Image recognition-based glass spacer paper defect detection method and system

Abstract

The present application relates to the technical field of data analysis, and is a glass spacer paper defect detection method and system based on image recognition, specifically comprising: constructing a glass spacer paper stable laying evolution model to obtain a stable laying reference evolution state of a current paper sample to be detected; analyzing the disturbance influence of residual winding memory release on the image recognition result of the current paper sample to be detected; analyzing the additional influence of the delayed morphology evolution caused by the time-delayed rebound evolution process and the in-plane abnormal expansion process of the current paper sample to be detected after tension unloading on the image recognition result; constructing an image judgment correction model to output a corrected defect recognition result, and evaluating whether the glass protection effectiveness of the current glass spacer paper to be detected meets the use requirements according to the corrected defect recognition result. The present application solves the problem that there is a lack of effective early warning for delayed and evolution-type hidden defects in the prior art.

Description

Technical Field

[0001] This invention relates to the field of data analysis technology, and is a method and system for detecting defects in glass spacer paper based on image recognition. Background Technology

[0002] Glass spacers, as a crucial paper layer used for surface protection during the stacking, transportation, and storage of glass sheets, protect architectural glass, photovoltaic glass, coated glass, and electronic display glass, which require extremely high surface flatness and smoothness. Unlike ordinary packaging spacers, glass spacers are typically stored, transported, and supplied in rolls. During prolonged winding, residual bending stress accumulates within the paper fiber layers, and variations in moisture content distribution occur along the radial direction of the roll. Consequently, the paper does not immediately exhibit a stable morphology after unwinding; instead, it undergoes a morphological evolution process governed by the paper's viscoelastic relaxation, environmental humidity rebalancing, and tension release history. In actual production lines, after unwinding and laying, the surface of the glass spacer may initially appear relatively flat. However, as the unwinding tension is released and the resting time lengthens, the latent residual stress gradually releases, manifesting as delayed morphological changes such as the paper edges curling back, the appearance of shadowed boundaries near the edges, the restoration of local ripples, or the slow expansion of in-plane micro-arches. Once such morphological changes occur in a glass stacking system, they transform the surface contact between glass sheets into localized point or line contact, causing stress concentration and leading to indentations, scratches, haze, or film damage on the glass surface. Existing methods for detecting defects in glass spacer paper generally rely on single-moment, single-image machine vision inspection schemes, acquiring only a bright-field or oblique-light image immediately after the paper is laid out, and using image processing algorithms to identify static appearance defects such as stains, holes, creases, tears, or obviously curled edges. This detection mode implicitly assumes that all paper defects that could potentially damage the glass are fully visible at the time of image capture. However, for the aforementioned latent defects driven by winding memory release and moisture rebalancing, these defects are often still latent or only slightly manifest at the time of image capture, and their appearance characteristics are difficult to distinguish from normal paper texture fluctuations. Existing single-moment image-based detection methods cannot effectively identify them. Because existing technologies lack observation of the temporal morphological evolution process after unwinding, there is a disconnect between the test results and the protective effectiveness of glass spacers in actual stacking, transportation, and storage scenarios, making it impossible to provide effective early warning for the aforementioned delayed and evolutionary latent defects. Summary of the Invention

[0003] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0004] The technical problem to be solved by this invention is that the existing technology lacks an effective early warning for late-onset and evolutionary latent defects. A method and system for detecting defects in glass spacers based on image recognition is proposed.

[0005] To achieve the above objectives, the technical solution of the glass spacer paper defect detection method based on image recognition of the present invention includes the following steps: Collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, obtain the stable laying reference evolution state of the current paper pattern to be inspected. Among them, the historical paper pattern image data includes image data collected at least at three time points: after unwinding and initial spreading, after unwinding traction tension is unloaded, and after tension is unloaded and after passing through a preset static window. The historical glass consequence data is used to mark whether each historical paper pattern actually induces glass surface abnormalities in the glass usage scenario. Based on the current working condition driving data, the disturbance effect of residual winding memory release on the current paper sample image recognition result is analyzed, including comparing the deviation between the actual edge recovery degree of the current paper sample after tension unloading and the corresponding allowable recovery upper limit in the stable laying reference evolution state. Based on the time-series image changes of the current paper sample under inspection at three time points, we analyze the additional impact of the delayed morphological evolution caused by the delayed rebound manifestation process and the in-plane anomaly expansion process after tension unloading on the image recognition results. An image judgment correction model is constructed to convert the disturbance effect and additional effect into the correction amount of the original defect identification result of the current paper sample to be inspected, so as to obtain the corrected defect identification result. Based on the corrected defect identification result, the glass protection effectiveness of the current glass spacer paper to be inspected is evaluated to see if it meets the usage requirements.

[0006] Preferably, the historical operating condition driven data includes paper roll diameter position data, tension release data, static dwell time data, ambient humidity data, and paper roll storage humidity data; The current operating condition driving data includes paper roll diameter position data, tension release data, static dwell time data, ambient humidity data, and paper roll storage humidity data; Roll diameter location data is used to characterize the different radial layers from which the paper sample originates; tension release data is used to characterize the constraint change process of the paper sample transitioning from a traction-layout state to a free-layout state; static dwell time data is used to characterize the dwell time of the paper sample after tension unloading and before it is placed on the glass; ambient humidity data is used to characterize the relative humidity level and fluctuation of the paper sample during the development stage; and paper roll storage humidity data is used to characterize the initial moisture balance state of the paper roll before it enters the unwinding station.

[0007] Preferably, the stable laying reference evolution state includes, under the current working conditions, the upper limit of the allowable width of the paper edge rebound band, the upper limit of the allowable density of the shadow boundary of the near-edge main area, the upper limit of the allowable area of ​​the abnormal connected area, and the upper limit of the allowable recovery curvature at two time points: after the normal paper pattern is unloaded from the unwinding traction tension and after passing through the preset static window.

[0008] Preferably, the analysis of the disturbance effect of residual winding memory release on the current paper sample image recognition result includes: obtaining the actual edge recovery curvature of the current paper sample after tension unloading, and comparing it with the upper limit of the allowable recovery curvature at the corresponding moment in the stable laying reference evolution state to obtain the curvature recovery ratio; Obtain the ratio between the actual curvature increase of the current paper sample after tension unloading and after passing through the preset static window and the corresponding allowable increase. Multiplying the curvature recovery ratio by this ratio yields an index characterizing the impact of the disturbance.

[0009] Preferably, the additional impact of delayed morphological evolution on image recognition results is analyzed, including: combining the degree of delayed growth of the edge recovery curvature of the current test paper sample within a preset static window, the degree of coupling between the relative change of the shadow boundary density of the near-edge main area and the relative change of the edge curvature, and the product of the expansion of the area of ​​the in-plane abnormal connected area and the degree of advancement towards the central bearing area, and combining the three to obtain an index characterizing the additional impact.

[0010] Preferably, the disturbance effect and the additional effect are converted into a correction amount for the original defect identification result of the current paper sample to be inspected, including: determining the mapping coefficient of the disturbance effect and the additional effect on the defect score correction through statistical regression of historical data; The perturbation effect and the additional effect are converted into correction components with the same dimensions as the original defect identification result by using the mapping coefficient, and then superimposed to obtain the basic correction amount; Based on the location of the anomaly on the current paper sample to be inspected, the basic correction amount is multiplied by the area weighting factor to obtain the final correction amount.

[0011] Preferably, the evaluation of whether the glass protection effectiveness of the current glass spacer paper meets the usage requirements includes: classifying the current paper sample into one of the preset multiple protection risk levels based on the corrected defect identification results, as well as key process indicators characterizing the degree of delayed rebound, key process indicators characterizing the degree of edge-body coupling, and key process indicators characterizing the degree of in-plane abnormal expansion. Based on the surface sensitivity type of the target glass, determine the protection adaptation conclusion corresponding to the protection risk level; When the protection risk level and protection compatibility conclusion meet the release conditions set by the production line, the glass spacer paper to be inspected is determined to meet the usage requirements.

[0012] Preferably, the protection risk levels, from low to high, include at least: stable laying level, delayed early warning level, and subsidence risk level; the determination of the protection risk level adopts the priority matching principle from high to low. When the absolute value of the difference between the paper roll storage humidity and the average relative humidity in the preset static window exceeds the preset threshold, the judgment threshold for the corresponding delay rebound degree in the delay warning level is adjusted to the stricter direction.

[0013] Preferably, the paper pattern image data includes both bright field images and oblique light images. The bright field images are used to extract the paper pattern outline, paper edge lines, local brightness uniformity, and in-plane texture distribution features, while the oblique light images are used to enhance the display effect of paper edge lifting bands, wavy ridge lines, shadow boundaries, and local micro-arched areas. The stable deployment evolution model adopts a structure that combines a physics-inspired basic model with a data-driven correction model; The physical-inspired basic model is established based on the viscoelastic recovery law of paper and is used to give initial estimates of the upper limit of the allowable recovery curvature and the upper limit of the allowable rebound band width of the paper edge; The data-driven correction model is used to learn the upper limit of the allowable density of the shadow boundary of the near-edge subject region, the upper limit of the allowable area of ​​the abnormal connected region, and the correction term for the residual of the physical heuristic base model based on historical samples.

