A method, equipment, and medium for detecting foreign objects on fiberglass cloth.

Abstract

This invention discloses a method, device, and medium for foreign object detection on a glass fiber fabric surface, relating to the field of intelligent sensing technology. The method includes setting a lateral range based on a calibration value of the scanning lateral coverage width, extracting the minimum and maximum longitudinal coordinates of the update nodes, constructing reconstruction points, building a continuous foreign object state field on the reconstruction points using a benchmark difference, extracting a time window sequence based on the continuous foreign object state field, constructing an optimization objective function and performing iterative optimization to obtain the final order and diffusion intensity of each reconstruction point, and calculating the residuals. By introducing a time-series discrimination method based on fractional-order dynamic parameter inversion, transient disturbances and real foreign object events are effectively distinguished, reducing the false alarm rate and improving the stability of foreign object type determination. Furthermore, by using a stable interval determination mechanism based on the instantaneous velocity sequence of the fabric surface, the spatiotemporal consistency and reliability of foreign object detection results under high-speed production conditions are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent sensing technology, and in particular to a method, device and medium for detecting foreign objects on a glass fiber cloth surface. Background Technology

[0002] With the increasing demands for manufacturing precision and product consistency, foreign object detection technology for fiberglass fabrics focuses on surface defect detection methods based on visible light imaging, infrared imaging, or single-spectral information. It acquires fabric images using linear or area array cameras, and then combines threshold segmentation, texture feature extraction, or machine learning models to identify foreign objects. Furthermore, it introduces spectral sensors or polarization imaging devices to enhance the perception of differences in material composition or surface reflectivity, expanding the detection scope from simple geometric anomalies to material property anomalies.

[0003] However, existing technologies still have shortcomings. They rely on static thresholds or short-term statistical characteristics for judgment, failing to model the diffusion, retention, or penetration behavior of foreign objects over time. This makes it difficult to distinguish between transient disturbances and real foreign object events, resulting in high false alarm and false negative rates. Existing detection methods assume that the fabric surface is in a constant state and do not systematically model the instantaneous velocity fluctuations of the fabric surface and their impact on sampling timing and spatial mapping accuracy. This can easily introduce spatial mismatch errors during acceleration, deceleration, or mechanical disturbance phases, leading to foreign object positioning deviations. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method, device, and medium for detecting foreign objects on fiberglass fabric, which solves the problem that existing technologies rely on static thresholds or short-term statistical characteristics for judgment, failing to model the diffusion, retention, or penetration behavior of foreign objects over time, making it difficult to distinguish between transient disturbances and real foreign object events, resulting in high false alarm and false negative rates. Existing detection methods assume that the fabric surface is in a constant operating state and do not systematically model the instantaneous velocity fluctuations of the fabric surface and their impact on sampling timing and spatial mapping accuracy, which easily introduces spatial mismatch errors during acceleration, deceleration, or mechanical disturbance stages, leading to foreign object positioning deviation problems.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for detecting foreign objects on a glass fiber cloth surface, comprising,

[0008] The sampling time interval is obtained by configuring the frame rate of the spectrometer, the raw outputs of the spectral and polarization channels are collected, and a unified sensitivity is obtained by weighted summation. A continuous reference field is constructed, and the update node is set based on the scanning lateral coverage width and factory configuration. The raw output of the polarization channel is assigned to the update node, the observation value is calculated and subtracted from the reference value of the continuous reference field to obtain the reference difference.

[0009] The horizontal range is set based on the calibration value of the horizontal coverage width of the scan, and the minimum and maximum vertical coordinates of the update node are extracted to construct the reconstruction point. Combined with the benchmark difference, a continuous foreign object state field is constructed on the reconstruction point. The time window sequence is extracted based on the continuous foreign object state field, the optimization objective function is constructed and iterative optimization is performed to obtain the final order and diffusion intensity of each reconstruction point, and the residual is calculated.

[0010] The residuals are compared to generate contact indications, the cumulative contact time is calculated, the presence of foreign objects is determined, and the feature vector at the first trigger moment is extracted. The foreign object category label is obtained through K-means clustering and uploaded to the cloud platform for display.

[0011] As a preferred embodiment of the foreign object detection method for glass fiber fabric described in this invention, the method for constructing reconstruction points, combined with reference difference, constructs a continuous foreign object state field at the reconstruction points, including:

[0012] The horizontal range is set based on the calibration value of the horizontal coverage width of the scan, and the minimum and maximum vertical coordinates of the updated nodes are extracted to define the reconstructed points and construct a set of neighborhood points.

[0013] Based on the set of neighborhood points, a weight scalar is constructed for each neighborhood point, and a weight diagonal matrix is ​​created.

[0014] Subtracting the horizontal and vertical coordinates of the reconstructed point from the horizontal and vertical coordinates of the neighboring points respectively, we obtain the relative horizontal and vertical coordinates, construct a quadratic basis vector, and stack all the quadratic basis vectors column by column to obtain the polynomial basis matrix.

[0015] Construct the linear system matrix based on the polynomial basis matrix and the weighted diagonal matrix;

[0016] The coefficient column vectors of the linear system matrix are solved to obtain the reconstructed foreign object state field scalar. The reconstructed foreign object state field scalar is then arranged according to coordinates to obtain the continuous foreign object state field.

