Machine vision-based online detection method for glue thickness of spider silk protein

Abstract

The present application relates to the field of machine vision detection, and discloses a spider silk protein gluing thickness online detection method based on machine vision, which comprises the following steps: collecting multi-modal images before gluing and after gluing and drying to complete polarization Stokes parameter reconstruction and obtain corrected images; establishing tension normalized material coordinates and generating a registration image sequence; extracting reflection polarization differential features and near-infrared attenuation features from the registration image sequence to generate an optical feature vector, and synchronously collecting gluing and drying process parameters to generate a process state vector; inputting the optical feature vector into a thickness prediction model to obtain an initial equivalent dry glue thickness, correcting the initial thickness based on a linear compensation amount calculated from the process state vector to obtain a target equivalent dry glue thickness; calculating the comprehensive confidence of the target thickness, judging according to a threshold value and outputting a detection result. The present application offsets the interference of the substrate and the environment, realizes the alignment of the physical regions before gluing and after gluing and drying, and guarantees the accuracy and data reliability of online detection.

Description

Technical Field

[0001] This invention relates to the field of machine vision inspection, specifically to an online detection method for the thickness of spider silk protein coating based on machine vision. Background Technology

[0002] In the coating process of medical spider silk protein, the equivalent dry adhesive thickness is a key quality control indicator, and currently, visual technology is often used for online thickness detection. However, medical nonwoven fabric substrates have the physical characteristic of uneven substrate thickness. Existing visual detection methods are easily affected by substrate background interference and have difficulty overcoming the influence of sensor mode noise and abnormal surface illumination. This results in the inability to accurately separate physical information directly related to the adhesive layer thickness, leading to low thickness detection accuracy.

[0003] Meanwhile, detecting the adhesive coating thickness requires comparing image features before and after adhesive coating and drying. Because non-woven fabric is a flexible substrate, it is prone to stretching and localized non-uniform deformation during dynamic transport due to production line movement and tension changes. This deformation causes spatial misalignment in the images acquired before and after adhesive coating and drying. Existing detection methods lack deformation correction and coordinate mapping mechanisms for flexible substrates, making it impossible to achieve physical alignment of images before and after adhesive coating and drying, resulting in a loss of positional reference for subsequent feature difference calculations.

[0004] Furthermore, the actual coating thickness is not only affected by optical characteristics but also by non-optical factors such as environmental fluctuations or changes in the condition of coating equipment, resulting in process drift. Most current purely optical inspection methods fail to incorporate such process deviations into their correction scope. In industrial settings, existing methods also lack mechanisms for evaluating the validity of inspection data, failing to identify and filter invalid data under conditions of equipment malfunction or environmental interference. This leads to the stability and reliability of online inspection systems failing to meet production requirements. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides an online detection method for spider silk protein coating thickness based on machine vision. This method solves the problem that in existing medical spider silk protein coating processes, the flexible nonwoven substrate is prone to deformation under dynamic tension, and the optical properties of the coating before and after drying are greatly affected by environmental and process parameters, which makes it impossible for traditional single-modal vision detection methods to achieve accurate online detection and traceability of equivalent dry adhesive thickness.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an online detection method for spider silk protein coating thickness based on machine vision, comprising: Multimodal images of spider silk protein before and after coating and drying were acquired, and polarization Stokes parameters were reconstructed to obtain corrected images; Tension-normalized material coordinates are established to determine the material coordinates corresponding to the calibrated image, and a registration image sequence is generated based on the material coordinates; Optical feature vectors are generated by extracting reflection polarization difference features and near-infrared attenuation features from the registered image sequence, and coating process parameters and drying process parameters are acquired simultaneously to generate process state vectors. The optical feature vector is input into the thickness prediction model to obtain the initial equivalent dry adhesive thickness. The linear compensation amount is calculated based on the process state vector to perform residual correction on the initial equivalent dry adhesive thickness, and the target equivalent dry adhesive thickness is obtained. Calculate the overall confidence level for the target equivalent dry adhesive thickness. When the overall confidence level reaches the preset confidence threshold, output the target equivalent dry adhesive thickness; otherwise, output the thickness invalidation flag, failure reason code, and material coordinates.

[0007] This invention reduces the interference from substrate background and environmental fluctuations on single optical measurements by multimodal optical feature fusion and process state compensation. It uses tension-normalized coordinates to align the areas before and after coating and drying, and sets a comprehensive confidence level assessment to ensure the reliability of online detection data.

[0008] Furthermore, multimodal images of spider silk protein before and after coating and drying were acquired, and polarization Stokes parameters were reconstructed to obtain corrected images. Specifically, this included: acquiring reflected polarization images from the multimodal images using a micro-polarization camera; forming a four-dimensional column vector from the light intensity observations belonging to the four polarization directions, which had been normalized by exposure time but had not yet had their fixed dark field bias deducted; introducing a position-correlated instrument response matrix for pixel-level physical compensation decoupling; and reconstructing three-dimensional polarization Stokes parameters based on the four-dimensional column vector. The three-dimensional polarization Stokes parameters include components representing the local total reflected light intensity, components representing the horizontal and vertical polarization dominance, and components representing the 45° and 135° polarization dominance. Specifically, the decoupled and recovered three-dimensional polarization Stokes parameters are equal to the generalized inverse of the instrument response matrix at the spatial coordinate position multiplied by the difference between the four-dimensional light intensity observation column vector at the spatial coordinate position (normalized by exposure time but before deducting the fixed bias) and the four-dimensional bias column vector at the corresponding position (composed of dark current, fixed light leakage, and fixed readout bias). Based on the reconstructed three-dimensional polarization Stokes parameters, a corrected reflected polarization image is obtained, and the corrected reflected polarization image and the corrected near-infrared transmission image are integrated to generate a corrected image. Decoupling using the position-dependent instrument response matrix can compensate for sensor fixed-mode noise and the non-uniformity of the instrument polarization response at the pixel level, thereby obtaining an accurate physical polarization state.

[0009] Furthermore, tension-normalized material coordinates are established to determine the material coordinates corresponding to the corrected images, and a registration image sequence is generated based on the material coordinates. Specifically, this includes: receiving production line encoder counts, substrate linear velocity, tension sensor signals, and swing roller position signals to establish tension-normalized material coordinates; and using the tension-normalized material coordinates, the conveying direction coordinates of each imaging station relative to the reference position, and the dynamic path changes caused by the swing rollers to determine the material coordinates corresponding to each image. The medical nonwoven fabric substrate is divided into multiple fixed physical regions, each defined as a thickness estimation unit. A one-to-one candidate relationship is established for the same thickness estimation unit before and after adhesive application and drying, based on material coordinates. A registered image sequence that meets the physical region overlap requirement of the thickness estimation unit is then generated based on this one-to-one candidate relationship. Establishing tension-normalized material coordinates overcomes the positional deviation caused by asynchronous image acquisition, ensuring that the images before and after adhesive application and drying physically correspond in space, providing a coordinate reference for feature difference operations.

[0010] Furthermore, based on the tension sensor signal, the tension and elongation relationship of the nonwoven fabric model is used to establish tension-normalized material coordinates. After coarse material tracking is completed by combining the dynamic path changes caused by the position signal of the swing roller, a global two-dimensional affine transformation is obtained using the uncoated reference side strips on both sides. When the registration residual shows local non-uniform deformation, restricted local deformation correction is performed to generate a registration image sequence that meets the physical region overlap requirements of the thickness estimation unit. Combining tension and swing roller status for material tracking and correcting local non-uniform deformation can reduce image registration errors caused by nonlinear stretching of flexible nonwoven fabric during long-distance transport.

[0011] Furthermore, the reflection polarization difference features are extracted from the registered image sequence, specifically including: performing pixel-level physical anomaly filtering, removing overexposed pixels that reach the hardware highlight cutoff threshold in any original polarization channel, removing underexposed pixels whose local total reflected light intensity component is lower than the minimum effective light intensity threshold, and reconstructing polarization measurement units whose Stokes parameters do not meet the physical consistency requirements, and marking the location of the removed pixels as invalid; based on the reconstructed three-dimensional polarization Stokes parameters, the square root of the sum of the squares of the horizontal and vertical polarization dominance components and the squares of the 45° and 135° polarization dominance components is taken, and then divided by the local total reflected light intensity component, so as to calculate the degree of linear polarization at the corresponding positions of the pre-coating image and the post-drying image in the registered image sequence respectively; By subtracting the linear polarization degree of the corresponding position in the image before coating from the linear polarization degree of the corresponding position in the dried image, a reflection polarization difference feature matrix is ​​generated. When any corresponding position is marked as invalid, the corresponding reflection polarization difference feature is simultaneously marked as invalid, and the valid features in the reflection polarization difference feature matrix are used to construct the optical feature vector. By removing overexposed, underexposed, and abnormal pixels that do not meet physical consistency requirements, the difference in linear polarization degree before coating and after coating and drying is extracted to reflect the modulation change of polarization state by the coating microstructure, avoiding interference from abnormal lighting conditions.

[0012] Furthermore, near-infrared attenuation features are extracted from the registered image sequence, including: obtaining the normalized transmittance before coating the image before coating and the normalized transmittance after drying the image after drying in the registered image sequence respectively; and calculating the near-infrared attenuation features under the condition that the normalized transmittance before coating and the normalized transmittance after drying are both higher than the minimum effective transmittance, the corresponding original images are not saturated, and the spatial registration is effective. The near-infrared attenuation characteristic at a spatial coordinate location is equal to the natural logarithm of the quotient obtained by dividing the normalized transmittance before adhesive application at the spatial coordinate location by the normalized transmittance after drying at the spatial coordinate location. The mean and variance of the near-infrared attenuation characteristic are calculated separately. The mean and variance of the near-infrared attenuation characteristic are combined to form the optical feature vector. The near-infrared attenuation characteristic is calculated based on the logarithm of the transmittance ratio to offset the background interference caused by the uneven thickness of the nonwoven fabric substrate, thus obtaining attenuation information related to the adhesive layer thickness.

[0013] Furthermore, the optical feature vector is input into the thickness prediction model to obtain the initial equivalent dry adhesive thickness. Specifically, this includes: retrieving the optical feature vector containing the mean and variance of the near-infrared attenuation features, and introducing a constant bias term with a value of 1 to combine them into the input feature vector of the thickness prediction model. After obtaining the determined thickness prediction coefficient vector corresponding to the thickness prediction model, the thickness prediction coefficient vector is multiplied by the current input feature vector using an inner product operation. This product is then used as the output of the thickness prediction model to calculate and generate the initial equivalent dry adhesive thickness. A linear model with a constant bias term is used to establish the mapping relationship between multidimensional optical features and thickness, reducing computational complexity and preventing numerical divergence of the model under boundary conditions.

[0014] Furthermore, a linear compensation amount is calculated based on the process state vector to perform residual correction on the initial equivalent dry adhesive thickness, thereby obtaining the target equivalent dry adhesive thickness. This includes: calculating the process state deviation vector between the process state vector and the reference process vector corresponding to the current process interval; using the process correction coefficient vector to perform inner product mapping on the calculated process state deviation vector; applying a fixed amplitude boundary truncation limit; and generating a linear compensation amount. The linear compensation amount is equal to the result of multiplying the transpose of the process correction coefficient vector by the process state deviation vector, and then applying an upper and lower limit truncation function consisting of the maximum process perturbation truncation threshold and the lower truncation boundary limit. The initial equivalent dry adhesive thickness is added to the linear compensation amount to obtain the target equivalent dry adhesive thickness. The linear compensation amount is calculated using the process state deviation to correct the thickness deviation caused by non-optical process factors, while an amplitude truncation limit is set to prevent abnormal process data from causing excessive compensation.

