Recycled carbon fiber quality detection method based on machine vision

Abstract

The invention provides a recycled carbon fiber quality detection method based on machine vision, and the method comprises the steps: obtaining a surface image of a carbon fiber bundle through a machine vision method, and carrying out the preprocessing of the image, including image denoising and enhancement; for the preprocessed surface image of the carbon fiber bundle, extracting the morphological characteristics of the fiber bundle, including fiber bundle surface texture and fiber orientation distribution, for representing the morphological change of the fiber bundle under the reciprocating bending stress; acquiring a scattered light intensity distribution image on the surface of the carbon fiber bundle, and extracting spatial distribution characteristics of scattered light intensity through an image segmentation method, including uniformity and continuity of scattered light intensity distribution; if the fracture risk exists, positioning the identified high-risk area, and collecting image data of the high-risk area; according to the high-risk area image data, scattered light intensity distribution characteristics are extracted, a quantitative relation between scattered light intensity space distribution and the fiber bundle fracture risk is established, and a quantitative index for evaluating the fiber bundle fracture risk is obtained.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for detecting the quality of recycled carbon fiber based on machine vision. Background Art

[0002] In the quality inspection of recycled carbon fiber, the machine vision-based method faces a key technical challenge: how to accurately assess the fracture risk of carbon fiber bundles under reciprocating bending stress. When a carbon fiber bundle is subjected to periodic bending loads, its surface morphology changes significantly, and local high-intensity bending areas appear. These high-intensity bending areas will cause the spatial distribution of scattered light on the fiber bundle surface to change, thus affecting the image acquisition and analysis of the machine vision system. Specifically, when high-intensity bending occurs on the surface of the carbon fiber bundle, the spatial distribution of scattered light becomes more uneven and complex. This change in scattered light distribution is closely related to the location of stress concentration and potential fracture points inside the fiber bundle. However, existing machine vision algorithms have difficulty in accurately capturing and quantifying the subtle changes in this scattered light distribution, resulting in the inability to effectively assess the fracture risk of the fiber bundle. In addition, the surface morphology and scattered light distribution of the carbon fiber bundle are constantly changing during dynamic bending. This dynamic change further increases the difficulty of real-time tracking and analysis by the machine vision system. How to accurately extract the key features of scattered light distribution under dynamic conditions and establish a reliable mapping relationship between it and the risk of fiber bundle fracture is a technical challenge that needs to be solved urgently. To solve this problem, we need to conduct in-depth research on the morphological evolution of carbon fiber bundles under reciprocating bending stress, develop highly sensitive scattered light collection and analysis technology, and establish a quantitative correlation model between scattered light distribution and the risk of fiber bundle breakage. Only by overcoming these technical difficulties can we achieve a breakthrough in the quality detection method of recycled carbon fiber based on machine vision and provide reliable quality assurance for carbon fiber recycling. Summary of the invention

[0003] The present invention provides a method for detecting the quality of recycled carbon fiber based on machine vision, which mainly includes:

[0004] The machine vision method is used to obtain the surface image of the carbon fiber bundle, and the image is preprocessed, including image denoising and enhancement;

[0005] For the surface image of the pretreated carbon fiber bundle, the fiber bundle morphological characteristics are extracted, including the fiber bundle surface texture and fiber orientation distribution, which are used to characterize the morphological changes of the fiber bundle under reciprocating bending stress;

[0006] Obtaining the scattered light intensity distribution image of the carbon fiber bundle surface, and extracting the spatial distribution characteristics of the scattered light intensity, including the uniformity and continuity of the scattered light intensity distribution, by using the image segmentation method;

[0007] Based on the morphological characteristics of the fiber bundle and the spatial distribution characteristics of the scattered light intensity, determine whether there are high-intensity bending regions and fracture risks in the fiber bundle. If the fiber orientation distribution of the fiber bundle is inconsistent and the scattered light intensity distribution is uneven, it is determined that there are high-intensity bending regions and fracture risks in the fiber bundle;

[0008] If there is a fracture risk, locate the identified high-risk region and collect the image data of this high-risk region;

[0009] According to the image data of the high-risk region, extract the scattered light intensity distribution characteristics, establish a quantitative relationship between the spatial distribution of the scattered light intensity and the fracture risk of the fiber bundle, and obtain a quantitative index for evaluating the fracture risk of the fiber bundle;

[0010] Compare the quantitative index of the fiber bundle fracture risk with a preset threshold. If it exceeds the preset threshold, it is determined that the quality of the recycled carbon fiber bundle is unqualified and there is a fracture risk, and it needs to be removed. Otherwise, it is determined that the quality of the carbon fiber bundle is qualified and can be used for recycling.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0012] The present invention discloses a method for detecting the quality of recycled carbon fiber based on machine vision. This method obtains the surface image of the carbon fiber bundle, extracts the morphological characteristics of the fiber bundle and the scattered light intensity distribution characteristics, and determines whether there are high-intensity bending regions and fracture risks in the fiber bundle. For the regions with risks, detailed image acquisition and analysis are carried out, and a quantitative relationship model between the spatial distribution of the scattered light intensity and the fracture risk is established. Finally, the quantitative index of the fracture risk is compared with the preset threshold to determine whether the quality of the carbon fiber bundle is qualified. The present invention realizes the automatic and quantitative evaluation of the quality of the recycled carbon fiber bundle through image processing and machine learning technologies, improves the accuracy and efficiency of the evaluation, and provides effective technical support for the quality control and reuse of recycled carbon fiber. Description of the Drawings

[0013] Figure 1 It is a flowchart of a method for detecting the quality of recycled carbon fiber based on machine vision according to the present invention. Detailed Embodiments

[0014] To further understand the content of the present invention, the present invention will be described in detail in combination with the drawings and embodiments. The following further details the present application in combination with the drawings and embodiments. It can be understood that the specific embodiments described herein are only used to explain the related invention, rather than limiting the invention. In addition, it should be noted that only the parts related to the invention are shown in the drawings for the convenience of description.

[0015] Such as Figure 1, a quality inspection method for recycled carbon fiber based on machine vision in this embodiment may specifically include:

[0016] S101. Obtain the surface image of the carbon fiber bundle by machine vision method, and preprocess the image, including image denoising and enhancement.