[0017] As a preferred embodiment of the foreign object detection method for glass fiber fabric described in this invention, the step of extracting a time window sequence based on a continuous foreign object state field includes:

[0018] For each reconstruction point, a window sequence is extracted based on the continuous foreign object state field. The mean within the time window length is calculated using the arithmetic mean method for each reconstruction point. The window sequence is then demeaned to obtain the demeaned sequence.

[0019] For each candidate order of the demeaned sequence, fractional Fourier transform coefficients are calculated using fractional Fourier transform. The energy of the fractional Fourier transform coefficients is calculated using the complex modulus square and then normalized to obtain normalized energy. The normalized energy is then arranged horizontally to obtain a discrete set of values ​​with a length equal to the window length.

[0020] As a preferred embodiment of the foreign object detection method for glass fiber fabric described in this invention, the method of constructing and optimizing the objective function and performing iterative optimization to obtain the final order and diffusion intensity of each reconstruction point includes:

[0021] The arithmetic mean and unbiased sample variance method are used to extract the mean and variance of the discrete value set, construct the optimization objective function, and obtain the probability mass and Lagrange function values.

[0022] The probability mass is iteratively updated using the Newton-Raphson method to obtain the updated probability mass, and the updated Lagrange function value is recalculated using the updated probability mass.

[0023] The updated Lagrange function values ​​are accumulated to obtain the cumulative distribution function value. The tail residual is calculated, the logarithm of the tail residual is taken, and the slope is obtained by fitting it with a least squares straight line. The slope is mapped to fractional order candidates, and the fractional order candidates are pruned to obtain the final order.

[0024] As a preferred embodiment of the foreign object detection method for glass fiber fabric described in this invention, the calculation of residuals includes:

[0025] Based on the final order and the coordinates of the reconstructed points, a continuous field of order is constructed using the moving least squares method.

[0026] Based on historical experimental experience, minimum diffusion intensity, maximum diffusion intensity, and discrete step size of diffusion intensity are set and arranged horizontally to obtain a candidate set. The candidate set is then input into a fractional-order discrete predictor to obtain discrete sampling times. The predicted value;

[0027] Based on the continuous field, the Lagrangian function values ​​of the final order and the optimal fractional domain order of each reconstruction point are extracted. The weighted error is calculated by combining the predicted values, the final diffusion intensity is constructed, and the intensity continuous field is constructed by combining the coordinates of the reconstruction points and using the moving least squares method.

[0028] Based on the intensity continuous field, the final diffusion intensity of each reconstruction point is extracted, and combined with the final order input to a fractional-order discrete predictor to obtain the discrete sampling time. The predicted values ​​are used to calculate the residuals.

[0029] As a preferred embodiment of the foreign object detection method for glass fiber fabric described in this invention, the step of obtaining foreign object category labels through K-means clustering and uploading them to a cloud platform for display includes:

[0030] If the absolute value of the residual is less than the residual threshold, the contact indication is set to 1; otherwise, it is set to 0. The cumulative contact time is calculated. If the cumulative contact time is greater than or equal to the trigger threshold, a foreign object is detected and marked as 1; otherwise, no foreign object is detected and marked as 0. (The last part, "if discrete sampling time...", appears to be an error and doesn't translate directly.) and Then Marked as the first trigger time;

[0031] The absolute values ​​of cumulative contact time, final order, final diffusion intensity, and material response offset are arranged horizontally and normalized to obtain the feature vector at the first trigger moment. The feature vector is then input into K-means for clustering to obtain the foreign object category label.

[0032] All coordinates marked as 1 are used to extract the boundaries of the foreign object region using the connected component extraction method, and the area is calculated using the polygon area calculation method.

[0033] The boundaries, area, and category labels of the foreign object area are uploaded to the cloud platform via API for display.

[0034] As a preferred embodiment of the foreign object detection method for glass fiber cloth surface described in this invention, the step of calculating the observed value and subtracting the reference value from the reference value of the continuous reference field to obtain the reference difference includes:

[0035] Obtain the horizontal coverage width of the scan line, and construct a node set based on the factory configuration setting for the number of horizontal nodes;

[0036] Based on the instantaneous velocity and sampling time interval in the instantaneous velocity sequence, the periodic displacement is obtained through matrix integration, and the node set is updated to obtain the updated node set;

[0037] The two polarization intensities in the original output are assigned to the update nodes in each update node set. For the coordinates of each update node, the reference value of the corresponding coordinate is extracted on the continuous reference field. The reference value is then subtracted from the observed value to obtain the reference difference.

[0038] As a preferred embodiment of the foreign object detection method for glass fiber fabric described in this invention, the method includes: collecting the raw outputs of the spectral channel and polarization channel, weighting and summing them to obtain a unified sensitivity quantity, and constructing a continuous reference field, comprising:

[0039] The sampling time interval is obtained by configuring the frame rate of the spectrometer to extract the stable interval. Within the stable interval, two-channel data are collected through the API interface, and a unified sensitivity value is obtained by weighted summation.

[0040] The fabric movement distance at each sampling time is extracted and set as the longitudinal coordinate. The transverse coordinate of the scan line is also extracted to obtain the reference position. A uniform sensitivity value is assigned to each reference position to obtain discrete reference points. Two-dimensional interpolation is performed on all discrete reference points to obtain a continuous reference field.