[0015] Furthermore, the comprehensive confidence level is calculated for the target equivalent dry adhesive thickness. Specifically, this includes calculating the light intensity effectiveness, polarization physical consistency, spatial registration quality, feature space proximity, and near-infrared feature distribution consistency corresponding to the generation of the target equivalent dry adhesive thickness. The five indicators are mapped to confidence levels in the range of 0 to 1, with higher values ​​indicating higher confidence. The minimum value of the light intensity effectiveness confidence level, polarization physical consistency confidence level, spatial registration quality confidence level, feature space proximity confidence level, and near-infrared feature distribution consistency confidence level is taken to form the comprehensive confidence level.

[0016] The overall confidence level is calculated and the minimum value of the five indicators is taken to evaluate the current testing status, invalid test results caused by filtration equipment malfunction or environmental interference.

[0017] In a preferred embodiment of the present invention, the method further includes: constructing an incremental optimization objective function in the calibration database that includes historical parameter retention constraints, retaining the historical parameter retention constraints and incorporating the newly added validation sample used in this parameter update; wherein, the incremental optimization objective function is set to minimize the sum of the weighted residual squares corresponding to the independent variables and the historical parameter retention constraint penalty term, and the weighted residual squares are obtained by subtracting the square of the residuals obtained by subtracting the inner product prediction result of the input feature vector and the regression coefficient independent variables from the equivalent dry adhesive thickness labels corresponding to all samples participating in the update, and then multiplying by the corresponding weight coefficients and summing them up; The historical parameter preservation constraint penalty term is equal to the pre-set parameter preservation penalty coefficient multiplied by the square of the L2 norm of the difference between the regression coefficient independent variable and the historical thickness prediction coefficient vector, where the historical thickness prediction coefficient vector is the thickness prediction coefficient vector before the update. The updated thickness prediction coefficient vector is obtained by solving the incremental optimization objective function. Incremental updates using an objective function that includes historical parameter preservation constraints allow the prediction model to retain historical feature parameters while fitting new sample data, reducing model fluctuations caused by parameter updates.

[0018] This invention provides a machine vision-based online detection method for spider silk protein coating thickness. It offers the following advantages: 1. This invention generates an optical feature vector by extracting reflection polarization difference features and near-infrared attenuation features, and combines them with a pixel-level physical anomaly filtering mechanism to offset background interference caused by uneven thickness of the nonwoven substrate, reduce the influence of sensor mode noise and surface illumination anomalies, obtain physical information related to the adhesive layer thickness, and improve the accuracy of thickness detection.

[0019] 2. This invention establishes tension-normalized material coordinates by introducing the production line motion state and tension signal, and generates a registered image sequence by combining edge tracking and local deformation correction. This solves the problem of image misalignment caused by tensile deformation of flexible nonwoven substrate during dynamic transportation, and realizes physical alignment of images before and after gluing and drying, providing a positional reference for subsequent feature difference calculation.

[0020] 3. This invention corrects the initial thickness by generating a linear compensation amount through calculation of process state deviation, and uses multi-dimensional detection indicators to calculate the comprehensive confidence level to compensate for process drift caused by non-optical factors such as changes in environment or coating equipment. At the same time, it filters invalid data under conditions of equipment abnormality or environmental interference, ensuring the stability and reliability of online detection results. Attached Figure Description

[0021] Figure 1 This is an architecture diagram of the online detection system for spider silk protein coating thickness based on machine vision according to the present invention; Figure 2 This is a flowchart of the online detection method for spider silk protein coating thickness based on machine vision according to the present invention; Figure 3 This is a flowchart of the multimodal optical imaging parameter calibration and system decoupling of the present invention; Figure 4 This is a schematic diagram illustrating the principle of flexible substrate spatial tracking and image registration according to the present invention. Figure 5 This is a flowchart illustrating the construction process of the cross-domain multidimensional state feature space of the present invention. Figure 6 This is a schematic diagram illustrating the principle of the combined process perturbation thickness regression solution and defect mapping of the present invention. Figure 7 This is a comparison chart of the model predictions and validation labels of the present invention; Figure 8 This is a response curve diagram of the online monitoring and closed-loop control of adhesive coating thickness according to the present invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figure 1 This invention provides an online detection system for spider silk protein coating thickness based on machine vision, including a substrate detection station, a coating detection station, and a data processing device.

[0024] The system uses equivalent dry adhesive thickness as the core evaluation index, which is the thickness evaluation value obtained by converting the mass of spider silk protein dry adhesive per unit area corresponding to the thickness estimation unit according to the preset apparent dry adhesive density. The calibration process of apparent dry adhesive density includes: preparing dry adhesive samples from the spider silk protein adhesive solution of the current production formula under the same drying conditions; after the dry adhesive samples reach a stable state, determining the apparent dry adhesive density according to the ratio of the dried mass of the dry adhesive sample to the actual volume, and writing the measurement equipment information, sample preparation conditions, number of repeated measurements, and allowable error into the apparent dry adhesive density calibration version.

[0025] Each adhesive formulation and main drying conditions correspond to at least one apparent dry adhesive density calibration version; if the solid content of the adhesive, formulation composition, or main drying conditions exceed the original calibration range, recalibration is performed. Apparent dry adhesive density is saved as a fixed conversion parameter for the corresponding process range.

[0026] The data processing equipment divides the medical nonwoven fabric substrate into multiple fixed physical regions according to the pre-written longitudinal and transverse dimensions of the current production formula. Each fixed physical region is defined as a thickness estimation unit. The size of the thickness estimation unit is pre-determined based on the camera object resolution, encoder trigger spacing, number of transverse partitions of the substrate, and minimum number of effective pixels within the unit, and remains unchanged within the same production batch. The data processing equipment generates an initial equivalent dry adhesive thickness and a target equivalent dry adhesive thickness for each thickness estimation unit. Image areas smaller than the thickness estimation unit are used for local defect localization.

[0027] The substrate inspection station is located upstream of the coating mechanism, including an adjacent first polarization station and a first transmission station, used to acquire reflected polarization images and near-infrared transmission images of the medical nonwoven fabric substrate before adhesive coating, respectively. The coating inspection station is located downstream of the coating and drying mechanisms and upstream of the winding mechanism, including an adjacent second polarization station and a second transmission station, used to acquire reflected polarization images and near-infrared transmission images of the medical nonwoven fabric substrate after adhesive coating and drying, respectively. The images acquired by both the substrate inspection station and the coating inspection station are correlated with the material coordinates of the corresponding thickness estimation unit.

[0028] The coating unit applies adhesive according to the effective coating width set in the current production formula, and retains continuous uncoated reference strips on both sides of the medical nonwoven fabric substrate in the transverse direction. These uncoated reference strips are located in the reserved trimming area or the non-effective coating area of ​​the product, and their preset width meets the requirements for positioning, feature extraction, and lateral offset tolerance, remaining unchanged within the same production batch. The data processing equipment saves the preset width of the reference strips, the allowable width deviation, and the minimum number of effective feature points as parameters of the current production formula.

[0029] The data processing equipment establishes a unified path coordinate system based on a predetermined reference position on the material conveying path, and records the conveying direction coordinates of the first polarization station, the first transmission station, the second polarization station, and the second transmission station relative to the predetermined reference position. The difference in the conveying direction path between the first polarization station and the first transmission station is the first path distance, and the difference in the conveying direction path between the second polarization station and the second transmission station is the second path distance. The data processing equipment combines the conveying direction path differences between each pair of imaging stations with the dynamic path changes caused by the oscillating rollers to construct a real-time path model for image pairing and material tracking. The data processing equipment establishes image data transmission channels with the substrate detection station and the coating detection station, and interacts with the coating mechanism and the drying mechanism for process parameters.

[0030] Reference Figure 2 This invention provides a machine vision-based online detection method for spider silk protein coating thickness, comprising the following steps: S10. Multimodal Image Acquisition and Hardware Calibration. During equipment installation, nonwoven fabric model switching, spider silk protein adhesive formula switching, imaging hardware replacement or recalibration, the system performs optimal working incident angle and physical calibration of the image acquisition instrument, and saves the calibration parameters. During normal production, the system maintains a constant working incident angle and acquires multimodal images under a unified hardware clock and material coordinate reference. The multimodal images include at least a reflected polarization image and a near-infrared transmission image. After applying exposure constraints, the calibration parameters are called to complete polarization signal decoupling and polarization Stokes parameter reconstruction to obtain the corrected reflected polarization image, and dark field subtraction, flat field correction, exposure normalization, and reference transmittance calibration to obtain the corrected near-infrared transmission image. In this invention, the corrected reflected polarization image and the corrected near-infrared transmission image are collectively referred to as the corrected image.

[0031] S20. Material Tracking and Spatial Image Registration. For the corrected reflected polarization image and the corrected near-infrared transmission image, the data processing equipment receives the production line encoder count, substrate linear velocity, tension sensor signal, and swing roller position signal, and establishes tension-normalized material coordinates by calling the tension-elongation relationship of the corresponding nonwoven fabric model. The data processing equipment uses the tension-normalized material coordinates, the conveying direction coordinates of each imaging station relative to the reference position, and the dynamic path changes caused by the swing roller to determine the material coordinates corresponding to each image, and establishes a one-to-one candidate relationship for the same thickness estimation unit before and after adhesive application and drying. After completing coarse material tracking, a global two-dimensional affine transformation is obtained using the uncoated reference side strips on both sides; when the registration residual shows local non-uniform deformation, restricted local deformation correction is performed to generate a registration image sequence that meets the physical region overlap requirements of the thickness estimation unit.

[0032] S30. Optical Feature Extraction and Process Status Acquisition. The data processing equipment extracts reflection polarization difference features and near-infrared attenuation features from the registered image sequence, and performs physical anomaly filtering based on polarization Stokes parameters to generate an optical feature vector. Simultaneously, the data processing equipment acquires coating and drying process parameters, including the spider silk adhesive flow rate. It performs validity assessment and normalization on pre-selected process parameters to generate a process status vector, and then fuses the optical feature vector and the process status vector to construct a joint feature set.

[0033] S40. Joint Regression Solution and Local Defect Location. The data processing equipment determines the current process range based on the nonwoven fabric type, spider silk adhesive formulation, and drying conditions. It then calls the calibrated and validated thickness prediction model within this range, inputting the optical feature vector of the thickness estimation unit into the model to obtain the initial equivalent dry adhesive thickness. When the process state is within the validated small perturbation range, a linear compensation amount is calculated based on the deviation between the process state vector and the baseline process vector. This residual is used to correct the initial equivalent dry adhesive thickness, yielding the target equivalent dry adhesive thickness. Simultaneously, a local defect anomaly map is generated using the reflection polarization difference feature map before statistical dimensionality reduction and after registration, along with the near-infrared attenuation feature map. This map is used to locate suspected missed coating and stripe areas. Furthermore, production line anomaly alarms are triggered based on the defect area, morphological characteristics, and the continuous extension range of the anomaly area in adjacent material coordinates or adjacent thickness estimation units.

