Carbon fiber quality analysis method in combination with surface topography

Abstract

The invention provides a carbon fiber quality analysis method combined with surface topography, which comprises the following steps: acquiring the surface topography of recycled carbon fibers by adopting a scanning electron microscope, identifying microstructure images of the surfaces of the carbon fibers at different recycling stages, and judging the stripping degree of the surfaces of the carbon fibers according to the layered stripping characteristics of the surfaces of the carbon fibers in the images; identifying the axial tensile strength corresponding to the carbon fibers with different stripping depths through a tensile test, and analyzing the correlation between the surface stripping depths of the carbon fibers and the axial tensile strength; key indexes representing the quality of the carbon fibers are screened out according to the correlation between the surface stripping depth and the axial tensile strength of the carbon fibers and the mapping relation between the stripping depth and the graphite lattice orientation degree; and by taking the accuracy rate of quality grading of the recycled carbon fibers as an optimization target, optimizing parameters of the support vector machine classifier by adopting a genetic algorithm to obtain an optimized support vector machine classifier.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a carbon fiber quality analysis method combined with surface morphology. Background Art

[0002] Fiber materials are key structural materials in modern industry, and their mechanical properties are crucial for aerospace, automotive, and composite materials. The axial tensile strength of fibers directly determines their reliability under high-load conditions, making in-depth research on the relationship between fiber microstructure and properties of these materials of strategic importance. However, existing research methods for analyzing the evolution of graphite lattice orientation on the fiber surface are often limited to static characterization, lacking a systematic analysis of the dynamic debonding process. In particular, when the fiber surface undergoes layered debonding due to mechanical or chemical reactions, the correlation between debonding depth and lattice orientation, and its impact on tensile strength, remains underdeveloped. This limitation results in inaccurate fiber performance prediction models, which in turn impacts the development of quality grading standards for high-performance fibers. A key challenge lies in accurately quantifying the interaction between debonding depth and graphite lattice orientation, and how this interaction further influences the fiber's axial tensile strength. First, increasing debonding depth can lead to disordered surface lattice structure, reducing the overall stiffness of the fiber, but the specific mechanisms of this evolution remain unclear. Second, the dynamic evolution of orientation during debonding is difficult to capture in real time, and existing detection technologies struggle to achieve both high resolution and large-scale analysis. These factors make the degradation mechanism of fiber mechanical properties after stripping unclear, limiting the scientific basis for quality grading standards. Therefore, systematically revealing the correlation between fiber surface stripping depth and graphite lattice orientation, and clarifying its specific impact on axial tensile strength, has become a key issue in developing high-precision quality grading standards. Summary of the Invention

[0003] The present invention provides a carbon fiber quality analysis method combined with surface morphology, which mainly includes:

[0004] Scanning electron microscopy was used to obtain the surface morphology of the recycled carbon fiber, identify the microstructure images of the carbon fiber surface at different recycling stages, and determine the degree of carbon fiber surface peeling based on the layered peeling characteristics of the carbon fiber surface in the image;

[0005] Determine the orientation of the graphite lattice on the surface of the recycled carbon fiber, obtain the orientation diffraction patterns of the graphite lattice on the surface of the carbon fiber at different recycling stages, perform background subtraction and normalization on the orientation diffraction patterns, obtain the change pattern of the orientation diffraction peak parameters in different recycling stages, and determine the change pattern of the orientation of the graphite lattice on the surface of the carbon fiber;

[0006] According to the variation law of carbon fiber surface peeling degree and graphite lattice orientation, an exponential function model between carbon fiber surface peeling depth and graphite lattice orientation was established. By fitting the exponential function model, the mapping relationship between peeling depth and graphite lattice orientation was simulated.

[0007] The axial tensile strength of carbon fibers at different peeling depths was identified through tensile testing, and the correlation between the peeling depth of the carbon fiber surface and the axial tensile strength was analyzed.

[0008] The key indicators characterizing carbon fiber quality were screened out through the correlation between carbon fiber surface peeling depth and axial tensile strength, as well as the mapping relationship between peeling depth and graphite lattice orientation.

[0009] The key indicators of recycled carbon fiber are trained using a support vector machine algorithm. The axial tensile strength of the recycled carbon fiber is predicted using a support vector machine classifier. The quality of the recycled carbon fiber is graded according to the axial tensile strength, and a recycled carbon fiber quality grading standard is obtained.

[0010] Taking the accuracy of recycled carbon fiber quality classification as the optimization goal, the genetic algorithm was used to optimize the parameters of the support vector machine classifier to obtain the optimized support vector machine classifier.