[0041] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the foreign object detection method for glass fiber cloth as described in the first aspect of the present invention.

[0042] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the foreign object detection method for glass fiber cloth as described in the first aspect of the present invention.

[0043] The beneficial effects of this invention are as follows: This invention sets the lateral range based on the calibration value of the scanning lateral coverage width, extracts the minimum and maximum longitudinal coordinates of the update node, constructs reconstruction points, and, combined with the benchmark difference, constructs a continuous foreign object state field on the reconstruction points. Based on the continuous foreign object state field, it extracts the time window sequence, constructs an optimization objective function and performs iterative optimization to obtain the final order and diffusion intensity of each reconstruction point, and calculates the residual. This effectively distinguishes transient disturbances from real foreign object events, reduces the false alarm rate, improves the stability of foreign object type determination, and enhances the spatiotemporal consistency and reliability of foreign object detection results under high-speed production conditions. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of the foreign object detection method for the glass fiber cloth surface in Example 1.

[0046] Figure 2This is a schematic diagram of the diffusion intensity inversion and residual calculation of the foreign matter detection method on the glass fiber cloth surface in Example 1. Detailed Implementation

[0047] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0048] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0049] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0050] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides a method for detecting foreign objects on a glass fiber cloth surface, including the following steps:

[0051] S1. Obtain the sampling time interval through the frame rate configuration of the spectrometer, collect the raw outputs of the spectral channel and polarization channel, and obtain a unified sensitivity quantity by weighted summation. Construct a continuous reference field, set the update node based on the scanning lateral coverage width and factory configuration, allocate the raw output of the polarization channel to the update node, calculate the observation value and subtract it from the reference value of the continuous reference field to obtain the reference difference.

[0052] Specifically, the raw outputs of the spectral and polarization channels are collected, weighted, and summed to obtain a unified sensitivity quantity, thus constructing a continuous reference field, including:

[0053] The sampling time interval is obtained by configuring the frame rate of the spectrometer;

[0054] Define the fabric coordinate system, including setting the horizontal direction of the fabric as the x-axis, the direction of fabric movement as the y-axis, and setting the intersection of the left edge of the fabric and the scan line as the origin;

[0055] The cumulative pulse count is collected through a data acquisition system (such as a microcontroller). The cumulative pulse count is converted into instantaneous velocity using a fixed pulse interval method. The instantaneous velocities are arranged horizontally according to time to obtain an instantaneous velocity sequence. The time interval between instantaneous velocities in the instantaneous velocity sequence is extracted. If k consecutive time intervals are all greater than the velocity fluctuation tolerance (based on the factory-set tolerance band setting), the corresponding time interval is marked as a stable interval, and the start and end times of the stable interval are recorded.

[0056] Within the stable range, two-channel data are collected via API interface, including the raw output of the spectral channel and the raw output of the polarization channel;

[0057] The raw output of the spectral channel is normalized to obtain a normalized reflectance spectrum. The chemical sensitivity quantity is extracted from the normalized reflectance spectrum using the in-band integration method and then normalized to obtain a normalized chemical sensitivity quantity.

[0058] Extract the two polarization intensities from the raw output of the polarization channel, which represent the sensor readings for different polarization directions. Calculate the polarization contrast of the two polarization intensities using a polarization contrast formula (e.g., the Stokes parameter method) and normalize it to obtain the normalized polarization contrast.

[0059] The normalized chemical sensitivity and the normalized polarization contrast are weighted and summed to obtain a unified sensitivity.

[0060] Extract the fabric movement distance (obtained by laser rangefinder) at each sampling time, set it as the longitudinal coordinate, and extract the transverse coordinate of the scan line to obtain the reference position. Assign a uniform sensitivity value to each reference position to obtain discrete reference points.

[0061] Two-dimensional interpolation is performed on all discrete reference points to obtain a continuous reference field.

[0062] By configuring the frame rate of the spectrometer to obtain the sampling time interval and defining the fabric coordinate system, all subsequent sampling data are described under a unified time scale and a unified spatial reference, reducing positioning errors caused by clock drift of different devices and fluctuations in fabric operation from the source. The data acquisition system collects cumulative pulse counts and converts them into instantaneous velocity sequences. Then, by filtering the stable interval using velocity fluctuation tolerance, spectral and polarization disturbances in the unsteady phase of the fabric can be excluded from modeling. This makes the original outputs of the two channels more comparable and repeatable under stable conditions. Within the stable interval, the original outputs of the spectral and polarization channels are acquired, and normalized reflectance spectra and normalized chemical sensitivity quantities are constructed, as well as polarization contrast is normalized. This allows for the identification of common scales related to changes in illumination intensity and sensor gain drift. The absorption process is normalized to enhance the sensitivity to changes in material chemical response and surface microstructure scattering caused by foreign objects. The normalized chemical sensitivity quantity and the normalized polarization contrast are weighted and summed to obtain a unified sensitivity quantity, which essentially forms a cross-mechanism information fusion quantity. This allows weak features at the same location under different modes to complement each other, improving both the ability to identify penetrating foreign objects and the ability to distinguish attached particulate foreign objects. Furthermore, the fabric movement distance and the transverse coordinate of the scan line obtained by the laser rangefinder are combined to construct the reference position and generate discrete reference points. Finally, a continuous reference field is formed through two-dimensional interpolation, which improves the output from discrete sampling to a continuous spatial reference. This facilitates consistent reference reading and differential comparison on any spatial coordinate, thus providing support for stable and interpretable background baseline modeling.