[0034] S50. Confidence Assessment and Detection Data Output. For the target equivalent dry adhesive thickness, the data processing equipment calculates light intensity effectiveness, polarization physical consistency, spatial registration quality, feature space proximity, and near-infrared feature distribution consistency. These five indicators are then mapped to a unified confidence level, with higher values ​​indicating greater confidence, thus forming a comprehensive confidence level. When the comprehensive confidence level falls below a preset confidence threshold, an invalid thickness flag, a failure cause code, and material coordinates are output.

[0035] Each thickness estimation unit is bound to a unique tension-normalized material coordinate range during material tracking, and the material coordinates remain unchanged as the substrate passes through each station. The data processing equipment records the timestamps of the thickness estimation unit passing through the coating mechanism, arriving at the coating inspection station, and outputting the inspection results, and determines the material coordinates currently being processed by the coating mechanism at the time of output. The external control system adjusts the material coordinate ranges of subsequent materials that have not yet passed through the coating mechanism based on the thickness deviation corresponding to the material coordinate ranges that have already been inspected.

[0036] If the proportion of effective thickness estimation units within a material control segment is lower than the effective control proportion, or if there is a process range out-of-bounds error, encoder malfunction, sudden tension change, imaging sensor failure, or communication failure, the data processing equipment sends a control freeze or degradation signal. When the linear speed is stable, the external control system maintains the most recently verified unit area glue supply setting; when the linear speed changes, the glue supply amount is adjusted according to the verified linear speed and glue supply feedforward relationship. The data processing equipment uses a weighing verification sample with traceable material coordinates to verify model errors, and only performs model parameter maintenance after the equipment status check and independent verification are passed.

[0037] The following section will elaborate on the core steps of the machine vision-based online detection method for spider silk protein coating thickness, taking into account specific implementation details.

[0038] Reference Figure 3 For the multimodal image acquisition and hardware calibration in step S10, the data processing equipment controls the substrate detection station and coating detection station to perform physical-level parameter configuration and signal decoupling, including the following sub-steps: S11. Determine the working incident angle of the reflection polarization imaging system. The data processing equipment uses multiple sets of physical calibration samples covering the target unit area dry adhesive mass range, and acquires images of the substrate before adhesive application and the samples after drying at multiple candidate working incident angles. For each candidate working incident angle, the data processing equipment extracts the main polarization features and calculates and generates a comprehensive evaluation result based on the slope, monotonicity, and standard deviation of the main polarization features as a function of unit area dry adhesive mass, as well as the proportion of effective pixels, the proportion of overexposed pixels, and the proportion of low-signal pixels.

[0039] In this embodiment, the first polarization station and the second polarization station use the same working incident angle. The data processing device selects the candidate angle with the best comprehensive evaluation result from the candidate angles that simultaneously meet the minimum signal-to-noise ratio, effective pixel ratio, and repeatability requirements of both polarization stations. After the angle is determined, the camera and illumination components are mechanically locked to maintain a fixed working incident angle within the calibration validity period. If the two polarization stations use different incident angles, feature calibration models should be established for the corresponding angles separately, and the difference in polarization features under different incident angles is not directly calculated.

[0040] S12. Triggering and Exposure Control. The four imaging stations use a unified hardware clock reference to record trigger timestamps and trigger at equal material length intervals based on the production line encoder pulses. Since each station is located in a different physical position, each station is not required to acquire the same thickness estimation unit at the same time. Instead, each image frame is bound to the corresponding encoder count, material coordinates, station number, and trigger timestamp.

[0041] The data processing equipment determines the upper limit of the exposure time based on the object pixel size, the allowable material displacement during exposure, and the real-time linear velocity. Specifically, the global shutter exposure time of each camera must be less than or equal to the quotient obtained by dividing the product of a preset tolerance factor and the object pixel physical size by the substrate linear velocity. The tolerance factor, set between 0.1 and 0.3, limits the proportion of the substrate physical displacement relative to the pixel physical size within the exposure window. In one embodiment, a fixed exposure time that meets the highest production line speed requirements is used within the same stable production phase, and the camera's automatic gain is turned off. When it is necessary to adjust the exposure time according to the line speed, the data processing device records the exposure time and gain status of each frame of the image, converts the original light intensity into the corrected light intensity under a unified exposure benchmark, and then extracts features. If the line speed is lower than the preset effective operating limit, the encoder signal is abnormal, or the exposure time cannot simultaneously meet the anti-blurring requirements and the minimum signal-to-noise requirements, the system will mark the corresponding image as invalid.

[0042] S13. Polarization Decoupling and Reconstruction. The imaging modules of the first and second polarization stations employ area array micro-polarization cameras. The image sensor surface of the area array micro-polarization camera integrates an array of polarization filter units. Four adjacent micro-polarization pixels, corresponding to polarization transmission directions of 0°, 45°, 90°, and 135° respectively, collectively form a polarization measurement unit. The polarization image coordinates are the coordinates of the polarization measurement unit, not the coordinates of a single micro-polarization pixel.

[0043] The data processing equipment constructs a four-dimensional column vector from the light intensity observations in four polarization directions, which have been normalized to the exposure time but have not yet had their fixed dark field bias subtracted. The equipment then incorporates a position-dependent instrument response matrix for pixel-level physical compensation decoupling, and reconstructs the three-dimensional polarization Stokes parameters based on the four-dimensional column vector. ; in, This represents the spatial coordinate position of the current polarization measurement unit in the two-dimensional photosensitive array matrix; Represents spatial coordinate position The three-dimensional polarization Stokes parameters obtained by decoupling and recovery are represented by the components of the local total reflected light intensity. , representing the dominant polarization components in the horizontal and vertical directions. And representing the polarization dominance components at 45° and 135°. Form a three-dimensional column vector; Represents spatial coordinate position The generalized inverse of the instrument response matrix at the location; Represents spatial coordinate position The four-dimensional light intensity observation column vector has been normalized by exposure time but has not yet had its fixed bias deducted; This represents a four-dimensional bias column vector at the corresponding location, consisting of dark current, fixed light leakage, and fixed readout bias.

[0044] The position-dependent instrument response matrix is ​​obtained during the offline calibration phase. The data processing equipment utilizes a reference light source capable of outputting multiple known linear polarization states to acquire pixel responses under different known polarization directions and unpolarized states. Based on the correspondence between the known input Stokes parameters and the responses of the four micro-polarization channels, the equipment calculates the instrument response matrix and its generalized inverse matrix for each polarization measurement unit. The data processing equipment uses the decoupled and recovered three-dimensional polarization Stokes parameters to replace the correction light intensity for feature calculation.

[0045] After reconstruction, the data processing equipment was inspected. Whether it exceeds the minimum effective light intensity threshold, which is preset based on the average background noise of the sensor in a dark environment and the preset signal-to-noise ratio requirement. Also check whether the synthesized amplitude of the linearly polarized component is within the allowable error range and does not exceed [a certain value]. Polarization measurement units that fail to meet physical constraints, have any original polarization channel saturated, or have any channel below the noise lower limit are marked as invalid and removed from the subsequent polarization difference statistical process. The noise lower limit is the combined effective value of dark current bias and readout noise.

[0046] S14. Near-infrared transmission imaging wavelength selection, radiometric calibration, and cross-station benchmark unification. Both the first and second transmission stations include a near-infrared backlight and a monochrome camera, using the same near-infrared center wavelength and the same nominal spectral bandwidth. The center wavelength of the near-infrared light source is selected within the range of 780nm to 940nm. The light source, filters, and camera response bands used in both stations should cover the same target spectral range. The target wavelength should avoid overexposure and low signal, have a repeatable response to the target equivalent dry adhesive thickness range, remain stable to normal batch differences in nonwoven fabric, and avoid indistinguishable dominant responses caused by residual moisture changes. Data collected under backlight-free conditions is used to determine the dark current bias, readout noise, and low-end quantization noise of the monochrome camera. When the transmitted light intensity signal after coating is lower than the noise and quantization lower limits, it is judged as a low-signal failure, and the corresponding feature extraction process is terminated.

[0047] Each transmission station acquires a dark-field reference image with the near-infrared light source off, and an unobstructed transmission reference image under stable illumination and without substrate obstruction. Standard references with known transmittance can also be periodically used for verification. The data processing equipment uses the dark-field and unobstructed transmission references to perform dark-field subtraction and normalization, and combines flat-field calibration to compensate for spatial inhomogeneities of the light source and lens vignetting, converting the original grayscale into dimensionless normalized transmittance. The outputs from the first and second transmission stations are only integrated into subsequent cross-station transmission feature calculations after their respective dark-field subtraction, flat-field correction, exposure normalization, and reference transmittance calibration have been completed. Automatic gain control of the cameras at both stations remains off, and the center wavelength, exposure parameters, and reference calibration version are saved along with the image data.

[0048] The data processing equipment integrates the reflected polarization image and near-infrared transmission image obtained from the substrate inspection station and coating inspection station after hardware decoupling and reference correction, providing corrected reflected polarization and near-infrared transmission images for subsequent image registration and calculation. The corrected reflected polarization image refers to the image generated after decoupling the original reflected polarization image from the instrument response matrix and reconstructing it using Stokes parameters; the corrected near-infrared transmission image refers to the image generated after dark field subtraction, flattening correction, exposure normalization, and reference transmittance calibration of the original near-infrared transmission image. In this invention, the corrected reflected polarization image and the corrected near-infrared transmission image are collectively referred to as the corrected image.

[0049] Reference Figure 4 Regarding the material tracking and spatial image registration in step S20, the data processing device performs physical tracking and image alignment based on a real-time path model for the same thickness estimation unit in motion, including the following sub-steps: S21. Establish material coordinates based on the calibration relationship between encoder length and measured tension and elongation. For each nonwoven fabric model and specified tension range, the correspondence between tension and longitudinal elongation ratio is obtained in advance through tensile calibration tests and stored as a calibration curve or segmented lookup table. The data processing equipment calls the corresponding elongation correction value according to the real-time tension, converts the encoder's cumulative length into the cumulative length of the material under the reference tension state, and forms tension-normalized material coordinates.

[0050] Young's modulus is used for simplification only when calibration tests confirm that the material can be approximated linearly within the specified tension range. When the online tension exceeds the calibration range, the rate of tension change exceeds the allowable value, or roller slippage is detected, the corresponding area is marked as a low-confidence tracking state. The encoder cumulative length can be periodically corrected using identifiable natural textures within the uncoated reference edge strip or dedicated positioning marks set within the reserved trimming area. The encoder cumulative length is corrected only when the feature matching results meet the specified number of matching points and residual requirements. The data processing equipment uses the starting point of roll production, the roll changeover joint, or the preset material origin of the current batch as the coordinate zero point, and uses the calculated tension-normalized material coordinates as a unified material coordinate reference in the longitudinal direction of the continuous substrate within the current production batch.

[0051] S22. Coarse tracking of thickness estimation units is completed using material coordinates and a real-time path model. The data processing equipment writes a unique tracking number, tension-normalized material coordinate range, and acquisition timestamp to each thickness estimation unit acquired at the substrate inspection station. The data processing equipment keeps the already determined tension-normalized material coordinate range of the thickness estimation unit unchanged, and predicts the time window for the thickness estimation unit to reach the coating inspection station and the corresponding candidate coverage area of ​​the downstream image based on the fixed path coordinates of each station, the real-time position of the swing roller, and the calibrated internal path difference of the station.