[0014] In addition, the image recognition-based glass spacer paper defect detection system of the present invention includes the following modules: Stable laying evolution prediction module, winding memory perturbation analysis module, delayed manifestation analysis module, identification result correction and evaluation module; The stable laying evolution prediction module is used to collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, the stable laying reference evolution state of the current paper pattern to be inspected is obtained. The winding memory disturbance analysis module analyzes the disturbance effect of residual winding memory release on the image recognition result of the current paper sample under inspection based on the current working condition driving data, including comparing the deviation relationship between the actual edge recovery degree of the current paper sample under inspection after tension unloading and the corresponding allowable recovery upper limit in the stable laying reference evolution state. The delayed manifestation analysis module analyzes the delayed morphological evolution of the current paper sample after tension unloading and the in-plane abnormal expansion process based on the temporal image changes of the current paper sample at three time points, and analyzes the additional impact of the delayed morphological evolution on the image recognition results. The identification result correction and evaluation module is used to construct an image judgment correction model, convert the disturbance effect and additional effect into the correction amount of the original defect identification result of the current paper sample to be inspected, obtain the corrected defect identification result, and evaluate whether the glass protection effectiveness of the current glass spacer paper to be inspected meets the usage requirements based on the corrected defect identification result.

[0015] Compared with the prior art, the technical effects of the present invention are as follows: The beneficial effects of this invention lie in transforming the defect detection of glass spacer paper from the existing static appearance conformity evaluation to a predictive assessment of service risks. By introducing a multi-time-node time-series image acquisition and analysis mechanism, latent defects such as delayed rebound, which are not fully manifested in the initial stage after unwinding and laying, can be effectively identified. This overcomes the inherent limitation of existing single-moment image detection methods, which can only capture fully manifested appearance defects and cannot predict the morphological evolution trend during subsequent static placement. This establishes a more accurate correspondence between the detection conclusions and the actual protective effectiveness of the glass spacer paper on the glass surface in actual stacking, transportation, and storage scenarios. In addition, this invention can significantly reduce the risk of quality loss such as indentations, scratches, and fogging on the glass surface caused by the failure to detect latent defects in the paper in a timely manner. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a schematic flowchart of a glass spacer paper defect detection method based on image recognition according to the present invention; Figure 2 This is a schematic diagram of the structure of a glass spacer paper defect detection system based on image recognition according to the present invention. Detailed Implementation

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0018] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0019] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0020] Example 1: like Figure 1 As shown in the figure, an embodiment of the present invention provides a method for detecting defects in glass spacer paper based on image recognition, such as... Figure 1 As shown, the specific steps include the following: Step S1: Collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, obtain the stable laying reference evolution state of the current paper pattern to be inspected. Step S2: Analyze the impact of residual winding memory release on the image recognition result of the current paper sample based on the current working conditions and driving data analysis. This step targets the mechanical response of the paper immediately after it is unrolled from the roll and the tension is released, distinguishing which edge curling is normal and within a limited range of elastic recovery, and which has exceeded the normal range.

[0021] Step S3: Based on the temporal image changes of the current paper sample to be inspected, simulate the delayed rebound manifestation process and in-plane anomaly expansion process that occur during the static period after the tension is unloaded, and analyze the additional impact of delayed morphological evolution on the image recognition results. This step further observes whether the deformation continues to develop and spread inwards during the period when the paper is left to stand and awaits its online application.

[0022] Step S4: Construct an image judgment correction model to correct the original defect identification results of the current paper sample output by the single image detection system based on the disturbance effect of residual winding memory release and the additional effect of delayed morphological evolution. Step S5: Based on the corrected identification results, assess whether the glass protection effectiveness of the current glass spacer paper meets the usage requirements.

[0023] To make the technical background of this solution easier to understand, the application scenarios and special characteristics of glass spacer paper are further explained here. Glass spacer paper is a surface protective paper layer used during the stacking, transfer, storage, or automatic paper-laying process between glass sheets. Its fundamental difference from ordinary packaging spacer paper lies in the fact that it protects glass sheets, which are extremely sensitive, rigid, and require extremely high flatness. On glass production lines, this paper is usually stored and supplied in roll form. After being wound for a long time, the paper roll inevitably develops residual winding memory, specifically manifested as accumulated bending stress within the paper fiber layer, differences in stress recovery ability between the edge and middle areas, and uneven moisture content distribution along the radial direction of the paper roll.

[0024] When a paper roll is unwound and cut into individual sheets, its surface may initially appear relatively flat. However, as the unwinding tension decreases, the resting time increases, and the paper's moisture content gradually rebalances with the surrounding humidity, the latent residual stress is slowly released. This release manifests on the paper's surface morphology as follows: the paper edges curl upwards again; shadow boundaries appear in the inner areas near the edges that were not previously present; weak wavy patterns reappear on the paper surface; or a slowly expanding localized arch appears in the center of the paper. These subtle deformations, barely perceptible to the naked eye, can transform uniform surface contact into point or line contact within a glass stack, causing localized stress concentration. Ultimately, this results in indentations, scratches, or haze on the glass surface. These defects cannot be fully detected by existing single-moment, single-image surface defect detection methods.

[0025] Existing detection methods mostly rely on a fixed-position industrial camera to take a bright-field or oblique-light image immediately after the paper sample is laid flat. Then, image processing algorithms are used to identify stains, holes, creases, tears, or obviously curled edges on the paper surface. This detection mode implicitly assumes that all defects that may cause glass damage are fully revealed at the moment the paper sample is photographed. However, as the above analysis shows, for glass spacer paper, latent defects caused by delayed rebound and slow stress release may still be in a latent state at the moment of photographing, appearing completely normal or with only minor abnormalities. Existing algorithms cannot distinguish them from harmless normal paper texture. This is the key technical problem that this solution aims to solve: how to enable the detection results to predict those latent defects that have not yet fully appeared but will gradually develop into glass trapping sources during subsequent use.

[0026] The overall approach to solving this problem in this embodiment is to no longer use the flatness of the paper pattern as the sole criterion, but instead to establish a predictive judgment framework based on whether the paper pattern will evolve into a risk of glass collapse under the current working conditions.

[0027] To this end, three core innovative methods were introduced: First, instead of relying solely on images at a single moment, images were sampled at at least three time points to capture the morphological evolution trend; second, instead of analyzing only the paper sample itself, the winding history of the paper roll, the unwinding tension release process, the resting time, and the environmental humidity were incorporated as driving factors into the model input; third, instead of relying solely on the defect features extracted by the image processing algorithm, the actual damage consequences of historical glass were used as labels for training, enabling the model to learn what kind of evolutionary trends would actually lead to glass accidents. The specific implementation methods of each step will be explained in detail below.

[0028] Step S1: Collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, obtain the stable laying reference evolution state of the current paper pattern to be inspected. In this embodiment, step S1 includes the following specific steps: Step S11: Collect historical paper pattern image data, historical glass consequences data, and historical working condition driven data, and perform time alignment, abnormal jump point removal, missing value imputation and standardization on these three types of data, and finally output a historical paper pattern evolution dataset with a single paper pattern as the basic recording unit. Step S12: Construct a stable laying evolution model. The purpose of this model is not to directly determine whether the paper sample is qualified or not, but to build a basis for defect identification: Under the given paper roll diameter position, unwinding tension unloading degree, resting time, ambient humidity fluctuation range within the resting window, average relative humidity within the resting window, and paper roll storage humidity conditions, what kind of change pattern should the morphological characteristics of the edge area, near-edge main body area, and central bearing area of ​​a glass spacer with normal protective performance follow between the three time nodes, and what is the maximum allowable change range? Step S13: Using the historical paper pattern evolution dataset obtained in step S11, train the stable laying evolution model constructed in step S12 so that the model learns the various undetermined parameters in the above-mentioned change law. Step S14: Acquire the image data and working condition driving data of the current paper sample to be inspected, and input them into the trained stable laying evolution model. The model outputs the stable laying reference evolution state corresponding to the current paper sample to be inspected. The stable laying reference evolution state includes at least the upper limit of the allowable recovery curvature at the second and third time nodes. , The maximum allowable width of the paper edge springback band , Upper limit of allowable density of shadow boundary in near-edge main area , and the maximum allowable area of ​​abnormal connected regions , This state serves as the reference benchmark for judging abnormal exceedances in steps S2 and S3.

[0029] The following is a detailed explanation of each of the four sub-steps mentioned above: Step S11: Collect historical paper pattern image data, historical glass consequences data, and historical working condition driving data; Historical paper pattern image data refers to image records of multiple batches of glass spacer paper used in the past, taken at multiple time points after unwinding and laying according to a unified standard. This needs to include image data at least three time points: the first time point is the moment of image acquisition after unwinding and initial laying; the second time point is the moment of image acquisition after unwinding tension is released; and the third time point is the moment of image acquisition after tension is released and the paper has passed through a preset settling window. It should be noted that the reason for requiring no less than three time points is that images with only two time points can usually only reflect the changes before and after, making it difficult to distinguish whether the change mainly occurred at the moment of tension unloading or mainly occurred during the subsequent resting period, and it is also difficult to determine whether the change tends to converge. Three images can distinguish the normal situation where the glass rebounds as soon as the tension is removed, but quickly stabilizes, from the abnormal situation where it seems fine at the time, but becomes more and more serious after a few hours. The two have completely different risks of glass sinking.

[0030] The first time point records the forced flattening state of the paper sample immediately after it is released from the roll but is still pulled by traction force. Its image reflects the initial morphology of the paper sample under the combined action of winding history and traction tension. The second time point records the instantaneous rebound state of the elastically deformed parts that were originally flattened after the paper sample loses traction force. This change is directly related to the stress accumulation in the core area of ​​the paper roll. The third time point records the delayed manifestation state of the paper sample during the resting period. At this time, not only elastic recovery plays a role, but also the viscoelastic creep of the fiber layer, the rebalancing of moisture content and ambient humidity, and the difference in recovery rate of different areas of the paper surface. This is the main manifestation window of the delayed rebound type of latent defects that this invention focuses on.

