of 0. This is mainly used to complete information fusion and dimensionality adjustment between feature channels. The weight and bias parameters inside all depthwise separable convolution units are trained and converged offline using 24,000 sets of adhesive samples containing surface and subsurface defects. The training phase has a base learning rate of 0.01, 300 iterations, and 16 batches of samples. After training, all weight and bias values ​​are fixed and the calculation is performed directly using the fixed parameters during the online execution phase.

[0085] (2) Complete the network layer channel number and activation function parameter settings. Specifically, the number of input feature channels of the first layer of the depthwise separable convolution of the decoder matches the number of multi-optical path fusion feature tensor channels of the upstream output, and is fixed at 192 channels. The number of output feature channels of the first layer of the depthwise separable convolution is set to 256 channels. The depthwise separable convolutions from the 2nd to the 8th layers follow the channel number increment rule. The number of output channels of each layer increases by 32 channels on the basis of the previous layer. The number of output channels of the 8th layer of the depthwise separable convolution reaches 480 channels. The number of output channels of the 9th to the 15th layers of the depthwise separable convolution remains stable at 480 channels. Only feature depth analysis is completed and the channel size is not adjusted. After all convolution operations in the network, a non-linear activation function is used. The threshold parameter inside the activation function remains at the default standard configuration to introduce non-linear feature expression capability and distinguish the feature differences between normal and defective regions.

[0086] (3) The decoder has a preset defect feature discrimination threshold. In this embodiment, it is determined by a large number of sample statistics and is fixed at 0.55. If the single-point feature response value obtained by the network parsing is greater than the defect feature discrimination threshold, the corresponding pixel position is determined to belong to the defect area. If it is less than or equal to the defect feature discrimination threshold, the corresponding pixel position is determined to belong to the normal area. At the same time, the output data format rules are defined in advance, and the complete set of defect spatial scale parameters that need to be output synchronously include the surface and subsurface defect segmentation mask, the pixel statistics results of the defect area, and the equivalent diameter of the defect are determined to be completely consistent with the size of the mask image and the size of the original imaging image.

[0087] Furthermore, after the multi-branch feature fusion network completes feature concatenation and generates a complete multi-optical path fusion feature tensor, it transmits them sequentially to the input of the defect detection decoder. The defect detection decoder first performs standardization preprocessing, which includes: the input of the defect detection decoder monitors the data transmission channel in real time. When it detects that a complete set of multi-optical path fusion feature tensors has been transmitted, it first verifies whether the spatial size and number of channels of the multi-optical path fusion feature tensor are consistent with the network's preset input standard. In the current scene, the spatial height of the input tensor is fixed at 512 pixels, the spatial width is fixed at 512 pixels, and the number of feature channels is fixed at 192 channels. After the verification is passed, the set of features is officially received. If the multi-path fusion feature tensor fails verification, abnormal data is automatically discarded and the system waits for the next set of normal data input. Next, all floating-point feature values ​​within the multi-path fusion feature tensor are uniformly mapped to a value range of 0 to 1. The normalization process uses a linear mapping method, first extracting the global maximum and global minimum feature values ​​within the current multi-path fusion feature tensor, and then scaling the values ​​of the entire set of data according to a fixed transformation rule. This ensures that the value distribution range of each set of input data remains consistent, preventing fluctuations in the original feature values ​​from affecting the convolution operation and defect discrimination results. After all preprocessing is completed, the standardized multi-path fusion feature tensor is officially fed into the first layer of the depthwise separable convolutional unit of the defect detection decoder.

[0088] Furthermore, the standardized multi-path fusion feature tensor sequentially passes through 15 depthwise separable convolutional layers and two ordinary convolutional transition layers within the defect detection decoder. Layers 1 to 4, the depthwise separable convolutional layers, serve as the shallow layers of the network, primarily analyzing basic visual features of the image. These four convolutional layers, relying on fixed 3×3 depth convolutional kernels and 1×1 pointwise convolutional kernels, progressively traverse all spatial locations and feature channels of the feature tensor, focusing on extracting fundamental features such as the texture variations, edge contours, and local brightness differences of the PVG overlay adhesive surface material, while simultaneously filtering out over-image processing. In the process, minor noise, dust interference, and other invalid features are processed by activation functions after each convolutional operation to amplify the response intensity of effective features and weaken useless interference features. After the four shallow convolutional layers are completed, a shallow feature map fused with basic visual information is output. The 5th to 8th depthwise separable convolutional layers focus on completing the fusion of locally related features. The four middle-layer convolutional layers expand the feature receptive field based on the shallow features, and perform correlation analysis on the feature information of adjacent pixels and adjacent regions, and begin to distinguish between normal flat areas and areas with scratches and bumps. The network first identifies anomalous areas with localized defects and then initially separates the surface imaging features from the subsurface stress features corresponding to the polarization optical path. During this stage, each convolutional layer continuously increases the number of output channels. After the 8th convolutional layer finishes running, the number of feature channels reaches the preset 480 channels, completing the mid-level feature parsing. The 9th to 15th layers of deep separable convolutions serve as the deep structure of the network, specifically extracting high-dimensional features unique to defects. These seven convolutional layers do not adjust the number of feature channels, maintaining a stable 480 channels throughout, further exploring the latent features hidden within the feature tensor, and focusing on strengthening three types of core defect features: surface adhesive texture defects corresponding to bright-field images, edge bubble diffraction defects corresponding to dark-field images, and subsurface stress distribution defects corresponding to polarization images. After repeated parsing and feature enhancement through multiple layers of deep separable convolutions, the network can clearly distinguish the feature patterns corresponding to normal cover adhesive areas, surface defect areas, and subsurface defect areas. Finally, the 15th deep separable convolutional layer outputs an integrated global defect feature map, which centrally carries the defect feature response information of all locations of the current cover adhesive.

[0089] Furthermore, after the defect detection decoder obtains the global defect feature map output by the deep network, it initiates pixel-level traversal and discrimination. Combined with a pre-set defect feature discrimination threshold of 0.55, it distinguishes between intact and defective areas of the PVG overlay adhesive pixel by pixel. The entire traversal process is executed sequentially from top to bottom and from left to right, covering all pixels within a 512-pixel height and 512-pixel width range. Specifically: first, the traversal rules and reading method are determined. The spatial dimensions of the global defect feature map are also 512 pixels high and 512 pixels wide. Each pixel position corresponds to a set of feature response values, which represent the significance of the defect feature at the current pixel position. The system follows a row-first order, starting from the [number]th [pixel] in the global defect feature map... Starting with the first pixel in a row, the system sequentially reads the feature response value corresponding to each pixel, processing only a single pixel at a time. After processing, it jumps to the next adjacent pixel, completing the full image traversal row by row and column by column. Next, it performs single-pixel defect type discrimination. For each read feature response value, it compares it with the defect feature discrimination threshold. If the feature response value of the current pixel is greater than 0.55, the system determines that there are surface bubbles / scratches or sub-interface delamination / micropore defects at the PVG cover adhesive position corresponding to that pixel. If the feature response value of the current pixel is less than or equal to 0.55, the system determines that the adhesive at that position is intact and is a defect-free pixel. After all pixels are discriminated, the system outputs a surface and sub-surface defect segmentation mask adapted to the PVG cover adhesive.

[0090] Furthermore, after the segmentation mask is generated, the total number of pixels with a value of 1 is counted. This count represents the pixel area of ​​the defect region. The specific process is as follows: First, the generated and cached surface and subsurface defect segmentation masks are read, and the traversal is still performed in the pixel order from top to bottom and from left to right. At the same time, a statistical counter is initialized, with its initial value set to zero, specifically for accumulating the number of defect pixels. Second, the mask values ​​are read pixel by pixel and the counting is completed. During the traversal, the assignment result of each individual pixel in the mask is read one by one. When a pixel value of 1 is read, it means that the current pixel belongs to the defect region, and the value of the statistical counter is automatically incremented by 1. When a pixel value of 0 is read, it means that the current pixel belongs to the non-defect region, and the counter value remains unchanged. The traversal continues to the next pixel. The entire traversal process is executed continuously until all pixels in the mask image have been read and judged. When the entire mask image is traversed, the final value of the statistical counter is the total pixel area of ​​the overall defect region of the PVG conveyor belt cover rubber in the current segment.

[0091] Furthermore, a connected component detection algorithm is used to segment the surface and subsurface defect mask. The connected component detection algorithm sets the connectivity judgment rule to four-neighbor connectivity, that is, only the four adjacent pixels above, below, left, and right of the current pixel are judged, and diagonally adjacent pixels are not included in the connectivity range. According to this rule, the surface and subsurface defect mask image is traversed, and pixels that are connected to each other and have a value of 1 are divided into the same independent defect region. Finally, all independent defect blocks in the image are distinguished, and each independent block corresponds to an actual PVG conveyor belt cover rubber defect. Next, for each segmented independent defect region, the total number of pixels with a value of 1 in the independent defect region is counted separately to obtain the local pixel area corresponding to a single defect. Then, according to the calculation logic of equivalent circle area, the diameter of a standard circle that is exactly equal to the area of ​​this irregular defect region is calculated based on the local pixel area. This standard circle diameter is the defect equivalent diameter corresponding to a single defect. For images with multiple independent defect regions, the equivalent diameter of each defect is calculated in turn, while retaining the position information of each defect.

[0092] S1.4: Synchronously collect real-time process timing parameters of the PVG cover rubber extrusion production line, perform sliding window mean filtering and timestamp alignment interpolation processing to obtain a standardized process parameter vector that corresponds one-to-one with each frame of multi-optical path synchronous imaging image sequence in the time dimension; the real-time process timing parameters include extruder screw speed, rubber extrusion feed speed, die extrusion gap and temperature-controlled cooling water flow rate.

[0093] Furthermore, the specific steps in S1.4 include:

[0094] (1) Dedicated data acquisition terminals are deployed in the central control system of the entire PVG cover rubber extrusion production line. These terminals are connected to four signal output ports: the extruder spindle encoder, the traction feed servo drive, the die lip gap laser displacement sensor, and the cooling water turbine flow meter. The spindle encoder outputs the actual value of the extruder screw speed, and the sampling frequency is set to 1000Hz, that is, one extruder screw speed data point is collected every 1ms. The traction feed servo drive feeds back the actual value of the rubber extrusion feed speed, and the sampling frequency is also set to 1000Hz. The die lip gap laser displacement sensor collects the actual value of the die head extrusion gap and is installed next to the extrusion die head outlet to detect the opening and closing gap of the PVG cover rubber forming die lip in real time online, and the sampling frequency is set to 500Hz. The cooling water turbine flow meter collects the actual value of the temperature-controlled cooling water flow rate and is installed on the main cooling pipe shared by the die head and the shaping section to collect the cooling water flow rate, and the sampling frequency is set to 100Hz.

[0095] (2) All sensor data streams are aligned through a time synchronization protocol. The precise time protocol, namely the IEEE1588 protocol, is used to ensure that signals from different sampling frequencies are recorded under a unified system time reference. The data acquisition terminal packages the collected extruder screw speed sequence, rubber extrusion feed speed sequence, die extrusion gap sequence and temperature control cooling water flow rate sequence together with their respective timestamps and pushes them to the data processing server in real time.

[0096] (3) After receiving the four original process parameter sequences, the data processing server first performs sliding window mean filtering to eliminate sensor noise and instantaneous fluctuations. The specific parameter settings for sliding window mean filtering are as follows: Since the extruder screw speed sequence and rubber extrusion feed speed sequence have relatively stable signals due to their high sampling frequency, the window width is set to 50 sampling points, which corresponds to a time length of 0.05s; The die extrusion gap sequence and temperature control cooling water flow rate sequence are easily affected by rubber melt pressure and water pulsation, so the window width is set to 20 sampling points, where the single window of the die extrusion gap sequence corresponds to 0.04s, and 0.2s for a single window of the temperature-controlled cooling water flow rate sequence; during the filtering calculation, for each sampling point, the arithmetic mean of all sampling points within half the width of the window before and after it is taken as the filtered value of that point. When the number of sampling points at the window boundary is insufficient, the data is supplemented by mirror filling, that is, the average value is calculated based on the actual sampling points within the window; after completing the sliding window mean filtering, four filtered process parameter sequences are obtained, which are respectively denoted as the extruder screw speed filtering sequence, the rubber extrusion feed speed filtering sequence, the die extrusion gap filtering sequence, and the temperature-controlled cooling water flow rate filtering sequence;

[0097] (4) Read the timestamp of each frame image from the multispectral synchronous trigger acquisition device and record it as the image frame time. Then, for each image frame time, find two sampling points adjacent to the image frame time in the extruder screw speed filtering sequence, rubber extrusion feed speed filtering sequence, die extrusion gap filtering sequence, and temperature-controlled cooling water flow rate filtering sequence, respectively, and record them as the previous sampling point and the next sampling point. The timestamp of the previous sampling point is recorded as... The corresponding process parameter values ​​are denoted as The timestamp of the next sampling point is recorded as The corresponding process parameter values ​​are denoted as If the image frame time is exactly equal to the timestamp of any sampling point, the value of that sampling point is directly taken as the interpolation result; if the image frame time is between two sampling points, the interpolation result is calculated according to the linear interpolation formula; the extruder screw speed filtering sequence, rubber extrusion feed speed filtering sequence, die extrusion gap filtering sequence, and temperature-controlled cooling water flow rate filtering sequence are searched and calculated respectively to obtain four interpolation results corresponding to the image frame time, which are denoted as extruder screw speed interpolation, rubber extrusion feed speed interpolation, die extrusion gap interpolation, and temperature-controlled cooling water flow rate interpolation respectively;

[0098] (5) The extruder screw speed interpolation, rubber extrusion feed speed interpolation, die extrusion gap interpolation and temperature control cooling water flow rate interpolation are combined into a four-dimensional vector according to the fixed order of extruder screw speed, rubber extrusion feed speed, die extrusion gap and temperature control cooling water flow rate. This vector is the standardized process parameter vector that corresponds one-to-one with the frame image in the time dimension.