[0052] After images are acquired at the coating inspection station, the data processing equipment establishes a one-to-one upstream and downstream image candidate relationship within the prediction time window, based on the optimal overlap of material coordinates and an error less than the length matching tolerance. The length matching tolerance is set comprehensively based on the physical resolution of the production line encoder and the maximum permissible dynamic elongation error of the medical nonwoven fabric substrate under the current tension. When multiple approximate candidates exist, no candidates exist, or the matching error exceeds the preset length matching tolerance, the corresponding area is marked as coarse tracking failure and prohibited from entering thickness regression.

[0053] S23. Global registration is completed using bilateral reference sidebands, and local deformation correction is performed as needed. After coarse tracking and locking are completed, the data processing device locates the left and right physical edges of the medical nonwoven fabric substrate in the image before adhesive coating based on grayscale or gradient differences, and determines the bilateral reference sidebands before adhesive coating by combining the preset width of the reference sidebands; at the same time, the left and right physical edges and the actual boundary of the coating in the dried image are located to determine the actual uncoated reference sideband after drying, and the actual sideband width is compared with the preset width.

[0054] Substrate edges and coating boundaries can be obtained through grayscale gradient threshold segmentation. When a single threshold segmentation cannot reliably distinguish boundaries, constraints can be imposed by combining edge continuity, preset coating width, and boundary positions in adjacent images. If the reference edge width is less than the allowable value, the reference edge is contaminated by adhesive or impurities and the contaminated area exceeds the preset edge area tolerance limit, the boundary is discontinuous, or the positions of the left and right edge bands are significantly inconsistent with the production formula, the data processing equipment will mark the corresponding thickness estimation unit as a reference edge failure and terminate the spatial registration process.

[0055] After positioning is complete, the data processing equipment uses scale-invariant feature transformation or accelerated robust feature algorithms to extract natural fiber texture features, local corner points, or spot features within the edge band. The data processing equipment performs coordinate matching and alignment of the edge band feature points between the pre-coating image and the post-drying image, and obtains the global two-dimensional affine transformation matrix using a feature matching method with anomaly matching rejection capabilities to compensate for overall translation, rotation, lateral offset, and scale changes.

[0056] When the matching residuals of the two sidebands indicate local non-uniform deformation of the substrate, the data processing equipment also performs local correction using segmented affine mapping, mesh deformation, or other continuous local mapping methods based on the sideband control points obtained segmentally along the conveying direction. Local deformation correction is limited by maximum displacement and deformation continuity to prevent erroneous feature points from causing image folding.

[0057] The data processing equipment uses the material coordinate grid corresponding to the dried image as the target coordinate grid. Based on the global affine transformation and local deformation mapping relationship, it performs inverse resampling on the matching pre-coating image, so that the physical boundaries of the same thickness estimation unit in the pre-coating and post-coating and drying images coincide in the target coordinate grid. The reflected polarization image and the near-infrared transmission image are registered according to the material coordinates and geometric calibration relationship of the corresponding station, respectively. When different modal images need to be converted to the same material grid, the data processing equipment calls the lateral coordinate transformation parameters obtained in advance through the calibration plate.

[0058] Only when the registration residual is less than a preset registration residual threshold, the number of effective matching points reaches a preset matching point number threshold, and the bilateral sideband consistency reaches a preset sideband consistency threshold, can a registered and aligned image sequence suitable for thickness statistics be generated; otherwise, the thickness estimation unit is marked as spatial registration failure. The aforementioned registration residual threshold, matching point number threshold, and sideband consistency threshold are extracted from the statistical distribution of normal samples during the offline registration testing phase and bound to the current process interval for storage.

[0059] Reference Figure 5 Regarding the optical feature extraction and process status acquisition in step S30, after acquiring the registered and aligned image sequence, the data processing device establishes a cross-domain data association from the underlying physical pixels to the macroscopic fluid state, including the following sub-steps: S31. Extract the reflected polarization difference features and perform signal validity filtering. The data processing equipment calculates the local total reflected light intensity component. Horizontal and vertical polarization dominant components and the 45° and 135° polarization dominance components Calculate the degree of linear polarization. To ensure the physical validity of the division operation and avoid amplifying local noise, pixel-level physical anomaly filtering is performed before calculation: overexposed pixels that reach the hardware highlight cutoff threshold in any original polarization channel are removed, and local total reflected light intensity components are removed. Underexposed pixels below the minimum effective light intensity threshold, and polarization measurement units whose reconstructed Stokes parameters do not meet physical consistency requirements. Highlight determination must be based on the maximum response value among the four original polarization channels before decoupling. The relevant high / low light thresholds and consistency tolerances are determined during the instrument calibration phase and are tied to the current camera configuration.

[0060] For the retained data after physical anomaly filtering, the data processing device, based on the reconstructed 3D polarization Stokes parameters, sums the squares of the horizontal and vertical polarization dominance components with the squares of the 45° and 135° polarization dominance components, takes the square root, and then divides by the local total reflected light intensity component. This is used to calculate the degree of linear polarization of the images before and after coating. After calculating the degree of linear polarization, the data processing device subtracts the degree of linear polarization of the corresponding position in the image before coating from the degree of linear polarization of the corresponding position in the image after drying to generate a reflection polarization difference feature matrix. When any corresponding position is marked as invalid, its corresponding reflection polarization difference feature is simultaneously marked as invalid to remove it from subsequent statistics of the thickness estimation unit. The specular highlight and physical consistency filtering steps are used to eliminate the interference of specular highlights and anomalous polarization responses on the reflection polarization difference features.

[0061] Because the surface of the spider silk protein glue has micro-undulations after drying, the local bright reflection usually corresponds to direct specular reflection of light without subsurface scattering. This local bright reflection area has been eliminated by the high-brightness filtering mechanism based on the original channel, thereby ensuring that the output differential feature matrix truly reflects the polarization information inside the material.

[0062] S32. Extract near-infrared attenuation features and perform statistical dimensionality reduction within the thickness estimation unit. The data processing equipment obtains the normalized transmittance before coating from the first transmission station and the normalized transmittance after drying from the second transmission station. Under the conditions that both the normalized transmittance before coating and the normalized transmittance after drying are higher than the minimum effective transmittance, the corresponding original images are not saturated, and spatial registration is effective, the near-infrared attenuation features are calculated according to the following formula: ; in, Represents spatial coordinate position Near-infrared attenuation characteristics at that location; Represents spatial coordinate position Normalized transmittance before coating; Represents spatial coordinate position Normalized transmittance after drying.

[0063] The data processing equipment separately calculates the effective pixel ratio of the reflection polarization difference feature and the near-infrared attenuation feature within each thickness estimation unit. When the effective pixel ratio of any mode is lower than the corresponding feature confidence lower limit threshold, the corresponding mode is marked as feature missing. The feature confidence lower limit threshold is determined based on the robustness boundary test of the thickness prediction model for local feature missing, and is set to 60% to 80% of the total pixels as effective pixels. When the thickness prediction model requires both modes to exist simultaneously, the entire thickness estimation unit is marked as feature missing.

[0064] For modes that pass the validity assessment, the mean and variance are calculated within their respective valid pixel sets. The mean is used to characterize the overall level of the macroscopic equivalent dry adhesive distribution within the thickness estimation unit, while the variance is used to help characterize the optical spatial non-uniformity of the spider silk protein adhesive's penetration and distribution between nonwoven fibers. The combined mean and variance of the data processing equipment constitute an optical feature vector.

[0065] S33. Synchronously collect data on the coating and drying process status and calculate the nominal dry adhesive supply. The data processing equipment shall collect or retrieve at least the volumetric adhesive supply flow rate, adhesive supply pressure, adhesive temperature, adhesive density, adhesive solid content, substrate linear velocity, and actual coating width, and collect at least one of the following factors affecting the final state of the dry adhesive: drying zone temperature, airflow, and exhaust status. The adhesive density and adhesive solid content shall be obtained through offline sampling measurement of the current adhesive batch, or by a verified online density and concentration sensor, and shall be linked to the adhesive batch information.

[0066] For adhesive supply systems with return lines, the net volumetric flow rate entering the coating die is used. When the net volumetric flow rate cannot be directly measured, the pump output flow rate corrected by the effective deposition coefficient calibrated through offline weighing tests is used as the equivalent net flow rate. The data processing equipment calculates the dry matter supply mass per unit time by combining the above net volumetric flow rate or equivalent net flow rate, adhesive density, and solid content. Simultaneously, it calculates the increase in coating area per unit time by combining the substrate linear velocity and the actual coating width. Finally, the nominal dry adhesive supply mass is obtained by dividing the dry matter supply mass per unit time by the increase in coating area per unit time.

[0067] The actual coating width is given by the current production formula and verified using the actual coating boundary identified by the coating inspection station; if it exceeds the allowable deviation, the corresponding process state is marked as out of bounds. When any of the following occurs: linear velocity below the effective calculation lower limit, adhesive supply in a start-stop transition, formula not yet stable, or flow signal failure, the data processing equipment terminates the calculation of the nominal dry adhesive supply and marks the corresponding process state as invalid. The flow signal should also be time-aligned with the corresponding thickness estimation unit according to the actual transmission lag of the adhesive supply pipeline and die head.

[0068] S34. Construct a joint feature set. The data processing equipment uses a pre-determined normalization method for the current model version to process the process status data. The mean, standard deviation, or upper and lower limits used for normalization are obtained from the calibration database and are fixed with the model version. During normal production, the parameters used for normalization remain fixed. Data exceeding the normalization calibration range retains an out-of-bounds flag for reliability assessment and operating condition freezing. The optical feature vector and the process status vector are stored separately and together form a joint feature set for subsequent thickness calculation and operating condition validity assessment.

[0069] Reference Figure 6 Regarding the joint regression calculation and local defect localization in step S40, the data processing equipment, based on the extracted joint feature set, introduces a process perturbation mechanism to perform regression calculation of the target equivalent dry adhesive thickness, and simultaneously outputs the spatial coordinates of the underlying visual defect distribution at the original image resolution, including the following sub-steps: S41. Establish process intervals, calibration samples, and equivalent dry adhesive thickness labels. Each process interval corresponds to at least a defined nonwoven fabric type, adhesive formulation, solid content range, apparent dry adhesive density calibration value, and drying condition range. The calibration database stores the substrate coordinates, dry adhesive mass per unit area, apparent dry adhesive density, equivalent dry adhesive thickness label, optical features, process status, registration quality, and image validity information of the physical samples. The input for the thickness prediction model is the optical feature vector; the process status vector is used for small-range residual compensation, condition out-of-range judgment, and reliability assessment.

[0070] S42. In the offline calibration stage, the thickness prediction coefficient vector is obtained, and the initial equivalent dry adhesive thickness is calculated in the linear inference stage. The data processing device retrieves the optical feature vector containing the mean and variance from the joint feature set, and introduces a constant bias term with a value of 1 to form the input feature vector. In the offline calibration stage, the data processing device divides the calibration samples into model training samples and independent validation samples. The model training samples are used to solve the optimization objective function composed of the weighted residual sum of squares and the regularization penalty term. The sample weights are determined based on the entity weighing uncertainty, process stability, effective pixel ratio, and spatial registration quality, and are limited to a preset range. The regularization parameter is determined through cross-validation within the model training samples. For the specific mathematical solution process of the optimization objective function, those skilled in the art can use partial derivative calculation or gradient descent algorithm. The parameter optimization logic is a well-known technology in this field and will not be elaborated here.