[0011] Furthermore, a scanning electron microscope was used to capture the surface morphology of recycled carbon fibers, identifying microstructural images of carbon fibers at different recycling stages. The degree of surface delamination was determined based on the layered delamination characteristics of the carbon fibers in the images. This process involved calculating surface roughness parameters based on the grayscale distribution curve in the recycled carbon fiber sample surface morphology images, performing edge enhancement on the image by establishing a surface grayscale histogram matrix, and extracting a set of coordinates of microstructural feature points on the carbon fiber surface from the enhanced image. The minimum enclosing rectangle algorithm was used to calculate the orientation angle and area of each feature point region. A surface morphology feature vector matrix was constructed based on the orientation angle and area, and the feature vectors were classified using a density clustering method. Based on the feature vector classification results, the layered structure of the carbon fiber surface was identified. The Sobel operator was used to calculate the grayscale gradient between adjacent layers, and a quantitative index of the interlayer delamination degree was obtained by setting a gradient threshold. Morphological processing was performed on the identified layered structure regions using opening and closing operations, setting the structural element size to a 3×3 pixel matrix, and extracting the coordinate sequence of the carbon fiber surface wrinkle contours. The least squares method was used to fit a cubic spline curve to the wrinkle contour coordinate sequence. The curvature of the fitted curve at the sampling point was calculated, and the curvature change rate was used to obtain a quantitative indicator of the surface wrinkle degree. The characteristic points of the carbon fiber fracture edge were extracted based on the surface wrinkle contour. The carbon fiber fracture contour was constructed using the Canny edge detection operator. The degree of carbon fiber fracture was determined based on the mean fiber length in a standard carbon fiber sample image library.

[0012] Furthermore, the orientation of the graphite lattice on the surface of the recycled carbon fiber is measured, the orientation diffraction patterns of the graphite lattice on the surface of the carbon fiber at different recycling stages are obtained, the orientation diffraction patterns are background-subtracted and normalized, the variation pattern of the orientation diffraction peak parameters in different recycling stages is obtained, and the variation pattern of the orientation of the graphite lattice on the surface of the carbon fiber is determined, including: obtaining the original diffraction intensity matrix according to the diffraction point distribution in the diffraction pattern of the graphite lattice on the surface of the recycled carbon fiber, performing frequency domain decomposition on the diffraction intensity matrix using Fourier transform, and obtaining a frequency domain signal matrix. The power spectrum density distribution is calculated for the frequency domain signal matrix, the background noise area is identified based on the power spectrum density threshold, the background noise is suppressed using Gaussian filtering, and the diffraction peak signal matrix after noise reduction is obtained. The intensity distribution curve is constructed for the diffraction peak signal matrix after noise reduction, the diffraction peak profile is fitted using the least squares method, the peak shape symmetry is calculated based on the ratio of the curve areas on the left and right sides of the peak position, and the standardized diffraction peak shape characteristic value is obtained through normalization. Based on the standardized diffraction peak shape characteristics, a sequence of relative intensity ratios between adjacent diffraction peaks was extracted. The intensity ratio distribution was fitted using a binormal distribution function, and the graphite lattice orientation angle distribution function was calculated using the normal distribution parameters. The angle standard deviation of the graphite lattice orientation angle distribution function was calculated, and a dispersion reference value was set based on the angle standard deviation of a standard carbon fiber sample. The orientation quantification parameter was obtained by comparing the measured value with the reference value. Based on the orientation quantification parameter, a database of samples from different recycling stages was established. The sample data was grouped using a hierarchical clustering method, and the variation pattern of the graphite lattice orientation on the carbon fiber surface was determined through between-group analysis of variance.

[0013] Furthermore, based on the variation in carbon fiber surface peeling degree and graphite lattice orientation, an exponential function model was established to correlate the peeling depth with the graphite lattice orientation. By fitting the exponential function model, the mapping relationship between peeling depth and graphite lattice orientation was simulated. The method involved calculating a peeling depth distribution map based on the grayscale gradient matrix in the carbon fiber surface topography image, performing a three-layer decomposition of the peeling depth distribution map using a multiscale wavelet transform, and reconstructing the peeling depth feature sequence through wavelet coefficient reconstruction. An orientation distribution map was calculated based on the graphite lattice diffraction peak intensity matrix in the diffraction pattern. Features were extracted from the orientation distribution map using singular value decomposition, and an orientation feature sequence was constructed based on the principal eigenvalues. A corresponding point set was established based on the peeling depth feature sequence and the orientation feature sequence. The sparsely sampled points were numerically interpolated using a cubic spline function, and a densely sampled data set was obtained using a local weighted average method. For the densely sampled data set, an exponential function model y = aexp(-bx) was constructed: where y represents the graphite lattice orientation, x represents the peeling depth, and a and b are unknown coefficients. Initial values of the coefficients were determined using the least squares method. The sum of squared residuals was calculated based on the exponential function model, and the model coefficients were iteratively optimized using gradient descent. The optimization termination condition was determined by setting a convergence threshold to obtain the optimized exponential function model parameters. Cross-validation was performed on the optimized exponential function model, using a holdout method to randomly partition the training and validation sets. The training set was resampled using the bootstrap method, and confidence intervals for the model coefficients were calculated based on the resampled data.