[0063] Furthermore, the observed values ​​are calculated and subtracted from the reference values ​​of the continuous reference field to obtain the reference difference, including:

[0064] The horizontal coverage width of the scan line is obtained through the API interface (based on the horizontal travel calibration of the scanning mechanism), and the number of horizontal nodes is set according to the factory configuration. The horizontal coverage width is divided into equal intervals to obtain the horizontal coordinates of the horizontal nodes and initialize the vertical coordinates to 0. All horizontal nodes are arranged vertically to obtain the node set.

[0065] Based on the instantaneous velocity and sampling time interval in the instantaneous velocity sequence, the periodic displacement is obtained by matrix integration;

[0066] Based on the node set, if the periodic displacement is equal to 1, the vertical coordinate of the horizontal node is set to 0; if the periodic displacement is greater than 1, the vertical coordinate of the horizontal node at time t is set to the sum of the vertical coordinate of the horizontal node at time t and the periodic displacement, thus obtaining the updated node set.

[0067] The two polarization intensities in the original output are assigned to each update node in the update node set. This includes reading the lateral position coordinates of the scanning mechanism at time t and assigning them to the lateral coordinates of the nearest update node using the nearest neighbor mapping method. This means that the two polarization intensities in the original output of the scanning mechanism are assigned to the update node. If the same update node is assigned multiple times, the two polarization intensities in the original output assigned to the update node are aggregated using the sample mean method to obtain the two polarization intensities of the update node. Otherwise, they are directly set as the two polarization intensities of the update node. A unified sensitivity value is calculated for the two polarization intensities of the node and marked as the observation value of the update node.

[0068] For each updated node coordinate, extract the corresponding reference value on the continuous reference field, and subtract the reference value from the observed value to obtain the reference difference.

[0069] By calibrating the lateral coverage width and generating a node set according to the factory configuration, observations within the scan line coverage area are structurally mapped to fixed spatial nodes, avoiding sample scattering and spatial sparsity caused by relying solely on instantaneous scan positions. Periodic displacement is obtained using instantaneous velocity sequences and sampling time intervals to update the node longitudinal coordinates, introducing continuous fabric movement into the node evolution rules. This allows the same physical region to correspond to continuous node trajectories at different sampling times, thereby reducing misalignment caused by slight changes in fabric velocity. The two polarization intensities of the polarization channel are assigned to the update nodes according to nearest neighbor mapping, and the sample mean method is used to aggregate when assigning multiple times to the same update node. This can suppress random noise and occasional pulse interference without increasing the complexity of the model, improving the statistical robustness of node observations. The benchmark value is extracted from the update node on the continuous benchmark field, and the benchmark difference is obtained by subtracting the benchmark value from the observation value. This transforms the output from an absolute quantity to a relative quantity, significantly reducing the system bias caused by material batch differences, background texture, and slow changes in illumination. The benchmark difference is closer to the real disturbance components caused by foreign objects, improving the reliability of subsequent reconstruction and judgment.

[0070] S2. Based on the calibration value of the scanning horizontal coverage width, set the horizontal range, extract the minimum and maximum vertical coordinates of the update node, construct the reconstruction point, combine the benchmark difference, construct the continuous foreign object state field on the reconstruction point, extract the time window sequence based on the continuous foreign object state field, construct the optimization objective function and perform iterative optimization to obtain the final order and diffusion intensity of each reconstruction point, and calculate the residual.

[0071] Specifically, a reconstruction point is constructed, and combined with the baseline difference, a continuous foreign object state field is built at the reconstruction point, including:

[0072] The horizontal range is set based on the calibration value of the horizontal coverage width of the scan (obtained through the API interface), and the minimum and maximum vertical coordinates of the update node are extracted to construct the computational domain.

[0073] Defining a reconstruction point within the computational domain includes dividing the horizontal spacing obtained by equally dividing the horizontal coverage width into equal intervals based on the number of horizontal nodes set by the factory configuration into 2 to obtain the horizontal step size, sorting the periodic displacements in ascending order, selecting the smallest periodic displacement as the vertical step size, and setting the horizontal step size and vertical step size as the horizontal and vertical coordinates of the reconstruction point to obtain the reconstruction point.

[0074] The support domain radius is set to n times the horizontal spacing. The Euclidean distance between the updated node and the reconstructed node is calculated using the Euclidean distance formula. If the Euclidean distance is less than or equal to the support domain radius, the updated node is set as a neighbor node. All reconstructed points are traversed to obtain all neighbor points. If the absolute value of the number of neighbor points is less than the minimum number of nodes (based on the fixed threshold method), the support domain radius is increased by one horizontal spacing. The above steps are repeated until the absolute value of the number of neighbor points is greater than or equal to the minimum number of nodes. The obtained neighbor points are arranged vertically to obtain the neighbor point set.