[0071] Once the model parameters are determined, the model is evaluated using independent validation samples that were not involved in model training and regularization parameter selection. Only when the independent validation results meet the preset requirements for average error, maximum error, output stability, and abnormal output ratio are the corresponding thickness prediction coefficient vectors registered as valid model parameters for the current process range. Acceptance requirements are jointly determined by the product thickness tolerance and offline weighing measurement uncertainty, and are saved with the model version.

[0072] During the online linear inference phase, after acquiring the determined thickness prediction coefficient vector, the data processing device determines whether the current input feature vector is within the model calibration feature range. If the current input feature vector exceeds the model calibration feature range, a feature space out-of-bounds flag is generated and retained for subsequent comprehensive confidence assessment; if it does not exceed the range, it is multiplied by the current input feature vector using an inner product operation to calculate the initial equivalent dry adhesive thickness of the thickness estimation unit. The online calculation formula is as follows: ; in, Represents the current thickness estimation unit The initial equivalent dry adhesive thickness; Represents the corresponding process range The thickness prediction coefficient vector; Represents the matrix transpose operator; This represents the input feature vector after introducing a constant bias term.

[0073] S43. Perform constrained process state perturbation correction and calculate the target equivalent dry adhesive thickness. To compensate for non-optically induced coating amount offsets caused by transient physical fluctuations in fluid dynamic parameters such as adhesive temperature and supply pump pressure, as well as other process state parameters, the data processing equipment introduces a process state vector for correction and compensation. The process correction coefficient vector is pre-obtained through the regression relationship between the residual of the entity thickness label minus the thickness prediction model's predicted value and the process state deviation in the calibration database, and confirmed through independent validation samples.

[0074] During online operation, the data processing equipment retrieves the process state vector from the joint feature set, calculates the process state deviation vector between the process state vector and the reference process vector corresponding to the current process interval, and performs an inner product mapping on the calculated process state deviation vector using the process correction coefficient vector. Subsequently, a fixed amplitude boundary truncation constraint is applied to generate a linear compensation amount for correcting the initial equivalent dry adhesive thickness, calculated using the following formula: ; in, Represents the corresponding spatial grid index The linear compensation amount; Represents the upper and lower limit cutoff function; Represents the corresponding process range The pre-set vector of process correction coefficients; Represents the matrix transpose operator; Represents the process state vector; Represents the corresponding process range The reference process vector within; Represents the corresponding process range Maximum perturbation cutoff threshold allowed; This represents the lower boundary limit of the cutoff. The maximum process perturbation cutoff threshold is obtained by combining the statistical analysis of extreme values ​​of historical production line process fluctuations with the sensor noise tolerance. Its value range is set to 5% to 15% of the nominal equivalent dry adhesive thickness in the current process range to prevent excessive correction errors caused by sensor anomalies.

[0075] After generating the linear compensation amount, the initial equivalent dry adhesive thickness is corrected for process perturbation using the linear compensation amount to calculate the target equivalent dry adhesive thickness of the thickness estimation unit under the current spatial grid index. To prevent physical thickness logic errors due to large negative correction magnitudes, the linear compensation amount is only used when the process state is within a verified small perturbation range, the thickness prediction model is reliable, and the relevant process sensors are effective. The data processing device adds the initial equivalent dry adhesive thickness to the linear compensation amount, using this as the target equivalent dry adhesive thickness. The maximum process perturbation truncation threshold and the allowed number of consecutive truncation boundaries or duration are determined during the calibration phase of the corresponding process interval and saved with the model version. When the number of consecutive truncation boundaries reached by the compensation amount exceeds the preset number, or the proportion of truncation boundaries reached within the preset operating condition evaluation time window exceeds the preset upper limit of the truncation proportion, the data processing device marks the corresponding results as process compensation failure and operating condition out of bounds, and stops using the truncated compensation amount for normal thickness output.

[0076] After adding the initial equivalent dry adhesive thickness to the linear compensation amount, if the sum is less than zero, the data processing device directly configures the overall confidence level of the sum to be lower than the confidence threshold used for closed-loop control, sets a physical boundary cutoff flag, and marks the corresponding target equivalent dry adhesive thickness as invalid.

[0077] After completing the calculation of each thickness estimation unit, the data processing equipment generates the target equivalent dry adhesive thickness matrix, the thickness valid state matrix, and the failure cause code matrix according to the spatial grid index. The external control system only uses the target equivalent dry adhesive thickness marked as valid in the valid state matrix.

[0078] S44. Parallel processing of low-level pixel features generates local defect anomaly maps to locate surface imperfections. To compensate for the compression of local spatial distribution details during statistical dimensionality reduction, while performing thickness regression calculations, the data processing device extracts reflection polarization difference feature maps and near-infrared attenuation feature maps in parallel, before and after registration. Pixel-level processing refers to processing at the effective sampling unit scale of the corresponding feature map, calculating the background mean and local standard deviation of local features within the thickness estimation unit to establish a local background model, thereby characterizing optical defects that are not thickness mapping properties.

[0079] In the reflection polarization differential feature, each pixel is judged to determine whether the absolute deviation of the feature value relative to the background mean exceeds a preset feature deviation threshold (i.e., whether the absolute difference is greater than the product of a preset fluctuation judgment factor and the local standard deviation), thereby simultaneously identifying positive and negative anomalies. The fluctuation judgment factor is generally in the range of 2 to 3, and its specific value is set according to the normal distribution confidence interval principle and the allowable false alarm rate on site. After the anomalous pixels are filtered by minimum connected area, aspect ratio, and extension direction, the data processing equipment determines whether the suspected anomalous area extends continuously in adjacent images or adjacent thickness estimation units based on material coordinates. Only when the material coordinate range of the anomalous area is continuous or adjacent, and the lateral position deviation is within the preset position deviation threshold range, is it judged as a suspected stripe anomaly with continuity. The anomaly connectivity judgment is based on the aligned material coordinates.

[0080] Subsequently, the data processing device aggregates adjacent connected anomalous pixels into geometric regions and marks these geometric regions as suspected stripe anomalies. For the connected domain analysis and morphological processing of aggregating adjacent connected anomalous pixels into geometric regions, those skilled in the art can employ region growing methods or morphological dilation and erosion algorithms. The connectivity judgment logic is well-known in the field and will not be elaborated upon here.

[0081] Meanwhile, areas of missed coating are determined based on the similarity between the normalized transmittance before and after coating. The data processing equipment identifies continuous areas in the near-infrared attenuation feature map where the normalized transmittance after drying is still within the corresponding substrate background tolerance range before coating, and the near-infrared attenuation feature is close to zero. These areas are marked as suspected areas of missed coating.

[0082] The data processing equipment generates a local defect anomaly map by combining the material coordinates of suspected stripe anomalies and suspected missed coating areas. When the total area of ​​the defective areas marked on the local defect anomaly map exceeds the preset tolerance area limit as a percentage of the physical area of ​​the current thickness estimation unit, a production line anomaly alarm is triggered to guide manual equipment inspection and cleaning. The tolerance area limit is set between 1% and 5% of the physical area of ​​the current thickness estimation unit, with the specific value determined based on the physical tolerance of downstream medical-grade products for minor local defects and quality inspection standards. The local defect threshold and tolerance area are jointly calibrated using normal samples and manually confirmed defect samples.

[0083] Regarding the reliability assessment and detection data output in step S50, the data processing equipment performs data validity verification and maps the time delay caused by physical transmission and computation to a specific material length domain space, thereby providing a reliable feedback benchmark to the external control system, including the following sub-steps: S51. Calculate the overall confidence level of the unified dimensions and output the valid state. For the target equivalent dry adhesive thickness output by the regression solution, the data processing equipment calculates the overall confidence level. The data processing equipment maps light intensity effectiveness, polarization physical consistency, spatial registration quality, feature space proximity, and near-infrared feature distribution consistency to confidence levels ranging from 0 to 1, with higher values ​​indicating more reliable results. Near-infrared feature distribution consistency is determined by comparing the deviation between the current near-infrared attenuation characteristics and the corresponding process interval calibration distribution, and the deviation is inversely mapped to the near-infrared feature distribution consistency confidence level. The near-infrared feature distribution consistency confidence level is used to determine whether the current near-infrared optical response is within the calibrated range.

[0084] The confidence level of light intensity effectiveness is formed based on the proportion of effective pixels that have not experienced hardware exposure saturation or underexposure masking; the confidence level of polarization physical consistency is formed based on the legal proportion of Stokes parameters that satisfy physical law constraints; the confidence level of spatial registration quality is generated based on the matching residuals of the uncoated reference sidebands on both sides, the proportion of effective matching points, and the consistency mapping of the left and right sides; the confidence level of feature space proximity is formed based on the degree of deviation of the extracted joint feature set vector from the center of the multidimensional spatial distribution of the reference state features in the calibration database (i.e., the regularized feature covariance and Mahalanobis distance); and the confidence level of near-infrared feature distribution consistency is formed based on the deviation of the current transmission features from the corresponding calibration distribution.

[0085] Each confidence level has a normal boundary and a failure boundary. When the corresponding evaluation value is within the normal boundary, the confidence level is close to 1; when it reaches or exceeds the failure boundary, the confidence level is 0; when it is between the normal boundary and the failure boundary, the confidence level is determined according to a preset monotonic mapping relationship. In a specific embodiment, the monotonic mapping relationship adopts a piecewise linear mapping formula: For evaluation indicators where a larger value indicates a better result, when the actual evaluation value lies between the normal boundary and the failure boundary, the corresponding formula for calculating the individual confidence level is: ; For evaluation indicators where smaller values ​​indicate better results, when the actual evaluation value lies between the normal boundary and the failure boundary, the corresponding formula for calculating the individual confidence level is: .

[0086] in, This represents the individual confidence level generated after the piecewise linear formula mapping, and its value ranges from 0 to 1. The specific numerical value of the evaluation quantity actually collected or calculated by the data processing equipment for the current indicator; This represents the pre-defined critical value of the normal boundary in a fully trustworthy state for this indicator. This represents the critical failure boundary value that is preset for this indicator to reach a completely unreliable state. The normal boundary, failure boundary, mapping direction, and mapping parameters are determined by normal calibration samples, manually confirmed abnormal samples, and independent verification samples, and are stored together with the corresponding process range and model version.

[0087] The overall confidence level is the minimum of the confidence levels for light intensity effectiveness, polarization physical consistency, spatial registration quality, feature spatial proximity, and near-infrared feature distribution consistency. The data processing device compares the overall confidence level with a preset confidence threshold. When the overall confidence level is lower than the preset confidence threshold, the target equivalent dry adhesive thickness calculation result of the thickness estimation unit is deemed invalid. At this time, the data processing device outputs a thickness invalidation flag, a failure cause code, and the corresponding material coordinates, while retaining the identified valid target equivalent dry adhesive thickness in the data stream. Simultaneously, when the proportion of valid thickness estimation units within a material control segment is lower than the preset effective control proportion, the external control system suspends closed-loop integral term updates and forces the system into a control freeze or degraded state. The preset effective control proportion is set based on the disturbance rejection requirements of the closed-loop control system, typically ranging from 50% to 70%, while the confidence threshold is set between 0.75 and 0.85, determined jointly by expert evaluation during the offline calibration phase and historical sample statistical distribution.