[0014] Furthermore, tensile testing was used to identify the axial tensile strength of carbon fibers with different peel depths and analyze the correlation between the surface peel depth and the axial tensile strength. The method involved: Based on the raw stress and strain data collected by the tensile test sensor, the signal was decomposed in the frequency domain using Fourier transform. High-frequency noise was removed using a power spectral density threshold, and a smoothed stress-strain curve was obtained using an inverse Fourier transform. The smoothed stress-strain curve was numerically differentiated, and the slope of the stress-strain curve was calculated using a five-point difference method. The yield strength was determined at the point of maximum slope, and the axial tensile strength of the carbon fiber was obtained based on the stress value at the fracture point. A grayscale gradient matrix was constructed based on the carbon fiber surface topography image. The peeled region was segmented using a region growing algorithm. The peel boundary contours were extracted using an edge detection operator, and the peel depth values were calculated based on the perpendicular distance between the contours. Statistical analysis of the peel depth value sequence was performed, and a depth distribution function was constructed using a histogram method. The peel depth characteristic parameters were obtained using a density estimation method, and the normalized depth parameters were obtained through normalization. A sample data matrix was established based on the axial tensile strength and normalized depth parameters. A strength prediction function was constructed using support vector regression. The kernel function parameters were optimized through cross-validation, and the optimal mapping relationship was selected based on the validation error. The Pearson correlation coefficient was calculated for the optimal mapping relationship, the eigenvector was extracted using principal component analysis, the main influencing factors were determined through eigenvalue decomposition, and the quantitative relationship between peeling depth and axial tensile strength was obtained based on the factor weights.

[0015] Furthermore, through the correlation between carbon fiber surface peeling depth and axial tensile strength and the mapping relationship between peeling depth and graphite lattice orientation, key indicators characterizing carbon fiber quality were screened out, including: constructing a parameter matrix based on the carbon fiber surface peeling depth and axial tensile strength data set, calculating the correlation coefficient using the Pearson correlation analysis method, screening significantly correlated parameters using a significance level of 0.05, and constructing a feature data set based on the significantly correlated parameters. The peeling depth and graphite lattice orientation mapping data set was standardized, the principal component analysis method was used to calculate the eigenvalues and eigenvectors, the main components were screened by cumulative variance contribution rate greater than 0.85, and the parameter importance sequence was constructed based on the principal component weights. An evaluation matrix was constructed based on the feature data set and parameter importance sequence, the parameter correlation coefficient was calculated using the gray correlation algorithm, strongly correlated indicators were screened by setting the correlation coefficient threshold to 0.8, and the independence of the indicators was judged based on the variance inflation factor being less than 5. A regression equation was constructed for the strongly correlated indicators, the confidence interval of the regression coefficient was calculated using the leave-one-out cross-validation method, the parameter stability was judged by the interval width, and a reliable parameter set was screened based on the stability index. Single-factor sensitivity was calculated based on a reliable parameter set, and a parameter importance score was constructed using the random forest regression method. Key parameters were screened by having an importance score greater than the mean, and parameter reliability was verified based on bootstrap sampling. A comprehensive evaluation function was established for the selected key parameters, and a weighted summation method was used to calculate the parameter comprehensive score. The set of carbon fiber quality characterization indicators was determined by setting a score threshold.

[0016] Furthermore, a support vector machine (SVM) algorithm was used to train key indicators of recycled carbon fiber. The SVM classifier was used to predict the axial tensile strength of recycled carbon fiber. The quality of the recycled carbon fiber was then graded based on the axial tensile strength, resulting in a grading standard for the recycled carbon fiber quality. This approach involved constructing a feature matrix based on three key indicators: peel depth, orientation, and tensile strength. The data was normalized to the interval [0, 1] using the maximum-minimum normalization method. The sample probability distribution was obtained using kernel density estimation. A training set and validation set were divided into two groups using a stratified random sampling method with an 8:2 ratio. A three-class SVM classifier was constructed for the training set data, using a radial basis kernel function. A grid search was used to optimize the kernel parameter γ in the interval [0.01, 100] and the penalty factor C in the interval [0.1, 1000]. The optimal parameter combination was selected based on the five-fold cross-validation error. The SVM classifier was trained based on the optimal parameter combination, using a one-vs-all strategy for multi-class training. The optimal separating hyperplane was solved using a sequential minimum optimization algorithm, and a nonlinear classification boundary was constructed using the kernel technique. The validation set prediction results were calculated for the trained classifier. The true positive rate and false positive rate were calculated using a confusion matrix. The optimal classification threshold was determined through receiver operating characteristic curve analysis, and a threshold combination was selected based on the Youden index. Axial tensile strength grading intervals were established based on the validation set prediction results. The strength values were divided into three levels using the K-means clustering method. The strength boundary values were determined using the inter-group variance maximization criterion, and a quality grading function was established based on these boundary values. The stability of the quality grading function was verified, and a validation dataset was generated using the bootstrap resampling method. The grading performance was evaluated by calculating the classification accuracy, precision, and recall rates. The grading criteria were determined based on the evaluation results.