[0075] Based on the set of neighborhood points, a weight scalar is constructed for each neighborhood point, which includes dividing the square of the Euclidean distance by the square of the support domain radius obtained after stopping, taking the negative value of the division result, and using the exponential function to obtain the weight scalar for the negative value.

[0076] Construct a weight diagonal matrix by placing the weight scalars in the order of their neighborhood points on the diagonal;

[0077] Subtracting the horizontal and vertical coordinates of the reconstructed point from the horizontal and vertical coordinates of the neighboring points respectively yields the relative horizontal and vertical coordinates. This information is used to construct a quadratic basis vector. Stacking all the quadratic basis vectors column-by-column yields the polynomial basis matrix, as shown in the formula:

[0078] ,

[0079] in, are quadratic basis vectors. and To update nodes The relative x-coordinate and relative y-coordinate;

[0080] The linear system matrix is ​​constructed based on the polynomial basis matrix and the weighted diagonal matrix, as follows:

[0081] ,

[0082] ,

[0083] in, and For a linear system matrix, For polynomial basis matrices, This is a weighted diagonal matrix. For transpose, For the first The x and y coordinates of each reconstructed point;

[0084] The reconstructed foreign object state field scalar is obtained by solving the coefficient column vector based on the linear system matrix, as shown in the formula:

[0085] ,

[0086] ,

[0087] ,

[0088] in, For the coefficient column vector, The observed column vector is obtained by stacking the benchmark differences in the order of the neighborhood nodes. As basis vectors, For foreign object state field scalar;

[0089] Arrange the reconstructed foreign object state field scalars according to coordinates to obtain a continuous foreign object state field.

[0090] By constructing a computational domain through scanning the lateral coverage width and updating the vertical range of nodes, reconstruction is performed only within actually observable areas with sufficient data support, avoiding artifacts caused by extrapolation in areas without data support. Reconstruction points are defined within the computational domain, and regular sampling with lateral and vertical step sizes enables spatially controllable and scale-consistent reconstruction meshes, making the spatial resolution of the continuous heterogeneous state field engineering-adjustable. An adaptive expansion mechanism for the support domain radius and neighborhood point set is employed, gradually expanding the support domain when the number of neighborhood points is insufficient, effectively preventing underdeterminism caused by local sparsity. This approach ensures that each reconstructed point has a stable neighborhood that meets the minimum number of nodes requirement. Distance attenuation weighting is achieved through weight scalars and weight diagonal matrices, making neighboring nodes contribute more to the reconstruction and distant nodes contribute less. This preserves local details while suppressing long-distance noise propagation. A polynomial basis matrix is ​​constructed and the linear system is solved to obtain the foreign object state field scalar. Essentially, this achieves smooth reconstruction and continuous expression of the benchmark difference, transforming the output from discrete difference to a continuous foreign object state field. This benefits both the consistency of subsequent time window sequence extraction and the stability of spatial morphology analysis such as connected component extraction.

[0091] Furthermore, the time window sequence is extracted based on the continuous foreign object state field, including:

[0092] The time window length is read from the factory-set fixed parameter table via the API interface. The discrete sampling time is obtained by multiplying the sampling time interval and the number of samplings. For each reconstruction point, a window sequence is extracted based on the continuous foreign object state field within the time window length. ,in, For a continuous foreign object state field, and For reconstruction point x and y coordinates For discrete sampling times, Index representing the current sampling number Subtract backtracking steps The obtained historical index;

[0093] For each reconstruction point, the mean within the time window is calculated using the arithmetic mean method. The window sequence is then demeaned to obtain the demeaned sequence.

[0094] For the demeaned sequence, fractional Fourier transform coefficients are calculated one by one on the candidate order set using fractional Fourier transform. The energy of the fractional Fourier transform coefficients is calculated using the complex modulus square and then normalized to obtain the normalized energy.

[0095] The candidate order set refers to the product of the time window length and the sampling time interval to obtain the time span, and then using twice pi divided by the time span and pi divided by the sampling time interval to obtain the minimum and maximum resolvable angular frequencies, thus generating the candidate order set. The formula is as follows:

[0096] ,

[0097] ,

[0098] ,

[0099] ,

[0100] ,

[0101] in, For rotation angle, Pi For order, and For the minimum and maximum resolvable angular frequencies, and Let the order of the minimum and maximum feasible fractional fields be denoted by . The number of candidate orders is set based on statistical analysis. For candidate order index, For the candidate order set;

[0102] By arranging the normalized energy horizontally, a discrete set of values ​​with a length equal to the window length is obtained.

[0103] By reading the time window length from the factory-set fixed parameter table and combining it with the sampling time interval and sampling number, discrete sampling times are obtained. This provides a traceable engineering basis for the length and sampling density of the time window sequence, ensuring comparable time analysis scales under different equipment or operating conditions. Within the time window length, the window sequence is extracted based on the continuous foreign matter state field. The spatial location of each reconstructed point and its temporal evolution can be bound to the same sequence, forming a spatiotemporally consistent observation for dynamic infiltration and diffusion processes. De-meaning the window sequence can eliminate the influence of slowly varying background and fixed bias on frequency domain characteristics, allowing subsequent energy analysis to focus more on the wave structure caused by foreign matter rather than the static baseline. A fractional Fourier transform is performed on the candidate order set to obtain normalized energy, which can map the complex non-stationary response within the time window into an order-related energy spectrum. This explicitly characterizes the memory and multi-scale diffusion characteristics of the sequence, providing sufficient statistical input for subsequent determination of the final order and diffusion intensity.