[0088] S52. Perform detection delay alignment based on unified material coordinates. For the retained target equivalent dry adhesive thickness, the data processing equipment performs reference alignment of the material's spatial position. The data processing equipment uses the same tension-normalized material coordinates as in steps S21 and S22, and a real-time path model including the fixed path coordinates of each station and the real-time position of the oscillating roller, to determine the material coordinates of the thickness estimation unit when it passed through the coating mechanism. The path length between the coating mechanism and the coating detection station includes the fixed path portion and the dynamic path changes caused by the oscillating roller position.

[0089] The system's computational latency is obtained by recording the timestamps of image triggering, image reception, feature calculation, thickness calculation, and data output using a unified hardware clock reference. The data processing equipment reads the historical timestamps of the thickness estimation unit as it passes through the coating mechanism, the image trigger timestamps when it is acquired at the coating inspection station, and the thickness result output timestamps, based on the unique tracking number and tension-normalized material coordinates of the thickness estimation unit.

[0090] The data processing equipment determines the physical transmission delay between the coating mechanism and the coating inspection station based on the real-time path model corresponding to the image acquisition time, and determines the material travel distance during system operation based on the tension-normalized material coordinate increment between the image trigger timestamp and the result output timestamp. The calculation of the material travel distance excludes dynamic path changes already included in the real-time path model. Therefore, the data processing equipment determines the historical coating coordinates, image acquisition coordinates, current material coordinates at the time of detection result output, and corresponding delay time corresponding to the target equivalent dry adhesive thickness. The resolution and timestamp recording error of the unified hardware clock should meet the time accuracy requirements for material coordinate alignment.

[0091] The data processing equipment uses the material delay length to perform coordinate backtracking within the tension-normalized cumulative length numerical sequence, generating historical coating coordinates of the medical nonwoven fabric substrate corresponding to the target equivalent dry adhesive thickness as it passes through the coating mechanism, and the current material coordinates of the detection result. Time alignment is used to indicate which historical coating segment the current detection result originates from and provides a basis for the controller to process pure time delays. Based on the historical coating coordinates and the target equivalent dry adhesive thickness after alignment with the time reference, the data processing equipment divides the continuously moving substrate into material control segments of preset length. The physical length of the material control segment is generally set to 2.0m to 5.0m, determined by a combination of the mechanical span and response time of the production line tension buffer zone.

[0092] S53. Output the detection results and execute closed-loop control with delay constraints and safety degradation. The data processing equipment outputs the effective thickness, overall confidence level, historical coating coordinates, current material coordinates, and status flags to the external control system. The external control system uses historical coating segments that have completed detection and timing alignment as feedback samples, and directly adjusts the glue supply amount of subsequent material segments that have not yet passed through the coating mechanism based on their thickness deviations. The external control system generates and executes incremental proportional-integral-derivative flow update commands to adjust the glue supply pump speed of the coating mechanism. The controller update cycle, proportional gain, and integral gain are tuned according to the total pure time delay from the coating mechanism to the detection station and the response time of the glue supply system.

[0093] The full-width glue supply pump control uses the weighted average value of the effective thickness estimation unit. The controller is equipped with an upper and lower output limit and an integral anti-saturation mechanism. When the glue supply output reaches the upper or lower output limit, the detection data is invalid, or the control is frozen, integral accumulation is paused. After control is restored, integral accumulation is reactivated using a non-disruptive switching method. For the specific control logic of performing closed-loop control of glue supply flow based on the target thickness deviation, those skilled in the art can use a proportional-integral-derivative control algorithm. The specific error tracking is a well-known technique in the field and will not be elaborated here.

[0094] If the coating mechanism is configured with a coating die head that has independent lateral partitions, the laterally independent partition die head will only make lateral fine-tuning when the detection data of each partition is valid and the preset offline influence matrix passes verification, and safety limits will be set for the adjustment amount of each partition, the difference between adjacent partitions, and the change per unit time. The offline influence matrix is ​​used to characterize the hydrodynamic coupling effect of the change in the amount of adhesive supplied to a single partition on the thickness of adjacent lateral regions.

[0095] When the overall confidence level indicator data is abnormal, the production line is in the acceleration / deceleration start / stop phase, a spider silk protein adhesive formula change is performed, a sudden change in tension occurs, or the optical sensor experiences a communication failure, the data processing equipment sends a control freeze signal to the external control system. Upon receiving the control freeze signal, the external control system suspends dynamic error adjustment and integral term response. In the event of a short-term detection failure and stable substrate line speed, the external control system maintains the most recently verified nominal adhesive supply setting per unit area; when the line speed changes, it adjusts the adhesive supply according to the verified feedforward relationship between the line speed and adhesive supply. In the event of an unclear formula, drastic tension changes, flow sensor failure, serious optical system malfunction, or freezing time exceeding the allowable limit, the system issues a shutdown or manual takeover request.

[0096] S54. Error verification is performed using a weighing verification sample with traceable material coordinates. Verification samples are obtained from the reserved sacrificial edge strip, automatic sampling area, or during roll change, and their start and end material coordinates, lateral position, and actual cut area are recorded. The sample is weighed after being placed under predetermined temperature and humidity conditions until its quality stabilizes. The actual dry adhesive mass is obtained by subtracting the corresponding substrate mass from the mass of the verification sample after adhesive application. The corresponding substrate mass can be obtained from upstream uncoated paired samples of the same batch and area, or from the average measured value of the nonwoven fabric basis weight at multiple points in the same batch and the actual cut area. When the substrate basis weight fluctuates significantly, the stable uncoated sample adjacent to the verification sample is used preferentially. After deducting the substrate mass, the dry adhesive mass per unit area is obtained based on the actual area, and then converted into an equivalent dry adhesive thickness label using the apparent dry adhesive density of the corresponding process interval.

[0097] The data processing equipment extracts online thickness results consistent with the coverage area based on the material coordinates of the verification sample. The weighing equipment number, calibration status, sample area, environmental conditions, substrate tare method, and measurement uncertainty are all saved with the label. The online verification error is calculated by subtracting the target equivalent dry adhesive thickness from the equivalent dry adhesive thickness label, and the equivalent dry adhesive thickness label is recorded as a new sample in the calibration database.

[0098] When the online verification error exceeds the model maintenance boundary within multiple consecutive detection cycles, a subsequent troubleshooting and update mechanism is triggered. The model maintenance boundary is set at 2% to 5% of the standard thickness target value, determined by the physical tolerance of downstream medical-grade products for minor local defects and quality inspection standards. After the error exceeds the boundary, the external control system triggers a sequential equipment status check mechanism, checking the light source attenuation status, polarization and transmission module calibration status, nonwoven fabric model and batch, spider silk protein adhesive liquid solid content, adhesive supply fluid pipeline, and drying exhaust actuator in sequence.

[0099] The data processing equipment intercepts unverified model correction commands; newly added validation sample data is divided into parameter update samples and independent validation samples. Historical data is used to update the thickness prediction coefficient vector and process correction coefficient vector in the thickness prediction model, dynamically compensating for systematic thickness drift. After the model update is complete, the error is first checked on validation samples not involved in the update, and then compared in parallel with the existing model in a parallel testing mode. Only if the new model continuously meets the acceptance criteria and does not increase the proportion of abnormal outputs can it replace the existing model; otherwise, the original model is retained and the reason for update failure is recorded. After the model parameter update is completed, the data processing equipment generates and archives a status tracking file in the storage medium. The status tracking file records the material properties, thermodynamic and fluid dynamic process status of the current batch, multimodal raw image data of the update cycle, and reference weighing values.

[0100] While performing credibility assessment and testing data output, the data processing equipment also establishes a dynamic maintenance mechanism to maintain calculation accuracy and system stability, including the following sub-steps: S55. Implement restricted hardware drift monitoring and radiation reference recalibration. To address the attenuation of luminous intensity and luminous efficiency of the illumination source and the thermal drift of the camera's photoelectric sensor, the data processing equipment continuously collects near-infrared transmitted light intensity and visible light reflected light intensity in the uncoated reference edge areas on both sides of the medical nonwoven fabric substrate. The data processing equipment prioritizes monitoring changes in the light source and camera response using stable reference components fixed at the edge of the field of view or periodically accessible to the field of view. The uncoated reference edge data is used only as an auxiliary monitoring quantity, and separate baselines are established according to the nonwoven fabric model and batch to avoid directly interpreting material batch variations as hardware attenuation.

[0101] For area array micro-polarization cameras, gain and dark field changes are monitored independently for the 0°, 45°, 90°, and 135° channels; for near-infrared transmission cameras, dark field, full-scale reference, and spatial flat field changes are monitored separately. Channel-specific correction is only performed when both the fixed reference and sideband auxiliary data indicate the presence of slow hardware drift.

[0102] The data processing equipment extracts light intensity data according to a preset time period. It calculates and updates the time-series moving average of the light intensity data by multiplying the light intensity data of the current acquisition period by a light intensity update smoothing coefficient, adding 1 to the product of the difference between the current light intensity data and the light intensity update smoothing coefficient, and multiplying this product by the moving average of the light intensity data from the previous period. This process eliminates high-frequency noise. The light intensity update smoothing coefficient is set to a range of 0.05 to 0.15, and its specific value is obtained through offline calibration combined with historical noise spectrum statistics.

[0103] When the fixed reference and sideband auxiliary data jointly indicate the existence of slow hardware drift, and the drift amount does not exceed the preset hardware attenuation safety tolerance, the correction coefficients generated by the sub-channels are used to perform low-level numerical compensation on the original light intensity values ​​collected by each channel of the polarization measurement unit and the near-infrared transmitted light intensity signal received by the monochrome camera. Based on the compensated data, the reflection polarization difference features and transmission features are re-extracted, and the high light cutoff threshold is updated synchronously to maintain the consistency of the numerical distribution of the input feature vector.

[0104] Numerical compensation is only permitted within the calibrated drift range. If the slow hardware drift exceeds the preset hardware attenuation safety tolerance, the inter-channel response inconsistency exceeds the preset response consistency threshold, or the reference characteristic distribution cannot be restored after compensation, the data processing equipment will stop outputting the corresponding thickness and send a maintenance and recalibration request. The response consistency threshold is determined by the inter-channel background difference tolerance during instrument factory calibration. The hardware attenuation safety tolerance is set to a fluctuation range of 10% to 15% of the absolute background light intensity, determined by a combination of the illumination source lifetime curve and sensor factory parameters.

[0105] S56. Incremental updates are performed while maintaining constraints based on historical parameters. Only after systematic thickness deviations are detected and investigations of hardware, material batches, adhesive solid content, adhesive supply lines, and drying conditions are completed can the data processing equipment initiate model updates. Updated data consists of newly added validation sample pieces and representative samples selected from the historical calibration database. The optimization process reduces prediction errors from both new and historical samples while constraining the parameter drift between the new and existing parameters.