[0017] Furthermore, with the accuracy of recycled carbon fiber quality classification as the optimization goal, the genetic algorithm is used to optimize the parameters of the support vector machine classifier to obtain the optimized support vector machine classifier, including: constructing an optimization parameter sequence according to the kernel function parameters and penalty factors of the support vector machine classifier, encoding the parameters using a 32-bit binary code, and setting the kernel function parameter encoding range.

[0018] A parameter encoding rule was generated using the range [0.01, 100] and the penalty factor encoding range [0.1, 1000]. A fitness function was constructed based on cross-validation accuracy. The genetic algorithm's initial population was constructed based on the parameter encoding sequence. A population size of 100 was used, and high-quality individuals were selected using a roulette wheel selection method. The top 20% of elite individuals were retained based on their fitness scores. Genetic operations were performed based on the elite individual sequence, using a single-point crossover method with a crossover probability of 0.8 and a bit-flip method with a mutation probability of 0.1. Individuals in the population were updated using an elite retention strategy. Individual fitness values were calculated for the updated population, using a maximum fitness change rate of less than 0.001 for 50 consecutive generations as the convergence criterion. Support vector machine parameters were obtained by decoding the optimal individual, and the classifier was retrained based on these parameters. The validation set classification performance of the trained classifier was calculated, using a confusion matrix to calculate the true positive rate and false positive rate. The optimization effect was evaluated by calculating classification accuracy, precision, and recall. The optimal classifier parameter combination was determined based on these evaluation metrics. A quality grading model was constructed based on the optimal parameter combination, the cross-validation method was used to verify the model stability, the parameter reliability was evaluated through confidence interval analysis, and the optimized carbon fiber quality grading model was obtained based on the evaluation results.

[0019] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0020] The present invention discloses a method for analyzing carbon fiber quality in combination with surface morphology. The method observes the surface morphology of carbon fibers using a scanning electron microscope, analyzes changes in graphite lattice orientation, and establishes an exponential function model of peeling depth and orientation. Combined with tensile test data, key indicators characterizing carbon fiber quality are screened out, a support vector machine grading model is constructed, and a genetic algorithm is used to optimize model parameters. The present invention achieves accurate grading of recycled carbon fiber quality, provides a scientific basis for carbon fiber recycling and reuse, helps to improve the utilization value of recycled carbon fibers, promotes the recycling of carbon fiber composite materials, and has important economic and environmental benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The present invention is a flow chart of a carbon fiber quality analysis method combined with surface morphology.

[0022] Figure 2 Schematic diagram of a carbon fiber quality analysis method combined with surface morphology according to the present invention. DETAILED DESCRIPTION

[0023] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] In this embodiment, a carbon fiber quality analysis method based on surface morphology may specifically include:

[0025] The embodiments of the present application do not impose excessive restrictions on the specific type of device. It can be a high-resolution microscope, a computing terminal, or other device that supports data processing, which is determined according to the actual scenario.

[0026] In this example, the analysis target is recycled carbon fiber, with the goal of quantifying the degree of debonding and changes in mechanical properties based on surface morphology and microstructural features, thereby enabling quality grading. By capturing surface topography images, extracting microfeatures, and measuring lattice orientation, the system reveals a correlation between debonding depth and axial tensile strength.

[0027] Figure 1 The implementation process of the carbon fiber quality analysis method combined with surface morphology provided in Example 1 of the present application is shown and is detailed as follows:

[0028] S101 uses high-resolution microscopy equipment to collect surface morphology images of recycled carbon fibers, extract microstructural features, and determine the degree of peeling of the carbon fiber surface based on the feature distribution.

[0029] In the embodiments of the present application, a high-resolution microscope, such as a scanning electron microscope, is used to capture images of the surface topography of the recycled carbon fibers. During the acquisition process, high-resolution images are obtained for carbon fiber samples at different stages of recycling to capture detailed changes in the surface microstructure. The image acquisition resolution can be set according to actual needs, for example, 1024×1024 pixels or higher can be used to ensure accurate feature extraction.

[0030] S1011 generates a grayscale distribution matrix based on the surface topography image and extracts microstructure feature points through edge enhancement processing.

[0031] In an embodiment of the present application, the collected surface morphology image is preprocessed to generate a grayscale distribution matrix to quantify the pixel intensity distribution of the image. The grayscale value range is set to 0 to 255, where the grayscale distribution of the recycled carbon fiber varies due to different degrees of peeling. The image is edge enhanced using the Gaussian Laplace operator to highlight microstructural feature points, such as wrinkles, cracks, or peeling edges. After enhancement, a set of feature point coordinates is extracted for subsequent feature analysis. For a group of typical samples, the grayscale value fluctuates significantly in the range of 100 to 160, and the feature points are more prominent after edge enhancement. The coordinate set contains approximately 500 to 800 feature points.

[0032] S1012 constructs a surface morphology feature vector based on the feature point coordinate set and quantifies the degree of peeling through cluster analysis.