[0099] S1.5: Calculate the Pearson correlation coefficient between the equivalent diameter of the defect and the extruder screw speed and rubber extrusion feed speed in the standardized process parameter vector to obtain the first correlation feature value;

[0100] Furthermore, the specific steps in S1.5 include:

[0101] (1) Extract the equivalent diameter value of the defect corresponding to the current frame image from the output result of the defect detection decoder. The equivalent diameter of the defect is defined as the diameter of the equivalent circle with the same pixel area as the defect area. The value is in millimeters.

[0102] (2) Read the standardized process parameter vector corresponding to the current frame image timestamp from the data processing server. The standardized process parameter vector is a four-dimensional vector, which contains the extruder screw speed value, rubber extrusion feed speed value, die extrusion gap value and temperature control cooling water flow rate value in a fixed order. Here, only the extruder screw speed value and rubber extrusion feed speed value are extracted. The unit of extruder screw speed is rad / min, and the unit of rubber extrusion feed speed is mm / min.

[0103] (3) Set the width of the sliding window to 60 segments of cover rubber collected continuously. For the 60 segments of cover rubber in the window, extract the defect equivalent diameter, extruder screw speed and rubber extrusion feed speed of each segment of cover rubber to form three sets of sequences with a length of 60. The first set is the defect equivalent diameter sequence, the second set is the extruder screw speed sequence, and the third set is the rubber extrusion feed speed sequence.

[0104] (4) Calculate the Pearson correlation coefficient between the defect equivalent diameter sequence and the extruder screw speed sequence, and the Pearson correlation coefficient between the defect equivalent diameter sequence and the rubber extrusion feed rate sequence, respectively;

[0105] Furthermore, when calculating the Pearson correlation coefficient between the defect equivalent diameter sequence and the extruder screw speed sequence, the calculation process for the Pearson correlation coefficient is as follows:

[0106] Calculate the arithmetic mean of the defect equivalent diameter sequence and the arithmetic mean of the extruder screw speed sequence;

[0107] The covariance between the defect equivalent diameter sequence and the extruder screw speed sequence is calculated. Specifically, for each segment of cover rubber numbered from 1 to 60, the difference between the defect equivalent diameter and the average defect equivalent diameter of the segment of cover rubber, and the difference between the extruder screw speed and the average extruder screw speed of the segment of cover rubber are calculated. These two differences are multiplied to obtain the product value of the segment of cover rubber. The product values ​​of the 60 segments of cover rubber are then summed to obtain the cumulative sum. Finally, the cumulative sum is divided by 60 to obtain the covariance.

[0108] Calculate the standard deviation of the defect equivalent diameter sequence and the standard deviation of the extruder screw speed sequence;

[0109] The Pearson correlation coefficient between the defect equivalent diameter and the extruder screw speed is obtained by dividing the covariance between the defect equivalent diameter and the extruder screw speed by the product of the standard deviation of the defect equivalent diameter and the standard deviation of the extruder screw speed. This coefficient is denoted as the extruder screw speed correlation coefficient and its value ranges from -1 to 1.

[0110] Furthermore, similarly, the Pearson correlation coefficient between the defect equivalent diameter sequence and the rubber extrusion feed rate sequence can be calculated. This simply requires replacing the extruder screw speed sequence with the rubber extrusion feed rate sequence. Specifically, this involves calculating the arithmetic mean of the rubber extrusion feed rate sequence, calculating the covariance between the defect equivalent diameter sequence and the rubber extrusion feed rate sequence, calculating the standard deviation of the rubber extrusion feed rate sequence, and finally dividing the covariance by the product of the standard deviation of the defect equivalent diameter and the standard deviation of the rubber extrusion feed rate to obtain the Pearson correlation coefficient between the defect equivalent diameter and the rubber extrusion feed rate, denoted as the rubber extrusion feed rate correlation coefficient.

[0111] (5) The Pearson correlation coefficient between the obtained defect equivalent diameter and the extruder screw speed, and the Pearson correlation coefficient between the defect equivalent diameter and the rubber extrusion feed rate are weighted and fused according to the preset weights. The preset weights are set according to process experience, with the weight of the extruder screw speed being 0.6 and the weight of the rubber extrusion feed rate being 0.4. The calculation formula for the weighted fusion is that the first correlation feature value is equal to the extruder screw speed correlation coefficient multiplied by 0.6 plus the rubber extrusion feed rate correlation coefficient multiplied by 0.4.

[0112] S1.6: Calculate the mutual information entropy of the pixel area of ​​the defect region with the die extrusion gap and temperature-controlled cooling water flow rate in the standardized process parameter vector to obtain the second associated feature value;

[0113] Furthermore, the specific steps in S1.6 include:

[0114] (1) Extract the pixel area value of the defect region corresponding to the current frame image from the output result of the defect detection decoder. The pixel area of ​​the defect region is defined as the total number of pixels with a mask value of 1 in the surface and subsurface defect segmentation mask. The value is expressed in units of pixels.

[0115] (2) Read the standardized process parameter vector corresponding to the timestamp of the current frame image from the data processing server. The standardized process parameter vector is a four-dimensional vector that contains the extruder screw speed value, the rubber extrusion feed speed value, the die extrusion gap value and the temperature control cooling water flow rate value in a fixed order. Here, only the die extrusion gap value and the temperature control cooling water flow rate value are extracted. The unit of the die extrusion gap is mm and the unit of the temperature control cooling water flow rate is L / min.

[0116] (3) Set the width of the sliding window to 120 segments of cover adhesive collected continuously. For the 120 segments of cover adhesive in the window, extract the pixel area of ​​the defect area, the die extrusion gap and the temperature control cooling water flow rate of each segment of cover adhesive to form three sets of sequences with a length of 120. The first set is the pixel area sequence of the defect area, the second set is the die extrusion gap sequence and the third set is the temperature control cooling water flow rate sequence.

[0117] (4) Set the number of intervals to 20, find the minimum and maximum values ​​in the defect area pixel area sequence, divide the minimum to the maximum value into 20 intervals, determine which interval each value in the defect area pixel area sequence falls into, and replace the value with the interval number, numbered from 1 to 20. After discretization, the original defect area pixel area sequence is converted into a sequence of 120 discrete values ​​composed of interval numbers.

[0118] (5) Find the minimum and maximum values ​​in the die extrusion gap sequence, divide the minimum to the maximum value into 20 intervals, and then for each value in the die extrusion gap sequence, determine which interval it falls into and replace the value with the interval number, numbered from 1 to 20. After discretization, the die extrusion gap sequence is also converted into a sequence of 120 discrete values ​​composed of interval numbers.

[0119] (6) Similarly, find the minimum and maximum values ​​in the temperature-controlled cooling water flow rate sequence, and divide the minimum to the maximum value into 20 intervals. Then, for each value in the temperature-controlled cooling water flow rate sequence, determine which interval it falls into and replace the value with the interval number. After discretization, the temperature-controlled cooling water flow rate sequence is converted into a sequence of 120 discrete values ​​composed of interval numbers.

[0120] (7) Calculate the mutual information entropy between the pixel area of ​​the defect region and the die extrusion gap;

[0121] Furthermore, the calculation process for mutual information entropy is as follows:

[0122] The joint probability distribution of defect region pixel area and die extrusion gap is constructed as follows: A 20×20 joint counting matrix is ​​created, with row indices corresponding to the interval numbers of defect region pixel area from 1 to 20, and column indices corresponding to the interval numbers of die extrusion gap from 1 to 20. All elements in the matrix are initialized to zero. The 120 PVG conveyor belt cover rubber samples within the sliding window are traversed, and the discrete interval numbers of defect region pixel area of ​​each cover rubber sample are extracted as row indices and discrete interval numbers of die extrusion gap as column indices. The corresponding matrix position count value is incremented by 1. After traversing all 120 cover rubber samples, each element in the joint counting matrix represents the frequency of the corresponding combination. Each frequency in the joint counting matrix is ​​divided by 120 to obtain the value of each element in the joint probability distribution matrix.

[0123] The specific process for calculating the edge probability distribution of the pixel area of ​​the defect region is as follows: For each interval number of the pixel area of ​​the defect region from 1 to 20, the joint probability values ​​of all columns in the row corresponding to the interval number are added together to obtain the edge probability value of the interval number. After calculation, 20 edge probability values ​​are obtained, each value representing the probability that the pixel area of ​​the defect region falls into the corresponding interval.

[0124] The marginal probability distribution of the die extrusion gap is calculated as follows: For each interval of the die extrusion gap, numbered from 1 to 20, the joint probability values ​​of all rows in the column corresponding to the interval number are added together to obtain the marginal probability value of the interval number. After calculation, 20 marginal probability values ​​are obtained, each value representing the probability that the die extrusion gap falls into the corresponding interval.

[0125] For each non-zero element in the joint probability distribution matrix, extract the joint probability value at that position, along with the edge probability values ​​of the defect region pixel area interval represented by the row index corresponding to that position, and the edge probability values ​​of the die extrusion gap interval represented by the column index corresponding to that position. Calculate the joint probability value multiplied by a base-2 logarithm, which is obtained by dividing the joint probability value by the product of the edge probability values ​​and then taking the base-2 logarithm. Sum the calculation results of all non-zero elements, and the sum is the mutual information entropy value between the defect region pixel area and the die extrusion gap, denoted as the die extrusion gap mutual information entropy value, with the unit being bits.

[0126] Furthermore, similarly, the die extrusion gap sequence is replaced with a temperature-controlled cooling water flow rate sequence. The mutual information entropy between the defect region pixel area and the temperature-controlled cooling water flow rate is calculated. Specifically, this includes constructing a joint probability distribution of the defect region pixel area and the temperature-controlled cooling water flow rate, calculating the edge probability distribution of the defect region pixel area, calculating the edge probability distribution of the temperature-controlled cooling water flow rate, and finally calculating the mutual information entropy value in the same way to obtain the mutual information entropy value between the defect region pixel area and the temperature-controlled cooling water flow rate, which is denoted as the temperature-controlled cooling water flow rate mutual information entropy value.

[0127] (8) The mutual information entropy values ​​between the pixel area of ​​the defect region and the die extrusion gap, and the mutual information entropy values ​​between the pixel area of ​​the defect region and the temperature-controlled cooling water flow rate are weighted and fused according to the preset weights. The preset weights are set according to process experience, with the weight of the die extrusion gap being 0.7 and the weight of the temperature-controlled cooling water flow rate being 0.3. The weighted fusion calculation method is that the second associated feature value is equal to the mutual information entropy value of the die extrusion gap multiplied by 0.7 plus the mutual information entropy value of the temperature-controlled cooling water flow rate multiplied by 0.3.

[0128] S1.7: The first associated feature value and the second associated feature value are weighted and fused according to the preset defect type weight to construct the defect-process coupling mapping coefficient matrix.

[0129] Furthermore, the specific steps in S1.7 include:

[0130] (1) Read the most recently calculated first and second correlation feature values ​​from the data cache;

[0131] (2) Obtain the defect type label corresponding to the current frame image. The defect type label is output by the defect detection decoder at the same time as outputting the defect spatial scale parameter and defect segmentation mask. The defect type label is divided into three categories according to the morphological and positional characteristics of the defect: the first category is surface bubble scratch defects, the second category is surface uneven defects, and the third category is sub-interface layered micropore defects. The defect detection decoder outputs a type label for each detected defect area. If there are multiple defects in a section of cover adhesive, the frequency of occurrence of each of the three types of defects is counted, and the defect type with the highest frequency is taken as the defect type label of this section of cover adhesive. For example, if 5 bubble scratches, 2 uneven thicknesses, and 1 layered micropore are detected in a single section of cover adhesive in the current frame image, then the defect type label of this section is determined to be surface bubble scratches.

[0132] (3) Obtain the preset defect type weight vector according to the defect type label. The defect type weight vector is a three-dimensional vector. The three dimensions correspond to the weight values ​​of surface bubble scratch defects, surface unevenness defects and sub-interface layered micropore defects, respectively. The weight vector values ​​corresponding to different defect types are different. The weight vector is set according to the degree of hazard of PVG cover glue defects and the priority of extrusion process control. In this embodiment, for surface bubble scratch defects, the weight vector is set to the weight of the first associated feature value as 0.3 and the weight of the second associated feature value as 0.7; for surface unevenness defects, the weight vector is set to the weight of the first associated feature value as 0.6 and the weight of the second associated feature value as 0.4; for sub-interface layered micropore defects, the weight vector is set to the weight of the first associated feature value as 0.8 and the weight of the second associated feature value as 0.2; Since the defect type label of the current frame image is surface bubble scratch defects, the weight vector corresponding to surface bubble scratch defects is adopted, that is, the weight of the first associated feature value is 0.3 and the weight of the second associated feature value is 0.7.

[0133] It is important to understand that for surface bubble scratches, the scratches and tiny surface bubbles are mostly caused by die head material accumulation, cooling water flow rate oscillation, and die lip gap burrs. Die head and cooling-related parameters are the dominant factors in defect generation. Screw speed and feed rate only indirectly affect the total number of bubbles. Therefore, the weight of the first associated feature is reduced to 0.3, and the weight of the second associated feature is increased to 0.7. Surface bubble scratches only damage the appearance and do not damage the internal skeleton structure of the conveyor belt. They can be slightly repaired and have the lowest degree of harm. Their control priority is lower than that of delamination and unevenness defects. For surface unevenness defects, the causes are affected by two types of processes in a balanced manner. Abnormal screw feed rate will cause fluctuations in melt delivery flow, and die head gap deviation and uneven cooling flow will directly change the forming thickness of the preform. The two types of process parameters contribute similarly to the generation of unevenness defects, and there is no single dominant process variable. Therefore, the weights are evenly allocated, with the weight of the first associated feature value being 0.6 and the weight of the second associated feature value being 0.4. Surface unevenness will cause conveyor belt deviation and local stress concentration, and its degree of harm is moderate. Process control needs to simultaneously consider the extrusion conveying and die head cooling processes. For sub-interface delamination micropore defects, these defects are caused by mismatch between extrusion screw speed and rubber feed rate, resulting in insufficient plasticizing and shearing of the rubber. The equivalent diameter of the defect directly determines the delamination extension range and the degree of micropore aggregation. Die clearance and cooling water flow rate only play a secondary role in aggravating the problem. The root cause of the defect is highly correlated with screw and feed process parameters. Therefore, the weight of the first correlation feature value is increased to 0.8, and the weight of the second correlation feature value is weakened to 0.2. At the same time, sub-interface micropore delamination is a hidden defect inside the conveyor belt and cannot be repaired by grinding later. It directly reduces the core performance of PVG conveyor belt in terms of tear resistance and delamination resistance. Process control has the highest priority, and the coupling correlation influence weight of screw and feed speed should be increased first.