[0104] Furthermore, an optimization objective function is constructed and iteratively optimized to obtain the final order and diffusion intensity of each reconstruction point, including:

[0105] The arithmetic mean and unbiased sample variance method are used to extract the mean and variance of the discrete value set, construct the objective function, and obtain the Lagrange function value, as shown in the formula:

[0106] ,

[0107] ,

[0108] in, To optimize the objective function, For fractional field indexing, from 0 to Counting by loop yields the result. The time window length, The entropy order is defined based on empirical rules. For probability mass, For normalized energy, This is the average value. Candidates for the order of the fractional domain are obtained by enumerating the candidate set. For variance, , as well as For Lagrange multipliers, The value of the Lagrange function;

[0109] The probability mass is iteratively updated using the Newton-Raphson Method until the maximum number of iterations is reached, resulting in the updated probability mass. The updated Lagrange function value is then recalculated using the updated probability mass.

[0110] The updated Lagrange function values ​​are accumulated to obtain the cumulative distribution function value. The cumulative distribution function value is then subtracted from 1 to obtain the tail residual. The logarithm of the tail residual is taken and fitted using a least-squares line to obtain the slope. This slope is mapped to fractional-order candidates, and these candidates are then pruned to obtain the final order. The formula is as follows:

[0111] ,

[0112] ,

[0113] ,

[0114] in, This is the remaining amount at the tail end. The slope The regression intercept is obtained from least squares linear regression. Candidates for fractional order. For the final order, and The upper and lower bounds of the order are defined based on rules of thumb. To determine the optimal fractional order, we select the fractional order candidate with the largest objective function value from all candidate fractional orders and set it as the optimal fractional order.

[0115] By extracting the mean and variance of the discrete value set and constructing an optimization objective function, the concentration and dispersion of energy distribution can be unified into the same evaluation framework. This transforms the selection of the final order from relying on subjective thresholds into a repeatable optimization solution. Iterative updates of probabilistic quality and stopping at the maximum number of iterations ensure convergence and provide a definite upper bound for computation, making it suitable for online detection and real-time processing scenarios. The updated Lagrangian function values ​​are accumulated and the tail residue is calculated. The slope is then obtained through least-squares linear fitting and mapped to fractional-order candidates. The sensitivity of tail behavior to anomalous perturbations can be used to capture the dynamic characteristics of sub-diffusion or super-diffusion, making the final order more consistent with the mechanism of foreign matter penetration and diffusion. By pruning the fractional-order candidates and determining the optimal fractional-domain order, overfitting or numerical instability caused by extreme orders can be avoided, while ensuring that the final order has an engineering-usable range, thus laying the foundation for robust inversion of subsequent diffusion intensity.

[0116] Finally, calculate the residuals, including:

[0117] Based on the final order and the coordinates of the reconstructed points, a continuous field of order is constructed using the moving least squares method.

[0118] Based on historical experimental experience, minimum diffusion intensity, maximum diffusion intensity, and discrete step size of diffusion intensity are set and arranged horizontally to obtain a candidate set. The candidate set is then input into a fractional-order discrete predictor (e.g., GL or L1) to obtain discrete sampling times. The predicted value;

[0119] Based on the continuous field, the Lagrangian function values ​​of the final order and the optimal fractional order at each reconstruction point are extracted. Combined with the predicted values, the weighted error is calculated to construct the final diffusion intensity. Then, combined with the reconstruction point coordinates, a continuous intensity field is constructed using the moving least squares method. The formula is as follows:

[0120] ,

[0121] ,

[0122] ,

[0123] in, For window sample weights, For the final order The optimal probability quality. For weighted error, The candidate diffusion intensity is determined by the candidate set. Enumeration yields, For the final diffusion intensity, Candidate diffusion intensity The predicted value;

[0124] Based on the intensity continuous field, the final diffusion intensity of each reconstruction point is extracted, and combined with the final order input to a fractional-order discrete predictor to obtain the discrete sampling time. The predicted value is used to calculate the residual, and the formula is:

[0125] ,

[0126] in, For residuals, These are predicted values.

[0127] Constructing a continuous order field based on the final order and reconstructed point coordinates elevates the point-level order estimation to a spatially continuous parameter, allowing the differences in infiltration and diffusion at different locations to be expressed continuously. This facilitates the formation of a visual and spatially consistent judgment criterion. Setting a candidate set of diffusion intensity and inputting it into a fractional-order discrete predictor yields predicted values, transforming the solution for diffusion intensity from an uncontrollable continuous search to a controllable discrete enumeration. This reduces computational complexity and improves the reproducibility of results. Combining the Lagrange function value under the optimal fractional order to calculate the weighted error and determine the final diffusion intensity, and then constructing a continuous intensity field, allows the diffusion intensity to simultaneously reflect the fitting error and statistical reliability, avoiding bias caused by selecting based on a single error criterion. Obtaining predicted values ​​and calculating residuals under the final order and final diffusion intensity enables the output to simultaneously possess the ability to predict mechanisms and characterize deviations. The residuals thus become direct evidence for determining foreign body mutations and local anomalies, and provide core input for subsequent contact indication and cumulative contact time calculations.