[0033] In an embodiment of the present application, the minimum enclosing rectangle algorithm is used to calculate the orientation angle and area of each feature area for the extracted feature point coordinate set. The orientation angle reflects the direction of carbon fiber arrangement, and the area reflects the size of the peeling area. Taking a group of recycled carbon fiber samples as an example, the orientation angle is distributed between 70 and 90 degrees, and the area is distributed between 60 and 100 square pixels, indicating that the fiber surface structure is relatively regular. A feature vector matrix is constructed based on the orientation angle and area, and each vector contains two components: angle and area. Based on the feature vector matrix, a density clustering method is used for classification, and the cluster radius is set to 8, resulting in 3 to 4 feature categories. Based on the classification results, the Sobel operator is used to calculate the grayscale gradient value of adjacent feature areas, and the gradient threshold range is set to 25 to 45. When the gradient value exceeds 45, it is determined to be a severely peeling area. The peeling area is further subjected to morphological processing, and 3×3 pixel structure elements are used for opening and closing operations to remove noise and fill holes to obtain a continuous sequence of wrinkle contour line coordinates. For the contour line coordinates, the curvature change rate is calculated by cubic spline curve fitting as a quantitative indicator of the degree of peeling. For lightly recovered samples, the curvature change rate ranges from 0.01 to 0.04; for heavily recovered samples, the curvature change rate can reach 0.07 to 0.10.

[0034] S1013 extracts the fracture edge features based on the wrinkle contours and determines the fracture degree based on the standard sample data.

[0035] In an embodiment of the present application, the Canny edge detection operator is used to extract the carbon fiber fracture edge features from the wrinkle contour coordinate sequence and construct the fracture contour. Based on the mean fiber length in the standard carbon fiber sample image library, for example, 7500 microns, the ratio of the fracture contour length to the standard length is calculated. If the ratio is less than 0.4, it is determined to be a severe fracture. Analysis found that the degree of peeling is positively correlated with the degree of fracture. When the quantitative index of the degree of peeling exceeds 0.5, the probability of fracture increases significantly. This correlation provides an important basis for subsequent quality grading.

[0036] In the examples of this application, surface morphology images were collected using high-resolution microscopy equipment, and combined with grayscale distribution analysis, edge enhancement, and feature clustering, the degree of stripping of recycled carbon fibers was accurately quantified. Compared to traditional methods, this application can dynamically capture surface microstructural changes and extract multidimensional feature vectors, significantly improving the accuracy and reliability of stripping assessments and laying the foundation for subsequent lattice orientation and mechanical property analysis. Through the comprehensive analysis of wrinkle contours and fracture edge features, a scientific microscopic basis is provided for carbon fiber quality grading.

[0037] S102 determines the orientation of the graphite lattice on the surface of the recycled carbon fiber through lattice diffraction analysis, obtains diffraction pattern data at different recycling stages, and determines the change pattern of the graphite lattice orientation through signal processing and fitting analysis.

[0038] In an embodiment of the present application, an X-ray diffraction device or other lattice analysis instrument is used to collect the diffraction pattern of the graphite lattice on the surface of the recycled carbon fiber to determine the orientation. For carbon fiber samples at different recovery stages, high-resolution diffraction patterns are obtained and the intensity distribution of the diffraction signal is recorded. The resolution of the pattern acquisition can be set according to actual needs, for example, 512×512 pixels, to ensure the accuracy of signal processing. Through background subtraction and normalization processing, the characteristic parameters of the diffraction peak are extracted, and then the dynamic change law of the graphite lattice orientation is analyzed to provide a microscopic basis for quality grading.

[0039] like Figure 2 , S1021 generates the original intensity matrix according to the diffraction pattern, and extracts clear diffraction peak signals through frequency domain transformation and noise suppression.

[0040] In an embodiment of the present application, the original intensity matrix is extracted from the diffraction pattern, and the matrix records the diffraction signal intensity of each pixel. Taking a group of typical samples as an example, the matrix size is 512×512 pixels, and the main diffraction peaks are distributed in the concentric ring area with a radius of 70 to 110 pixels. The intensity matrix is fast Fourier transformed and converted into a frequency domain signal matrix. The signal energy is mainly concentrated in the frequency range of 0.04 to 0.12. The power spectral density distribution of the frequency domain signal is further calculated, and the density threshold is set to 0.015. The area below the threshold is determined to be background noise. A Gaussian filter is used for noise suppression, and the filter window size is 7×7 pixels to obtain the diffraction peak signal matrix after noise reduction. The noise reduction process significantly improves the signal-to-noise ratio. For example, after noise reduction, the power spectral density value of the main diffraction peak is increased from 0.04 to 0.06, and the density value of the background noise area is reduced to below 0.005.

[0041] S1022 constructs an intensity distribution curve based on the noise reduction signal and extracts the standardized diffraction peak characteristics through curve fitting.