[0134] (4) Multiply the read first associated feature value by the determined first associated feature value weight to obtain the first weighted component; multiply the second associated feature value by the second associated feature value weight to obtain the second weighted component; add the first weighted component and the second weighted component to obtain the fusion feature value of the current frame image;

[0135] (5) Determine the overall structure of the defect-process coupling mapping coefficient matrix. The defect-process coupling mapping coefficient matrix is ​​a 3×2 matrix. The three rows correspond to three types of defects: surface bubble scratch defects, surface unevenness defects, and sub-interface layered micropore defects. The two columns correspond to the first associated eigenvalue channel and the second associated eigenvalue channel, respectively. Each element in the matrix represents the coupling mapping coefficient between a defect type and an associated eigenvalue channel. When the matrix is ​​initialized, all elements are set to zero.

[0136] (6) Construct or update the first row of the defect-process coupling mapping coefficient matrix. The first row corresponds to the surface bubble scratch defect. Whenever a cover adhesive with a defect type label of surface bubble scratch defect is detected and the fusion feature value is calculated, the fusion feature value is decomposed into the first associated feature value contribution component and the second associated feature value contribution component, which are used to update the first row of the matrix. The specific update method adopts the moving average method. The sliding window width is set to 50 surface bubble scratch defect cover adhesives. A queue with a length of 50 is maintained to store the original values ​​of the first associated feature value and the second associated feature value of the most recent 50 surface bubble scratch defect cover adhesives. Whenever a new surface bubble scratch defect cover adhesive arrives, its first associated feature value and second associated feature value are added to the queue. At the same time, the oldest value in the queue is removed. The arithmetic mean of the 50 first associated feature values ​​in the queue is calculated and the average value is used as the new value of the first row and first column of the matrix. The arithmetic mean of the 50 second associated feature values ​​in the queue is calculated and the average value is used as the new value of the first row and second column of the matrix.

[0137] (7) Similarly, construct or update the second row of the defect-process coupling mapping coefficient matrix. For example, the second row corresponds to surface unevenness defects. Maintain a queue of length 50 to store only the data related to the cover adhesive with thick and thin unevenness as the main defect. For each new corresponding sample, add its first and second related feature values ​​to the queue and remove the oldest value in the queue. Calculate the arithmetic mean of the 50 first related feature values ​​in the queue and use this mean as the new value of the first column of the second row of the matrix. Calculate the arithmetic mean of the 50 second related feature values ​​in the queue and use this mean as the new value of the second column of the second row of the matrix.

[0138] (8) Similarly, construct or update the third row of the defect-process coupling mapping coefficient matrix. For example, the third row corresponds to the sub-interface layered micropore defect. Maintain a queue of length 50, which only includes the cover adhesive data with layered micropore as the main defect. For each new corresponding sample, add its first and second associated feature values ​​to the queue, and remove the oldest value in the queue. Calculate the arithmetic mean of the 50 first associated feature values ​​in the queue, and use this average as the new value of the first column of the third row of the matrix. Calculate the arithmetic mean of the 50 second associated feature values ​​in the queue, and use this average as the new value of the second column of the third row of the matrix.

[0139] (9) The three-row, two-column matrix constructed by (6)-(8) serves as the defect-process coupling mapping coefficient matrix at the current moment. The row index of the matrix represents three types of defects, and the column index represents two types of associated feature value channels. Each value in the matrix represents the average coupling strength between the corresponding defect type and the corresponding associated feature value channel.

[0140] S2: Construct a multi-process decoupling control model based on the defect-process coupling mapping coefficient matrix. Take the defect spatial scale parameter and the good cover rubber benchmark corresponding to the surface and subsurface defect segmentation mask as the optimization target. Iteratively optimize the process variable correction coefficient of the multi-process decoupling control model until the difference between the total proportion of defects in the PVG conveyor belt cover rubber collected in real time and the proportion of defects in the good cover rubber benchmark is less than the preset first threshold. Output the process pre-correction ready signal.

[0141] The specific steps of S2 include:

[0142] S2.1: Based on the defect-process coupling mapping coefficient matrix, a multi-process decoupling control model is constructed; the multi-process decoupling control model adopts a non-negative matrix decomposition structure based on sparse constraints, which decomposes the defect-process coupling mapping coefficient matrix into the product of the defect basis matrix and the process contribution coefficient matrix;

[0143] It is important to understand that the core structure of the multi-process decoupling control model is based on non-negative matrix decomposition with sparse constraints. Non-negative matrix decomposition means decomposing the original non-negative matrix into the product of two non-negative matrices, while sparse constraints require that, during the decomposition process, one of the matrices obtained should have as many zero elements as possible, thereby enhancing the interpretability of the model and highlighting the most important process influencing factors.

[0144] Furthermore, the construction process of the multi-process decoupling control model includes:

[0145] (1) Set the defect-process coupling mapping coefficient matrix as the original data matrix, denoted as matrix V. The size of the matrix is ​​D×P, where D represents the number of defect feature types. In this embodiment, D is 2, which represents the coupling feature value corresponding to the equivalent diameter of the defect and the coupling feature value corresponding to the pixel area of ​​the defect region, respectively. P represents the number of process parameters under consideration. According to the acquisition process, P is 4, which corresponds to the screw speed of the extruder, the extrusion feed speed of the rubber material, the extrusion gap of the die head and the flow rate of the temperature-controlled cooling water, respectively. Therefore, in this embodiment, matrix V is a 2×4 matrix.

[0146] (2) Set the decomposition target, that is, find a defect basis matrix of size D×K columns, denoted as matrix W, and a process contribution coefficient matrix of size K×P columns, denoted as matrix H, such that matrix V is approximately equal to the product of matrix W and matrix H. Here, K is a preset number of latent factors, representing the number of potential and independent process defect action modes. In this embodiment, according to the actual complexity of the multi-process production line, K is set to 3. This means that the system assumes that there are three basic process action modes. Each mode may affect multiple defect features at the same time and is contributed by multiple process parameters. Each column of matrix W represents a defect basis. A defect basis is a two-dimensional vector, and its value represents the relative probability of different defect features being triggered under this process action mode. Each row of matrix H corresponds to a defect basis, and each column corresponds to a process parameter. Its value represents the contribution weight of each process parameter under this process action mode.

[0147] (3) Based on the traditional nonnegative matrix factorization loss function, a regularization term for matrix H is added. This regularization term uses the L1 norm, that is, the absolute values ​​of all elements in matrix H are summed and multiplied by a sparsity penalty coefficient, denoted as . In this embodiment, The value is set to 0.1. This setting will cause most elements in matrix H to tend to zero, retaining only a few key process parameter contribution coefficients that are most relevant to the formation of the current defect, thereby achieving a sparse expression of process contribution.

[0148] (4) Initialize all elements of matrix W and matrix H to be non-negative random decimals with values ​​between 0 and 1. Then set the maximum number of iterations to 500 and the convergence threshold to 0.0001.

[0149] (5) In each iteration, a multiplication update rule is used to alternately update matrix W and matrix H. This rule can automatically ensure that the elements of the updated matrix are non-negative. Specifically, the current matrix W is fixed first, and matrix H is updated according to a formula. The formula for updating matrix H is: ,in, This represents the process contribution coefficient matrix for the (t+1)th iteration. This represents the process contribution coefficient matrix after the t-th iteration following the update. Represents element-wise product. Let W be the transpose of W, and a be the sparsity penalty constant, which is set to 0.1 in this embodiment to prevent the denominator from being 0 and to achieve L1 sparsity constraint. After each update of matrix H, the system will force all its elements less than zero to be set to zero, and further compress smaller coefficients to zero through the sparsity penalty term. Specifically, for each element in matrix H, if its value is less than a preset sparsity threshold such as 0.01, the element is directly set to zero. Then the system fixes the updated matrix H and updates matrix W in a symmetric manner, also using the multiplication update rule, and ensures that all elements of matrix W are non-negative.

[0150] (6) After each iteration, the system calculates the current loss function value, which consists of two parts. The first part is the reconstruction error, which is the square of the norm of the difference between the original matrix V and the products of matrices W and H, used to measure the accuracy of the decomposition. The second part is the sparsity constraint term, i.e. Multiply by the sum of the absolute values ​​of all elements of matrix H; the system subtracts the loss function value of the previous iteration from the loss function value obtained in this calculation, and then divides this difference by the absolute value of the loss function value of the previous iteration. If this ratio is less than 0.0001, or the number of iterations has reached 500, then the iteration stops.

[0151] (7) Output the defect basis matrix W and process contribution coefficient matrix H obtained by sparse constraint nonnegative matrix decomposition, where the defect basis matrix W is a 2×3 matrix and the process contribution coefficient matrix H is a 3×4 matrix.

[0152] S2.2: Load the pre-stored PVG cover rubber good product benchmark database, extract the benchmark defect spatial scale parameter vector and benchmark surface and subsurface defect segmentation mask corresponding to the good product cover rubber with the same formula and extrusion batch as the PVG conveyor belt cover rubber to be inspected, and merge them as the cover rubber good product benchmark.

[0153] Furthermore, the construction process of the good cover rubber benchmark database is as follows: First, when the production line is stable and producing recognized qualified cover rubber, operators select at least 100 segments of cover rubber that have been precisely tested offline and confirmed to be free of any surface or subsurface defects. These cover rubbers belong to the same formulation and extrusion batch as the PVG conveyor belt cover rubber to be inspected, in order to eliminate benchmark deviations caused by inherent product differences. The system sends these 100 segments of good cover rubber one by one into a multi-optical-path synchronous imaging device to obtain the defect spatial scale parameters and surface and subsurface defect segmentation masks corresponding to each good cover rubber. For a perfect good cover rubber, the equivalent diameter of the defect in its defect spatial scale parameters should be close to zero, and its defect segmentation mask should be a completely black image with all pixel values ​​being zero. However, considering the slight noise in the actual imaging and extraction process, the system does not use... Instead of using the ideal value for a single segment of cover adhesive, the system calculates the arithmetic mean of the defect spatial scale parameters of these 100 good-quality cover adhesive segments element-wise along the vector dimension to obtain a baseline defect spatial scale parameter vector. Simultaneously, the system calculates the probability average of the binary masks for the defect segmentation masks of these 100 good-quality cover adhesive segments at the same pixel position, i.e., it counts how many masks in the 100 defect segmentation masks have a value of 1 at each pixel position. If this proportion exceeds 50%, the final value of the baseline defect segmentation mask at that position is set to 1; otherwise, it is set to zero. This generates a baseline surface and subsurface defect segmentation mask. The system stores the calculated baseline defect spatial scale parameter vector and the baseline surface and subsurface defect segmentation mask, along with the corresponding formulation and extrusion batch, into a structured database, which is the good-quality cover adhesive baseline database.

[0154] Furthermore, when the system reaches the point where a good cover rubber benchmark is required, it uses the formula and batch information of the PVG conveyor belt cover rubber to be inspected as search keywords to quickly match and extract the corresponding benchmark defect spatial scale parameter vector and benchmark surface and subsurface defect segmentation mask in the cover rubber good benchmark database. These two parts of data are then merged and uniformly referred to as the good cover rubber benchmark.

[0155] S2.3: Set the optimization objective function; the optimization objective function is composed of a weighted sum of the defect scale error term, the defect mask coincidence error term, and the process variable smoothing constraint term. The defect scale error term is obtained based on the defect spatial scale parameter and the reference defect spatial scale parameter vector, and the defect mask coincidence error term is obtained based on the surface and subsurface defect segmentation mask and the reference surface and subsurface defect segmentation mask.

[0156] Furthermore, the defect scale error term measures the difference between the defect spatial scale parameters extracted from the current cover adhesive and the corresponding baseline defect spatial scale parameter vector in the good cover adhesive benchmark. Assuming that in the current iteration, the actual defect equivalent diameter extracted by the system from the multi-path image is 1.5 mm, while the baseline defect equivalent diameter in the good cover adhesive benchmark database is 0.2 mm, the system uses the squared error as the difference measure, that is, it calculates the square of the difference between the actual defect equivalent diameter and the baseline defect equivalent diameter. Therefore, the calculated value is 1.69.

[0157] Furthermore, the defect mask overlap error term measures the degree of overlap between the surface and subsurface defect segmentation mask output by the defect detection decoder of the current cover adhesive and the reference surface and subsurface defect segmentation mask in the good product benchmark of the cover adhesive. Both masks are binary images with the same size. Assuming the image resolution is 512×512 pixels, there are a total of 262,144 pixels. Pixels with a value of 1 in the mask indicate that the position belongs to the defect area, and pixels with a value of 0 indicate that the position belongs to the non-defect area. The system uses the intersection-union ratio (IUU) based on the confusion matrix to quantify the difference between the two masks. The specific calculation process is as follows: First, four intermediate variables are counted, namely the number of true cases, that is, the number of pixels in the actual mask and the reference mask where both are 1 at the same pixel position. Let the current result be 3500; the number of false positives, i.e., the number of pixels where the actual mask is 1 and the reference mask is 0, is assumed to be 450; the number of false negatives, i.e., the number of pixels where the actual mask is 0 and the reference mask is 1, is assumed to be 280; the number of true negatives, i.e., the number of pixels where both the actual mask and the reference mask are 0; the number of true negatives is divided by the sum of the number of true negatives, false positives, and false negatives to obtain the intersection-union ratio (IUGR), which, after substituting the value, yields 0.827, representing the overlap of the two masks. The closer the value is to 1, the better the overlap; the defect mask overlap error term needs to be converted into an error value that increases as the overlap decreases. Therefore, the system subtracts this IUGR from 1 to obtain 0.173, which is the defect mask overlap error term.