[0128] S3. Compare the residuals to generate contact indications, calculate the cumulative contact time, determine whether a foreign object exists, extract the feature vector at the first trigger moment, obtain the foreign object category label through K-means clustering, and upload it to the cloud platform for display.

[0129] Specifically, foreign object category labels are obtained through K-means clustering and uploaded to the cloud platform for display, including:

[0130] If the absolute value of the residual is less than the residual threshold (set based on empirical rules), the contact indication is determined to be 1; otherwise, it is 0. The cumulative contact time is calculated using the following formula:

[0131] ,

[0132] in, To calculate the cumulative contact time, The sampling period is and To update nodes x and y coordinates For contact indication;

[0133] Define a trigger threshold. If the cumulative contact time is greater than or equal to the trigger threshold, a foreign object is determined to be present and marked as 1; otherwise, no foreign object is determined to be present and marked as 0. If the sampling time is discrete... and Then Marked as the first trigger time;

[0134] The cumulative contact time, final order, final diffusion intensity, and absolute values ​​of material response offset (obtained by subtracting the baseline sensitivity value (based on historical experimental experience analysis) from the normalized chemical sensitivity value) are arranged horizontally and normalized to obtain the feature vector at the first trigger moment.

[0135] The number of clusters is set based on the elbow rule. The feature vector is input into K-means for clustering to obtain the cluster number. For example, when the cluster number is equal to 1, it corresponds to the permeation and diffusion class. When the cluster number is equal to 2, it corresponds to the attached particle class, thus obtaining the foreign matter category label.

[0136] All coordinates marked as 1 are used to extract the boundaries of the foreign object region using the connected component extraction method, and the area is calculated using the polygon area calculation method.

[0137] The boundaries, area, and category labels of the foreign object area are uploaded to the cloud platform via API for display.

[0138] By generating contact indications based on residual thresholds and accumulating contact time, instantaneous noise triggers can be transformed into time-consistent, continuous evidence, effectively reducing false alarms caused by occasional disturbances and making the judgment more consistent with the physical fact that foreign objects on the fabric are continuous in space and time. The trigger threshold determines the first trigger time, enabling the system to locate the key time point when foreign objects begin to affect the fabric, facilitating subsequent traceability and coordinated handling. By constructing feature vectors from accumulated contact time, final order, final diffusion intensity, and material response offset and performing clustering, the persistence, dynamic memory, diffusion ability, and material response differences of foreign objects can be integrated into a separable feature space, thereby improving the interpretability and cross-batch stability of foreign object category labels. By extracting connected components and calculating the area of ​​the coordinates of foreign objects, point-like judgments can be elevated to regional morphological quantification, giving the output boundary and scale information that can be used for process decision-making. Uploading the foreign object region boundary, area, and foreign object category label to the cloud platform for display expands the detection results from a local signal processing closed loop to a visualized and manageable production closed loop, facilitating quality traceability, threshold calibration, and multi-line collaborative control.

[0139] This embodiment also provides a computer device applicable to the foreign object detection method for fiberglass fabric, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the foreign object detection method for fiberglass fabric as proposed in the above embodiment.

[0140] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0141] This embodiment also provides a storage medium storing a computer program, which, when executed by a processor, implements the foreign object detection method for glass fiber cloth as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0142] In summary, this invention sets the lateral range based on the calibration value of the scanning lateral coverage width, extracts the minimum and maximum longitudinal coordinates of the update nodes, constructs reconstruction points, and, combined with benchmark difference, constructs a continuous foreign object state field on the reconstruction points. Based on the continuous foreign object state field, it extracts the time window sequence, constructs an optimization objective function and performs iterative optimization to obtain the final order and diffusion intensity of each reconstruction point, and calculates the residual. This effectively distinguishes transient disturbances from real foreign object events, reduces the false alarm rate, improves the stability of foreign object type determination, and enhances the spatiotemporal consistency and reliability of foreign object detection results under high-speed production conditions.