[0042] In an embodiment of the present application, an intensity distribution curve is generated for the diffraction peak signal matrix after noise reduction to reflect the profile characteristics of the diffraction peak. The least squares method is used to fit the curve, and 60 data points on both sides of the peak are selected to calculate the peak symmetry and peak area ratio. Taking a group of recycled carbon fiber samples as an example, the peak symmetry of the main diffraction peak is 0.90, indicating that the graphite lattice structure is relatively regular. The fitting curve is normalized to obtain a standardized diffraction peak shape characteristic value, which ranges from 0 to 1, which is convenient for cross-sample comparison. The relative intensity ratio sequence between adjacent diffraction peaks is further extracted, and the sequence reflects the orientation distribution law of the graphite lattice. For complete carbon fibers, the intensity ratio sequence shows periodic fluctuations, while the amplitude of the sequence fluctuation of the recycled sample increases, indicating that the orientation degree is reduced.

[0043] S1023 fits the intensity ratio sequence through binormal distribution, calculates the orientation angle distribution, and quantifies the orientation degree parameter.

[0044] In an embodiment of the present application, a double normal distribution function is used to fit the intensity ratio sequence to extract the orientation angle distribution function of the graphite lattice. During the fitting process, the means of the two normal distributions are set to correspond to the main orientation directions of the graphite lattice, such as 40 degrees and 130 degrees. For a group of samples, the fitting results show that the mean of the first normal distribution is 42 degrees and the standard deviation is 8 degrees; the mean of the second normal distribution is 132 degrees and the standard deviation is 9 degrees. The orientation angle standard deviation is calculated according to the distribution function as the core indicator of orientation quantification. The angle standard deviation of the standard carbon fiber sample is usually within 8 degrees, which is set as the reference value. The ratio of the standard deviation of the measured sample to the reference value is defined as the orientation quantification parameter. The closer the parameter is to 1, the higher the orientation degree. The measured data show that the orientation degree parameter of the lightly recovered sample is between 0.85 and 0.95, the moderately recovered sample is between 0.65 and 0.85, and the heavily recovered sample is below 0.65.

[0045] S1024 establishes a sample database based on the orientation parameters and determines the orientation variation pattern through cluster analysis.

[0046] In an embodiment of the present application, a sample database of different recovery stages is constructed according to the orientation quantification parameters, and the diffraction peak characteristics and orientation data of each sample are recorded. The sample data are grouped by a hierarchical clustering method, and the cluster distance threshold is set to 0.1 to obtain 3 to 5 orientation levels. The significant differences in orientation between the levels are verified by inter-group variance analysis. Taking a group of databases as an example, the mean orientation parameter of the lightly recovered samples is 0.90, and the inter-group variance is 0.02; the mean value of the heavily recovered samples is 0.60, and the inter-group variance is 0.03, indicating that the orientation difference is statistically significant. Further analysis of the correlation between orientation and other parameters shows that for every 0.1 decrease in the orientation parameter, the half-height width of the diffraction peak increases by 12% to 18%, and the peak symmetry decreases by 0.04 to 0.07, reflecting the systematic degradation of the lattice structure.

[0047] In the examples of the present application, the precise determination of the orientation of the graphite lattice of the recycled carbon fiber is achieved through lattice diffraction analysis and multi-stage signal processing. Compared with traditional methods, this application significantly improves the accuracy and robustness of orientation quantification through frequency domain transformation and binormal distribution fitting. The dynamic change law of the orientation parameter provides reliable data support for the subsequent correlation analysis between stripping depth and mechanical properties. Finally, through database construction and cluster analysis, the evolution trend of the graphite lattice orientation during the recycling process is systematically revealed, providing a scientific microscopic basis for carbon fiber quality grading.

[0048] S103 integrates the carbon fiber surface peeling degree and graphite lattice orientation data to construct an exponential function model between peeling depth and orientation, and quantifies the mapping relationship between the two.

[0049] In the examples of this application, mathematical modeling was used to establish a correlation model between the degree of exfoliation and graphite lattice orientation data obtained previously. An exponential function was used to simulate the effect of exfoliation depth on orientation. Feature extraction and numerical interpolation techniques were used to enhance data density, and model accuracy was ensured through iterative optimization. Cross-validation was used to confirm the model's stability and predictive power, providing reliable theoretical support for carbon fiber quality grading.

[0050] S1031 extracts the peeling depth distribution from the surface topography image and generates a peeling depth feature sequence through multi-scale wavelet transform.

[0051] In an embodiment of the present application, a peeling depth distribution map is calculated based on the grayscale gradient matrix of the carbon fiber surface morphology image to reflect the spatial characteristics of the peeling area. Taking a 512×512 pixel image as an example, the grayscale gradient value of the peeling area is usually between 60 and 160, while that of the non-peeling area is less than 25. The distribution map is decomposed into three layers using a multi-scale wavelet transform, and the Daubechies wavelet basis function is used to capture the peeling characteristics of different scales. The first layer of decomposition extracts micro-scale features, with coefficients ranging from 0.08 to 0.25; the second layer of decomposition reflects the meso-scale features, with coefficients ranging from 0.15 to 0.45; the third layer of decomposition corresponds to the macro-scale features, with coefficients ranging from 0.35 to 0.75. The peeling depth feature sequence is reconstructed by wavelet coefficients. The sequence contains about 800 data points, which retains the spatial distribution information of the peeling depth. Compared with traditional methods, wavelet transform can effectively separate multi-scale features, enhance the quantification accuracy of peeling depth, and provide high-quality input data for model construction.