[0158] Furthermore, the process variable smoothing constraint term is used to prevent drastic changes in the process variable correction coefficients during the iteration process, thus avoiding impact on production. The process variable correction coefficient is a vector with the same dimension as the process parameters. In this embodiment, the dimension is 4, corresponding to the correction coefficients of the extruder screw speed, the rubber extrusion feed speed, the die extrusion gap, and the temperature-controlled cooling water flow rate, respectively. The Euclidean distance is used to calculate the degree of difference between the process variable correction coefficient vector obtained in the previous iteration and the newly generated candidate process variable correction coefficient vector in the current iteration. Specifically, the correction coefficients at each corresponding position are subtracted and then squared. The sum of the four squared values ​​is the process variable smoothing constraint term.

[0159] In this embodiment, since the defect scale is directly related to whether the cover adhesive is qualified, it is given the highest weight, and the weight of the defect scale error term is set to 0.5; the defect mask coincidence reflects whether the shape and distribution position of the defect are consistent with the good product, and its importance is second, so the weight of the defect mask coincidence error term is set to 0.3; the process variable smoothing constraint term is mainly to ensure production stability, and its importance is relatively low, so the weight of the process variable smoothing constraint term is set to 0.2.

[0160] S2.4: Initialize the process variable correction coefficients as unit vectors, and use a gradient descent-based optimization algorithm to iteratively update the initial process variable correction coefficients. After each iteration, calculate the updated process variable correction coefficients, and multiply the updated process variable correction coefficients element by element with the standardized process parameter vector to generate a virtual corrected extrusion process parameter sequence.

[0161] Furthermore, gradient descent is an optimization method that finds the minimum value by repeatedly adjusting variables along the direction of the fastest descent of the objective function. The system needs to set several key algorithm hyperparameters. The first is the learning rate, which controls the step size of the adjustment of the correction coefficient in each iteration. Setting the learning rate too large will cause the optimization process to oscillate and fail to converge, while setting it too small will make the convergence speed very slow. In this embodiment, the learning rate is set to 0.01. The second is the maximum number of iterations, which is used to prevent the optimization process from going on indefinitely. In this embodiment, the maximum number of iterations is set to 500. The third is the convergence threshold, which is used to determine whether the optimization has reached a sufficiently good level. When the rate of change of the objective function value between two adjacent iterations is less than this convergence threshold, the system considers that the optimal solution has been found and can stop the iteration in advance. In this embodiment, the convergence threshold is set to 0.0001.

[0162] S2.5: Input the virtual modified extrusion process parameter sequence into the digital twin simulation platform to obtain the simulation defect spatial scale parameters and the simulation surface and subsurface defect segmentation mask, and substitute them into the optimization objective function to calculate the current objective function value;

[0163] Furthermore, in each iteration, the system performs the following operations: First, based on the current process variable correction coefficients, a set of virtual corrected extrusion process parameter sequences is generated. This is achieved by element-wise multiplying the process variable correction coefficient vector of the current iteration with the standardized process parameter vector. After generating the virtual corrected extrusion process parameter sequence, it is input into the digital twin simulation platform. The platform simulates the defects of the cover adhesive produced under these parameters, outputting the simulated defect spatial scale parameters and the simulated surface and subsurface defect segmentation mask. Then, the system substitutes these simulation outputs into the constructed optimization objective function to calculate an objective function value, denoted as the current objective function value. The gradient of the objective function relative to the current process variable correction coefficients is calculated. Specifically, the system sequentially applies a small perturbation to each process variable correction coefficient. The magnitude of the perturbation is usually related to the value of the correction coefficient itself. In this embodiment, the perturbation is set to 1% of the current correction coefficient value, but this is to ensure computational stability. Simultaneously, a minimum disturbance is set. When 1% is less than 0.001, it is set to 0.001. For example, for the extruder screw speed correction coefficient, the current value is 1.0, and 1% of it is 0.01, which is greater than the minimum disturbance. Therefore, the disturbance is set to 0.01. The system temporarily increases the extruder screw speed correction coefficient from 1.0 to 1.01 by 0.01, while keeping the other three correction coefficients unchanged: the compound extrusion feed rate correction coefficient is 1.0, the die extrusion gap correction coefficient is 1.0, and the temperature-controlled cooling water flow rate correction coefficient is 1.0. Then, this... The new set of process variable correction coefficients is multiplied element-wise with the standardized process parameter vector to generate a new set of virtual corrected extrusion process parameter sequences. This sequence is then input into the digital twin simulation platform, and the corresponding objective function value is calculated, which is the second objective function value. The second objective function value is subtracted from the current objective function value to obtain the change in the objective function value. This change is then divided by the disturbance to obtain the approximate partial derivative of the objective function with respect to the extruder screw speed correction coefficient. The system calculates the partial derivatives of the other three correction coefficients in the same way, and these four partial derivatives form the gradient vector.

[0164] Furthermore, the system adjusts the process variable correction coefficients according to the gradient descent update rule, which is that the new process variable correction coefficient is equal to the current process variable correction coefficient minus the learning rate multiplied by the gradient.

[0165] Furthermore, the system checks two convergence conditions after each iteration: the first condition is whether the rate of change of the objective function value between two adjacent iterations is less than the convergence threshold; the second condition is whether the current iteration count has reached the maximum iteration count of 500. After stopping the iteration, the system outputs the process variable correction coefficient calculated in the last iteration as the current optimal process variable correction coefficient.

[0166] S2.6: When the rate of change of the current objective function value between two adjacent iterations is less than the preset convergence threshold, stop the iteration and take the updated process variable correction coefficient corresponding to the last iteration as the current optimal process variable correction coefficient.

[0167] S2.7: Apply the current optimal process variable correction coefficient to the real-time PVG cover rubber extrusion production line, continuously produce multiple sections of cover rubber and collect actual defect detection data, calculate the total percentage of actual defects for each section of cover rubber, and calculate the average of the total percentage of actual defects for multiple sections of cover rubber, which is recorded as the real-time collected total percentage of defects in cover rubber; calculate the difference between the real-time collected total percentage of defects in cover rubber and the percentage of defects in the good cover rubber corresponding to the good cover rubber benchmark, and when the difference is less than the preset first threshold, output the process pre-correction ready signal.

[0168] Furthermore, the specific steps in S2.7 include:

[0169] (1) The current optimal process variable correction coefficients are issued and applied to the controller of the real-time PVG cover rubber extrusion production line. Assuming that the specific values ​​of the current optimal process variable correction coefficient vector obtained through iterative optimization are: extruder screw speed correction coefficient 1.0025, rubber extrusion feed speed correction coefficient 0.9980, die extrusion gap correction coefficient 0.9972, and temperature-controlled cooling water flow rate correction coefficient 1.0015, the system multiplies these current optimal process variable correction coefficients element-wise with the currently collected standardized process parameter vector to generate the actual process parameters issued to the production line controller. The set values ​​are as follows: Assuming the current standardized process parameter vector is an extruder screw speed of 2500 rad / min, a compound extrusion feed rate of 0.15 mm / rad, a die extrusion gap of 1.2 mm, and a temperature-controlled cooling water flow rate of 5.0 L / min, the actual issued process parameter settings are an extruder screw speed of 2506.25 rad / min, a compound extrusion feed rate of 0.1497 mm / rad, a die extrusion gap of 1.19664 mm, and a temperature-controlled cooling water flow rate of 5.0075 L / min. The production line starts to continuously produce cover rubber according to this set of corrected process parameters.

[0170] (2) The system simultaneously starts online defect detection and performs actual defect detection on each continuously produced cover adhesive. That is, it acquires the bright field, dark field and polarization image sequence of the cover adhesive through a multispectral synchronous trigger acquisition device, inputs it into a multi-branch feature fusion network to extract features, and then outputs the surface and subsurface defect segmentation mask of each segment of cover adhesive through a defect detection decoder based on depth separable convolution. After the system completes the detection of each segment of cover adhesive, it immediately calculates the total percentage of actual defects of that segment of cover adhesive. The definition of the total percentage of actual defects is: in the defect segmentation mask of the current cover adhesive, the total number of pixels with a mask value of 1 accounts for the total number of pixels in the entire mask image. The system records this value as the total percentage of actual defects of the first segment of cover adhesive. Iteratively repeats. In this embodiment, 5 segments of cover adhesive are continuously acquired, and the total percentage of actual defects of each segment of cover adhesive is calculated to obtain 5 percentages.

[0171] (3) After completing the collection and statistics of the 5 sections of cover rubber, calculate the average of the total proportion of actual defects of the 5 sections of cover rubber to obtain the total proportion of defects of the real-time collected cover rubber, which represents the average defect severity of the cover rubber produced by the real-time PVG cover rubber extrusion production line under the current optimal process variable correction coefficient.

[0172] (4) Extract the defect ratio of the good cover adhesive from the good cover adhesive benchmark. The good cover adhesive benchmark is constructed by collecting at least 100 good cover adhesive segments in the offline stage. Its benchmark defect segmentation mask is a comprehensive binary image. The system counts the total number of pixels with a mask value of 1 in the good cover adhesive benchmark defect segmentation mask in the same way. Since the good cover adhesive has almost no defects, but considering the imaging noise and small errors in image processing, there may still be a very small number of isolated pixels marked as 1 in the good cover adhesive benchmark mask. Assuming that after statistics, the total number of pixels with a value of 1 in the good cover adhesive benchmark defect segmentation mask is 50, then the defect ratio of the good cover adhesive benchmark is 50÷262144×100%≈0.019%, which represents the background defect ratio of the cover adhesive under the ideal good condition. It can be understood as the unavoidable small noise level allowed by the system.

[0173] (5) Calculate the difference between the total percentage of defects in the real-time collected cover adhesive and the percentage of defects in the good cover adhesive benchmark. This represents the percentage of the additional defect area of ​​the workpiece produced on the current production line compared to the perfect good cover adhesive benchmark.

[0174] (6) Set a first preset threshold. In this embodiment, according to product design specifications and quality control requirements, the first preset threshold is set to 0.05%.

[0175] (7) Compare the difference calculated in (5) with the first threshold. When the difference is less than 0.05%, it is considered that the current process correction has made the product quality close to the good product level. The pre-adjustment stage of process correction has been completed, and the system outputs a process pre-correction ready signal.

[0176] S3: After receiving the process pre-correction ready signal, fine-tune the multi-process decoupling control model, collect full-dimensional defect statistics of PVG conveyor belt cover rubber segment by segment, and after continuously statistically analyzing N segments of cover rubber, calculate the defect month-on-month growth rate, comprehensive process drift rate and its coupling ratio based on the full-dimensional defect statistics.

[0177] The specific steps of S3 include:

[0178] S3.1: Receive the process pre-correction ready signal and activate the fine-tuning mode controller;

[0179] S3.2: The fine-tuning mode controller loads the current optimal process variable correction coefficient as the initial value for fine-tuning, and loads the defect basis matrix and process contribution coefficient matrix as the objects to be fine-tuned. It uses a Bayesian optimization-based hyperparameter fine-tuning algorithm to update the defect basis matrix and process contribution coefficient matrix online, and obtains the updated defect basis matrix and process contribution coefficient matrix, which is denoted as the fine-tuned multi-process decoupling control model.

[0180] It should be noted that after the system outputs the process pre-correction ready signal, it signifies that the coarse-tuning stage based on digital twin simulation and offline optimization has been completed, and the current process parameters of the production line have brought the product quality initially close to the good benchmark for cover adhesive. However, there are still unavoidable differences between the offline simulation model and the actual production line, and the production line itself will also slowly drift over time. Therefore, the system needs to enter an online fine-tuning stage, using real-time collected actual defect data to finely adjust the multi-process decoupling control model.

[0181] Furthermore, the fine-tuning mode controller reads three key data objects from the system memory. The first is the current optimal process variable correction coefficient, which the fine-tuning mode controller loads as the initial value for fine-tuning. Assuming the specific values ​​of the current optimal process variable correction coefficient vector are: extruder screw speed correction coefficient 1.0025, rubber extrusion feed speed correction coefficient 0.9980, die extrusion gap correction coefficient 0.9972, and temperature-controlled cooling water flow rate correction coefficient 1.0015; the second and third data objects are the defect basis matrix and the process contribution coefficient matrix, respectively; the fine-tuning mode controller uses the defect basis matrix and the process contribution coefficient matrix as the objects to be fine-tuned.

[0182] Furthermore, after loading is complete, the fine-tuning mode controller initiates a Bayesian optimization-based hyperparameter fine-tuning algorithm. In this embodiment, the objects to be fine-tuned are all elements in the defect basis matrix and the process contribution coefficient matrix. These elements are the hyperparameters that Bayesian optimization seeks to search for. However, these two matrices contain a total of 18 elements, and directly performing Bayesian optimization on 18 dimensions results in an excessively large search space and very low computational efficiency. Therefore, the system optimizes a set of hyperparameters for the overall structure of the control matrix. This set of hyperparameters includes a sparsity adjustment factor, a matrix update learning rate, and a reconstruction error tolerance. The sparsity adjustment factor is a real number between 0 and 1. The larger the value, the more zero elements there are in the process contribution coefficient matrix, and the stronger the interpretability of the model; the matrix update learning rate is a real number between 0.001 and 0.1, which controls the step size when updating matrix elements each time based on the hyperparameter combination recommended by Bayesian optimization; the reconstruction error tolerance is a positive real number that defines the maximum deviation between the product of the defect basis matrix and the process contribution coefficient matrix and the original defect-process coupling mapping coefficient matrix allowed during fine-tuning; in this embodiment, the sparsity adjustment factor is set to 0.05~0.95, the matrix update learning rate is 0.001~0.1, and the reconstruction error tolerance is 0.01~0.5.