[0017] Obtain the first image of the surface of the carbon fiber bundle by using a high-resolution industrial camera, and obtain the second image through Gaussian kernel filtering and median filtering according to the gray value in the first image; perform histogram equalization processing on the second image to obtain the gray peak position, and perform binary segmentation and morphological operation on the second image according to the gray peak position to obtain the third image; extract the set of edge region pixel points according to the third image, enhance the pixel point set through the Sobel operator, and use the double-threshold method to extract the complete contour line to obtain the preprocessed surface image of the carbon fiber bundle.

[0018] Specifically, a high-resolution industrial camera is used to capture the original image of the carbon fiber bundle surface. The dark areas in the original image with gray values less than 250 are processed by 3×3 Gaussian kernel smoothing filtering. The obtained image is denoised by the 5×5 window median filtering algorithm to obtain a denoised image. Histogram equalization is performed on the denoised image, the position of the maximum peak is obtained from the gray histogram, and the image is binarized according to the gray value corresponding to the peak to obtain a binary image. Three iterative erosion and dilation operations are performed on the binary image to eliminate residual noise points to obtain a fiber morphology image. According to the fiber morphology image, the set of pixel points in the edge region of the carbon fiber bundle is extracted, and the 3×3 Sobel operator is used to enhance the edge pixel points, and the complete contour line is extracted by the double-threshold method to obtain a contour feature image. The local binary pattern feature vector is extracted from the contour feature image, a carbon fiber surface texture feature library is established, and the surface energy, contrast, and entropy value features of the carbon fiber bundle are calculated using the gray-level co-occurrence matrix. According to the texture feature library, a random forest classifier training sample set is constructed, and the surface of the carbon fiber bundle is segmented by the recursive region growing algorithm to obtain pixel blocks in the suspected defect area. For the pixel blocks in the suspected defect area, three types of feature parameters, namely shape features, gray features, and texture features, are extracted, and the defect type is finally determined based on the feature parameters. When the high-resolution industrial camera shoots the surface of the carbon fiber bundle, the camera resolution reaches 2448×2048 pixels, and the light intensity is maintained at about 800 lux to obtain a clear original image. Due to uneven illumination in the original image, there are dark areas with low gray values in some regions. By setting the gray threshold of 250 as the judgment standard, these dark areas are processed by 3×3 Gaussian kernel smoothing filtering, and the standard deviation of the Gaussian kernel is taken as 0.5. There is still salt-and-pepper noise in the smoothed image, and the 5×5 window median filtering is continued for denoising. After the denoised image is processed by histogram equalization, the gray distribution is more uniform, which is beneficial to subsequent image segmentation. In the equalized gray histogram, the gray values corresponding to the carbon fiber bundle area are mainly concentrated between 180 and 220, forming an obvious main peak. The binarization is performed with the gray value of 200 corresponding to the peak as the segmentation threshold. In order to eliminate the isolated noise points and burrs in the binary image, a circular structure element with a radius of 2 pixels is used for 3 iterative erosion and dilation operations to obtain a relatively complete fiber morphology. Based on the obtained fiber morphology image, the edge contour features of the carbon fiber bundle are extracted. The 3×3 Sobel operator is used to enhance the edge pixels, and the gradient values of the edge pixel points in the enhanced image are significantly increased. By setting a double threshold, a low threshold of 40 and a high threshold of 100 for edge detection, a continuous and complete contour line is obtained. For the contour feature image, when extracting the local binary pattern features, the sampling radius is set to 2 and the number of sampling points is 8, and a 256-dimensional feature vector is calculated.At the same time, texture features are calculated in four directions of 0 degrees, 45 degrees, 90 degrees and 135 degrees by using the gray-level co-occurrence matrix. Among them, energy reflects the uniformity of the texture, contrast represents the severity of gray-level changes, and entropy represents the complexity of the texture. These features constitute a feature library for subsequent defect recognition. The random forest classifier uses 100 decision trees, and the maximum depth of each tree is set to 10, and it is trained with the samples in the feature library. When detecting the suspected defect area, the growth threshold of the recursive region growing algorithm is set to 15, and the minimum area of the region is set to 50 square pixels. For the detected suspected defect area, shape features, area, perimeter, circularity, gray-level features, average gray-level value, standard deviation and texture features, energy, contrast, entropy are extracted, and these feature parameters are integrated to determine the defect type, so as to achieve an accurate evaluation of the surface quality of the carbon fiber bundle.

[0019] S102. For the preprocessed surface image of the carbon fiber bundle, extract the fiber bundle morphology features, including the surface texture of the fiber bundle and the fiber orientation distribution, to characterize the morphological changes of the fiber bundle under the reciprocating bending stress.

[0020] Calculate the gray-level co-occurrence matrix according to the preprocessed surface image of the carbon fiber bundle by using a sliding window, obtain the contrast matrix, entropy matrix and correlation matrix from the gray-level co-occurrence matrix, and get the texture feature dataset; perform window gradient operator calculation and Hough transform detection on the texture feature dataset, statistically analyze the line segment inclination angle distribution from the Hough transform detection results, and obtain the fiber orientation angle frequency distribution curve; calculate the fiber distribution histograms in four directions according to the fiber orientation angle frequency distribution curve, extract the fiber number ratios in each direction from the fiber distribution histograms, and generate a fiber direction feature vector; process the fiber direction feature vector by using the maximum inter-class variance method to obtain fiber spacing data, and perform Fourier transform and wavelet decomposition on the fiber spacing data and the fiber direction feature vector to generate a deformation feature descriptor.