[0031] Historical paper pattern image data preferably includes both bright field images and oblique light images. Bright field images use diffuse surface light sources for uniform illumination, resulting in balanced overall brightness of the paper surface. This is suitable for extracting the paper pattern's outline, edge lines, local brightness uniformity, and in-surface fiber texture distribution characteristics. Oblique light images use low-angle directional light sources that sweep across one side of the paper surface. They utilize the shadow effect caused by the slight undulations on the paper surface to enhance the display effect of raised edges, wavy ridges, shadow boundaries, and local micro-arched areas. It should be noted that the oblique light angle is preferably between 15 and 30 degrees. Within this range, it is beneficial to enhance the contrast of light and dark boundaries formed by the slight undulations on the paper surface, thereby improving the sensitivity of identifying raised edges, wavy ridges, and local micro-arched areas.

[0032] For the same pattern, three shots must be taken from the same camera position, with the same field of view, the same light source angle, and the same magnification to ensure pixel-by-pixel comparability between images taken at different times.

[0033] The steps for acquiring the first time node image are as follows: after the unwinding and traction device completes the initial spreading of the paper pattern, and before the traction tension is completely unloaded, the line array camera or area array camera is triggered to capture images at the preset position. The illumination methods include bright field diffuse illumination and low-angle oblique light illumination. Bright field images and oblique light images are acquired respectively to record the forced flattening state of the paper pattern under the combined action of winding memory and external traction. The steps for acquiring the second time node image are as follows: After the unwinding traction tension is completely unloaded, keep the paper pattern in a free-laying state on the air-floating platform or reference plane, and trigger the image acquisition again under the same camera position and the same lighting conditions to record the immediate recovery state of the paper pattern after losing the traction and flattening effect. The steps for acquiring the third time node image are as follows: After the tension is unloaded, the paper sample is kept in a free-laying state within a preset resting window. The length of the resting window is determined by the production line cycle time, usually 2 to 24 hours. After the resting period, the image is captured again to record the delayed manifestation state of the paper sample under the combined effects of the continued release of residual winding memory, moisture rebalancing, and asynchronous regional recovery.

[0034] In another preferred embodiment, for scenarios where the production line cycle time does not allow for a complete resting window, an accelerated resting alternative is provided: a temperature and humidity-controlled accelerated recovery chamber is configured in the unwinding area. After tension is unloaded, the paper sample is placed inside the chamber and subjected to accelerated resting at 40±2°C and controlled humidity. The accelerated resting time is converted to the equivalent time of natural resting at room temperature using a pre-calibrated equivalence relationship between temperature and humidity conditions and morphological evolution. This equivalence relationship is preferably established through parallel comparative experiments of the same batch of paper samples under natural and accelerated resting conditions. The parameters of the accelerated recovery chamber need to be calibrated through offline comparative experiments to ensure that the morphological evolution trajectory of the paper sample under accelerated conditions is statistically equivalent to that under natural resting conditions. This alternative makes the method feasible on fast-paced production lines, avoiding the time conflict between resting waiting and production cycle time.

[0035] Historical glass consequence data refers to the surface quality results recorded during the glass inspection process for the batch of glass sheets corresponding to the aforementioned historical paper samples, after stacking, transportation, storage, or online placement. Subsequent glass surface results include at least one or more of the following: indentations, scratches, localized haze, embossing, film damage, or localized friction marks. The acquisition process involves: at the glass washing machine exit or pre-coating inspection station, using automated optical inspection equipment or trained quality inspectors, inspecting each glass sheet individually, categorizing and recording various anomalies according to a preset severity level, and creating a glass consequence label associated with each historical paper sample. This label aims to indicate whether the glass, separated by this paper, ultimately had any problems.

[0036] It should be noted that all subsequent model training will use this label as the final basis for the truth.

[0037] It is important to note that the causes of glass damage are complex, and paper pattern defects may be confounded by other factors such as rough handling and overloading during stacking. To eliminate the interference of such confounding factors on model training, when constructing the historical dataset, priority is given to samples where the handling process was controlled, the number of stacked layers and the load conformed to standard process specifications, and the glass damage can be clearly attributed to the contact area of ​​the paper pattern.

[0038] Samples with normal paper pattern evolution but glass damage caused by factors other than the paper pattern are removed. Samples whose glass damage cannot be clearly attributed to abnormal paper pattern contact are preferably excluded from the training set as uncertain samples or placed in a separate manually reviewed sample pool. This screening step ensures that the label signals in the training set mainly originate from the quality differences of the paper patterns themselves, rather than external noise.

[0039] Historical working condition driven data refers to the process and environmental parameters that can be objectively recorded during the unwinding and laying of each historical paper pattern. In this embodiment, it includes at least five types of data, as follows: One is the roll diameter position data, which is obtained by recording the local winding radius corresponding to the roll layer where the paper pattern is located from the paper roll production management system or the winding machine encoder. And convert it into paper curvature. The greater the curvature of the paper roll, the closer the paper sample is to the core, and the more severe the accumulation of residual bending stress.

[0040] Secondly, tension release data is obtained by reading the tension-time curve of the paper pattern during the unwinding and spreading process from the control system of the unwinding and traction device, and recording the initial traction tension before tension unloading. And the residual tension of the paper pattern after unloading Take the difference As the tension release amplitude.

[0041] Thirdly, static dwell time data is obtained by recording the cumulative time between the moment the paper pattern completes tension unloading and the moment the paper pattern is taken and the first piece of glass is placed in the production line management system. Accurate to the minute.

[0042] Fourthly, environmental humidity data is acquired through the following steps: relative humidity is continuously collected using humidity sensors placed in the unwinding area and the resting area, and the relative humidity fluctuation range within the resting window is recorded. Simultaneously record the average relative humidity within the stationary window. It is used to assess the magnitude of the driving force for the rebalancing of the moisture content of paper samples.

[0043] Fifthly, the paper roll storage humidity data is obtained by taking the average relative humidity of the paper roll in the storage or temporary storage environment before it enters the unwinding station from the storage environment monitoring system or the paper roll entry and exit records. This average relative humidity is used as the paper roll storage humidity to characterize the initial moisture balance state of the paper roll.

[0044] After collecting the above three types of data, the historical image data, glass consequence labels, and working condition driving data of each paper pattern are time-aligned and associated and stitched together using the paper pattern number as the primary key to form a complete sample record. Subsequently, the set of all sample records is processed by removing abnormal transition points and imputing missing values, and then standardized to ensure that features of different dimensions have comparable numerical scales in subsequent model training. The dataset obtained after the above processing is the historical paper pattern evolution dataset.

[0045] It is important to note that the labeling logic of this dataset is fundamentally different from that of traditional appearance detection datasets. Traditional datasets label whether there are creases or holes in the image, while this dataset labels whether the paper, under the given conditions of roll diameter, tension, humidity, and static conditions, ultimately caused anomalies on the glass surface. In other words, samples with slight edge changes in the paper appearance but normal glass consequences are labeled as stable evolution samples; while samples with relatively flat paper appearance but abnormal glass consequences are labeled as deviation samples. This consequence-based labeling method is a key prerequisite for the model to learn to distinguish between stable recovery and latent instability.

[0046] Step S12: Construct a stable laying evolution model. This model is based on a fundamental physical fact: the residual bending stress accumulated by the paper roll in the wound state cannot be completely dissipated immediately after unwinding, but rather relaxes gradually at a rate governed by the viscoelasticity of the paper fibers. For a glass spacer paper with normal protective performance, the changes in paper morphology caused by this relaxation should follow certain spatiotemporal laws. That is, in the short period of time after the tension is unloaded, the edges and local areas may show a certain degree of restorative warping or ripples, but as the resting time increases, the rate of change will gradually slow down and eventually stabilize. At the same time, this normal change is usually mainly limited to the edge area or non-load-bearing area of ​​the paper sample in a statistical sense, and should not show an evolutionary trend of continuous expansion towards the central load-bearing area in the middle of the paper surface.

[0047] Therefore, the stable laying evolution model is not designed to directly determine whether the current paper sample to be inspected is qualified or unqualified. Instead, it exists as a reference system for steps S2 and S3. The input of this model includes the temporal morphological features extracted from historical paper sample images, as well as the paper curl rate in historical working condition driving data. Tension release amplitude Standing time Ambient humidity fluctuation range Average relative humidity inside the still window And the humidity of paper roll storage.

[0048] Preferred temporal morphological features include: paper edge springback width The measurement was taken from the widest point of the shadow band along the edge of the paper in the oblique light image; the degree to which the paper edge contour deviates from a straight line. The maximum deviation distance is obtained by fitting a straight line to the paper edge contour in the bright field image, which helps to determine the uniformity of edge warping; the shadow boundary density of the near-edge main area is also analyzed. Within a pre-defined width range from the edge of the paper, the total length of shadow boundary segments per unit area is obtained by performing grayscale gradient enhancement, edge detection, and length accumulation on the oblique light image of the near-edge main body region; the area of ​​abnormal connected regions is also calculated. The area of ​​abnormal grayscale textures or wavy regions in the brightfield image is calculated after connected component segmentation. The segmentation threshold for these abnormal grayscale textures or wavy regions is preferably determined by comparing and calibrating historical normal paper patterns with abnormal paper patterns to distinguish between normal paper texture fluctuations and abnormal morphological disturbances. The flatness disturbance characteristics of the central bearing area are determined using the grayscale standard deviation of the brightfield image of the central bearing area. Characterization, used to determine the slight wavy trend in the center of the paper.