[0106] An incremental optimization objective function containing historical parameter preservation constraints is constructed in the corresponding calibration database to incorporate the data used in this parameter update. The calculation formula is as follows: ; in, This represents the thickness prediction coefficient vector generated after incremental learning. Operators that solve for the independent variable that minimize the objective function; The independent variables representing the regression coefficients in the numerical optimization process; This represents the summation operator over all samples involved in the update, where all samples involved in the update include newly added validation sample samples and representative samples selected from the historical calibration database; This represents the total number of samples participating in this parameter update; Representing the The weight coefficient corresponding to each sample is inversely proportional to the variance of the fluid state parameters at the time of sampling, so as to reduce the weight of unstable condition samples. Representing the The equivalent dry adhesive thickness label corresponding to each sample; Represents the matrix transpose operator; Representing the The input feature vector corresponding to each sample; The default parameter represents the penalty coefficient, which is set to a value range of 0.1 to 0.5, and is preset through cross-validation and empirical values. This represents the vector of historical thickness prediction coefficients that the system was using before this update operation; The square of the L2 norm represents the difference between the independent variables of the regression coefficients and the vector of historical thickness prediction coefficients.

[0107] The thickness prediction coefficient vector and the process correction coefficient vector are updated separately. The thickness prediction coefficient vector corresponds only to the bias term and optical characteristic coefficients of the thickness prediction model; the process correction coefficient vector is separately calibrated based on the relationship between the thickness residual and the process state deviation. After the new parameters are generated, the data processing equipment saves the old parameter version and compares it using independent verification samples and parallel test data. Only when the overall error, maximum error, abnormal output ratio, and control stability of the new parameters all meet the acceptance criteria within the preset verification period can the system switch to the official version. If an increase in error, a rise in the abnormality rate, or control oscillation occurs after the switch, the system automatically rolls back to the previous valid version.

[0108] S57. Establish candidate process intervals and activate them after calibration and verification. Due to seasonal changes in the workshop's base temperature or batch changes in the rheological properties of spider silk adhesive, the real-time ambient temperature and adhesive pump pressure during actual production line operation may deviate from the baseline process vector corresponding to the originally defined process intervals. When the process status continuously exceeds the current process interval within a preset time window, but remains within the physical safety range allowed by equipment and materials, and sensor malfunctions, fluid pipeline blockages, formulation errors, and drying anomalies have been ruled out, the data processing equipment registers the current deviation as a candidate new process interval. The preset time window is set to a continuous 24-hour production cycle, determined based on the daily temperature variation cycle of the workshop and the fluid thermal equilibrium time.

[0109] After the candidate process intervals are established, the system collects multimodal images, process parameters, and physical weighing labels covering the state range of the candidate process intervals in the debugging batch, verification batch, or reserved non-product material segment. Variable parameter tests must be subject to equipment safety boundaries, product isolation rules, and manual approval. After reaching the predetermined sample size and covering the main state range of the candidate intervals, candidate baseline process vectors, feature distributions, thickness prediction models, and process correction coefficient vectors are established respectively. For the specific iteration and averaging logic of calculating the cluster centers in the multidimensional feature space, those skilled in the art can use the K-means algorithm; its data dimension aggregation rules are well-known in the field and will not be elaborated here.

[0110] Candidate process intervals are only registered as official process intervals after passing independent sample verification and confirming that thickness error, reliability judgment, and control stability meet acceptance requirements. Before verification is completed, the system uses the safety feedforward mode, manual setting mode, or stops production, all permitted by the original valid process intervals.

[0111] Each time a process interval is established, a model is updated, optical recalibration is performed, or parameters are rolled back, the data processing equipment generates a versioned operation and maintenance file. The file includes at least the version number, trigger reason, approval record, material and adhesive batch number, hardware calibration version, training and validation samples, model parameters, validation results, and rollback pointer, and is written to a storage medium with append-only and tamper-proof capabilities. The generation of this file data establishes a complete quality traceability chain and provides data recording support for subsequent process status rollbacks.

[0112] See attached document Figure 7 and Figure 8 This paper presents a specific application example of the online detection system and method for spider silk protein coating thickness based on machine vision according to the present invention.

[0113] For the production process of medical spider silk wound dressings, the medical nonwoven fabric base uses a polyester fiber mesh structure with a width of 600mm. The current production formula for the spider silk adhesive has a solid content of 15% and a standardized apparent dry adhesive density of 1.25g / cm³. 3 The effective coating width of the coating mechanism is set to 550 mm, with continuous uncoated reference strips of 25 mm width retained on both sides of the substrate in the transverse direction. The rated linear speed of the production line is set to 30 m / min. The data processing equipment divides the medical nonwoven fabric substrate into fixed physical regions of 20 mm × 20 mm, with each fixed physical region defined as a thickness estimation unit. The standard thickness target value is set to 45 μm.

[0114] During the multimodal image acquisition and hardware calibration phase, the working incident angle of the reflective polarization imaging system was determined to be 45°. The global shutter exposure time of the area array micro-polarization camera used in the first and second polarization stations was set to 200 μs. The center wavelength of the near-infrared light source in both the first and second transmission stations was uniformly set to 850 nm, with a spectral bandwidth of 20 nm. After the data processing equipment acquired the images, it used the position-correlation instrument response matrix obtained in the offline calibration phase to compensate for the light intensity of the original polarization channel, and reconstructed the three-dimensional polarization Stokes parameters based on the compensated values. The calculation formula is as follows: .

[0115] Physical anomaly filtering is performed on the reconstructed three-dimensional polarization Stokes parameters. Polarization measurement units that do not meet physical constraints, have saturated any original polarization channel, or have any channel below the noise lower limit are marked as invalid and removed from the subsequent polarization difference statistical process. The noise lower limit is taken as the effective value of the combined dark current bias and readout noise.

[0116] In the material tracking and spatial image registration stage, an absolute encoder positioned at the end of the guide roller shaft acquires cumulative length pulses. Combined with the real-time absolute tension value measured by a tension sensor, the linear tension-elongation calibration curve of the corresponding polyester medical nonwoven fabric substrate is retrieved to calculate tension-normalized material coordinates. After downstream images are acquired at the coating inspection station, data processing equipment extracts the natural fiber texture features within a 25mm wide uncoated reference side strip. Feature matching algorithms align the side strip feature points before and after coating, and a global two-dimensional affine transformation matrix incorporating translation, rotation, and scaling parameters is calculated.

[0117] During the optical feature extraction and process status acquisition stages, the data processing equipment, based on the reconstructed three-dimensional polarization Stokes parameters, calculates the degree of linear polarization of the pre-coating and post-drying images within the thickness estimation unit, and subtracts the two to generate a reflection polarization difference feature matrix. Simultaneously, the near-infrared attenuation characteristics are calculated using the normalized transmittance before coating and after drying output from the first and second transmission stations, as shown in the following formula: ; in, Represents spatial coordinate position Near-infrared attenuation characteristics at that location; Represents spatial coordinate position Normalized transmittance before coating; Represents spatial coordinate position Normalized transmittance after drying.

[0118] The mean and variance of the reflection polarization difference characteristics and near-infrared attenuation characteristics within the thickness estimation unit are statistically analyzed and combined to form an optical feature vector. Simultaneously, the net volumetric flow rate or equivalent net flow rate of the adhesive supply system, the absolute pressure of the adhesive supply pipeline, and the current substrate linear velocity are collected. Combined with the 15% solid content set in the current production formula, the adhesive density of the current batch, and the 550mm effective coating width, the ratio of dry matter supply mass per unit time to the newly added coating area per unit time is calculated to obtain the nominal dry adhesive supply. Subsequently, the above parameters, including the nominal dry adhesive supply, are uniformly evaluated for validity and normalized to generate the corresponding process state vector.

[0119] During the joint regression solution and local defect localization stage, the data processing equipment inputs an input feature vector containing constant bias terms, mean, and variance into the thickness prediction model for the current process range. An inner product multiplication operation is performed between the input feature vector and the offline-validated thickness prediction coefficient vector to obtain the initial equivalent dry adhesive thickness. The online calculation formula is as follows: .

[0120] Calculate the process state deviation vector between the current process state vector and the baseline process vector. Then, use a preset process correction coefficient vector to perform an inner product mapping on the process state deviation vector to generate a linear compensation amount. Set the maximum process perturbation truncation threshold for the current process interval. The value is 4.5 μm. After applying a fixed amplitude boundary truncation limit to the linear compensation amount (the calculation formula is as follows): ; in, Represents the corresponding spatial grid index The linear compensation amount; Represents the upper and lower limit cutoff function; Represents the corresponding process range The pre-set vector of process correction coefficients; Represents the matrix transpose operator; Represents the process state vector; Represents the corresponding process range The reference process vector within; Represents the corresponding process range Maximum perturbation cutoff threshold allowed; This represents the lower boundary constraint that is truncated. Adding this to the initial equivalent dry adhesive thickness, the target equivalent dry adhesive thickness is calculated.

[0121] Simultaneously, before statistical dimensionality reduction, the local standard deviation within the thickness estimation unit is calculated based on the differential polarization feature map after registration. Pixel by pixel, the absolute deviation of the feature value relative to the background mean is judged to see if it exceeds the preset feature deviation threshold. When the material coordinate range of the abnormal area is continuous or adjacent, and the lateral position deviation is within the preset position deviation threshold range, adjacent connected abnormal pixels are aggregated into a geometric region, and the geometric region is marked as a suspected stripe anomaly. In the near-infrared attenuation feature map, continuous areas where the normalized transmittance after drying is still within the corresponding substrate background tolerance range before coating, and the near-infrared attenuation feature is close to zero are identified and marked as suspected uncoated areas.

[0122] In the confidence assessment and test data output stage, the data processing equipment calculates multiple confidence components and generates a comprehensive confidence score. When the comprehensive confidence score is greater than the preset confidence threshold of 0.80, the test result is considered valid. Based on the fixed path distance from the coating mechanism to the coating inspection station, the dynamic path changes caused by the position of the swing roller, and the tension-normalized material coordinate increments generated during operation, the retained target equivalent dry adhesive thickness is spatially and temporally aligned. The external control system receives the aligned test results, updates the integral term of the proportional-integral control law using the weighted average deviation of the effective thickness estimation unit, and adjusts the actual speed of the adhesive pump to maintain the standard thickness target value of 45μm.

[0123] When the online verification error obtained by subtracting the output target equivalent dry adhesive thickness from the equivalent dry adhesive thickness measured by physical cutting and weighing exceeds the 2.0μm model maintenance boundary for three consecutive cycles, the parameter incremental update mechanism is activated. The thickness prediction coefficient vector is recalibrated using an incremental optimization objective function that incorporates historical parameter constraints. The formula for calculating the objective function is as follows: ; in, This represents the thickness prediction coefficient vector generated after incremental learning. Operators that solve for the independent variable that minimize the objective function; The independent variables representing the regression coefficients in the numerical optimization process; This represents the summation operator over all samples involved in the update. This represents the total number of samples participating in this parameter update; Representing the The weight coefficient corresponding to each sample is inversely proportional to the variance of the fluid state parameters at the time of sampling, so as to reduce the weight of unstable condition samples. Representing the The equivalent dry adhesive thickness label corresponding to each sample; Represents the matrix transpose operator; Representing the The input feature vector corresponding to each sample; The default parameter represents the penalty coefficient, which is set to a value range of 0.1 to 0.5, and is preset through cross-validation and empirical values. This represents the vector of historical thickness prediction coefficients that the system was using before this update operation; The square of the L2 norm represents the difference between the independent variables of the regression coefficients and the vector of historical thickness prediction coefficients.