[0021] Specifically, a sliding window with a size of 32x32 pixels is constructed for the preprocessed image, the window stride is set to 16 pixels, the gray-level co-occurrence matrix within each window is calculated, and by extracting the contrast matrix, entropy matrix, and correlation matrix, a texture feature dataset of the fiber bundle surface area is obtained. Based on the texture feature dataset, the window gradient operator is calculated, the Hough transform is used to detect edge line segments, and according to the statistical distribution of the line segment inclination angles, the fiber orientation angle frequency distribution curve is obtained, and the main orientation angle data is extracted for the peak positions of the curve. According to the main orientation angle data, the fiber distribution histograms in the four directions of 0 degrees, 45 degrees, 90 degrees, and 135 degrees are calculated, the fiber number distribution ratios in each direction are obtained, and the fiber direction feature vector is generated. A detection window with a sliding step of 8 pixels is used to perform a horizontal scan on the area where the fiber direction feature vector is located, and the fiber spacing threshold is calculated by the Otsu method, and the fiber arrangement spacing data and the winding tightness parameter are extracted. The Fourier transform is performed on the fiber bundle surface texture feature dataset, the amplitude spectrum and phase spectrum are extracted, the high-frequency coefficients and low-frequency coefficients are obtained based on four-layer wavelet decomposition, and the fiber surface periodic feature vector is constructed. The fiber direction feature vector, spacing data, and periodic feature vector are combined to form a deformation feature descriptor, which is input into the radial basis kernel support vector machine for deformation type recognition and degree determination. The surface image of the preprocessed carbon fiber bundle presents complex texture features. The local features of the image are extracted through a sliding window with a size of 32×32 pixels, and the window moves 16 pixels each time, ensuring a 50% overlapping area between adjacent windows. This overlapping design helps to capture the continuous change features of the fiber bundle surface. Within each sliding window, the calculated gray-level co-occurrence matrix reflects the spatial correlation between pixel points, where the numerical value of the contrast matrix between 0.8 and 1.2 indicates obvious texture changes on the fiber bundle surface, and the entropy matrix between 2.5 and 3.5 reflects a relatively high texture complexity. Based on the obtained texture feature dataset, the edge information is extracted by applying the window gradient operator, and the line segments detected by the Hough transform are mainly distributed at ±5 degrees,

[0022] At typical orientation angles such as ±45 degrees and 90 degrees. In the frequency distribution curve obtained by statistics, the peak height at the 45-degree position reaches 0.35, indicating that this direction is the main arrangement direction of the fiber bundle. This orientation characteristic directly affects the mechanical properties of the carbon fiber bundle, especially the deformation behavior under reciprocating bending stress. The four-direction distribution histogram constructed based on the main orientation angle data shows that the proportion of fiber quantity in the 0-degree direction is 15%, 40% in the 45-degree direction, 25% in the 90-degree direction, and 20% in the 135-degree direction. This distribution characteristic is characterized by generating a 16-dimensional direction feature vector. The 4 components in each direction of the vector respectively correspond to the fiber quantity, arrangement regularity, continuity, and local aggregation degree in that direction. The fiber spacing is measured using a detection window with an 8-pixel step size. The fiber spacing threshold determined by the maximum inter-class variance method is 12 pixels. The measured average adjacent fiber spacing is 15 pixels, and the standard deviation is 2.3 pixels. The winding tightness parameter is defined as the number of fiber intersection points per unit area, and the typical value is between 0.08 and 0.12. These parameters together reflect the internal structural characteristics of the fiber bundle. For the extraction of the periodic characteristics on the fiber surface, first, the texture feature data is subjected to Fourier transform. Significant peaks are observed at spatial frequencies of 0.15 and 0.3 in the amplitude spectrum, indicating the existence of regular texture arrangement. The high-frequency coefficients obtained by four-layer wavelet decomposition reflect the local change characteristics of the fiber surface, and the coefficient values are distributed between 0.5 and 2, while the low-frequency coefficients reflect the overall morphological characteristics, with values varying between 0.2 and 0.8. The constructed 48-dimensional deformation feature descriptor is input into the support vector machine and mapped through the radial basis kernel function. The kernel parameter is selected as 0.8, achieving accurate classification of different degrees of bending deformation.

[0023] S103. Obtain the scattered light intensity distribution image on the surface of the carbon fiber bundle, and extract the spatial distribution characteristics of the scattered light intensity through image segmentation methods, including the uniformity and continuity of the scattered light intensity distribution.

[0024] An optoelectronic detector is used to obtain the original image of the scattered light on the surface of the carbon fiber bundle. Based on the original image, a light intensity gradient feature image is obtained through Gaussian kernel filtering and wavelet transform. For the light intensity gradient feature image, a light intensity threshold is set according to the bimodal positions of the gray histogram for binary segmentation to obtain a light intensity region map, and the watershed algorithm is used to obtain the boundary curve from the light intensity region map. According to the statistical feature quantity of the light intensity of the marked region corresponding to the boundary curve, a light intensity distribution histogram is constructed, and the k-means clustering method is used to divide the brightness level regions from the light intensity distribution histogram. The light intensity attenuation curve is obtained by least squares fitting. For the brightness level regions, a set of edge pixel points is extracted, and the regional integrity parameter is obtained according to the Euclidean distance between the edge pixel points. The regional integrity parameter and the characteristics of the light intensity attenuation curve are input into a support vector regressor to train the light intensity spatial distribution prediction model, and the uniformity evaluation index is extracted from the prediction results.