[0049] In this embodiment, the core relationship of the stable laying evolution model is based on the physical inspiration of the viscoelastic recovery of paper. The reason why the edge of a piece of paper unrolled from a paper roll curls up is fundamentally driven by the curvature of the paper roll. The curved history it represents.

[0050] At the moment of unwinding, the tension is... When paper is forcibly stretched flat, its actual curvature approaches zero. After the tension is unloaded, the stored bending elastic stress within the paper begins to release, pushing the paper surface to deform in the direction of restoring its curl. Since the unwinding process of the paper pattern is essentially a mechanical process of the paper material recovering from a forced bending state to its natural state, the degree of recovery depends primarily on the original curvature of the paper roll, and is also affected by the degree of tension release. That is, the more fully the tension is released, the stronger the tendency of the paper pattern to recover to a state close to its original bending state. As a first-order approximation based on this physical mechanism, the restored curvature... The following relationship can be preferred to describe it: ; Among them, time variable It is preferable to take the moment when the paper pattern completes tension unloading as the starting point for time zero; The tension release influence coefficient is used to comprehensively reflect the promoting effect of the degree of tension release during unwinding on elastic recovery, and its value is between 0 and 1. The first-order principal driving force characterizing the winding history on the restoration of curvature, while the function Introduced in China The relevant items describe the effect of the coupling modulation between winding curvature and tension release degree, and paper bending stiffness on the recovery ratio. The function form is preferably established based on the combination of various input variables, wherein the input variables include the tension release amplitude. Bending stiffness of paper and paper curl Each input quantity is normalized relative to a reference operating condition value, nominal material parameter value, or equipment calibration benchmark value to ensure... The coefficient is limited to a value between 0 and 1; The specific function form was determined through finite element simulation combined with physical calibration, including the bending stiffness of the paper material. The preferred input is not used as a direct input for online testing, but is determined during the offline calibration stage through standard bending tests, finite element fitting, or historical sample parameter inversion, and then incorporated into the final result. In the function parameters. This represents the stress relaxation time constant of paper fibers, reflecting the rate of viscoelastic recovery of the paper. Under fixed-angle oblique light illumination, the width of the paper edge rebound band... The actual height of the paper edge Within the effective working range, a monotonically corresponding relationship is preferred; within a small range, it can be approximated as a linear relationship. This mapping relationship can be established through offline calibration; while the tilt height... And with the restoration of curvature Unconstrained length of washi paper edge It is proportional to the square of, that is .

[0051] The above relationship is preferably applicable to engineering conditions where the unconstrained section of the paper edge has small deflection and approximately constant curvature. When the paper edge deflection exceeds the preset small deflection range, the relationship can be corrected in segments through offline high-order geometric calibration.

[0052] The length of the unconstrained section of the paper edge The effective free span between the edge of the paper pattern and the position where it is first subjected to guiding pressure, planar support, or continuous bonding constraint can be pre-calibrated by the equipment structural parameters or calculated by combining online images with the geometric relationship of the mechanism. Therefore, the stable paper pattern at any time The maximum allowable width of the paper edge springback band It can be given by the following formula: ; in It is the inverse function of the calibration mapping function from the height of the raised area to the width of the shaded area. The allowable curvature is calculated by substituting the working parameters of the stable paper pattern into the curvature restoration equation; the mapping function Within the effective working interval corresponding to online detection, it is preferable to maintain monotonicity to ensure its inverse function. It is unique; when the full-interval fitting does not satisfy monotonicity, it is preferable to use a piecewise monotonic calibration method to establish the mapping relationship.

[0053] Meanwhile, the change in shadow boundary density in the near-edge main area is also related to the moisture expansion of paper fibers caused by humidity fluctuations. When the ambient humidity changes, the moisture content of different areas of the paper responds at different speeds. That is, the edge area responds faster due to its larger exposed area, while the central area responds slower. This asynchronous regional response will generate additional moisture-induced curvature in the paper, which in turn affects the distribution of shadow boundary density.

[0054] In summary, the model outputs the permissible occurrences of a normal paper pattern at three time points under the current operating conditions. , and The upper limit range, and the corresponding upper limit of the allowed curvature recovery. and .

[0055] Among them, the upper limit of curvature that can be restored and the maximum allowable width of the paper edge springback band The upper limit of the allowable density of the shadow boundary in the near-edge subject region is preferably determined based on the above-mentioned physical heuristic relationship. and the maximum allowable area of ​​abnormal connected regions The preferred approach is to input the temporal morphological features and the operating condition-driven data into the training model, and learn the statistically permissible range of the model under the corresponding operating conditions from historical normal paper pattern samples.

[0056] Based on this, the aforementioned deviation of the paper edge outline from a straight line Its purpose is: when When the third time node increases significantly relative to the second time node, it indicates the presence of non-uniform local warping concentrations at the paper edge. This feature will be used as the coupling response index in step S32. The amplitude modulation factor is used in the calculation, thereby incorporating the judgment of whether the edge warping is uniform into the coupled response evaluation system.

[0057] Meanwhile, when the average relative humidity inside the stationary window When the difference between the humidity of the paper roll storage and the humidity exceeds a preset threshold, it indicates that the driving force for moisture rebalancing is enhanced, the humidity gradient between different areas of the paper sample is larger, and the risk of delayed rebound is correspondingly increased. At this time, the delayed rebound manifestation index in step S31 is... The threshold for judgment should be tightened accordingly. The specific tightening range is determined by the statistical correlation between humidity difference and excessive delayed rebound in historical data. This is existing technology and will not be elaborated further.

[0058] In addition, it should be noted that the flatness disturbance characteristics (grayscale standard deviation) of the central bearing area In step S33, the function is specifically as follows: if the abnormal connected region has invaded the central bearing area, the rate of change of the grayscale standard deviation is used as an auxiliary criterion. That is, when the relative growth rate of the grayscale standard deviation exceeds a preset threshold, even if the area of ​​the abnormal connected region has not significantly expanded, the in-plane abnormal expansion index should be adjusted. The value of is adjusted upwards to reflect the trend of deteriorating micro-smoothness in the center of the paper.

[0059] Step S13: Train a stable laying evolution model based on the historical paper pattern evolution dataset; After the model structure is determined, the historical paper pattern evolution dataset established in step S11 needs to be used to determine the undetermined parameters in the model. In a preferred embodiment, the stable laying evolution model adopts a structure that combines a physical heuristic basic model and a data-driven correction model: the former is used to give the initial estimate of the upper limit of the allowable recovery curvature and the upper limit of the allowable width of the paper edge springback band, while the latter is used to learn the upper limit of the allowable density of the shadow boundary of the near-edge main area, the upper limit of the allowable area of ​​the abnormal connected area, and the correction term for the residual of the former based on historical samples.

[0060] The parameters that need to be determined include the relaxation time constant of the paper. Tension release influence coefficient The functional form, the moisture expansion response coefficient, and the length of the unconstrained paper edge related to the equipment structure and laying condition. Calibration value.

[0061] The dataset is randomly divided into a training set and a validation set at a ratio of 80% and 20%, respectively. The training set is used to learn the parameters, and the validation set is used to monitor the generalization ability.

[0062] The training process can employ ensemble regression models such as gradient boosting trees, or directly use least squares fitting based on physical models, depending on the scale of the paper sample data. When using gradient boosting trees as the data-driven correction model, corresponding regressors can be established for each target output. The initial output of each regressor is preferably set to the mean of the corresponding target quantity in the normal samples after glass formation in the training set. Then, the residuals are fitted round by round, and the error changes are monitored on the validation set. In each subsequent iteration, the residual between the current model prediction and the measured values ​​of each paper sample is calculated first, and then a new regression tree is trained to fit this residual. For each leaf node of the regression tree, the mean of the residuals within the node is taken as the output value, and this output is multiplied by a learning rate less than 1 and accumulated into the existing model. The mean squared error is monitored on the validation set, and training is stopped when the error no longer decreases in multiple consecutive iterations.

[0063] During the training process, paper samples with normal glass consequences are used as stable evolution samples, and their morphological evolution trajectories are used to learn the allowable variation boundaries under normal operating conditions; paper samples with abnormal glass consequences are used as deviation samples to identify which evolution trajectories will exceed the allowable boundaries.

[0064] Step S14: Obtain the stable laying reference evolution state corresponding to the current paper sample to be inspected; Based on the trained model, for a paper sample to be inspected on the current production line, according to step S11, image data of three time points are collected, and the corresponding working condition driving data of the paper sample is obtained from the production line system, including paper roll curvature, tension release amplitude, resting time, ambient humidity fluctuation amplitude, average relative humidity within the resting window, and paper roll storage humidity.

[0065] Inputting this data into the trained stable laying evolution model, the model will output the stable laying reference evolution state that the current paper sample should correspond to under the current working conditions, specifically the upper limit of the allowable width of the paper edge rebound band at the second and third time nodes. Upper limit of allowable density of shadow boundary in near-edge main area Maximum allowable area of ​​abnormal connected regions And the maximum curvature allowed to be restored and .

[0066] in, and It can be obtained directly from the model output, or by combining the intermediate parameters output by the model with the restoration curvature equation for inverse calculation.

[0067] The aforementioned stable laying reference evolution state can be understood as the reference recovery trajectory that a normal paper sample should exhibit after tension unloading and during static rest under the current working conditions. The subsequent steps S2 and S3 are used to determine, based on this reference trajectory, whether the morphological changes of the current paper sample to be inspected belong to a convergent normal recovery or an abnormal evolution with a continuous amplification trend.