[0052] S1032 extracts the orientation distribution from the diffraction pattern and generates an orientation characteristic sequence through singular value decomposition.

[0053] In an embodiment of the present application, an orientation distribution map is calculated based on the intensity matrix of the graphite lattice diffraction pattern to reflect the spatial heterogeneity of the lattice orientation. Taking a 2048×2048 pixel pattern as an example, the main diffraction peak intensity values are distributed between 1200 and 4500, showing anisotropic characteristics. The singular value decomposition method is used to extract features from the distribution map. The decomposition results show that the first three eigenvalues account for more than 88% of the total energy, of which the first eigenvalue is between 2200 and 3200, reflecting the average orientation information; the second and third eigenvalues reflect the orientation discreteness, with values between 800 and 1500. An orientation feature sequence is constructed based on the main eigenvalues. The sequence contains about 600 data points, which retains the main distribution characteristics of the orientation. This method reduces the data complexity through dimensionality reduction processing, while retaining the core information of the lattice orientation, providing a reliable basis for the construction of the mapping model.

[0054] S1033 generates a densely sampled data set through numerical interpolation and constructs an exponential function model to fit the relationship between peeling depth and orientation.

[0055] In an embodiment of the present application, a point set corresponding to the space is constructed based on the stripping depth feature sequence and the orientation feature sequence, and the initial point set contains about 250 sampling points. The point set is interpolated using a cubic spline function to generate about 1200 dense sampling points. The local weighted average method is used in the interpolation process, and the weight decays exponentially with distance. The attenuation coefficient is set to 0.1 to ensure that the interpolation result is smooth and the local features are preserved intact. For the densely sampled data set, an exponential function model y=a*exp(-b*x) is constructed, where y represents the orientation, x represents the stripping depth, and a and b are the initial orientation and attenuation coefficient, respectively. The initial values of the coefficients are calculated using the least squares method, for example, the initial value of a is 0.95, and the initial value of b is 0.15. The coefficients are further optimized by the gradient descent method, the iteration step size is set to 0.001, the convergence threshold is set to 0.0005, and after about 80 iterations, the residual sum of squares converges to below 0.0008, obtaining the optimized model parameters. Measured data show that the degree of orientation decays exponentially with every 10 microns increase in peeling depth. When the light peeling depth is within 15 microns, the degree of orientation decreases by about 15%; when the heavy peeling depth exceeds 40 microns, the degree of orientation can decrease by up to 55%.

[0056] S1034 evaluates the stability and prediction accuracy of the exponential function model through cross-validation and resampling methods.

[0057] In an embodiment of the present application, the optimized exponential function model was cross-validated, and the data set was divided into 80% training set and 20% validation set by the holdout method. The training set was resampled 1200 times by the Bootstrap method, and the 95% confidence intervals of the model parameters a and b were calculated. The interval width of a was 0.04 to 0.07, and the interval width of b was 0.01 to 0.03, indicating that the model parameters have high stability. The prediction error of the model was further evaluated, and the in-sample fitting error was controlled within 4%, and the validation set prediction error was within 7%. In another implementation of the exponential function model y=a*exp(-b*x), the exponential form of the model is not only computationally efficient and suitable for processing large-scale data, but also has clear physical meanings in the parameters: a reflects the orientation in the unpeeled state, and b quantifies the intensity of the influence of the peeling depth on the orientation. This quantitative mapping relationship provides a scientific basis for evaluating the microscopic damage of recycled carbon fiber and significantly improves the accuracy of quality analysis.

[0058] In the examples of this application, a high-quality dataset of peel depth and orientation was generated through multi-scale feature extraction and numerical interpolation. The exponential function model constructed accurately describes the nonlinear relationship between the two. Compared with traditional methods, this application enhances the robustness of feature extraction through wavelet transform and singular value decomposition, and improves the model's fitting accuracy through gradient descent optimization. Cross-validation confirmed the stability and reliability of the model, providing a solid data foundation for mechanical property analysis.

[0059] S104 quantifies the axial tensile strength of carbon fiber at different peeling depths through tensile testing, and combines surface morphology analysis and regression modeling to reveal the quantitative relationship between peeling depth and tensile strength.

[0060] In this application's examples, high-precision tensile testing equipment was used to collect stress and strain data for recycled carbon fibers at different peel depths. Signal processing and feature extraction were used to determine the axial tensile strength. Furthermore, peel depth characteristics were analyzed using surface topography images. Support vector regression and principal component analysis were used to establish a mapping model between peel depth and tensile strength. This accurately quantified the correlation between the two, providing a key indicator for quality grading.