[0183] Furthermore, the Bayesian optimization process includes:

[0184] (1) Initialize a Gaussian process surrogate model. A Gaussian process is a statistical model used to describe the distribution of functions. It does not require knowledge of the mathematical form of the objective function. It only needs to continuously sample to infer the possible values ​​of the objective function under various hyperparameter combinations. In the initial stage, the system has no information about the relationship between hyperparameters and fine-tuning effects. Therefore, it is necessary to randomly generate several sets of initial sampling points. The number of initial sampling points is usually set to about ten times the hyperparameter dimension. In this embodiment, the hyperparameter dimension is 3, so the number of initial sampling points is set to 30. The system randomly generates 30 sets of hyperparameter combinations. Each set of combinations includes a sparsity adjustment factor value, a matrix update learning rate value, and a reconstruction error tolerance value.

[0185] (2) For each set of randomly generated hyperparameter combinations, the system performs the following fine-tuning operations: First, the system starts with the currently loaded defect basis matrix and process contribution coefficient matrix; then, according to the sparsity adjustment factor, the process contribution coefficient matrix is ​​sparsified. Specifically, for each element in the process contribution coefficient matrix, if the absolute value of the element is less than the maximum value of the absolute value of all elements in the current matrix multiplied by the sparsity adjustment factor, the element is set to zero. Assuming that the maximum value of the absolute value of the elements in the current process contribution coefficient matrix is ​​0.92, if the sparsity adjustment factor in any set of hyperparameters is set to 0.3, the threshold is 0.3 × 0.92 = 0.276, and all elements with an absolute value less than 0.276 will be set to zero.

[0186] (3) The fine-tuning mode controller collects defect data and corresponding process parameters from continuously produced workpieces in real time, and reconstructs a defect-process coupling mapping coefficient matrix at the current moment, denoted as the real-time observation matrix; calculates the difference between the real-time observation matrix and the reconstruction matrix obtained by multiplying the current defect basis matrix by the current process contribution coefficient matrix. The calculation process of multiplying the current defect basis matrix by the current process contribution coefficient matrix is ​​as follows: multiply the first row of the defect basis matrix by the first column of the process contribution coefficient matrix and sum them, and so on, to obtain a reconstruction matrix of two rows and four columns; the system calculates the difference between the real-time observation matrix and the reconstruction matrix at each position to obtain an error matrix; then multiplies each element in the error matrix by the matrix update learning rate to obtain the adjustment amount of each matrix element; updates the corresponding elements in the defect basis matrix and the process contribution coefficient matrix synchronously according to the adjustment amount, so that the reconstruction matrix moves closer to the real-time observation matrix; after the update is completed, the system again performs sparsification processing on the process contribution coefficient matrix according to the sparsity adjustment factor, and sets the elements less than the threshold in (2) to zero;

[0187] (4) After the update is completed, the system calculates the reconstruction error after the current fine-tuning. The reconstruction error is defined as the square of the F norm between the real-time observation matrix and the reconstruction matrix obtained by multiplying the current defect basis matrix by the current process contribution coefficient matrix, that is, the sum of the squares of the differences of all corresponding elements. If the reconstruction error is greater than the tolerance of 0.2, it means that the fine-tuning effect is not up to standard.

[0188] (5) Repeat (2)-(4) for each set of randomly initialized hyperparameter combinations to obtain the reconstruction error value corresponding to each combination. The smaller the reconstruction error, the more accurately the model obtained by fine-tuning under the control of the set of hyperparameters can reflect the current production line status. After 30 sets of initial sampling, the system obtained 30 hyperparameter combinations and their corresponding reconstruction error values.

[0189] (6) The system uses 30 data points that have been collected. Each data point is a three-dimensional hyperparameter vector and its corresponding reconstruction error value to train the Gaussian process surrogate model. The Gaussian process surrogate model will output a prediction function. For any set of hyperparameter combinations that have not yet been evaluated, the function can predict its possible reconstruction error value and also give the degree of uncertainty of the prediction. Based on this prediction function, the system constructs an acquisition function. The commonly used acquisition function is the expected improvement amount. The expected improvement amount takes into account the predicted reconstruction error value and the prediction uncertainty. It tends to select hyperparameter combinations with small prediction error or large uncertainty for actual evaluation. The system finds the next set of hyperparameter combinations that maximizes the expected improvement amount in the current search space through numerical optimization methods.

[0190] (7) Take out this set of newly recommended hyperparameter combinations, apply sparsity, update matrix elements according to the real-time observation matrix and matrix update learning rate, calculate reconstruction error and compare with tolerance to obtain the true reconstruction error value corresponding to this set of hyperparameters, then add this new data point to the existing dataset, retrain the Gaussian process surrogate model, and find the next set of hyperparameter combinations that maximize the expected improvement; the system repeats this process, generating a new set of hyperparameters and evaluating them in each iteration; the system sets the maximum number of iterations for Bayesian optimization to 50, that is, in addition to the initial 30 samplings, 50 more iterations of optimization are performed; after completing all iterations, the system selects the set of hyperparameters with the smallest reconstruction error from all 80 sets of evaluation results, and this set of hyperparameters is the optimal hyperparameter combination found by Bayesian optimization.

[0191] S3.3: Collect full-dimensional defect statistics of PVG conveyor belt cover rubber segment by segment; the full-dimensional defect statistics include defect type label sequence, defect spatial scale parameter vector, defect depth distribution histogram, defect location spatial coordinate set, and standardized process parameter vector output at the corresponding time for each segment of cover rubber.

[0192] Furthermore, the process of collecting the defect type label sequence is as follows: First, connected component analysis is performed on the PVG cover adhesive defect segmentation mask. The connected component adopts the eight-connectivity judgment rule. If any pixel in any of the eight adjacent points (top, bottom, left, right, and diagonal) is the same as 1, it is classified into the same independent defect block. Assuming that the current segment of the cover adhesive mask is divided into three independent defect regions after parsing, they are labeled as region A, region B, and region C respectively. For each independent defect region, the multi-class probability vector output by the decoder classification head is retrieved. The classification head outputs a total of 5 types of typical PVG cover adhesive defects, namely surface scratches, adhesive pits, sub-interface delamination cracks, internal pores, and surface adhesive adhesion. Taking the probability distribution of each region as an example: Region A scratch 7 The system assigns labels to areas A, B, C, and D, based on the probability values ​​of 8%, 12%, delamination cracks, 3%, pores, and adhesive residue. Area B is defined as a dent (15%), 11%, 73%, 4%, pores, and 3%. Area C is defined as a sub-interface delamination crack (8%), 11%, 73%, 5%, and 3%. The system organizes the labels according to the spatial arrangement of the defect blocks in the image from top to bottom and left to right, generating an ordered defect type label sequence. For example, the output sequence is: [scratches, dents, interface delamination cracks]. If the adhesive mask covering this segment has no defect pixels (i.e., all pixel values ​​are 0), the defect type label sequence is empty.

[0193] Furthermore, the acquisition process of the defect spatial scale parameter vector is as follows: the system counts the total number of pixels contained in the defect region, denoted as the defect region pixel area; the diameter of the equivalent circle with the same area as the defect region pixel area is calculated, where the formula for the area of ​​the equivalent circle is... R is the radius. The value is pi; in addition to the equivalent diameter, the system also calculates the maximum length and maximum width of the defect, which are determined based on the minimum bounding rectangle of the defect; the system arranges the scale parameter vectors of all defect regions in the same order as the defect type label sequence to form the defect spatial scale parameter vector.

[0194] Furthermore, regarding the defect depth distribution histogram, the system's polarization imaging optical path can provide subsurface stress distribution information. Combining the light intensity distribution of bright-field and dark-field images, the defect detection decoder includes a depth estimation branch. This branch outputs the estimated depth value corresponding to each defect pixel location, in μm. For each defect region, the system extracts the depth estimates of all pixels within the region. Assuming region A contains 1200 pixels, each pixel has a depth estimate, and these values ​​are assumed to be between 0 and 200 μm. The system divides the depth range into multiple consecutive intervals, each called a depth box. In this embodiment, the number of depth boxes is set to 10, and the width of each box is 20 μm, i.e., the first box is 0 to 20 μm, the second box is 20 μm, and the third box is 20 μm. The depth range is 20-40 μm, and so on up to 180-200 μm. The system iterates through 1200 pixels in region A. For each pixel's depth estimate, it determines which depth bin the estimate falls into and increments the corresponding depth bin's count variable by 1. After the iteration, it obtains the pixel count values ​​corresponding to each of the 10 depth bins. These 10 pixel count values ​​correspond to 10 depth intervals. Then, it divides the pixel count value of each depth bin by the total number of pixels in the region, 1200, to obtain the pixel proportion of each depth interval, i.e., the frequency. The 10 frequency values ​​are then arranged in order of depth bin from shallowest to deepest. The resulting one-dimensional ordered frequency array is the defect depth distribution histogram of the current defect region. The system generates a depth distribution histogram for each defect region.

[0195] Furthermore, for the spatial coordinate set of defect locations, the system needs to record the precise position of each pixel within each defect region in the image to analyze the spatial clustering and distribution patterns of defects. For an image with a resolution of 512×512 pixels, the system uses a pixel coordinate system with the origin located at the top left corner of the image, and row coordinates increasing downwards and column coordinates increasing to the right. For the 1200 pixels of region A, the system records the row and column coordinates of each pixel; for example, the first pixel is located at row 150, column 80, the second pixel is located at row 150, column 81, and so on. Since directly storing the coordinates of each pixel would occupy a large amount of storage space, the system uses... Run-length encoding is used for compressed storage. The basic principle of run-length encoding is: for each row of the image, the starting column and continuous length of consecutively occurring identical mask values ​​are recorded. For example, if in row 150, the mask values ​​of 71 consecutive pixels from column 80 to column 150 are all 1, then the system records a record with row number 150, starting column 80, and continuous length 71. In this way, the pixel coordinates of the entire region A can be represented by dozens of such records. The system collects all run-length encoded records of each defect region to form the spatial coordinate set of the defect location of that region, and then merges the coordinates of all regions to form the spatial coordinate set of the defect location of the entire cover adhesive.

[0196] Furthermore, for the standardized process parameter vector output at the corresponding time, during the synchronous period of imaging detection of the cover rubber of each PVG conveyor belt, the system synchronously collects real-time process timing parameters, including extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate. After these raw collected values ​​are processed by sliding window mean filtering and timestamp alignment interpolation, a standardized process parameter vector is generated that corresponds one-to-one with the imaging time of each cover rubber segment. Taking the real-time working condition of this embodiment as an example, the specific values ​​of the standardized process parameter vector matched in the current detection period are: extruder screw speed 2510 rad / min, rubber extrusion feed speed 0.1498 mm / rad, die extrusion gap 1.197 mm, and temperature-controlled cooling water flow rate 5.009 L / min.

[0197] S3.4: Store the full-dimensional defect statistics of each section of the cover adhesive into a circular cache queue in the order of collection time;

[0198] S3.5: When the number of cover adhesive segments stored in the circular cache queue reaches a preset number N, extract the full-dimensional defect statistics of the most recent N consecutive cover adhesive segments from the circular cache queue, and then divide the N cover adhesive segments into the first N / 2 groups and the last N / 2 groups according to the collection time order. Calculate the total number of defects in the first N / 2 groups and the last N / 2 groups respectively. Divide the total number of defects in the last N / 2 groups by the total number of defects in the first N / 2 groups and subtract 1 to obtain the month-on-month defect growth rate. In this embodiment, the preset segmented statistical sample size N is an even number, with a value range of 10≤N≤40. In this embodiment, N=20 is preferred.

[0199] S3.6: Based on the extracted full-dimensional defect statistics of N segments of continuous cover rubber, extract the actual values ​​of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate from the standardized process parameter vector corresponding to each segment of cover rubber. These constitute four actual value sequences. Linear fitting is performed on each actual value sequence to obtain the slope of the fitted line. The absolute value of the slope of the fitted line is then divided by the arithmetic mean of the actual value sequences to obtain the drift rate of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate, respectively. These drift rates are then weighted according to the preset process parameter weights to obtain the comprehensive process drift rate.

[0200] In this embodiment, the process parameter weights are set using a min-max normalization method. Specifically, the mean values ​​of the extruder screw speed drift rate, the rubber extrusion feed speed drift rate, the die extrusion gap drift rate, and the temperature-controlled cooling water flow rate drift rate are calculated, along with the sum of the four mean values. The ratio obtained by dividing the mean value of each parameter by the sum of the four mean values ​​is the process parameter weight.

[0201] S3.7: Calculate the ratio of the year-on-year growth rate of defects to the overall process drift rate, denoted as the coupling ratio.

[0202] S4: If the coupling ratio is less than the preset second threshold, the process variable correction coefficient and segmented correction strategy are used to maintain the continuous extrusion production of the cover rubber; if the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and segmented correction strategy are updated respectively based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics.

[0203] If the coupling ratio is less than the preset second threshold, the current production line is determined to be in a stable process state. The current optimal process variable correction coefficient remains unchanged, and the current segmented correction strategy remains unchanged. The segmented correction strategy is divided into a plasticizing section, a die extrusion molding section, and a cooling and shaping section according to the extrusion process of the cover rubber. Each process stage corresponds to a set of sub-vectors of process variable correction coefficients. The segmented correction strategy is set to be a unit vector for each process stage sub-vector during initialization.

[0204] Furthermore, the coupling ratio is the ratio of the defect growth rate to the overall process drift rate. This ratio reflects the relative relationship between defect growth and process drift. When the coupling ratio is small, it indicates that the defect growth is mainly caused by process drift, and the process drift itself is still within a controllable range. The system can cope with this using the existing correction coefficients. When the coupling ratio is large, it indicates that the defect growth rate far exceeds the process drift rate, and there may be other abnormal factors, requiring more aggressive adaptive correction. Based on actual production experience, when the ratio of the defect growth rate to the process drift rate is close to or exceeds 2, it usually means that a non-linear defect deterioration trend has occurred. Therefore, the second threshold is set to 2.0.