[0068] Step S2: Analyze the impact of residual winding memory release on the image recognition result of the current paper sample based on the current working conditions and driving data analysis. In this embodiment, step S2 includes the following specific steps: Step S21: Extract parameters directly related to the residual winding memory release intensity from the current operating condition drive data, including paper curvature. Tension release amplitude The actual edge recovery curvature was calculated by using the width of the paper edge springback band measured by the oblique light image at the second time node. The paper curl rate and tension release amplitude As the driving parameter for the current operating condition, the upper limit of the allowable recoverable curvature obtained through step S1 is used. Participate in subsequent comparisons; Step S22: Construct a model of the impact of winding memory disturbance. The core logic of this model is to quantitatively compare the deviation between the actual degree of curvature recovery and the upper limit of curvature recovery allowed by normal paper pattern under the same working conditions given in step S1. Step S23: Output the disturbance effect value of residual winding memory release. The magnitude of this value directly reflects whether the elastic recovery exhibited by the paper pattern in the initial recovery stage after tension unloading exceeds the range that can be explained by normal working conditions.

[0069] The three sub-steps described above will be explained in detail below: Step S21: Obtain the winding memory release strength characterization parameters; It should be noted that the parameters characterizing the residual winding memory release intensity were extracted from the collected operating condition data and image data, among which, paper roll curvature... and tension release amplitude The data acquisition step in step S11 has already been completed, and the additional quantity to be extracted from the image is the actual edge restoration curvature. The method for obtaining this value is as follows: the measured value of the paper edge springback band width is obtained from the oblique light image at the second time node. Then, based on the geometric relationship of the oblique light projection, the corresponding actual warp height and restoration curvature are calculated.

[0070] The calibration method for the geometric relationship of the oblique light projection is as follows: Under the same camera position, light source angle, and magnification as the online detection, a standard sample with a known tilt height (the tilt height is precisely controlled by a feeler gauge or a micro-displacement stage) is used. Record the width of the shadow band point by point. With the height of the upturn The correspondence is established by using polynomial fitting to create a mapping function from the width of the shadow band to the height of the raised area. Then its inverse function The width of the shaded band is derived by inversely calculating the height of the raised section. Furthermore, this is based on the small deflection bending geometry of the elastic beam. Complete from arrive The reverse calculation.

[0071] It should be noted that this relationship is preferably applicable to engineering conditions where the unconstrained section of the paper edge has small deflection and approximately constant curvature. The unconstrained length of the paper edge refers to the effective free span between the edge of the paper pattern and the position where it first experiences guiding pressure, planar support, or continuous bonding constraint. Its value is pre-calibrated by the equipment structural parameters during paper pattern laying. Therefore, in this embodiment... The acquisition of [these] is feasible in engineering.

[0072] Step S22: Construct a model of the impact of winding memory perturbation, as follows: First, calculate the curvature recovery ratio. It is equal to the measured recovery curvature at the second time point. Divided by the upper limit of curvature recovery allowed at the same moment in the stable laying reference evolution state. : ; when This indicates that the elastic recovery amplitude exhibited by the paper sample in the initial recovery stage after tension unloading is still within the allowable range of normal paper samples under the current working conditions. This degree of rebound can usually be explained by the winding history of the paper roll and the tension release process under the current working conditions, and does not constitute an abnormal disturbance.

[0073] when This indicates that the actual recovery of the paper pattern has exceeded the maximum recovery that a normal paper pattern should have under the same working conditions. This excess recovery cannot be explained solely by the normal release process of residual winding memory, suggesting that there may be additional stress concentration sources or fiber structure defects inside the paper pattern.

[0074] when Less than the preset lower limit of curvature At that time, it is preferred to use The lower limit of the denominator is used in the calculation to avoid numerical instability of the curvature recovery ratio.

[0075] It should be noted that a single ratio at the second time point is insufficient to fully assess the severity of the problem. Some paper samples rebound sharply immediately after the tension is released, but tend to stabilize after a short period of relaxation. Such paper samples may not cause actual problems when stacking glass later. A higher-risk scenario is when the rebound exceeds the limit at the second time point and continues to increase at the third time point. Another scenario that requires attention is when the rebound does not significantly exceed the limit at the second time point, but the growth rate from the second to the third time point is significantly higher than expected. This type of delayed manifestation risk will be further assessed in step S3 in conjunction with the delayed rebound manifestation index.

[0076] Therefore, the recovery and development factor was introduced. Defined as: ; The meaning of this factor is as follows: the numerator is the actual curvature increase of the paper sample from the second to the third time point, and the denominator is the allowable curvature increase of a normal paper sample within the same time period. For paper samples with normal protective performance, the allowable curvature increase between the second and third time points is usually a small non-negative value; when this value is too small and affects the calculation stability, it is preferable to use a preset stability lower limit. As the lower limit of the denominator.

[0077] if This indicates that the rate of increase in the curvature of the paper pattern is not higher than expected, and even if the rebound at the second time point is large, its subsequent development is within a controllable range.

[0078] Conversely, if This means that the paper sample not only rebounds more when the tension is unloaded, but also continues to rebound at a rate higher than that of a normal paper sample during the resting period. This indicates that the residual stress release process in the paper has deviated from the normal viscoelastic relaxation law and is evolving towards an uncontrolled delayed instability.

[0079] Preferably, when At that time, development factors will be restored. We count it as zero so that it only represents the abnormal recovery risk that continues to grow after the second time point, without counting the subsequent decline or convergence process as a negative disturbance.

[0080] Multiplying the two factors above, we obtain the perturbation effect value of residual wound memory release. Preferably, the curvature recovery ratio and recovery and development factors All values ​​are treated as non-negative, and when the corresponding calculation result is less than zero, it is counted as zero, thus reducing the disturbance impact value. The product form is always used to represent the intensity of non-negative anomalous disturbances because these two factors have a synergistic amplification effect. If a paper sample only rebounds slightly at the second time point but then quickly stabilizes, then... Slightly greater than 1 The product of the two values ​​is close to 1, remaining within a controllable low range and unlikely to trigger an excessive alarm. However, if a paper sample rebounds beyond the limit and continues to grow, then... and At the same time, if the product is greater than 1, it will be amplified sharply, thus significantly distinguishing this paper sample with dual danger signals from those paper samples that only have a momentary rebound but can converge. This is beneficial to improving the identification sensitivity of paper samples with both momentary exceedance and subsequent continuous growth characteristics, thereby reducing the probability of missing high-risk paper samples.

[0081] Step S23: Output the disturbance effect of residual winding memory release; Based on the above model, the final output is the disturbance impact value. This value will be passed to step S4 as the first key input for correcting the defect identification result. This disturbance effect value is essentially used to characterize whether the edge curling of the paper after it is unrolled from the roll and the tension is removed is a normal recovery rebound or a signal that there is abnormal stress concentration or fiber structure instability inside.

[0082] Step S3: Simulate the delayed rebound manifestation process and in-plane anomaly expansion process of the current paper sample after tension unloading based on the temporal image changes of the current paper sample to be inspected, and analyze the additional impact of delayed morphological evolution on the image recognition results. The focus of step S2 is to observe the immediate response of the paper sample after the force is released, while the focus of step S3 is to observe whether the morphological changes of the paper sample continue to evolve in the direction of increasing risk within the static window after the tension is unloaded.

[0083] In this embodiment, step S3 includes the following specific steps: Step S31: Quantify the degree of delayed rebound of the paper pattern within the static window by utilizing the edge image feature changes of the third time node relative to the second time node. Step S32: Analyze whether edge deformation has been transmitted to the near-edge main body region and formed a coupled response by utilizing the difference in shadow boundary density between the second and third time nodes; Step S33: Utilize the temporal changes in the area of ​​the abnormal connected region and its distance to the central bearing region to determine whether the paper anomaly is expanding from a local area to a large area within the surface. Step S34: Combine the analysis results from the above three dimensions and output the additional impact value of delayed morphological evolution on image recognition results. .

[0084] Step S3 is constructed because it is considered that not all deformations will lead to glass damage; only those deformations that continue to develop over time and spread inward in space pose a real risk of entrapment.

[0085] The following is a detailed explanation of each of the four sub-steps mentioned above: Step S31: Obtain the degree of delayed rebound manifestation; The quantitative index for the degree of delayed rebound manifestation uses the recovery and development factor calculated in step S22. Here it is denoted as the delayed rebound manifestation index. ,Right now : ; The meaning of this index is that: under normal conditions, the growth rate of the edge curvature of a paper pattern should gradually slow down within the resting window, and the allowable growth should be decreasing. If the actual growth of a paper pattern exceeds the normal allowable growth, it means that its rebound process has not converged according to the normal law within the resting window, but instead shows an abnormal and continuous growth trend.

[0086] The larger the value, the more severe the delay rebound becomes. This index is one of the key features used in this scheme to characterize the risk of temporal evolution within the static window. It captures the static evolution process that is completely invisible in traditional single-image detection. Preferably, when At that time, We count it as zero so that it only represents the risk of delayed rebound that continues to grow within the resting window, without counting the subsequent decline or convergence process as a negative manifestation.

[0087] when Less than the preset stability lower limit At the same time, consistent with step S22, preferably using It is used as the lower limit of the denominator for calculation.

[0088] Step S32: Obtain the edge-subject coupling response degree; If the deformation of the edge region is always confined to the narrow non-load-bearing area near the edge of the paper, its threat to the glass is limited because the main contact surface of the glass when it is stacked is the central main area of ​​the paper pattern. However, when the edge deformation begins to affect the near-edge main area, its threat to the glass changes. Therefore, this step needs to assess the degree of coupling between the edge deformation and the near-edge main area anomaly.