[0124] in conclusion: Depend on Figure 7 The data distribution characteristics indicate that, for the data scatter points corresponding to the randomly selected verification samples, the predicted thickness output by the system is closely distributed on both sides of the ideal baseline where the predicted value is exactly equal to the weighing label, and all data scatter points fall within the allowable tolerance zone area composed of light gray polygonal bands (i.e., the set ±2.0μm strict tolerance boundary). The high degree of aggregation of the prediction results is related to the system's extraction and decoupling of the reflection polarization difference features, combined with the near-infrared transmitted light intensity signal to perform statistical dimensionality reduction characterization, and the introduction of process state vectors for micro-perturbation residual correction, ensuring that the coating prediction model continuously outputs high-precision results under micro-deformation conditions.

[0125] Depend on Figure 8 The measured operating curves show that in the first section marked as dark gray, the process transition and fluctuation area (specifically the open-loop debugging area for the first 20 cycles), the coating thickness deviates significantly due to fluctuations in fluid pumping pressure. When the system starts the closed-loop control logic in the 20th sampling cycle, the external control system dynamically adjusts the glue supply pump speed based on the alignment detection results from the high-frequency output of the machine, so that the measured operating curve and its contained data points converge quickly and stabilize within the allowable tolerance zone marked by the light gray background (i.e., near the target thickness setting line of 45μm).

[0126] Even if an external process disturbance is artificially introduced during the 60th sampling cycle, causing the system to enter the second phase—marked in dark gray—as a process transition and fluctuation region (specifically, the disturbance recovery zone), the system can still quickly identify deviations and perform pre-adjustment of the adhesive supply by relying on sensitive multimodal visual detection and a low-latency spatial-temporal alignment mechanism. After the controlled material crosses the pure hysteresis transmission segment, the actual thickness is re-constrained within the allowable tolerance zone. This process completely eliminates the control oscillations and integral saturation effects caused by the long-span physical transmission from the coating execution end to the optical detection end, maintaining the long-term control continuity and quality traceability of the adhesive coating process parameters in the medical dressing production line.

[0127] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result using substantially the same method falls within the scope of the present invention.

Claims

1. A machine vision-based online detection method for spider silk protein coating thickness, characterized in that, Includes the following steps: Multimodal images of spider silk protein before and after coating and drying were acquired, and polarization Stokes parameters were reconstructed to obtain corrected images; Tension-normalized material coordinates are established to determine the material coordinates corresponding to the corrected image, and a registration image sequence is generated based on the material coordinates; The optical feature vector is generated by extracting reflection polarization difference features and near-infrared attenuation features from the registered image sequence, and the coating process parameters and drying process parameters are collected simultaneously to generate a process state vector. The optical feature vector is input into the thickness prediction model to obtain the initial equivalent dry adhesive thickness. Based on the process state vector, a linear compensation amount is calculated to perform residual correction on the initial equivalent dry adhesive thickness, thereby obtaining the target equivalent dry adhesive thickness. Calculate the overall confidence level for the target equivalent dry adhesive thickness. When the overall confidence level reaches the preset confidence threshold, output the target equivalent dry adhesive thickness; otherwise, output the thickness invalidation flag, failure reason code, and material coordinates.

2. The online detection method for spider silk protein coating thickness based on machine vision according to claim 1, characterized in that, Multimodal images of spider silk proteins were acquired before and after coating and drying, and polarization Stokes parameters were reconstructed to obtain corrected images, including: The reflected polarization image in the multimodal image is acquired by a micro-polarization camera. The light intensity observation values ​​belonging to the four polarization directions, which have been normalized by the exposure time but have not yet had their fixed dark field bias deducted, are formed into a four-dimensional column vector. The position-related instrument response matrix is ​​introduced to perform pixel-level physical compensation decoupling. The three-dimensional polarization Stokes parameters are reconstructed based on the four-dimensional column vector. The three-dimensional polarization Stokes parameters include the local total reflected light intensity component, the polarization dominance components in the horizontal and vertical directions, and the polarization dominance components at 45° and 135°. The three-dimensional polarization Stokes parameter obtained by decoupling and recovery is equal to the generalized inverse of the instrument response matrix at the spatial coordinate position multiplied by the difference between the four-dimensional light intensity observation column vector at the spatial coordinate position (normalized by exposure time but before deducting fixed bias) and the four-dimensional bias column vector at the corresponding position (composed of dark current, fixed light leakage, and fixed readout bias). Based on the reconstructed three-dimensional polarization Stokes parameters, a corrected reflected polarization image is obtained, and the corrected reflected polarization image and the corrected near-infrared transmission image are integrated to generate the corrected image.

3. The online detection method for spider silk protein coating thickness based on machine vision according to claim 1, characterized in that, Establishing tension-normalized material coordinates to determine the material coordinates corresponding to the corrected image, and generating the registered image sequence based on the material coordinates, includes: The tension-normalized material coordinates are established by receiving the production line encoder count, substrate linear speed, tension sensor signal and swing roller position signal. The material coordinates corresponding to each image are determined by using the tension-normalized material coordinates, the conveying direction coordinates of each imaging station relative to the reference position and the dynamic path change caused by the swing roller. The medical nonwoven fabric substrate is divided into multiple fixed physical regions, and each fixed physical region is defined as a thickness estimation unit; Based on the material coordinates, a one-to-one candidate relationship is established for the same thickness estimation unit before and after adhesive application and drying, and the registration image sequence that satisfies the physical region overlap requirement of the thickness estimation unit is generated according to the one-to-one candidate relationship.

4. The online detection method for spider silk protein coating thickness based on machine vision according to claim 3, characterized in that, Based on the tension sensor signal, the tension and elongation relationship of the nonwoven fabric model is called to establish the tension-normalized material coordinates. After the material coarse tracking is completed by combining the dynamic path change caused by the swing roller position signal, the global two-dimensional affine transformation is obtained by using the uncoated reference side strips on both sides. When the registration residual shows local non-uniform deformation, a limited local deformation correction is performed to generate the registration image sequence that meets the physical region overlap requirements of the thickness estimation unit.

5. The online detection method for spider silk protein coating thickness based on machine vision according to claim 2, characterized in that, Extracting reflection polarization difference features from the registered image sequence includes: Perform pixel-level physical anomaly filtering, remove overexposed pixels where any original polarization channel reaches the hardware highlight cutoff threshold, remove underexposed pixels where the local total reflected light intensity component is lower than the minimum effective light intensity threshold, and reconstruct polarization measurement units where the Stokes parameters do not meet the physical consistency requirements, and mark the location of the removed pixels as invalid. Based on the reconstructed three-dimensional polarization Stokes parameters, the square root of the sum of the squares of the horizontal and vertical polarization dominance components and the squares of the 45° and 135° polarization dominance components is taken and then divided by the local total reflected light intensity component. In this way, the linear polarization degree of the corresponding positions of the image before coating and the image after drying in the registered image sequence is calculated respectively. The linear polarization degree of the corresponding position in the image before coating is subtracted from the linear polarization degree of the corresponding position in the dried image to generate a reflection polarization difference feature matrix. When any corresponding position is marked as invalid, the corresponding reflection polarization difference feature is simultaneously marked as invalid, and the valid features in the reflection polarization difference feature matrix are used to construct the optical feature vector.

6. The online detection method for spider silk protein coating thickness based on machine vision according to claim 1, characterized in that, Extracting near-infrared attenuation features from the registered image sequence includes: The normalized transmittance before coating and the normalized transmittance after drying of the image before coating and the image after drying in the registered image sequence are obtained respectively. Under the condition that the normalized transmittance before coating and the normalized transmittance after drying are both higher than the minimum effective transmittance, the corresponding original images are not saturated and the spatial registration is effective, the near-infrared attenuation characteristics are calculated. Wherein, the near-infrared attenuation characteristic at the spatial coordinate position is equal to the natural logarithm of the quotient obtained by dividing the normalized transmittance before coating at the spatial coordinate position by the normalized transmittance after drying at the spatial coordinate position. Calculate the mean and variance of the near-infrared attenuation characteristics respectively; The mean of the near-infrared attenuation feature and the variance of the near-infrared attenuation feature are combined to form the optical feature vector.

7. The online detection method for spider silk protein coating thickness based on machine vision according to claim 6, characterized in that, The optical feature vector is input into the thickness prediction model to obtain the initial equivalent dry adhesive thickness, including: The optical feature vector containing the mean of the near-infrared attenuation feature and the variance of the near-infrared attenuation feature is retrieved, and a constant bias term with a value of 1 is introduced to combine them to form the input feature vector of the thickness prediction model. After obtaining the determined thickness prediction coefficient vector corresponding to the thickness prediction model, the thickness prediction coefficient vector is multiplied by the currently input feature vector using an inner product operation, and this is used as the output of the thickness prediction model to calculate and generate the initial equivalent dry adhesive thickness.

8. The online detection method for spider silk protein coating thickness based on machine vision according to claim 1, characterized in that, Based on the process state vector, a linear compensation amount is calculated to perform residual correction on the initial equivalent dry adhesive thickness, resulting in the target equivalent dry adhesive thickness, including: Calculate the process state deviation vector between the process state vector and the reference process vector corresponding to the current process interval, perform inner product mapping on the calculated process state deviation vector using the process correction coefficient vector, apply a fixed amplitude boundary truncation limit, and generate the linear compensation amount. Wherein, the linear compensation amount is equal to the value obtained by multiplying the transpose of the process correction coefficient vector with the process state deviation vector and then applying an upper and lower limit truncation function composed of the maximum process perturbation truncation threshold and the lower truncation boundary limit. The initial equivalent dry adhesive thickness is added to the linear compensation amount to obtain the target equivalent dry adhesive thickness.

9. The online detection method for spider silk protein coating thickness based on machine vision according to claim 1, characterized in that, The overall confidence level for calculating the target equivalent dry adhesive thickness includes: The light intensity effectiveness, polarization physical consistency, spatial registration quality, feature space proximity, and near-infrared feature distribution consistency corresponding to the generation of the target equivalent dry adhesive thickness are calculated respectively. The five indicators are mapped to confidence levels in the range of 0 to 1, with higher values ​​indicating higher confidence. The minimum value of the light intensity effectiveness confidence level, polarization physical consistency confidence level, spatial registration quality confidence level, feature space proximity confidence level, and near-infrared feature distribution consistency confidence level is taken to form the comprehensive confidence level.

10. The online detection method for spider silk protein coating thickness based on machine vision according to claim 7, characterized in that, Also includes: An incremental optimization objective function containing historical parameter preservation constraints is constructed in the calibration database, retaining the historical parameter preservation constraints and incorporating the newly added verification sample used in this parameter update; The incremental optimization objective function is set to minimize the sum of the weighted residual squares corresponding to the independent variables and the historical parameter retention constraint penalty term. The weighted residual squares are obtained by subtracting the square of the prediction result of the inner product of the input feature vector and the regression coefficient independent variables from the equivalent dry glue thickness labels corresponding to all samples participating in the update, and then multiplying it by the corresponding weight coefficient and summing them up. The historical parameter retention constraint penalty term is equal to the pre-set parameter retention penalty coefficient multiplied by the square of the L2 norm of the parameter vector difference between the regression coefficient independent variable and the historical thickness prediction coefficient vector, wherein the historical thickness prediction coefficient vector is the thickness prediction coefficient vector before the update, and the updated thickness prediction coefficient vector is obtained by solving the incremental optimization objective function.

Images