[0205] Furthermore, the three stages—plasticizing, die extrusion, and cooling / setting—correspond to different process stages in the entire PVG cover rubber extrusion process, with significant differences in material state, process objectives, and control parameters at each stage: The plasticizing stage is the screw melt-mixing stage of the extruder, where the core function is to heat and shear the PVG granules, fully melt and mix them. This stage has high screw shear strength and large melt flow, with lower requirements for controlling minor defects on the cover surface, primarily avoiding large bubbles and clumps of impurities caused by uneven mixing; the die extrusion stage involves the melt being extruded through the die lip, initially... The transition process for forming the cover rubber preform balances rubber material conveying efficiency and surface forming quality, and is a high-incidence area for surface scratches and uneven thickness defects. The cooling and shaping section is the stage where the extruded rubber preform is cooled and solidified by water-cooled rollers, which directly determines the final thickness accuracy, surface smoothness and internal interlayer bonding state of the PVG cover rubber. Small cooling fluctuations can easily induce hidden defects such as sub-interface delamination and internal micropores. In the entire PVG cover rubber continuous extrusion production line, the rubber material flows through the three process zones of plasticizing, extruding and cooling in sequence. Therefore, the segmented correction sub-vectors are divided based on the complete extrusion cycle of a single section of cover rubber. This embodiment uses a percentage-based processing time division rule: the first 30% of the complete extrusion cycle is the rubber plasticizing section, the middle 40% is the die extrusion molding section, and the last 30% is the cooling and setting section. Taking a total extrusion time of 60s for a single-section cover rubber as an example, the segment intervals are as follows: the rubber plasticizing section is 0s~18s. During this stage, the extrusion screw speed is high and the die lip clearance is set relatively large. The process correction coefficient prioritizes controlling the uniformity of rubber melt melting and suppressing the generation of large-scale mixing defects. It does not impose strict constraints on minor surface defects. The die extrusion molding section is 19s~42s. The screw speed and die lip clearance are gradually reduced, while simultaneously taking into account the melt delivery efficiency and the surface forming effect of the rubber blank. The process correction coefficient needs to balance the production capacity and two types of indicators: surface scratches and thickness defects. The cooling and setting section is 43s~60s. The traction speed is slowed down and the cooling water flow rate is finely adjusted. The rubber completes the curing and setting. Small water flow and temperature fluctuations can induce layering micropores. The process correction coefficient in this section is highly sensitive to minor process drift and prioritizes controlling sub-interface hidden defects.

[0206] If the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and the segmented correction strategy are updated based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics, including:

[0207] S4.1: If the coupling ratio is greater than or equal to a preset second threshold, the current production line is determined to be in a process drift exceeding the limit state, triggering adaptive correction and extracting the latest full-dimensional defect statistics of the M-segment cover adhesive from the circular cache queue; where M is a preset positive integer less than or equal to N;

[0208] Furthermore, when the system determines that the coupling ratio is greater than or equal to the second threshold of 2.0, it triggers an adaptive correction process. The system first extracts the latest full-dimensional defect statistics of the M-segment cover adhesive from the circular cache queue, where M is a preset positive integer. In this embodiment, N is set to 20 and M is set to 10.

[0209] S4.2: Based on the extracted full-dimensional defect statistics of the M-segment cover adhesive, the local weighted regression algorithm is used to solve for the local fine-tuning correction amount; the local fine-tuning correction amount is an incremental vector with the same dimension as the correction coefficient of the current optimal process variable;

[0210] It should be explained that the core idea of ​​local weighted regression is that, for the target point to be predicted, only the training data points near it are used to fit a regression model, and the closer the data points are, the higher the weight is assigned. In this embodiment, the target point to be predicted is the current latest process parameter state, that is, the standardized process parameter vector of the 10th segment cover adhesive. Therefore, the system uses the standardized process parameter vector of the 10th segment cover adhesive as the regression center point.

[0211] Furthermore, the specific steps in S4.2 include:

[0212] (1) The system extracts the full-dimensional defect statistics of these 10 segments of cover adhesive from the circular cache queue in chronological order. Each record contains the defect type label sequence, defect spatial scale parameter vector, defect depth distribution histogram, defect location spatial coordinate set, and standardized process parameter vector at the corresponding time. The system needs to extract the independent and dependent variables for regression analysis from these data. The independent variable is the actual value of the process parameter, and the dependent variable is a comprehensive indicator representing the severity of the defect.

[0213] (2) For each section of cover adhesive, the system also needs to calculate a comprehensive score representing the severity of the defect, which serves as the dependent variable in the regression analysis. The calculation method for this comprehensive score is as follows: First, extract the defect spatial scale parameter vector from the full-dimensional defect statistics of the cover adhesive section. Assuming that three defect areas were detected in the first section of cover adhesive, with equivalent diameters of 0.391 mm, 0.597 mm, and 0.211 mm, the system calculates the sum of the squares of these three equivalent diameters. That is, the square of 0.391 is approximately 0.153, the square of 0.597 is approximately 0.356, and the square of 0.211 is approximately 0.045. The sum of the three is 0.554, which reflects the degree of defect severity. The system first assesses the overall size and severity of defects. Then, it extracts the depth hazard coefficient from the defect depth distribution histogram: the depth value at the midpoint of each depth interval is multiplied by the corresponding interval's pixel frequency, and the sum is obtained to get the average defect depth of a single segment of cover adhesive. Assuming the average defect depth of the first segment of cover adhesive is calculated to be 45μm, the system normalizes the average defect depth by dividing it by the maximum possible depth of 200μm, resulting in 0.225, which is the normalized depth hazard coefficient. Next, the system multiplies the overall size severity of defects (0.554) with the normalized depth hazard coefficient (0.225) to obtain the comprehensive defect score of that segment of cover adhesive (0.125). The system calculates the comprehensive defect score for each of the 10 segments of cover adhesive.

[0214] (3) Calculate the mean and standard deviation of the 10 cover adhesives in each process parameter dimension. Subtract the mean of the corresponding dimension from the original process parameter value of each cover adhesive and then divide by the standard deviation to obtain the standardized independent variable value, i.e. the standardized process parameter vector.

[0215] (4) For the j-th segment of the cover adhesive, the system calculates the Euclidean distance between its standardized process parameter vector and the standardized process parameter vector of the regression center point, i.e., the 10-th segment of the cover adhesive. The standardized process parameter vector of the regression center point itself is zero in each dimension because its original value minus the mean and then divided by the standard deviation results in zero. Therefore, the distance of the j-th segment of the cover adhesive is equal to the square root of the sum of the squares of each dimension of its standardized process parameter vector.

[0216] (5) After obtaining the distance, the system calculates a weight for each segment of the cover adhesive. The weight is a function of the distance. In this embodiment, a Gaussian kernel function is used, and the calculation formula is as follows: ,in, Let b represent the exponential function, and b represent the bandwidth. This represents the distance of the j-th segment of the adhesive covering. This represents the weight of the j-th segment of the cover adhesive. In this embodiment, b is set to 0.5, while the weight of data points that are more than twice the bandwidth away rapidly decays to near 0. In this embodiment, the cover adhesive segments that are closer to the regression center point have higher weights, while the cover adhesive segments that are farther away have almost zero weights.

[0217] (6) The system uses these weights to perform weighted linear regression, specifically: it is assumed that there is a linear relationship between the defect comprehensive score and the four process parameters in the local neighborhood of the regression center point, that is, the defect comprehensive score is equal to the intercept term plus the extruder screw speed coefficient multiplied by the extruder screw speed, plus the rubber extrusion feed speed coefficient multiplied by the rubber extrusion feed speed, plus the die extrusion gap coefficient multiplied by the die extrusion gap, plus the temperature control cooling water flow rate coefficient multiplied by the temperature control cooling water flow rate, plus an error term. In this embodiment, the error term is set to 0.015 based on the statistics of historical defect samples of the production line. The system needs to solve for these five coefficients, including an intercept and four slopes, that is, the corresponding four process parameter coefficients. The goal of the solution is to minimize the weighted residual sum of squares, that is, the square of the residual of each data point multiplied by the corresponding process parameter weight in S3.6, and then summed. The residual is the actual value of the defect comprehensive score minus the model prediction value.

[0218] (7) Let each row of the independent variable matrix X correspond to a segment of cover rubber, the first column be all 1s representing the intercept term, the second column be the extruder screw speed of the segment of cover rubber, the third column be the rubber extrusion feed rate, the fourth column be the die extrusion gap, and the fifth column be the temperature-controlled cooling water flow rate; each row of the dependent variable vector Y is the comprehensive defect score of the segment of cover rubber; the weight matrix Given a diagonal matrix, where the elements on the diagonal represent the weights of each covering adhesive segment, the weighted least squares solution is: Where T represents transpose. For coefficient vectors;

[0219] (8) Substitute specific values ​​for calculation. Assume that after matrix operation, the coefficient vector obtained is an intercept term of 0.185, the extruder screw speed coefficient is -0.00012, the rubber extrusion feed speed coefficient is 2.35, the die extrusion gap coefficient is 1.82, and the temperature control cooling water flow rate coefficient is 0.43. The meaning is: In the current local process area, for every 1 rad / min increase in extruder screw speed, the average defect comprehensive score decreases by 0.00012 points, that is, appropriately increasing the extruder screw speed is beneficial to reducing defects; for every 0.01 mm / rad increase in rubber extrusion feed speed, the average defect comprehensive score increases by 0.0235 points, that is, increasing the rubber extrusion feed speed will lead to the deterioration of defects; for every 0.01 mm increase in die extrusion gap, the average defect comprehensive score increases by 0.0182 points; for every 0.1 L / min increase in temperature control cooling water flow rate, the average defect comprehensive score increases by 0.043 points.

[0220] (9) Solve for the local fine-tuning correction amount. The local fine-tuning correction amount is an incremental vector with the same dimension as the current optimal process variable correction coefficient. Its function is to add a small adjustment amount based on the current correction coefficient so that the adjusted defect comprehensive score can be reduced. The system sets the target reduction of the defect comprehensive score to 1.5 times the standard deviation of the current defect comprehensive score. The current defect comprehensive score of the 10th segment cover adhesive is 0.182. The system hopes to reduce it to about 0.15. Based on the regression coefficient, the system can deduce the required process parameter adjustment amount. Due to the extruder screw speed system The score is negative, so increasing the extruder screw speed helps reduce the overall defect score. Therefore, the system sets the extruder screw speed adjustment to 15 rad / min. The extrusion feed rate coefficient is positive, so the extrusion feed rate needs to be reduced to lower the overall defect score. Therefore, the extrusion feed rate adjustment is set to -0.003 mm / rad. The die extrusion clearance coefficient is positive, so the die extrusion clearance needs to be reduced. The adjustment is set to -0.005 mm. The temperature-controlled cooling water flow rate coefficient is positive, so the temperature-controlled cooling water flow rate needs to be reduced. The adjustment is set to -0.02 L / min.

[0221] (10) Divide the absolute adjustment amount by the corresponding value of the current standardized process parameter vector to obtain the corresponding incremental correction coefficient.

[0222] S4.3: Based on the full-dimensional defect statistics of all N segments of cover adhesive in the circular buffer queue, the time-series prediction model is used to predict the defect evolution trend of P segments of cover adhesive, and the model prediction control algorithm is used to perform rolling optimization to obtain the global process compensation correction amount sequence; the global process compensation correction amount sequence contains P correction amount vectors, and each correction amount vector corresponds to the global compensation value of a segment of cover adhesive.

[0223] Furthermore, the system first extracts the full-dimensional defect statistics of all N segments of cover adhesive from the circular cache queue. In this embodiment, N is set to 20. These 20 segments of cover adhesive are arranged in order of collection time from earliest to latest, and are denoted as workpiece numbers from -19 to 0, where 0 represents the most recently produced segment of cover adhesive. The full-dimensional defect statistics of each segment of cover adhesive includes a comprehensive defect score. The system extracts the comprehensive defect scores of the 20 segments of cover adhesive in chronological order to form a time series of length 20.

[0224] Furthermore, the time series prediction model adopts a structure based on a long short-term memory network. The core of the long short-term memory network is a recurrent neural network unit containing three gating structures: a forget gate, an input gate, and an output gate. The forget gate determines which information to discard from the memory unit at the previous time step, the input gate determines the extent to which new information at the current time step is added to the memory unit, and the output gate determines which information to output from the memory unit to the hidden state at the current time step. This structure enables the network to learn long-term dependencies in the time series without being troubled by the gradient vanishing problem.

[0225] Furthermore, the specific structural configuration of the Long Short-Term Memory (LSTM) network model is as follows: The input layer dimension is set to 1 because the input to the network at each time step is a scalar value, namely the comprehensive defect score of the covering adhesive segment. The network contains two LSM layers, each with 50 hidden units. The first LSM layer receives each time step of the input sequence and outputs the complete hidden state sequence. The second LSM layer receives the output of the first layer as input, also with 50 hidden units, and outputs only the hidden state of the last time step. After the two LSM layers, a dropout layer is connected with a dropout rate of 0.2 to prevent overfitting. After the dropout layer, a fully connected layer is connected. The fully connected layer has an input dimension of 50 and an output dimension of 1, used to map the hidden states to the predicted comprehensive defect score. Finally, the output of the fully connected layer serves as the prediction result for the entire network.

[0226] Furthermore, the time-series prediction model training process includes: extracting a large number of continuous production PVG conveyor belt cover rubber defect comprehensive score sequences from the historical production database. Each sequence consists of 30 consecutive cover rubber segments. For each 30-segment sequence, the system uses the comprehensive defect scores of the first 20 cover rubber segments as input features and the scores of the last 10 cover rubber segments as prediction targets, forming a set of training samples. The system collected a total of 100,000 such training samples. During training, the mean squared error is used as the loss function, which is the average of the squares of the difference between the predicted score and the actual score. The optimizer uses the adaptive moment estimation algorithm with a learning rate set of 0.001. The system divides the training data into training and validation sets at a ratio of 80% and 20%, respectively. Parameters are updated on the training set, and the loss value is evaluated on the validation set after each round. The training is conducted for a total of 200 rounds. Training is terminated early when the loss value on the validation set no longer decreases for 10 consecutive rounds.