[0089] At the image processing level, the specific approach is as follows: A pre-defined near-edge subject region of a specified width is drawn inside the edge of the paper, and the shadow boundary density within this region is statistically analyzed at the second and third time points. and .

[0090] An increase in shadow boundary density indicates more local undulations, twists, or ripples on the paper surface in that area. These are direct visual evidence of inward transmission of edge deformation. To quantify the amplification of this anomalous transmission, a coupling response index is preferably constructed. This is the ratio of the relative increment of texture perturbation in the near-edge main region to the relative increment of edge curvature: ; The numerator of this formula is the relative growth rate of the shadow boundary density of the near-edge main region, and the denominator is the relative growth rate of the edge recovery curvature; the ratio of the two reflects the rate of deterioration of the main region relative to the rate of deterioration of the edge region.

[0091] if This indicates that the morphological anomalies in the main area are intensifying at a faster rate than edge deformation, and the initial anomalies at the edges are already transforming into a larger area inside the paper.

[0092] when or Each below the preset lower limit , In this case, it is preferable to replace the denominator with the corresponding lower limit value, or to use a value relative to the allowable upper limit. , The normalized form is used for calculation to avoid abnormal amplification of the coupling response exponent caused by the denominator approaching zero.

[0093] The preset lower limit and The value can be statistically determined based on the resolution of the camera used in the detection system and the noise level of the image processing algorithm. It is usually taken as three times the standard deviation of the corresponding measurement value on the stable paper sample.

[0094] Preferably, when the relative growth rate of edge recovery curvature is less than a preset lower limit for growth... At that time, with As the lower limit of the denominator of the ratio, or directly using The minimum value is set based on the absolute overlimit rule to avoid distortion of the coupling response exponent when the edge growth approaches zero or falls back.

[0095] Furthermore, based on the allowable upper limit threshold output in step S14, when the measured shadow boundary density at the third time node... It has already exceeded the allowed limit at this moment. At that time, regardless of the result calculated by the above formula Regardless of the value, it should be It should be set to at least 1.0 to ensure that paper samples that have exceeded the absolute threshold can be effectively captured by subsequent risk level classification. It should be noted that this minimum setting is used to avoid underestimating the coupling risk of paper samples that have exceeded the absolute limit in the near-edge main area when the relative incremental calculation is affected by the stability of the denominator.

[0096] At the same time, if the paper edge outline deviates from a straight line The relative growth rate at the third time point relative to the second time point A value exceeding 20% ​​indicates significant non-uniform warping concentration at the paper edge. At this point, the calculated... Multiply by an amplitude adjustment factor of 1.2 to further enhance the response to this type of localized stress concentration risk.

[0097] The 20% relative growth rate threshold and the amplitude adjustment coefficient of 1.2 are preferably determined based on statistical calibration of historical glass damage data, used to achieve a better differentiation effect between the concentrated samples of local non-uniform warping at the paper edge and the glass damage results; when Smaller than the preset contour deviation lower limit At that time, it is preferred to use Its relative growth rate is calculated using the lower limit of the denominator.

[0098] Step S33: Obtain the degree of in-plane anomaly expansion; Unlike step S32, which observes local regions near the edge, step S33 assesses the expansion trend of anomalies at the scale of the entire paper surface. It utilizes the connected component segmentation results of texture anomalies or wavy areas in the bright-field image to obtain the area of ​​the anomalous connected regions at the second and third time nodes. and And the shortest distance from the outer boundary of the abnormal connected region to the center bearing area on the paper. and .

[0099] Define the in-plane anomaly propagation index The product of the relative increase in the area of ​​the abnormally connected region and the degree of advancement towards the central carrying area: ; The first item is the abnormal area relative to the total area of ​​the effective detection surface. The increase in the proportion reflects the extent of the expansion of the anomalous region on an area scale; the second term is the inverse ratio of the distance from the anomalous boundary to the center, which is the ratio when the anomalous connected region is closer to the central carrying area at the third time node than at the second time node. This value is greater than 1, thus amplifying the index and reflecting the extent to which the abnormal area advances towards the central bearing area. The product of the two values ​​means that the index will only increase significantly when the area expands and advances towards the center occurs simultaneously. If only the abnormal area at the edge increases slightly but is still far from the central bearing area, the index remains low, indicating that the in-plane protection function has not been substantially threatened. It should be noted that this design avoids false alarms caused by normal small fluctuations in the edge area.

[0100] Preferably, when At that time, the area increment term is counted as zero so that the in-area abnormal expansion index only represents the incremental risk brought about by the expansion of abnormal areas, and does not count the abnormal contraction process as a negative risk contribution.

[0101] When the outer boundary of the abnormal connected region has intersected with or encroached upon the central carrying region, then... At that time, it is preferable to Set to the preset minimum distance lower limit Alternatively, the distance propagation term can be directly set to a preset saturation value to avoid distortion of the in-plane abnormal expansion exponent due to the denominator being zero.

[0102] To reinforce the effect of the absolute threshold, when the area of ​​the abnormal connected region at the third time node... It has directly exceeded the allowed upper limit given in step S14. hour, The value should not be lower than the preset base value of 0.03 to ensure that the paper sample is at least included in the consideration of the risk level of entrapment. The base value of 0.03 is preferably determined by calibration through historical glass consequence data to ensure that the paper sample whose abnormal connected area has exceeded the limit has at least the preset base extended risk contribution in the comprehensive risk assessment.

[0103] Furthermore, when the abnormal connected area has invaded the central bearing area, the rate of change of the gray standard deviation of the central bearing area described in step S12 is used as an auxiliary criterion: if the relative growth rate of the gray standard deviation exceeds the preset threshold, even if the area of ​​the abnormal connected area has not significantly expanded, it indicates that the micro-smoothness of the middle part of the paper is deteriorating.

[0104] The threshold for the relative growth rate of the grayscale standard deviation is preferably determined based on the statistical distribution calibration of historical glass consequences data. The value of is adjusted upwards, with an adjustment factor of . ,in The standard deviation of grayscale in the central bearing area, The sensitivity coefficient is determined through calibration using historical data; when Less than the preset grayscale fluctuation lower limit At that time, it is preferred to use Its relative growth rate is calculated using the lower limit of the denominator.

[0105] Step S34: Output the additional effects of delayed morphological evolution; Combining the indices from the three dimensions mentioned above yields the additional impact value of delayed morphological evolution. The synthesis method employs a product form to reflect the nonlinear synergistic amplification effect among the three: ; Preferably, the , and The negative part is truncated to zero, that is, when the corresponding calculation result is less than zero, it is counted as zero, thus reducing the additional impact value. Only the incremental risk brought about by delayed morphological evolution is characterized, without including the convergence improvement process in the negative risk contribution.

[0106] The use of a product form here has the engineering significance of the following: there is a progressive and interconnected causal relationship among these three factors. The continuous development of delayed rebound is a prerequisite, which provides a continuous driving force for subsequent coupling and expansion. The coupling between the edge and the near-edge main body area is a key turning point, which indicates that the deformation is no longer limited to the non-sensitive area at the edge, but begins to be transmitted to the more functionally sensitive interior of the paper. The abnormal expansion in the in-plane is the final morphological characteristic, which directly indicates that the paper has lost its uniform support capacity over a large area.

[0107] When all three occur simultaneously, the threat to the glass is far greater than that posed by any single factor alone. When expressed as a product, when... When the rebound is very small and tends to converge, even or There are slight fluctuations. The value will also be suppressed to a lower level, which aligns with the engineering intuition that if the rebound converges, the risk is limited. Conversely, when When it is already very large, the amplification effect of coupling and expansion will make The rapid increase enabled the effective interception of high-risk paper samples.

[0108] Step S4: Construct an image judgment correction model to correct the defect identification results of the paper sample under inspection based on the disturbance effect of residual winding memory release and the additional effect of delayed morphological evolution. In this embodiment, step S4 includes the following specific steps: Step S41: Based on the disturbance influence value obtained in step S2 The additional influence value obtained in step S3 Combined with the original defect identification results of the paper sample to be inspected by the traditional single-image detection system First calculate the basic correction amount. Then, the final comprehensive correction amount for the subsidence is obtained by combining the anomaly occurrence area. ; Step S42: Superimpose the original defect identification result with the final comprehensive defect correction amount, and output the corrected defect identification result. ; It should be noted that the significance of this correction is that it no longer merely reflects the degree of appearance defects of the paper pattern at the current moment, but comprehensively predicts the probability that the paper pattern will cause damage to the glass in subsequent use.

[0109] The two sub-steps mentioned above will be explained in detail below: Step S41: Obtain the comprehensive correction amount for the trapping effect; Original defect identification results of the paper sample to be inspected This can be provided by a traditional machine vision inspection system already deployed on the existing production line. In another embodiment, this system may already have the ability to detect paper stains, holes, creases, damage and obvious curling, and output a comprehensive appearance defect score, for example but not limited to the range of 0 to 100, with a higher score indicating a more serious appearance defect. The work in this step is not to replace the existing scoring system, but to add time dimension information that is completely invisible to the existing system, thus upgrading the static appearance score to a comprehensive score that incorporates the risk of temporal evolution.

[0110] Basic correction amount Depend on and Both factors determine that, due to and These are all ratio-based indices, and they correlate with defect scores. The units of measurement need to be converted using a conversion factor. This conversion factor is not arbitrarily set, but rather derived based on statistical regression of historical data. In another embodiment, the specific method is as follows: In the historical paper pattern dataset, the original defect identification results of each paper pattern are... Disturbance impact value and additional impact values A statistical correlation analysis was conducted with the severity of the corresponding glass-related consequences, and the results were determined through regression calibration. and Mapping coefficients for defect score correction and .