[0025] Specifically, a high-sensitivity photodetector is used to obtain the original image of the scattered light on the surface of the carbon fiber bundle. The original image is processed by three-by-three Gaussian kernel filtering to reduce noise. The light intensity gradient features are extracted through three-layer wavelet transform. The light intensity threshold is set according to the bimodal positions of the gray histogram of the scattered light image, and the image is binarized and segmented to obtain the light intensity region map. The connected components of the light intensity region map are labeled, and for each labeled region, four statistical feature quantities, namely the light intensity mean, variance, skewness, and kurtosis, are calculated. The watershed algorithm is used to extract the boundaries of the light intensity distribution region, and the standard deviation of the light intensity difference between adjacent pixel points is calculated based on the boundary curve to obtain the local continuity value. Based on the light intensity statistical feature quantities of the labeled regions, a light intensity distribution histogram is constructed. The k-means clustering method is used to divide the light intensity distribution into multiple brightness level regions. The light intensity attenuation curves of each region are fitted by the least squares method, and the attenuation coefficient and fitting error are obtained to form an eight-dimensional spatial distribution feature vector. According to the local continuity value, the overall light intensity distribution uniformity index is calculated. For each brightness level region, the set of edge pixel points is extracted, and the Euclidean distance between the edge pixel points is calculated to obtain the region integrity parameter. The spatial distribution feature vector is input into the support vector regressor, and the light intensity spatial distribution prediction model is trained. The uniformity evaluation index is extracted from the prediction results. The region growing algorithm based on pixel neighborhood relationship is used to analyze the connectivity of the light intensity distribution, and the shape factor and light intensity variance of each connected region are calculated to generate the continuity feature map of the scattered light distribution. In the original image of the scattered light obtained by the high-sensitivity photodetector, the image resolution reaches 2048×2048 pixels, the gray level is 256 levels, and the light intensity values are distributed between 0 and 255. Filtering and noise reduction are performed through a Gaussian kernel with a size of 3×3, and the standard deviation of the Gaussian kernel is set to 0.8, which effectively suppresses the random noise in the image while retaining the detailed features of the scattered light. The light intensity gradient features extracted through three-layer wavelet transform show that the gradient values are relatively large in the edge region of the fiber bundle, usually between 40 and 60, while the gradient values in the flat region are relatively small, between 5 and 15. The connected components in the light intensity region map usually contain multiple independent scattered light spots, and the area of each light spot region is between 1000 and 5000 pixels. Among the statistical features calculated for these regions, the light intensity mean reflects the overall level of the scattering intensity, with typical values between 150 and 200; the variance characterizes the degree of dispersion of the light intensity distribution, with values between 25 and 45; the skewness describes the asymmetry of the distribution, varying between -0.5 and 0.5; the kurtosis reflects the sharpness of the distribution, usually between 2.8 and 3.2. The boundary curve extracted by the watershed algorithm forms a closed contour around each scattered light spot, and the standard deviation of the light intensity difference between adjacent pixel points is between 8 and 12, indicating that the light intensity transition is relatively smooth. There are multiple peaks in the light intensity distribution histogram, and 4 to 6 brightness level regions are divided by k-means clustering, and the light intensity changes relatively uniformly within each region.During the fitting process of the light intensity attenuation curve, a quadratic polynomial model is adopted. The obtained attenuation coefficient is between 0.02 and 0.05, and the fitting error is controlled within 3%. These parameters together constitute an eight-dimensional vector describing the spatial distribution characteristics of the scattered light. The calculation of the overall light intensity distribution uniformity index is based on local continuity values. When the light intensity difference between adjacent regions is less than 20%, the transition is considered relatively smooth. In the set of edge pixel points, the average Euclidean distance between adjacent points is between 1.2 and 1.8 pixels, indicating that the edge contour of the light spot is complete. The support vector regressor adopts a radial basis kernel function, and the kernel parameter is set to 0.5. The cross-validation mean squared error is less than 0.02. The region growing algorithm starts from the seed point and expands. When the light intensity difference between adjacent pixel points is less than the preset threshold of 10, it is incorporated into the current region. The shape factor of the connected region is defined as the ratio of the square of the perimeter to the area. For an ideal circular light spot, this value is close to 12.56, and the actual measured value fluctuates between 13 and 15. The light intensity variance shows an increasing trend from the center to the edge within the connected region. The variance in the central region is between 5 and 8, and it increases to between 15 and 20 in the edge region. This change characteristic is completely recorded in the continuity characteristic map of the scattered light distribution.

[0026] S104. According to the morphological characteristics of the fiber bundle and the spatial distribution characteristics of the scattered light intensity, determine whether there are high-intensity bending regions and fracture risks in the fiber bundle. If the fiber orientation distributions of the fiber bundle are inconsistent and the scattered light intensity distribution is uneven, it is determined that there are high-intensity bending regions and fracture risks in the fiber bundle.

[0027] The maximum entropy method is used to calculate the orientation distribution probability of the fiber bundle in the main direction. The direction entropy value is obtained according to the orientation distribution probability. The orientation distribution difference degree is determined by the deviation rate between the direction entropy value and the reference sample entropy value. Uniform grids are divided according to the orientation distribution difference degree, and the ratio of the standard deviation of the light intensity in the grid to the average value of the light intensity in the adjacent grid is calculated. The light intensity distribution non-uniformity matrix is generated according to the ratio. A characteristic map is constructed according to the orientation distribution difference degree and the light intensity distribution non-uniformity matrix. A bending deformation recognition model is trained through a random forest classifier, and the position coordinates of the high-risk region are extracted from the recognition model. The local morphological curvature value and the light intensity gradient value are extracted for the position coordinates of the high-risk region. If the curvature value is greater than the reference curvature and the light intensity gradient is greater than the reference gradient, it is determined that there is a bending fracture risk in this region.

[0028] Specifically, for the fiber orientation data in the fiber bundle morphology characteristics, the maximum entropy method is used to calculate the orientation distribution probabilities in four main directions of 0 degrees, 45 degrees, 90 degrees, and 135 degrees. Based on the probability distribution, the direction entropy value is calculated, and the orientation distribution difference degree is determined by the deviation rate from the entropy value of the reference sample. The spatial distribution characteristic map of the scattered light intensity is divided into uniform grids of 32×32 pixel size, and the ratio of the standard deviation of the light intensity in each grid to the mean light intensity of the adjacent grid is calculated to generate the light intensity distribution non-uniformity matrix. The sliding window method is used to calculate the local area morphology curvature, the window size is set to 16×16 pixels, and the morphology deformation characteristic map is generated according to the curvature value, and the curvature abnormal area is marked. A two-dimensional feature map is constructed based on the orientation distribution difference degree and the light intensity distribution non-uniformity matrix, and the principal component analysis method is used to extract the main feature vectors to generate the risk assessment feature set. The random forest classifier is used to train the risk assessment feature set to obtain the bending deformation recognition model, and the position coordinates of the high-risk area are extracted from the recognition results. For the high-risk area, the local morphology curvature value and the light intensity gradient value are extracted. If the curvature value is greater than 1.5 times the reference curvature and the light intensity gradient is greater than 2 times the reference gradient, it is determined that there is a high-intensity bending and fracture risk in this area. In the orientation distribution analysis of the fiber bundle morphology characteristics, the orientation probability distributions of the normal samples in the four main directions are relatively uniform. The probability in the 0-degree direction is 0.26, in the 45-degree direction is 0.24, in the 90-degree direction is 0.25, and in the 135-degree direction is 0.25. The direction entropy value remains between 1.98 and 2.02. When local deformation occurs in the fiber bundle, the orientation distribution probability shows an obvious deviation. The dominant orientation probability in the deformed area rises above 0.35, the direction entropy value drops below 1.75, and the deviation rate from the entropy value of the reference sample exceeds 12%. After the scattered light intensity characteristic map is divided into grids of 32×32 pixels, the ratio of the mean light intensity between adjacent grids in the normal area fluctuates between 0.85 and 1.15, and the standard deviation of the light intensity inside the grid is less than 15% of the overall mean. For the area with bending deformation, there is a significant difference in the ratio of the mean light intensity between adjacent grids, the maximum ratio exceeds 1.5, and the standard deviation of the light intensity inside the grid increases to more than 25% of the overall mean. This non-uniform distribution characteristic is manifested as a local high-value area in the light intensity distribution non-uniformity matrix. The calculation of the local area morphology curvature uses a sliding window of 16×16 pixels, and the window moves 8 pixels each time to form an overlap of 50%. The local curvature values on the surface of the normal fiber bundle are distributed between 0.02 and 0.05, and the curvature changes gently. In the area where bending deformation occurs, the local curvature value surges above 0.08 and changes rapidly within a small range, forming a marked curvature abnormal area. The orientation distribution and light intensity distribution information are integrated in the feature map. The first two principal component feature vectors extracted by the principal component analysis contain more than 85% of the data variation information. The random forest classifier uses 200 decision trees, and the maximum depth of each tree is limited to 10 layers. The recognition accuracy obtained through cross-validation reaches 92%.When locating the identified high-risk areas, the local maximum suppression algorithm is adopted to retain the detection boxes with a confidence level greater than 0.85. In the refined analysis of high-risk areas, the reference curvature value of normal fiber bundles is 0.03, and the reference light intensity gradient value is 20. When the local topography curvature exceeds 0.045 and the light intensity gradient exceeds 40, it indicates that the area is subjected to excessive bending stress. The measured data shows that in the fracture risk area, the local curvature value often reaches above 0.06, and the maximum light intensity gradient value exceeds 50. This combined feature has a high correlation with the observed fiber bundle damage positions. From the statistical results of multiple test samples, when the curvature value exceeds 1.5 times the reference value and the light intensity gradient exceeds 2 times the reference value, the probability of fracture in this area during subsequent use increases significantly.