[0227] Furthermore, during online prediction, the system loads a pre-trained Long Short-Term Memory (LSTM) network and sequentially inputs the defect comprehensive score sequences of all 20 segments of overlay adhesive in the current recurrent cache queue into the LSM network in chronological order. For example, the system extracts 20 historical composite scores of cover adhesive defects, sorted by production sequence, from the circular cache queue, forming a one-dimensional time-series input sequence of length 20. In this embodiment, sample values ​​from the first, last, and middle segments of the sequence are used for calculation demonstration: the composite defect score at the first time step is 0.085, the score at the second time step is 0.088, and the score at the 20th time step is 0.170. The time-series iterative calculation process within the Long Short-Term Memory (LSTM) network is as follows: At the first time step, the time-series input value 0.085 is sent into the LSM network. The initialized hidden state and memory unit state are both all-zero vectors, each with a length of 50, and all elements are initialized to 0. The network sequentially updates its state through the input gate, forget gate, and output gate: the input gate controls the weight of storing the current 0.085 composite defect score information into the memory unit, the forget gate controls the proportion of discarding invalid historical memories, and the output gate combines the memory unit to generate the current hidden state. After completing the first time step calculation, the network refreshes the hidden state and memory unit state and caches them. The initial parameters for the next time step are used; then, the second time step is entered, and the second defect comprehensive score of 0.088 is read as input. The hidden state and memory unit state output from the previous time step are used as the initial state, and the three-gated state update logic is repeated; this time-series iterative process continues to loop, sequentially completing the forward operation of the comprehensive score of the covering adhesive defect for segments 3 to 19, until the comprehensive score of the input defect for the 20th time step, 0.170, is processed; after all 20 segments of historical time-series data have been processed, the second LSTM layer only outputs the hidden feature vector of the last time step with a length of 50; this vector is input to the dropout layer, which randomly masks 20% of the neuron outputs and divides the output values ​​of the remaining neurons by 0.8 for expectation compensation to avoid model overfitting; the 50-dimensional feature vector after being processed by the dropout layer is input to the fully connected layer, which is configured with a 50×1 weight matrix and a single-dimensional bias term to map the high-dimensional features into a single scalar. This scalar is the model's predicted comprehensive score of the defect for the next segment, i.e., segment 21, where no covering adhesive has been produced. In this embodiment, the output value of the fully connected layer is 0.178, which means that the system predicts that the severity of defects in the 21st stage of cover adhesive produced in the subsequent production will continue to increase.

[0228] In this embodiment, the controlled object is the entire extrusion process system of PVG conveyor belt cover rubber. The control input is the extrusion process variable correction coefficient, and the system output is the comprehensive defect score of a single section of cover rubber. The control objective is to dynamically adjust the process correction coefficient to reduce the comprehensive defect score to the level corresponding to the PVG cover rubber good product benchmark score. The comprehensive defect score of the good product benchmark is calculated from the standard defect segmentation mask in the good product benchmark database. In this embodiment, the comprehensive defect score of the good product benchmark is set to 0.02. The system requires a global process compensation correction amount sequence of length P. Each element in the sequence is an incremental vector with the same dimension as the process correction coefficient, corresponding to the process compensation adjustment amount of the subsequent P sections of cover rubber. In this embodiment, P=5. The system relies on a locally weighted regression model to establish a correlation prediction model between extrusion process correction coefficients and the comprehensive defect score of the cover rubber. Based on this, a model prediction control optimization objective function is built, which consists of three terms: The first part is the tracking error term: the sum of squares of the differences between the predicted comprehensive defect score of the subsequent 5 stages of cover rubber and the comprehensive defect score of the good product benchmark, with the core objective of reducing the gap between the defect level and the pass standard; The second part is the control increment penalty term: the sum of squares of the changes in the next 5 sets of process correction coefficients, multiplied by a penalty weight of 0.1, to avoid sudden changes in the magnitude of a single process adjustment and prevent drastic fluctuations in extrusion conditions; The third part is the terminal penalty term: the long-term defect state deviation after the 5th stage of cover rubber, multiplied by a terminal penalty weight of 0.5, to ensure the stability of extrusion quality in the long term. Simultaneously, process parameter solution constraints are set to ensure that the output correction coefficients are compatible with the safe operating range of the extrusion equipment: for four types of extrusion processes, including screw speed, feed rate, die clearance, and cooling water flow rate, the corresponding correction coefficient values ​​are limited to 0.8~1.2. Exceeding the range will cause the extruder, die lip, and cooling equipment to operate beyond their limits, posing a risk of equipment damage and batch defects; between two adjacent cover rubber sections, the absolute value of the single-step change of the correction coefficient of a single process parameter must not exceed 0.05, to prevent instantaneous large fluctuations in the melt pressure and cooling flow rate of the rubber compound, and to avoid the concentrated generation of batch bubbles and delamination defects.

[0229] Furthermore, given that the system currently has a defect comprehensive score of 0.170 for the current cover adhesive, and the current optimal process variable correction coefficients are: extruder screw speed 1.0025, adhesive extrusion feed rate 0.9980, die extrusion gap 0.9972, and temperature-controlled cooling water flow rate 1.0015, the system needs to calculate the changes in these correction coefficients at five different times, denoted as... arrive The correction coefficients at the five time points are expressed as the current correction coefficient plus the cumulative change.

[0230] Furthermore, the system employs a quadratic programming algorithm to solve this optimization problem. Because the objective function is quadratic and the constraints are linear, the standard form of quadratic programming is: The system arranges the variables to be solved into a 20-dimensional vector, since 5 time points multiplied by 4 process parameters equal 20 variables; the system calls a mature quadratic programming solver, setting the maximum number of iterations to 500 and the optimality tolerance to 10. -6 The solver, under the premise of satisfying all constraints, finds the combination of variables that minimizes the objective function. The resulting sequence is the global process compensation correction sequence, containing five correction vectors. Each vector corresponds to the global compensation value for a segment of the cover adhesive. The first correction vector is directly added to the current optimal process variable correction coefficient to obtain the updated correction coefficient used for the first segment of the cover adhesive. The second correction vector is further superimposed on this, and so on. It is worth noting that the system actually only implements the first correction vector. After the next segment of the cover adhesive is produced, the system will re-predict and optimize based on new measured data. Therefore, this sequence is dynamically updated.

[0231] S4.4: Add the obtained local fine-tuning correction amount to the current optimal process variable correction coefficient element by element to obtain the updated process variable correction coefficient;

[0232] S4.5: The obtained global process compensation correction sequence is averaged segment by segment according to the process stage division boundary of the defined segment correction strategy: the arithmetic mean of the correction vectors corresponding to the cover rubber covered by the rubber plasticizing section, the die extrusion molding section and the cooling and shaping section are calculated respectively, and the sub-vectors of the corresponding process stages in the current segment correction strategy are replaced with the three arithmetic means respectively to obtain the updated segment correction strategy.

[0233] After calculating the local fine-tuning correction amount and the global process compensation correction amount based on defect statistics, and updating the process variable correction coefficients and the segmented correction strategy respectively, the process also includes:

[0234] A1: The updated process variable correction coefficients and the updated segmented correction strategy are applied to the online extrusion production of the cover adhesive;

[0235] A2: Continuously collect full-dimensional defect statistics of the L-segment cover adhesive, where L is less than N;

[0236] A3: Extract full-dimensional defect statistics of the L-segment cover adhesive from the circular cache queue, calculate the actual total defect percentage of each segment of cover adhesive, calculate the average of the actual total defect percentage of the L-segment cover adhesive, and record it as the real-time total defect percentage. At the same time, calculate the difference between the real-time total defect percentage and the defect percentage of the cover adhesive good product benchmark corresponding to the cover adhesive good product benchmark, and record it as the first real-time difference.

[0237] A4: Extract the full-dimensional defect statistics of the last N segments of cover adhesive that are consecutively arranged before the execution of update operations S4.4 and S4.5 from the circular cache queue, calculate the actual total defect percentage of each segment of cover adhesive, calculate the average of the actual total defect percentage of the N segments of cover adhesive, and record it as the historical total defect percentage. Then calculate the difference between the historical total defect percentage and the defect percentage of the good product benchmark of cover adhesive, and record it as the second historical difference.

[0238] A5: If the first real-time difference is less than the second historical difference, and the first real-time difference is less than the preset first threshold, then the update is deemed valid, and the updated process variable correction coefficient and the updated segmented correction strategy are maintained.

[0239] A6: If the first real-time difference is greater than or equal to the second historical difference, or if the first real-time difference is greater than or equal to a preset first threshold, the update is deemed invalid, and the process variable correction coefficient and the segmented correction strategy before the update are rolled back. The process variable correction coefficient before the update is the current optimal process variable correction coefficient output in S2.6, and the segmented correction strategy before the update is either the maintained one or the segmented correction strategy before the update in S4.5.

[0240] A7: When the update is determined to be invalid, an alarm signal is triggered and the alarm signal is sent to the host computer system.

[0241] Furthermore, if the current extrusion process update is deemed invalid, the system simultaneously triggers a hardware and software linkage alarm mechanism: the on-site audible and visual alarm immediately activates, providing a clear indication of the failure of the adaptive adjustment of the PVG cover rubber extrusion process and the continued exceeding of material defects. Simultaneously, the system packages and uploads the alarm identifier, fault occurrence time, online detection station number, current section cover rubber full-dimensional defect statistics, and real-time extrusion process parameters to the central control computer of the entire extrusion production line. The central control computer centrally archives all alarm records, allowing maintenance personnel to remotely retrieve complete fault data and specifically assign personnel to investigate potential problems such as extruder screw wear, die head accumulation, cooling system flow drift, and visual inspection imaging malfunctions. After manually calibrating process parameters or repairing equipment, stable PVG cover rubber extrusion conditions are gradually restored.

[0242] In summary, this application addresses the continuous extrusion scenario of PVG conveyor belt cover rubber. It establishes a multi-process decoupled control model adapted to the extrusion process based on the defect-process coupling mapping coefficient matrix, and uses digital twin simulation to perform feedforward pre-correction, eliminating inherent baseline defect deviations in the rubber compound. During online production, the coupling ratio is calculated using the defect-year growth rate and the overall extrusion process drift rate to distinguish two types of abnormal operating conditions: a low coupling ratio indicates that the defect increment is caused only by small random fluctuations in the extrusion equipment, while a high coupling ratio indicates that defects such as bubbles and delamination are self-excitedly worsening with process drift. Local fine-tuning updates the global process correction coefficients, and segmented differentiated compensation updates the plasticizing / forming / cooling three-stage correction strategies are executed respectively, forming a complete closed-loop control system for PVG cover rubber extrusion defect identification, process adaptive adjustment, effect verification, and anomaly alarm.

[0243] Example 2:

[0244] Please see Figure 3 Another embodiment of the present invention provides: a multi-optical-path defect detection and process adaptive correction system for cover adhesive, comprising:

[0245] The data acquisition module 10 is used to simultaneously acquire multi-optical path synchronous imaging image sequences and real-time process timing parameters of the PVG conveyor belt cover rubber to be inspected, and to perform sliding window mean filtering and timestamp alignment interpolation on the process parameters to achieve synchronous acquisition and standardized preprocessing of image data and process parameters.

[0246] The coupling analysis module 20 is used to input the multi-optical path synchronous imaging image sequence into the multi-branch feature fusion network to generate a multi-optical path fusion feature tensor, which is then passed through a defect detection decoder based on depthwise separable convolution to output defect spatial scale parameters and surface and subsurface defect segmentation masks. The correlation calculation between defect parameters and process parameters is completed by Pearson correlation coefficient and mutual information entropy. After weighted fusion, a defect-process coupling mapping coefficient matrix is ​​constructed to establish a quantitative correlation between defects and process parameters.

[0247] The process pre-correction module 30 is used to build a multi-process decoupling control model based on sparse constraint non-negative matrix decomposition based on the defect-process coupling mapping coefficient matrix. It constructs an objective function with the good product benchmark of the cover adhesive as the optimization target, and uses the gradient descent algorithm to iteratively optimize the process variable correction coefficients. Combined with the digital twin simulation platform, it completes the iterative convergence judgment. By comparing the actual total defect ratio of the cover adhesive with the defect ratio of the good product benchmark of the cover adhesive, it determines and outputs the process pre-correction ready signal to complete the preliminary process optimization.

[0248] The state determination module 40 receives the pre-correction signal and performs Bayesian optimization fine-tuning on the multi-process decoupling control model. It collects full-dimensional defect statistics of the PVG conveyor belt cover rubber segment by segment and stores them in the circular cache queue. When the amount of data reaches the preset number, it calculates the defect month-on-month growth rate, the comprehensive process drift rate and the coupling ratio between the two. Based on the relationship between the coupling ratio and the second threshold, it determines whether the production line is in a stable process state or a process drift exceeding the limit state.

[0249] The process correction module 50, if it determines that the process drift exceeds the limit, retrieves the full-dimensional defect statistics of the corresponding PVG conveyor belt cover rubber, uses a local weighted regression algorithm to solve the local fine-tuning correction amount to update the process variable correction coefficient, and uses a time-series prediction model and model predictive control algorithm to obtain the global process compensation correction amount sequence. It then updates the segmented correction strategy according to the processing segment to achieve a layered process correction that combines local and global approaches.

[0250] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A method for detecting and correcting process self-adaption of multi-path defects of a cover glue, characterized in that, include: S1: Collect multi-path synchronous imaging image sequences and real-time process timing parameters of the PVG conveyor belt cover rubber to be inspected. Extract defect spatial scale parameters and surface and subsurface defect segmentation masks based on the multi-path synchronous imaging image sequences. Perform correlation analysis between the extracted defect spatial scale parameters and surface and subsurface defect segmentation masks and the synchronously acquired real-time process timing parameters to construct a defect-process coupling mapping coefficient matrix. S2: Construct a multi-process decoupling control model based on the defect-process coupling mapping coefficient matrix. Take the defect spatial scale parameter and the good cover rubber benchmark corresponding to the surface and subsurface defect segmentation mask as the optimization target. Iteratively optimize the process variable correction coefficient of the multi-process decoupling control model until the difference between the total proportion of defects in the PVG conveyor belt cover rubber collected in real time and the proportion of defects in the good cover rubber benchmark is less than the preset first threshold. Output the process pre-correction ready signal. S3: After receiving the process pre-correction ready signal, fine-tune the multi-process decoupling control model, collect full-dimensional defect statistics of PVG conveyor belt cover rubber segment by segment, and after continuously statistically analyzing N segments of cover rubber, calculate the defect month-on-month growth rate, comprehensive process drift rate and its coupling ratio based on the full-dimensional defect statistics. S4: If the coupling ratio is less than the preset second threshold, the process variable correction coefficient and segmented correction strategy will be used to maintain the continuous extrusion production of the cover rubber. If the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and the segmented correction strategy are updated respectively based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics.