[0111] Preferably, the above mapping can be completed using ordered risk regression, piecewise linear fitting, or other supervised learning calibration methods to find a set of conversion coefficients that optimizes the accuracy of glass damage risk prediction. The resulting basic correction calculation relationship is as follows: ; Preferably, the and All are treated as non-negative risk amounts, thereby ensuring and The domain is valid. The natural logarithm function is used here. The form of and The engineering considerations for compressing, rather than directly multiplying linearly, are: when or When the value of the defect score is very large, the correction of the defect score should show a marginal decreasing trend, that is, the increase of the correction amount gradually tends to be flat rather than infinitely linearly diverging. This is more in line with the engineering experience that there is an upper limit to the severity of defects in actual production.

[0112] Conversion factor and With For the same defect component, its specific value is automatically determined by the regression calibration process described above.

[0113] Furthermore, in actual production scenarios, anomalies in different areas of the paper pattern have drastically different consequences for glass denting. Minor warping or ripples in the pre-designed narrow edge area of ​​the outermost edge of the paper pattern are often located in the chamfer or edge of the glass plate when stacking glass, and have little actual impact. However, a weak ripple appearing in the central bearing area of ​​the paper pattern may directly lead to a visible indentation in the center of the glass.

[0114] Therefore, after calculating the basic correction amount Next, it is necessary to multiply by a region weighting factor based on the specific region where the anomaly occurred. The final comprehensive correction amount for the trap is obtained. The value of the regional weighting factor is determined by statistically analyzing the correlation strength between anomalies in different regions and glass consequences in historical data. Specifically, the ratio of the probability of glass damage when the same type of anomaly occurs in the edge region, the near-edge main body region, and the central bearing region is calculated, and the weighting factor of each region is obtained by normalization with the edge region as the benchmark.

[0115] The reference range for the aforementioned regional weighting factor is preferably determined by statistically calibrating the correlation strength between anomalies in different regions and the probability of glass damage in historical glass consequences data. An exemplary reference range is: 1.0 for the edge region, 1.3 to 1.5 for the near-edge main body region, and 1.6 to 2.0 for the central load-bearing region. Specific values ​​are determined comprehensively based on the paper pattern size, glass specifications, and customer quality standards. When anomalies involve multiple regions simultaneously, the regional weighting factor... The weighting factor is preferably selected from the area with the highest risk, or determined by weighted average based on the proportion of abnormal area in each area.

[0116] Step S42: Output the corrected defect identification results; Final output of corrected defect identification results This is the sum of the original score and the final comprehensive correction amount: ; When the corrected defect identification results When the score exceeds the upper limit of the preset scoring range, it is preferable to cut off at the upper limit or directly classify it into the highest risk level.

[0117] This revised scoring differs from the meaning of traditional appearance scoring. The score represents the paper's appearance and cleanliness at the current moment; while the corrected score... It represents the overall risk level of damage to the glass surface caused by the paper pattern in subsequent use, taking into account the source of the paper pattern winding, its immediate performance after unwinding, its evolution trend during resting, and its final position in the glass stack.

[0118] For example, in this embodiment, a piece of paper with a very low appearance score and a very flat appearance, if its and The values ​​are all very high, It may be corrected to a higher risk level, thus being correctly intercepted. This transformation from static appearance scoring to a comprehensive scoring that combines temporal evolution risk is an important distinguishing feature of this scheme from traditional single-moment appearance detection methods.

[0119] Step S5: Evaluate whether the glass protection effectiveness of the glass spacer paper to be inspected meets the usage requirements based on the corrected identification results.

[0120] In this embodiment, step S5 includes the following specific steps: Step S51: Based on the corrected defect identification results and the delayed rebound manifestation index Coupling response index The paper sample to be tested will be classified into one of the four preset protection risk levels; Step S52: Based on the intended use and surface sensitivity level of this batch of glass, determine the protection suitability of this risk level in the current application scenario; Step S53: Based on the target protection level requirements set by the production line, determine whether the glass spacer paper to be inspected can be released for use.

[0121] The three sub-steps described above will be explained in detail below: Step S51: Obtain the protection risk level of the paper sample to be inspected; For example, the protection risk level is divided into four levels, from low to high: stable laying level, delayed early warning level, sinking risk level and interception level.

[0122] The division of levels is not solely based on Instead of a single threshold, the assessment combines corrected scores and various process indicators to avoid the situation where an extreme anomaly in a single indicator is masked by the overall score. The determination of the above four risk levels preferably adopts a priority matching principle from high to low, that is, first determine whether the interception level condition is met, then determine whether the trapping risk level condition is met, and then determine whether the delay warning level condition is met. If none of them are met, it is determined to be the stable laying level. The specific classification logic is as follows: The stable laying grade is suitable for paper patterns with lower corrected scores and all process indicators within the normal range, meeting the requirements. Not higher than the first threshold, and at the same time Not exceeding 0.8, Not exceeding 1.0, Not exceeding 0.03; this level indicates that the paper sample's performance during tension unloading and resting is consistent with historical normal protective paper samples, and can be used as a priority release level under the corresponding process conditions and glass sensitivity level requirements.

[0123] The delayed warning level applies to cases where the corrected score is slightly higher than the first threshold but not exceeding the second threshold, or where the score is not high, but... The value has exceeded 0.8, indicating that although the paper sample has not yet shown coupling and in-plane expansion in the static window, the delayed rebound of its edges is already too fast, and there is a possibility of further deterioration; this grade of paper sample shall not be used for highly sensitive glass, but may be used for short-term storage of ordinary float glass under controlled conditions.

[0124] It should be noted that the humidity of the paper roll storage is similar to the average relative humidity inside the resting window. When the absolute value of the humidity difference exceeds 10%RH, the driving force for humidity rebalancing is significantly enhanced, and the risk of delayed rebound further increases. The 10%RH humidity difference threshold and its corresponding... The tightening threshold is preferably determined by calibration using historical paper samples under different humidity migration driving force conditions; under this condition, the above... The judgment threshold has been tightened from 0.8 to 0.6 to match a more stringent evaluation scale.

[0125] The "risk level" applies when the corrected score exceeds the second threshold, or It has exceeded 1.0. If any of the conditions in 0.03 are met, it indicates that the paper sample has shown clear signs of edge-to-body transmission or abnormal in-plane expansion, and its protective function is deteriorating; the paper sample of this grade is judged by default not to be released; in specific non-appearance applications with low requirements for glass surface quality, it can enter the manual review or special approval restricted use process.

[0126] The interception level applies when the corrected score has exceeded the third threshold, or , , Two or more of the three items exceed their respective high-risk thresholds simultaneously; The high-risk threshold is preferably determined by historical glass consequence data, corresponding to the risk boundary points that have a high degree of distinguishability for glass damage consequences, namely the delayed rebound manifestation index, coupling response index, and in-plane abnormal expansion index. Paper samples that meet the interception level conditions indicate that a complete trapping chain has been formed inside, from edge rebound to main body coupling and then to in-plane expansion. Under the default process rules, they are all judged to be unusable and preferably enter the scrapping or isolation disposal process.

[0127] The four levels mentioned above correspond to The first, second, and third thresholds are preferably determined by receiver operating characteristic (ROC) curve analysis of historical glass damage data, taking the dividing points that provide the best differentiation between each level of glass damage consequences.

[0128] Step S52: Calculate the glass protection adaptation result; The glass protection adaptation result is determined jointly based on the risk level and the surface sensitivity type of the target glass. For ordinary architectural float glass, its surface has a relatively high tolerance for minor indentations and abrasion marks. The paper samples of the stable lay-up level and the delayed warning level can be judged as acceptable, the dent risk level is judged as restricted use, and the interception level is judged as unacceptable. For coated glass and photovoltaic glass, its film surface or textured structure is more sensitive to local contact anomalies. It is preferable to judge only the stable lay-up level as acceptable, the delayed warning level as restricted use, and the dent risk level and the interception level as unacceptable. For electronic display glass and ultra-thin cover glass, its surface quality requirements reach the optical level. Any micron-level surface anomaly may cause the entire glass to be scrapped. Therefore, usually only the stable lay-up level is judged as acceptable, and the other three levels are judged as unacceptable.

[0129] This method of determining release based on the type of glass allows the same sample to yield different release conclusions for different downstream applications, thereby achieving graded release control that matches quality risks under different downstream application conditions.

[0130] Step S53: Determine whether the usage requirements are met; During production line operation, the detection system obtains the upper limit of the allowable risk level or the target protection adaptation rule corresponding to the current batch of glass from the manufacturing execution system. Only when the actual risk level of the paper sample to be inspected is not higher than the upper limit of the allowable risk level, and the protection adaptation conclusion obtained in step S52 is acceptable, will the system determine that the paper sample meets the usage requirements and control the paper laying machine to pick it up normally; otherwise, it will determine that it does not meet the usage requirements, trigger an audible and visual alarm, and automatically discharge the paper sample to the non-conforming product channel.

[0131] In summary, this embodiment expands the defect detection of glass spacer paper from traditional static appearance conformity inspection to predictive assessment combined with service risk. The detection system no longer judges solely based on the flatness of the paper sample at the moment of photographing, but integrates the winding memory of the paper roll, the release process of unwinding tension, the delayed evolution trend within the static window, and the final sensitivity level of the glass. It provides a risk assessment for each paper sample that is directly linked to the consequences of glass protection. This invention helps reduce the probability of surface damage to glass products due to spacer paper protection failure during stacking, storage, and transportation.