[0029] S105. If there is a fracture risk, locate the identified high-risk areas and collect the image data of these high-risk areas.

[0030] The region growing algorithm is used to construct the boundary contour of the risk area. The position coordinates and area data of the risk area are obtained through the calculation of the minimum bounding rectangle. According to the position coordinates and area data of the risk area, the boundary width is extended outward from the boundary of the risk area, and the image acquisition path planning data is generated through local sampling grids. Based on the image acquisition path planning data, the movement of the displacement platform is controlled to obtain a multi-angle image sequence of the risk area. The adaptive histogram equalization processing is performed on the multi-angle image sequence to obtain the target image, and the edge contour features of the target image are obtained through iterative optimal threshold segmentation.

[0031] Specifically, based on the identified high-risk area distribution map, the boundary contour of the risk area is constructed through the region growing algorithm. The position coordinates and area data of the risk area are calculated using the minimum bounding rectangle, and a risk area positioning map is generated. For the risk area positioning map, a rectangular area with a boundary width equal to the square root of the area of the region is extended outward from the boundary of the risk area, and a local sampling grid with a pixel pitch of half of the original image is constructed. According to the coordinate data of the sampling grid, an image acquisition path planning map is generated, recording the spatial position coordinates and shooting angle parameters of each sampling point. Based on the path planning map, the movement of the high-precision displacement platform is controlled, and an automatic focusing device is used to obtain a multi-angle image sequence of the risk area. The image sequence is processed by adaptive histogram equalization, and the image brightness, contrast, and sharpness parameters are extracted. The image with the optimal imaging quality is selected as the target image. For the target image, iterative optimal threshold segmentation is performed to extract the regional edge contour features, and morphological processing methods are used to remove local noise. Based on the edge contour features, a morphological descriptor is constructed, including four characteristic parameters: contour length, area ratio, roundness, and directionality. A support vector machine is used to classify and identify the regional morphological features. When the region growing algorithm processes the high-risk area distribution map, it starts to expand from the point with the highest risk intensity. When the difference in risk intensity between adjacent pixel points is less than the threshold of 0.2, they are included in the current region, and finally, the complete boundary of the risk area is obtained. Taking a risk area with an area of 2500 square pixels as an example, the aspect ratio of its minimum bounding rectangle is 1.8, and the coordinates of the four vertices of the rectangle are located at (1024, 768), (1280, 768), (1024, 1024), and (1280, 1024) in the original image coordinate system. When expanding the boundary, the square root of the area of the region is 50 pixels, and a sampling band with a width of 50 pixels is evenly expanded around the risk area. The pixel pitch of the sampling grid is set to half of the original image, that is, in the original image with a resolution of 2048×2048, the sampling grid pitch is 1024 pixels. This setting optimizes the data acquisition efficiency while ensuring image details. Each node in the sampling grid records the position offset and pitch angle value relative to the center point of the image. The image acquisition path planning adopts a serpentine scanning method, traversing row by row starting from the upper left corner. The movement speed between adjacent sampling points is controlled at 5 millimeters per second, and the repeat positioning accuracy of the displacement platform is better than 0.01 millimeters. At each sampling point position, the automatic focusing device adjusts the lens focal length to make the edge gradient value of the image reach the maximum. For a target area of 2500 square pixels, usually 25 to 36 images need to be acquired to achieve complete coverage. In the quality evaluation of the image sequence, for the image processed by adaptive histogram equalization, the brightness value is distributed between 0 and 255, and the mean value remains around 160. The image contrast is represented by the standard deviation of the gray difference between adjacent pixels. The larger the value, the clearer the image details, and the typical value is between 45 and 65.The clarity parameter is calculated by the Laplacian operator, and its value range is between 0.8 and 1.2. An image with a value greater than 1.0 is considered to have good imaging quality. During the iterative optimal threshold segmentation process, the initial threshold is set to the average gray value of the image. The average gray values of the target region and the background region are calculated in each iteration, and their mean value is taken as the new threshold. The iteration stops when the difference between the thresholds of two adjacent iterations is less than 1. For morphological processing, a circular structuring element with a radius of 3 pixels is used to perform opening and closing operations on the edge contour to eliminate noise regions with an area less than 20 pixels. In the finally constructed morphological descriptor, the contour length is usually between 200 and 300 pixels. The area ratio is defined as the ratio of the actual area to the area of the minimum bounding rectangle, and its normal range is between 0.7 and 0.9. The roundness parameter is used to describe the regularity of the region shape. The roundness of a complete circle is 1, and the actual measured value is between 0.6 and 0.8. The directional feature is represented by the angle between the major axis direction and the horizontal direction, which has an important indicative role in the classification of the fracture risk region.