2. The method of claim 1, wherein the method further comprises: determining the number of light paths of the coating glue based on the first image and the second image; and determining the number of light paths of the coating glue based on the third image and the fourth image. The specific steps of S1 include: S1.1: Based on the multispectral synchronous trigger acquisition device, acquire the bright field image sequence, dark field image sequence and polarization image sequence of the PVG conveyor belt cover rubber under preset lighting conditions, and merge them into a multi-optical path synchronous imaging image sequence. S1.2: Input the multi-optical path synchronous imaging image sequence into a pre-trained multi-branch feature fusion network to extract the surface rubber texture gradient feature map, edge bubble diffraction feature map and subsurface stress distribution feature map respectively, and cascade them in the channel dimension to generate a multi-optical path fusion feature tensor; S1.3: Input the multi-optical-path fusion feature tensor into a defect detection decoder based on depthwise separable convolution, and output defect spatial scale parameters and surface and subsurface defect segmentation masks; the surface and subsurface defect segmentation masks are binary masks, where a pixel with a mask value of 1 indicates that the corresponding pixel belongs to a defect region, and a pixel with a mask value of 0 indicates that the corresponding pixel belongs to a non-defect region; based on the surface and subsurface defect segmentation masks, count the total number of pixels contained in the defect region, and record it as the pixel area of ​​the defect region; the defect spatial scale parameters include at least the defect equivalent diameter, which is defined as the diameter of an equivalent circle with the same area as the pixel area of ​​the defect region.

3. The method of claim 2, wherein the method further comprises: determining the number of light paths of the coating glue based on the first image and the second image; and determining the number of light paths of the coating glue based on the third image and the fourth image. The specific steps of S1 also include: S1.4: Synchronously collect real-time process timing parameters of the PVG cover rubber extrusion production line, perform sliding window mean filtering and timestamp alignment interpolation processing to obtain a standardized process parameter vector that corresponds one-to-one with each frame of multi-optical path synchronous imaging image sequence in the time dimension; the real-time process timing parameters include extruder screw speed, rubber extrusion feed speed, die extrusion gap and temperature-controlled cooling water flow rate. S1.5: Calculate the Pearson correlation coefficient between the equivalent diameter of the defect and the extruder screw speed and rubber extrusion feed speed in the standardized process parameter vector to obtain the first correlation feature value; S1.6: Calculate the mutual information entropy of the pixel area of ​​the defect region with the die extrusion gap and temperature-controlled cooling water flow rate in the standardized process parameter vector to obtain the second associated feature value; S1.7: The first associated feature value and the second associated feature value are weighted and fused according to the preset defect type weight to construct the defect-process coupling mapping coefficient matrix.

4. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 3, characterized in that, The specific steps of S2 include: S2.1: Based on the defect-process coupling mapping coefficient matrix, a multi-process decoupling control model is constructed; the multi-process decoupling control model adopts a non-negative matrix decomposition structure based on sparse constraints, which decomposes the defect-process coupling mapping coefficient matrix into the product of the defect basis matrix and the process contribution coefficient matrix; S2.2: Load the pre-stored PVG cover rubber good product benchmark database, extract the benchmark defect spatial scale parameter vector and benchmark surface and subsurface defect segmentation mask corresponding to the good product cover rubber with the same formula and extrusion batch as the PVG conveyor belt cover rubber to be inspected, and merge them as the cover rubber good product benchmark. S2.3: Set the optimization objective function; the optimization objective function is composed of a weighted sum of the defect scale error term, the defect mask coincidence error term, and the process variable smoothing constraint term. The defect scale error term is obtained based on the defect spatial scale parameter and the reference defect spatial scale parameter vector, and the defect mask coincidence error term is obtained based on the surface and subsurface defect segmentation mask and the reference surface and subsurface defect segmentation mask.

5. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 4, characterized in that, The specific steps of S2 also include: S2.4: Initialize the process variable correction coefficients as unit vectors, and use a gradient descent-based optimization algorithm to iteratively update the initial process variable correction coefficients. After each iteration, calculate the updated process variable correction coefficients, and multiply the updated process variable correction coefficients element by element with the standardized process parameter vector to generate a virtual corrected extrusion process parameter sequence. S2.5: Input the virtual modified extrusion process parameter sequence into the digital twin simulation platform to obtain the simulation defect spatial scale parameters and the simulation surface and subsurface defect segmentation mask, and substitute them into the optimization objective function to calculate the current objective function value; S2.6: When the rate of change of the current objective function value between two adjacent iterations is less than the preset convergence threshold, stop the iteration and take the updated process variable correction coefficient corresponding to the last iteration as the current optimal process variable correction coefficient. S2.7: Apply the current optimal process variable correction coefficient to the real-time PVG cover rubber extrusion production line, continuously produce multiple sections of cover rubber and collect actual defect detection data, calculate the total percentage of actual defects for each section of cover rubber, and calculate the average of the total percentage of actual defects for multiple sections of cover rubber, which is recorded as the real-time collected total percentage of defects in cover rubber; calculate the difference between the real-time collected total percentage of defects in cover rubber and the percentage of defects in the good cover rubber corresponding to the good cover rubber benchmark, and when the difference is less than the preset first threshold, output the process pre-correction ready signal.

6. The method of claim 5, wherein the method further comprises: determining the number of light paths of the coating glue based on the first image and the second image; and determining the number of light paths of the coating glue based on the third image and the fourth image. The specific steps of S3 include: S3.1: Receive the process pre-correction ready signal and activate the fine-tuning mode controller; S3.2: The fine-tuning mode controller loads the current optimal process variable correction coefficient as the initial value for fine-tuning, and loads the defect basis matrix and process contribution coefficient matrix as the objects to be fine-tuned. It uses a Bayesian optimization-based hyperparameter fine-tuning algorithm to update the defect basis matrix and process contribution coefficient matrix online, and obtains the updated defect basis matrix and process contribution coefficient matrix, which is denoted as the fine-tuned multi-process decoupling control model. S3.3: Collect full-dimensional defect statistics of PVG conveyor belt cover rubber segment by segment; the full-dimensional defect statistics include defect type label sequence, defect spatial scale parameter vector, defect depth distribution histogram, defect location spatial coordinate set, and standardized process parameter vector output at the corresponding time for each segment of cover rubber. S3.4: Store the full-dimensional defect statistics of each section of the cover adhesive into a circular cache queue in the order of collection time.

7. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 6, characterized in that, The specific steps of S3 also include: S3.5: When the number of cover adhesive segments stored in the circular cache queue reaches a preset number N, extract the full-dimensional defect statistics of the most recent N consecutive cover adhesive segments from the circular cache queue, and then divide the N cover adhesive segments into the first N / 2 groups and the last N / 2 groups according to the collection time order. Calculate the total number of defects in the first N / 2 groups and the last N / 2 groups respectively. Divide the total number of defects in the last N / 2 groups by the total number of defects in the first N / 2 groups and subtract 1 to obtain the month-on-month growth rate of defects. S3.6: Based on the extracted full-dimensional defect statistics of N segments of continuous cover rubber, extract the actual values ​​of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate from the standardized process parameter vector corresponding to each segment of cover rubber. These constitute four actual value sequences. Linear fitting is performed on each actual value sequence to obtain the slope of the fitted line. The absolute value of the slope of the fitted line is then divided by the arithmetic mean of the actual value sequences to obtain the drift rate of extruder screw speed, rubber extrusion feed speed, die extrusion gap, and temperature-controlled cooling water flow rate, respectively. These drift rates are then weighted according to the preset process parameter weights to obtain the comprehensive process drift rate. S3.7: Calculate the ratio of the year-on-year growth rate of defects to the overall process drift rate, denoted as the coupling ratio.

8. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 7, characterized in that, If the coupling ratio is less than the preset second threshold, the current production line is determined to be in a stable process state. The current optimal process variable correction coefficient remains unchanged, and the current segmented correction strategy remains unchanged. The segmented correction strategy is divided into a plasticizing section, a die extrusion molding section, and a cooling and shaping section according to the extrusion process of the cover rubber. Each process stage corresponds to a set of sub-vectors of process variable correction coefficients. The segmented correction strategy is set to be a unit vector for each process stage sub-vector during initialization.

9. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 7, characterized in that, If the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and the segmented correction strategy are updated based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics, including: S4.1: If the coupling ratio is greater than or equal to a preset second threshold, the current production line is determined to be in a process drift exceeding the limit state, triggering adaptive correction and extracting the latest full-dimensional defect statistics of the M-segment cover adhesive from the circular buffer queue; where M is a preset positive integer less than or equal to N; S4.2: Based on the extracted full-dimensional defect statistics of the M-segment cover adhesive, the local weighted regression algorithm is used to solve for the local fine-tuning correction amount; the local fine-tuning correction amount is an incremental vector with the same dimension as the correction coefficient of the current optimal process variable; S4.3: Based on the full-dimensional defect statistics of all N segments of cover adhesive in the circular buffer queue, the time-series prediction model is used to predict the defect evolution trend of P segments of cover adhesive, and the model prediction control algorithm is used to perform rolling optimization to obtain the global process compensation correction amount sequence; the global process compensation correction amount sequence contains P correction amount vectors, and each correction amount vector corresponds to the global compensation value of a segment of cover adhesive.

10. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 9, characterized in that, If the coupling ratio is greater than or equal to the second threshold, the process variable correction coefficient and the segmented correction strategy are updated respectively based on the local fine-tuning correction amount and the global process compensation correction amount calculated according to the full-dimensional defect statistics. This also includes: S4.4: Add the obtained local fine-tuning correction amount to the current optimal process variable correction coefficient element by element to obtain the updated process variable correction coefficient; S4.5: The obtained global process compensation correction sequence is averaged segment by segment according to the process stage division boundary of the defined segment correction strategy: the arithmetic mean of the correction vectors corresponding to the cover rubber covered by the rubber plasticizing section, the die extrusion molding section and the cooling and shaping section are calculated respectively, and the sub-vectors of the corresponding process stages in the current segment correction strategy are replaced with the three arithmetic means respectively to obtain the updated segment correction strategy.

11. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 10, characterized in that, After updating the process variable correction coefficients and the segmented correction strategy by solving the local fine-tuning correction amount and the global process compensation correction amount based on the full-dimensional defect statistics, the method further includes: The updated process variable correction coefficients and the updated segmented correction strategy are applied to the online extrusion production of the cover adhesive; Continuously collect full-dimensional defect statistics of the L-segment cover adhesive, where L is less than N; Extract full-dimensional defect statistics of the L-segment cover adhesive from the circular cache queue, calculate the actual total defect percentage of each segment of cover adhesive, calculate the average of the actual total defect percentage of the L-segment cover adhesive, and record it as the real-time total defect percentage. At the same time, calculate the difference between the real-time total defect percentage and the defect percentage of the cover adhesive good product benchmark corresponding to the cover adhesive good product benchmark, and record it as the first real-time difference. Extract the full-dimensional defect statistics of the last N segments of cover adhesive before the update operation is performed from the circular cache queue. Calculate the actual total defect percentage of each segment of cover adhesive and the average of the actual total defect percentage of the N segments of cover adhesive, which is recorded as the historical total defect percentage. Then calculate the difference between the historical total defect percentage and the good product benchmark defect percentage of the cover adhesive, which is recorded as the second historical difference.

12. The method for multi-optical path defect detection and adaptive process correction of cover adhesive as described in claim 11, characterized in that, After updating the process variable correction coefficients and the segmented correction strategy by solving the local fine-tuning correction amount and the global process compensation correction amount based on the full-dimensional defect statistics, the method further includes: If the first real-time difference is less than the second historical difference, and the first real-time difference is less than the preset first threshold, then the update is determined to be valid, and the updated process variable correction coefficient and the updated segmented correction strategy are maintained. If the first real-time difference is greater than or equal to the second historical difference, or if the first real-time difference is greater than or equal to a preset first threshold, the update is deemed invalid, and the process variable correction coefficient and the segmented correction strategy before the update are rolled back. If the update is deemed invalid, an alarm signal is triggered and sent to the host computer system.

13. A system for detecting and adaptively correcting defects in cover adhesives via multiple optical paths, used to implement the method for detecting and adaptively correcting defects in cover adhesives via multiple optical paths as described in any one of claims 1-12, characterized in that, include: The data acquisition module is used to simultaneously acquire multi-path synchronous imaging image sequences and real-time process timing parameters of the PVG conveyor belt cover rubber to be inspected; The coupling analysis module is used to input multi-optical path synchronous imaging image sequences into a multi-branch feature fusion network to generate multi-optical path fusion feature tensors, and then output defect spatial scale parameters, surface and subsurface defect segmentation masks through a defect detection decoder; the correlation calculation between defect parameters and process parameters is completed by Pearson correlation coefficient and mutual information entropy, and the defect-process coupling mapping coefficient matrix is ​​constructed after weighted fusion. The process pre-correction module is used to build a multi-process decoupling control model based on the defect-process coupling mapping coefficient matrix, construct an optimization objective function with the good product benchmark of the cover adhesive as the optimization target, and determine and output the process pre-correction ready signal by comparing the actual total defect ratio of the cover adhesive with the defect ratio of the good product benchmark of the cover adhesive. The status determination module receives the process pre-correction ready signal and performs Bayesian optimization fine-tuning on the multi-process decoupling control model. It calculates the defect month-on-month growth rate, the comprehensive process drift rate and its coupling ratio to determine whether the production line is in a stable process state or a process drift exceeding the limit state. The process correction module is used to update the process variable correction coefficients, calculate the global process compensation correction sequence, and update the segment correction strategy according to the processing segment.