[0132] Example 2: like Figure 2 As shown in the figure, an image recognition-based glass spacer paper defect detection system according to an embodiment of the present invention is as follows: Figure 2 As shown, it includes the following modules: Stable laying evolution prediction module, winding memory perturbation analysis module, delayed manifestation analysis module, identification result correction and evaluation module; The stable laying evolution prediction module is used to collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, the stable laying reference evolution state of the current paper pattern to be inspected is obtained. The winding memory disturbance analysis module analyzes the disturbance effect of residual winding memory release on the image recognition result of the current paper sample under inspection based on the current working condition driving data, including comparing the deviation relationship between the actual edge recovery degree of the current paper sample under inspection after tension unloading and the corresponding allowable recovery upper limit in the stable laying reference evolution state. The delayed manifestation analysis module analyzes the delayed morphological evolution of the current paper sample after tension unloading and the in-plane abnormal expansion process based on the temporal image changes of the current paper sample at three time points, and analyzes the additional impact of the delayed morphological evolution on the image recognition results. The identification result correction and evaluation module is used to construct an image judgment correction model, convert the disturbance effect and additional effect into the correction amount of the original defect identification result of the current paper sample to be inspected, obtain the corrected defect identification result, and evaluate whether the glass protection effectiveness of the current glass spacer paper to be inspected meets the usage requirements based on the corrected defect identification result.

[0133] Example 3: This embodiment provides an electronic device, including: a processor and a memory, wherein the memory stores a computer program that can be called by the processor; The processor executes the aforementioned image recognition-based method for detecting defects in glass spacers by calling computer programs stored in memory.

[0134] The electronic device can vary considerably depending on its configuration or performance. It may include one or more processors (Central Processing Units, CPUs) and one or more memories, wherein the memory stores at least one computer program, which is loaded and executed by the processor to implement the image recognition-based glass spacer defect detection method provided in the above-described embodiment. The electronic device may also include other components for implementing its functions; for example, it may have wired or wireless network interfaces and input / output interfaces for data input and output. Details will not be elaborated upon in this embodiment.

[0135] Example 4: This embodiment proposes a computer-readable storage medium on which an erasable and rewritable computer program is stored. When the computer program runs on the computer device, it causes the computer device to perform the aforementioned image recognition-based method for detecting defects in glass spacers.

[0136] For example, computer-readable storage media can be read-only memory (ROM), random access memory (RAM), compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage devices.

[0137] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0138] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via a wired network and / or wireless network. A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives (SSDs).

[0139] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this invention can be implemented in electronic hardware or a combination of computer software and electronic hardware.

[0140] In the several embodiments provided by this invention, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only one method, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0141] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

Claims

1. A method for detecting defects in glass spacer paper based on image recognition, characterized in that, The method includes: Collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, obtain the stable laying reference evolution state of the current paper pattern to be inspected. Based on the current working condition driving data, the disturbance effect of residual winding memory release on the current paper sample image recognition result is analyzed, including comparing the deviation between the actual edge recovery degree of the current paper sample after tension unloading and the corresponding allowable recovery upper limit in the stable laying reference evolution state. Based on the time-series image changes of the current paper sample under inspection at three time points, we analyze the additional impact of the delayed morphological evolution caused by the delayed rebound manifestation process and the in-plane anomaly expansion process after tension unloading on the image recognition results. An image judgment correction model is constructed to convert the disturbance effect and additional effect into the correction amount of the original defect identification result of the current paper sample to be inspected, so as to obtain the corrected defect identification result. Based on the corrected defect identification result, the glass protection effectiveness of the current glass spacer paper to be inspected is evaluated to see if it meets the usage requirements.

2. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, Historical operating condition-driven data includes paper roll diameter position data, tension release data, static dwell time data, ambient humidity data, and paper roll storage humidity data; The current operating condition driving data includes paper roll diameter position data, tension release data, static dwell time data, ambient humidity data, and paper roll storage humidity data; Roll diameter location data is used to characterize the different radial layers from which the paper sample originates; tension release data is used to characterize the constraint change process of the paper sample transitioning from a traction-layout state to a free-layout state; static dwell time data is used to characterize the dwell time of the paper sample after tension unloading and before it is placed on the glass; ambient humidity data is used to characterize the relative humidity level and fluctuation of the paper sample during the development stage; and paper roll storage humidity data is used to characterize the initial moisture balance state of the paper roll before it enters the unwinding station.

3. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, The stable laying reference evolution state includes, under the current working conditions, the upper limit of the allowable width of the paper edge rebound band, the upper limit of the allowable density of the shadow boundary of the near-edge main area, the upper limit of the allowable area of ​​the abnormal connected area, and the upper limit of the allowable recovery curvature at two time points: after the unwinding traction tension of the normal paper pattern is unloaded and after passing through the preset static window.

4. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, The analysis of the impact of residual winding memory release on the image recognition result of the current paper sample to be inspected includes: obtaining the actual edge recovery curvature of the current paper sample to be inspected after tension unloading, and comparing it with the upper limit of the allowable recovery curvature at the corresponding moment in the stable laying reference evolution state to obtain the curvature recovery ratio; Obtain the ratio between the actual curvature increase of the current paper sample after tension unloading and after passing through the preset static window and the corresponding allowable increase. Multiplying the curvature recovery ratio by this ratio yields an index characterizing the impact of the disturbance.

5. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, The additional effects of delayed morphological evolution on image recognition results are analyzed, including: the degree of delayed growth of the edge recovery curvature of the current test paper sample within a preset static window, the degree of coupling between the relative change of the shadow boundary density of the near-edge main area and the relative change of the edge curvature, and the product of the expansion of the area of ​​the in-plane abnormal connected area and the degree of advancement towards the central bearing area. The three factors are combined to obtain an index characterizing the additional effects.

6. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, The disturbance and additional effects are converted into corrections to the original defect identification results of the current paper sample to be inspected, including: determining the mapping coefficients of the disturbance and additional effects on the defect score correction through statistical regression of historical data; The perturbation effect and the additional effect are converted into correction components with the same dimensions as the original defect identification result by using the mapping coefficient, and then superimposed to obtain the basic correction amount; Based on the location of the anomaly on the current paper sample to be inspected, the basic correction amount is multiplied by the area weighting factor to obtain the final correction amount.

7. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, Assess whether the glass protection effectiveness of the current glass spacer paper meets the usage requirements, including: classifying the current paper sample into one of the preset multiple protection risk levels based on the corrected defect identification results, as well as key process indicators characterizing the degree of delayed rebound, key process indicators characterizing the degree of edge-body coupling, and key process indicators characterizing the degree of in-plane abnormal expansion; Based on the surface sensitivity type of the target glass, determine the protection adaptation conclusion corresponding to the protection risk level; When the protection risk level and protection compatibility conclusion meet the release conditions set by the production line, the glass spacer paper to be inspected is determined to meet the usage requirements.

8. The method for detecting defects in glass spacer paper based on image recognition according to claim 7, characterized in that, The protection risk levels, from low to high, include at least: stable installation level, delayed warning level, and collapse risk level; the determination of the protection risk level adopts the priority matching principle from high to low. When the absolute value of the difference between the paper roll storage humidity and the average relative humidity in the preset static window exceeds the preset threshold, the judgment threshold for the corresponding delay rebound degree in the delay warning level is adjusted to the stricter direction.

9. The method for detecting defects in glass spacer paper based on image recognition according to claim 1, characterized in that, The paper pattern image data includes both bright field images and oblique light images. Bright field images are used to extract the paper pattern outline, paper edge lines, local brightness uniformity, and in-plane texture distribution features. Oblique light images are used to enhance the display effect of paper edge lifting bands, wavy ridges, shadow boundaries, and local micro-arched areas. The stable deployment evolution model adopts a structure that combines a physics-inspired basic model with a data-driven correction model; The physics-inspired basic model is established based on the viscoelastic recovery law of paper and is used to give initial estimates of the upper limit of the allowable recovery curvature and the upper limit of the allowable springback band width of the paper edge; The data-driven correction model is used to learn the upper limit of the allowable density of the shadow boundary of the near-edge subject region, the upper limit of the allowable area of ​​the abnormal connected region, and the correction term for the residual of the physical heuristic base model based on historical samples.

10. A glass spacer paper defect detection system based on image recognition, used to implement the glass spacer paper defect detection method based on image recognition as described in any one of claims 1-9, characterized in that, The system includes the following modules: Stable laying evolution prediction module, winding memory perturbation analysis module, delayed manifestation analysis module, identification result correction and evaluation module; The stable laying evolution prediction module is used to collect historical paper pattern image data, historical glass consequence data, and historical working condition driving data to construct a stable laying evolution model for glass spacer paper. Based on the current image data of the paper pattern to be inspected and the current working condition driving data, the stable laying reference evolution state of the current paper pattern to be inspected is obtained. The winding memory disturbance analysis module analyzes the disturbance effect of residual winding memory release on the image recognition result of the current paper sample under inspection based on the current working condition driving data, including comparing the deviation relationship between the actual edge recovery degree of the current paper sample under inspection after tension unloading and the corresponding allowable recovery upper limit in the stable laying reference evolution state. The delayed manifestation analysis module analyzes the delayed morphological evolution of the current paper sample after tension unloading and the in-plane abnormal expansion process based on the temporal image changes of the current paper sample at three time points, and analyzes the additional impact of the delayed morphological evolution on the image recognition results. The identification result correction and evaluation module is used to construct an image judgment correction model, convert the disturbance effect and additional effect into the correction amount of the original defect identification result of the current paper sample to be inspected, obtain the corrected defect identification result, and evaluate whether the glass protection effectiveness of the current glass spacer paper to be inspected meets the usage requirements based on the corrected defect identification result.

Images