[0032] S106. According to the high-risk region image data, extract the scattered light intensity distribution characteristics, establish a quantitative relationship between the spatial distribution of the scattered light intensity and the fiber bundle fracture risk, and obtain a quantitative index for evaluating the fiber bundle fracture risk.

[0033] According to the high-risk region image, the grid division method with a fixed pixel size is used to obtain the sub-region light intensity distribution data, and the spectral feature vector of the sub-region is obtained through Fourier transform; a spatial pyramid structure is constructed for the spectral feature vector, and the maximum pooling operation is used to obtain multi-layer feature maps, and the feature descriptor is obtained by extracting statistics from the feature maps; the principal component analysis dimensionality reduction process is performed on the feature descriptor, the random forest regressor is trained by the cross-validation method, and the risk prediction model is obtained; based on the risk prediction model, the correlation coefficient matrix between the feature vector and the fracture risk is calculated, and the regression equation is established by the partial least squares method to obtain the quantitative relationship between the spatial distribution of the scattered light intensity and the fiber bundle fracture risk.

[0034] Specifically, the detailed images of high-risk areas are divided into grids with a size of 32×32 pixels. A sliding window with a step size of 16 pixels is used to extract the light intensity distribution data in each sub-region. The light intensity spectrum features are obtained through Fourier transform. The amplitude spectrum and phase spectrum are extracted from the spectrum curve to form a light intensity feature vector. A three-layer spatial pyramid structure is constructed for the light intensity feature vector. 2×2 max-pooling operations are used to downsample successively to obtain feature maps, and the mean, variance, and kurtosis of each layer of feature maps are extracted as three statistics. A nine-dimensional feature descriptor is constructed based on the statistics of the three-layer feature maps. The principal component analysis method is used to reduce the dimensionality of the feature dimensions, and the principal components with a contribution rate greater than 85% are retained as the main feature vectors. Based on the main feature vectors and the pre-labeled fracture risk level data, a training sample set is constructed, and a random forest regressor is trained using the five-fold cross-validation method to obtain a risk prediction model. The correlation coefficient between the feature vector and the fracture risk is calculated using the risk prediction model to construct a correlation coefficient matrix, and the feature items with a correlation coefficient greater than the preset threshold are extracted. The partial least squares method is used to establish a regression equation for the selected feature items, and the regression coefficients are optimized through regularization methods to obtain a quantitative relationship between the light intensity distribution features and the fracture risk. Based on the quantitative relationship, the normalized risk index value is calculated. This index comprehensively considers the spatial inhomogeneity and local variability of the light intensity distribution to form the final fracture risk assessment index. After the high-risk area images are divided into grids with 32×32 pixels, each sub-region contains the light intensity data of 1024 pixel points. The 16-pixel overlap design between adjacent windows ensures the continuity of feature extraction. Among the spectrum features obtained through Fourier transform, the amplitude spectrum reflects the severity of the light intensity change. In the normal area, the amplitude spectrum shows a gentle exponential decay, and the energy proportion of the main frequency component is between 25% and 35%. In the abnormal area, multiple significant peaks appear in the amplitude spectrum, and the energy proportion of the main frequency component drops below 15%. In the constructed three-layer spatial pyramid structure, the first layer maintains the original resolution and contains all the texture details, and the size of the feature map is 32×32. After 2×2 max-pooling, the resolution of the second layer is reduced to 16×16, retaining the medium-scale structural features. The third layer is further downsampled to 8×8, reflecting the large-scale morphological features. The statistics of each layer of feature maps show that the variance value in the normal area decreases steadily with the increase of the layer, and the typical value drops from 0.45 in the first layer to 0.15 in the third layer. The variance value in the abnormal area fluctuates greatly between different layers, reflecting the instability of the scale structure. In the nine-dimensional feature descriptor, the mean of each layer of feature maps reflects the overall level of the light intensity distribution, the variance characterizes the degree of dispersion of the distribution, and the kurtosis describes the steepness of the distribution. The principal component analysis results show that the cumulative contribution rate of the first three principal components reaches 87%. Among them, the first principal component is highly correlated with the light intensity uniformity, with a contribution rate of 45%. The second principal component mainly reflects the change law of multi-scale features, with a contribution rate of 28%.The random forest regressor uses 200 decision trees with a maximum depth limit of 8 levels, and the standard deviation of the prediction error remains within 0.08 in the five-fold cross-validation. The correlation coefficient matrix analysis shows that there is a significant positive correlation between the spatial inhomogeneity of the light intensity distribution and the fracture risk, with a correlation coefficient reaching 0.82, and the correlation coefficient between the local variability and the fracture risk is 0.75. In the partial least squares regression, L1 regularization is used to control the model complexity, and the regularization coefficient is set to 0.1. The final obtained quantitative relationship contains 5 feature terms, corresponding to the uniformity index, local variability index, and scale cross-correlation index of the multi-scale features respectively. The value range of the normalized risk assessment index is between 0 and 1. When the index value exceeds 0.75, it indicates that there is a significant fracture risk in this area. The detection accuracy corresponding to this threshold reaches 92% in the experimental verification. This evaluation method based on multi-scale feature fusion comprehensively considers the local details and global features of the light intensity distribution, providing a reliable quantitative basis for the damage assessment of carbon fiber bundles.

[0035] S107. Compare the quantitative index of the fiber bundle fracture risk with a preset threshold. If it exceeds the preset threshold, it is determined that the quality of the recycled carbon fiber bundle is unqualified, there is a fracture risk, and it needs to be removed. Otherwise, it is determined that the quality of the carbon fiber bundle is qualified and can be used for recycling.

[0036] Perform min-max normalization on the quantitative index of the fiber bundle fracture risk, and use the kernel density estimation method to obtain the index distribution curve; calculate the mean and standard deviation for the index distribution curve, and obtain the risk level determination region through the standard deviation interval; select detection points on the surface of the fiber bundle, use the sliding window method to extract local features from the detection points, and obtain the feature statistics through the feature mean and variance; perform bootstrap resampling on the feature statistics to obtain multiple groups of sample data, calculate the upper and lower limits of the confidence interval through the sample data. If the upper limit of the confidence interval exceeds the high-risk threshold, it is marked as a non-conforming product. If the lower limit of the confidence interval is lower than the low-risk threshold, it is marked as a conforming product.