[0143] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for detecting foreign objects on a glass fiber cloth surface, characterized in that: include, The sampling time interval is obtained by configuring the frame rate of the spectrometer, the raw outputs of the spectral and polarization channels are collected, and a unified sensitivity is obtained by weighted summation. A continuous reference field is constructed, and the update node is set based on the scanning lateral coverage width and factory configuration. The raw output of the polarization channel is assigned to the update node, the observation value is calculated and subtracted from the reference value of the continuous reference field to obtain the reference difference. The horizontal range is set based on the calibration value of the horizontal coverage width of the scan, and the minimum and maximum vertical coordinates of the updated node are extracted. The reconstruction point is defined and a set of neighborhood points is constructed. Based on the set of neighborhood points, a weight scalar is constructed for each neighborhood point, and a weight diagonal matrix is ​​created. Subtracting the horizontal and vertical coordinates of the reconstructed point from the horizontal and vertical coordinates of the neighboring points respectively, we obtain the relative horizontal and vertical coordinates, construct a quadratic basis vector, and stack all the quadratic basis vectors column by column to obtain the polynomial basis matrix. Construct the linear system matrix based on the polynomial basis matrix and the weighted diagonal matrix; The coefficient column vector is solved based on the linear system matrix to obtain the reconstructed foreign object state field scalar. The reconstructed foreign object state field scalar is arranged according to coordinates to obtain a continuous foreign object state field. For each reconstruction point, a window sequence is extracted based on the continuous foreign object state field. The mean within the time window length is calculated using the arithmetic mean method for each reconstruction point. The window sequence is then demeaned to obtain a demeaned sequence. For the demeaned sequence, fractional Fourier transform coefficients are calculated one by one on the candidate order set using fractional Fourier transform. The energy of the fractional Fourier transform coefficients is calculated using the complex modulus square and then normalized to obtain the normalized energy. The normalized energy is arranged horizontally to obtain a discrete value set with a length equal to the window length. The mean and variance of the discrete value set are extracted using the arithmetic mean method and the unbiased sample variance method. An optimization objective function is constructed to obtain the probability mass and Lagrange function values. The probability mass is iteratively updated using the Newton-Raphson method to obtain the updated probability mass, and the updated Lagrange function value is recalculated using the updated probability mass. The updated Lagrange function values ​​are accumulated to obtain the cumulative distribution function value. The tail residual is calculated, the logarithm of the tail residual is taken and fitted with a least squares line to obtain the slope. The slope is mapped to fractional order candidates, the fractional order candidates are pruned to obtain the final order, and based on the final order and the coordinates of the reconstructed points, a continuous field of order is constructed using the moving least squares method. Based on historical experimental experience, minimum diffusion intensity, maximum diffusion intensity, and discrete step size of diffusion intensity are set and arranged horizontally to obtain a candidate set. The candidate set is then input into a fractional-order discrete predictor to obtain discrete sampling times. The predicted value; Based on the continuous field, the Lagrangian function values ​​of the final order and the optimal fractional domain order of each reconstruction point are extracted. The weighted error is calculated by combining the predicted values, the final diffusion intensity is constructed, and the intensity continuous field is constructed by combining the coordinates of the reconstruction points and using the moving least squares method. Based on the intensity continuous field, the final diffusion intensity of each reconstruction point is extracted, and combined with the final order input to a fractional-order discrete predictor to obtain the discrete sampling time. Calculate the residuals from the predicted values; The residuals are compared to generate contact indications, the cumulative contact time is calculated, the presence of foreign objects is determined, and the feature vector at the first trigger moment is extracted. The foreign object category label is obtained through K-means clustering and uploaded to the cloud platform for display.

2. The method for detecting foreign objects on a glass fiber cloth surface as described in claim 1, characterized in that: The process of obtaining foreign object category labels through K-means clustering and uploading them to the cloud platform for display includes: If the absolute value of the residual is less than the residual threshold, the contact indication is set to 1; otherwise, it is set to 0. The cumulative contact time is calculated. If the cumulative contact time is greater than or equal to the trigger threshold, a foreign object is detected and marked as 1; otherwise, no foreign object is detected and marked as 0. (The last part, "if discrete sampling time...", appears to be an error and doesn't translate directly.) and Then Mark as the first trigger time; The absolute values ​​of cumulative contact time, final order, final diffusion intensity, and material response offset are arranged horizontally and normalized to obtain the feature vector at the first trigger moment. The feature vector is then input into K-means for clustering to obtain the foreign object category label. All coordinates marked as 1 are used to extract the boundaries of the foreign object region using the connected component extraction method, and the area is calculated using the polygon area calculation method. The boundaries, area, and category labels of the foreign object area are uploaded to the cloud platform via API for display.

3. The foreign matter detection method for glass fiber cloth as described in claim 2, characterized in that: The calculation of the observed values ​​and subtraction of them with the reference values ​​of the continuous reference field yields the reference difference, including: Obtain the horizontal coverage width of the scan line, and construct a node set based on the factory configuration setting for the number of horizontal nodes; Based on the instantaneous velocity and sampling time interval in the instantaneous velocity sequence, the periodic displacement is obtained through matrix integration, and the node set is updated to obtain the updated node set; The two polarization intensities in the original output are assigned to the update nodes in each update node set. For the coordinates of each update node, the reference value of the corresponding coordinate is extracted on the continuous reference field. The reference value is then subtracted from the observed value to obtain the reference difference.

4. The foreign matter detection method for glass fiber cloth as described in claim 3, characterized in that: The raw outputs of the collected spectral and polarization channels are weighted and summed to obtain a unified sensitivity quantity, which is used to construct a continuous reference field, including: The sampling time interval is obtained by configuring the frame rate of the spectrometer to extract the stable interval. Within the stable interval, two-channel data are collected through the API interface, and a unified sensitivity value is obtained by weighted summation. The fabric movement distance at each sampling time is extracted and set as the longitudinal coordinate. The transverse coordinate of the scan line is also extracted to obtain the reference position. A uniform sensitivity value is assigned to each reference position to obtain discrete reference points. Two-dimensional interpolation is performed on all discrete reference points to obtain a continuous reference field.

5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the foreign object detection method for the glass fiber cloth surface according to any one of claims 1 to 4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the foreign object detection method for the glass fiber cloth surface as described in any one of claims 1 to 4.

Images