[0037] Specifically, the minimum-maximum normalization is performed on the quantitative index of the fiber bundle fracture risk. Based on the labeled sample dataset, a risk index distribution histogram is constructed, and the kernel density estimation method is used to fit the index distribution curve. According to the mean and standard deviation of the index distribution curve, three standard deviation intervals are set as warning thresholds to generate a risk level determination region, and the cross-validation method is used to verify the stability of the determination region. Detection points are evenly selected on the surface of each fiber bundle, and the sliding window method is used to extract local features at each detection point. The feature mean and variance are calculated to construct a feature statistic. Bootstrap resampling is performed on the feature statistic to generate multiple sets of sample data, and the upper and lower limits of the confidence interval of the sample mean are calculated to obtain the fluctuation range of the risk index. The support vector machine is used to classify the feature data, and the kernel function parameters are optimized through the grid search method to establish a risk level discriminator. Based on the risk level discriminator, the fiber bundles are classified. If the upper limit of the confidence interval of the risk index exceeds the high-risk threshold, it is marked as a defective product. If the lower limit of the confidence interval is lower than the low-risk threshold, it is marked as a qualified product. A majority vote is performed on the marking results, and the confidence score is calculated based on the determination results of all detection points. The weighted average method is used to obtain the final quality evaluation result. After normalization, the quantitative index of the fiber bundle fracture risk is distributed between 0 and 1. A total of 1,000 labeled fiber bundle samples are used to construct the index distribution histogram, and the kernel density estimation is performed using the Gaussian kernel function. The obtained distribution curve shows obvious bimodal characteristics, with two main peaks at 0.3 and 0.8, corresponding to the typical distributions of qualified and defective products respectively. Based on the statistical characteristics of the distribution curve, the three standard deviation intervals are set as follows: the low-risk interval is less than 0.4, the medium-risk interval is between 0.4 and 0.7, and the high-risk interval is greater than 0.7. The five-fold cross-validation method is used to verify the stability of these thresholds. The verification results show that the fluctuation of the determination accuracy of different batches of samples is less than 5%, indicating that the threshold setting has good robustness. A detection point is selected every 50 mm along the axial direction on the surface of the fiber bundle, and a 64×64 pixel sliding window is used to extract local features at each detection point. The feature statistic includes four indicators: the mean light intensity, variance, skewness, and kurtosis. Among them, the mean light intensity of qualified products is between 180 and 220, and the variance is less than 25. While the mean of defective products is usually lower than 160 or higher than 240, and the variance is greater than 40. Bootstrap resampling is performed 1,000 times on the feature data of each detection point, and the calculated confidence interval reflects the stability of the risk index. For typical qualified product samples, the upper limit of the 95% confidence interval of the risk index does not exceed 0.35, while the lower limit of the confidence interval of defective products is generally higher than 0.75. Samples in the middle region require a comprehensive judgment combining multiple detection points. The support vector machine uses the radial basis kernel function for classification, and the kernel parameter is optimized through the grid search within the range of 0.1 to 10. The finally selected parameter value is 1.5.The classification accuracy of the trained discriminator on the test set reaches 94%, and the recall rates of qualified and unqualified products are 96% and 92% respectively. The confidence score of the judgment result adopts the soft voting method, and the score weight of each detection point is inversely proportional to the confidence interval width of its risk index. In an actual case, the risk indexes of a fiber bundle at 5 detection points are 0.82, 0.79, 0.85, 0.76, and 0.81 respectively, and the confidence interval width is between 0.12 and 0.18. The final score after weighted average is 0.81, which is significantly higher than the high-risk threshold of 0.7, so it is judged as an unqualified product. The indexes of 5 detection points of another fiber bundle are 0.28, 0.32, 0.25, 0.31, and 0.29. The confidence interval width is less than 0.1, and the final score is 0.29, which is significantly lower than the low-risk threshold of 0.4, so it is judged as a qualified product. This method of multi-point joint judgment effectively avoids misjudgment caused by single-point anomalies.

[0038] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting the quality of recycled carbon fiber based on machine vision, characterized in that: The method comprises: The machine vision method is used to obtain the surface image of the carbon fiber bundle, and the image is preprocessed, including image denoising and enhancement; For the surface image of the pretreated carbon fiber bundle, the fiber bundle morphological characteristics are extracted, including the fiber bundle surface texture and fiber orientation distribution, which are used to characterize the morphological changes of the fiber bundle under reciprocating bending stress; Obtaining the scattered light intensity distribution image of the carbon fiber bundle surface, and extracting the spatial distribution characteristics of the scattered light intensity, including the uniformity and continuity of the scattered light intensity distribution, by using the image segmentation method; According to the fiber bundle morphology characteristics and the spatial distribution characteristics of scattered light intensity, it is judged whether the fiber bundle has a high-intensity bending area and a risk of fracture. If the fiber orientation distribution of the fiber bundle is uneven and the scattered light intensity distribution is uneven, it is judged that the fiber bundle has a high-intensity bending area and a risk of fracture; If there is a risk of fracture, the identified high-risk area is located and image data of the high-risk area is collected; Based on the image data of high-risk areas, the scattered light intensity distribution characteristics are extracted, and the quantitative relationship between the scattered light intensity spatial distribution and the fiber bundle fracture risk is established to obtain the quantitative index for evaluating the fiber bundle fracture risk. The quantitative index of the fiber bundle breakage risk is compared with the preset threshold. If it exceeds the preset threshold, the recycled carbon fiber bundle is judged to be of unqualified quality and has a risk of breakage and needs to be discarded. Otherwise, the carbon fiber bundle is judged to be of qualified quality and can be used for recycling.

2. The method according to claim 1, characterized in that The method of using machine vision to obtain the surface image of the carbon fiber bundle and preprocessing the image, including image denoising and enhancement, includes: A high-resolution industrial camera is used to obtain a first image of the surface of the carbon fiber bundle, and a second image is obtained by Gaussian kernel filtering and median filtering according to the gray value in the first image; Performing histogram equalization processing on the second image to obtain a grayscale peak position, and performing binary segmentation and morphological operation on the second image according to the grayscale peak position to obtain a third image; A pixel point set in the edge area is extracted according to the third image, the pixel point set is enhanced by a Sobel operator, and a complete contour line is extracted by a double threshold method to obtain a preprocessed carbon fiber bundle surface image.

3. The method according to claim 1, characterized in that The method extracts the fiber bundle morphological features, including the fiber bundle surface texture and fiber orientation distribution, from the pretreated carbon fiber bundle surface image to characterize the morphological changes of the fiber bundle under reciprocating bending stress, including: According to the pre-processed carbon fiber bundle surface image, a gray level co-occurrence matrix is ​​calculated using a sliding window, and a contrast matrix, an entropy matrix and a correlation matrix are obtained from the gray level co-occurrence matrix to obtain a texture feature data set; Performing window gradient operator calculation and Hough transform detection on the texture feature data set, and statistically analyzing the line segment inclination distribution from the Hough transform detection result to obtain a fiber orientation angle frequency distribution curve; Calculating fiber distribution histograms in four directions according to the fiber orientation angle frequency distribution curve, extracting the ratio of the number of fibers in each direction from the fiber distribution histogram, and generating a fiber direction feature vector; The fiber direction feature vector is processed by using the maximum inter-class variance method to obtain fiber spacing data, and the fiber spacing data and the fiber direction feature vector are subjected to Fourier transformation and wavelet decomposition to generate a deformation feature descriptor.

4. The method according to claim 1, characterized in that: The method of obtaining the scattered light intensity distribution image of the carbon fiber bundle surface and extracting the spatial distribution characteristics of the scattered light intensity, including the uniformity and continuity of the scattered light intensity distribution, by an image segmentation method includes: A photoelectric detector is used to obtain an original image of scattered light from the surface of the carbon fiber bundle, and a light intensity gradient characteristic image is obtained based on the original image through Gaussian kernel filtering and wavelet transformation; For the light intensity gradient feature image, a light intensity threshold is set according to the double peak position of the grayscale histogram to perform binary segmentation to obtain a light intensity area map, and a boundary curve is obtained from the light intensity area map using a watershed algorithm; A light intensity distribution histogram is constructed according to the light intensity statistical feature quantity of the marked area corresponding to the boundary curve, a k-means clustering method is used to divide the brightness level area from the light intensity distribution histogram, and a light intensity attenuation curve is obtained by least squares fitting; A set of edge pixel points is extracted for the brightness level area, and the regional integrity parameter is obtained according to the Euclidean distance between the edge pixel points. The regional integrity parameter and the light intensity attenuation curve characteristics are input into the support vector regressor to train the light intensity spatial distribution prediction model, and the uniformity evaluation index is extracted from the prediction result.

5. The method according to claim 1, characterized in that The method of judging whether the fiber bundle has a high-intensity bending area and a risk of fracture according to the fiber bundle morphology characteristics and the spatial distribution characteristics of the scattered light intensity, and judging whether the fiber bundle has a high-intensity bending area and a risk of fracture if the fiber orientation distribution of the fiber bundle is uneven and the scattered light intensity distribution is uneven, comprises: The maximum entropy method is used to calculate the orientation distribution probability of the fiber bundle in the main direction, and the directional entropy value is obtained according to the orientation distribution probability. The orientation distribution difference is determined by the deviation rate between the directional entropy value and the entropy value of the reference sample; Dividing uniform grids according to the orientation distribution difference, calculating the ratio of the light intensity standard deviation in the grid to the light intensity mean of adjacent grids, and generating a light intensity distribution non-uniformity matrix according to the ratio; Constructing a feature map according to the orientation distribution difference and the light intensity distribution unevenness matrix, obtaining a bending deformation recognition model through random forest classifier training, and extracting the position coordinates of the high-risk area from the recognition model; The local morphology curvature value and the light intensity gradient value are extracted for the position coordinates of the high-risk area. If the curvature value is greater than the reference curvature and the light intensity gradient is greater than the reference gradient, it is determined that there is a risk of bending and fracture in the area.

6. The method according to claim 1, characterized in that If there is a risk of fracture, the identified high-risk area is located and image data of the high-risk area is collected, including: The region growing algorithm is used to construct the boundary contour of the risk area, and the location coordinates and area data of the risk area are obtained by calculating the minimum circumscribed rectangle; According to the position coordinates and area data of the risk area, the boundary width is expanded outward from the boundary of the risk area, and image acquisition path planning data is generated through a local sampling grid; Control the movement of the displacement platform based on the image acquisition path planning data to obtain a multi-angle image sequence of the risk area; An adaptive histogram equalization process is performed on the multi-angle image sequence to obtain a target image, and edge contour features of the target image are obtained by iterative optimal threshold segmentation.

7. The method according to claim 1, characterized in that The method extracts scattered light intensity distribution characteristics based on the high-risk area image data, establishes a quantitative relationship between the scattered light intensity spatial distribution and the fiber bundle breakage risk, and obtains a quantitative index for evaluating the fiber bundle breakage risk, including: According to the high-risk area image, a grid division method with a fixed pixel size is used to obtain the sub-area light intensity distribution data, and the frequency spectrum feature vector of the sub-area is obtained by Fourier transform; Constructing a spatial pyramid structure for the spectral feature vector, using a maximum pooling operation to obtain a multi-layer feature map, and extracting statistics from the feature map to obtain a feature descriptor; Performing principal component analysis and dimensionality reduction processing on the feature descriptors, training a random forest regressor using a cross-validation method, and obtaining a risk prediction model; Based on the risk prediction model, the correlation coefficient matrix between the characteristic vector and the fracture risk is calculated, and the regression equation is established using the partial least squares method to obtain the quantitative relationship between the spatial distribution of scattered light intensity and the fiber bundle fracture risk.

8. The method according to claim 1, characterized in that The quantitative index of the fiber bundle breakage risk is compared with a preset threshold value. If the preset threshold value is exceeded, the quality of the recycled carbon fiber bundle is judged to be unqualified and there is a risk of breakage, and it needs to be discarded. Otherwise, the quality of the carbon fiber bundle is judged to be qualified and can be used for recycling, including: According to the quantitative index of fiber bundle fracture risk, the minimum and maximum normalization processing was performed, and the index distribution curve was obtained by using the kernel density estimation method; Calculate the mean and standard deviation of the indicator distribution curve, and obtain the risk level determination area through the standard deviation interval; Select detection points on the fiber bundle surface, use the sliding window method to extract local features from the detection points, and obtain feature statistics through feature mean and variance; The characteristic statistics are subjected to self-service resampling to obtain multiple groups of sample data, and the upper and lower limits of the confidence interval are calculated through the sample data. If the upper limit of the confidence interval exceeds the high-risk threshold, it is marked as a defective product, and if the lower limit of the confidence interval is lower than the low-risk threshold, it is marked as a qualified product.

Images