[0061] S1041 processes tensile test data and extracts yield strength and axial tensile strength using Fourier transform and numerical differentiation methods.

[0062] In an embodiment of the present application, raw data collected by a stress and strain sensor is received, and the mechanical response of the carbon fiber during the stretching process is recorded at a sampling frequency of 1000 Hz. The raw data often contains high-frequency noise, which affects the accuracy of the strength calculation. The signal is decomposed into the frequency domain using a fast Fourier transform, and it is found that the main mechanical signal energy is concentrated in the range of 0 to 25 Hz, and the signal above 40 Hz is mainly the noise component. The power spectrum density threshold is set to 0.008, and after filtering out the high-frequency noise, a smoothed stress-strain curve is generated by an inverse Fourier transform. For the smoothed curve, the slope is calculated using the five-point difference method. Five consecutive data points are selected to reduce numerical fluctuations. The maximum slope point corresponds to the yield strength, and the stress value at the breaking point is defined as the axial tensile strength. Taking a group of samples as an example, at a strain rate of 0.01 per second, the yield strength is approximately 3200 MPa, and the tensile strength is between 3400 and 3600 MPa. The curve shows linear elastic characteristics until it breaks near a strain of 0.018. Compared with traditional filtering methods, the Fourier transform combined with the five-point difference significantly improves the accuracy and stability of strength calculation.

[0063] S1042 extracts peeling depth features based on surface topography images and quantifies the peeling degree using region growing and edge detection algorithms.

[0064] In an embodiment of the present application, a grayscale gradient matrix is constructed based on a surface topography image of 2048×2048 pixels to reflect the spatial distribution of the peeling area. The grayscale gradient value of the peeling area is usually between 55 and 140, and that of the non-peeling area is less than 20. The peeling area is segmented using a regional growing algorithm, with pixels with a grayscale gradient greater than 60 as seed points, and extended to areas where the grayscale difference between adjacent pixels is less than 8 to generate a complete peeling area boundary. The Canny edge detection operator is further used to extract the boundary contour line, and a sequence of peeling depth values is generated by calculating the vertical distance between the contour line and the fiber surface. Taking a group of samples as an example, the depth sequence contains about 60 sampling points. The depth of the lightly peeled samples is concentrated in the range of 8 to 18 microns, the moderate peeling is 20 to 40 microns, and the severe peeling exceeds 45 microns. The kernel density estimation method is used to construct a depth distribution function and smoothness distribution characteristics to facilitate subsequent statistical analysis. This method significantly improves the spatial resolution and accuracy of peeling depth quantification by combining regional segmentation and edge detection.

[0065] S1043A quantitative mapping model between peeling depth and axial tensile strength was established through support vector regression and principal component analysis.

[0066] In an embodiment of the present application, a sample data matrix is constructed based on the peeling depth sequence and axial tensile strength value. The matrix contains about 100 sample points, each of which includes a normalized depth parameter and a corresponding strength value. The support vector regression method is adopted, and the radial basis kernel function is selected to construct a strength prediction model. The kernel parameters and regularization parameters are optimized by five-fold cross validation. Taking a set of optimization results as an example, when the kernel parameter is 0.08, the root mean square error of the training set is 110 MPa, and the error of the validation set is 140 MPa. The prediction accuracy meets the engineering requirements. The Pearson correlation coefficient is further calculated, and the result is -0.90, indicating that the peeling depth and tensile strength are strongly negatively correlated. Principal component analysis is used to extract the eigenvectors of the data matrix, and it is found that the contribution rate of the first principal component is 87%. Its eigenvector mainly reflects the dominant influence of the peeling depth on the strength. Quantitative analysis shows that for every 10 microns increase in peeling depth, the tensile strength decreases by an average of 450 to 550 MPa, and the correlation is particularly significant when the depth is less than 40 microns. Compared with linear regression, support vector regression can better capture nonlinear relationships and improve the generalization ability of the model.

[0067] In the examples of this application, the effect of peel depth on axial tensile strength was precisely quantified through tensile testing and surface topography analysis. The application of Fourier transform and region growing algorithms improved the reliability of data processing and feature extraction, while support vector regression combined with principal component analysis ensured the accuracy and interpretability of the mapping model. The resulting quantitative relationship provides a highly reliable reference for evaluating the mechanical properties of recycled carbon fiber, significantly improving the scientific nature and efficiency of material screening.

[0068] S105 integrates the correlation analysis of carbon fiber surface peeling depth, axial tensile strength and graphite lattice orientation to screen key indicators that characterize the quality of recycled carbon fiber and build a comprehensive evaluation system.

[0069] In this application's examples, statistical analysis and machine learning methods were used to extract key quality indicators based on the mapping relationship between peel depth, tensile strength, and orientation. Correlation analysis, principal component analysis, and random forest regression were used to quantify the impact of each parameter on quality. Multiple validations were performed to ensure the stability and reliability of the indicators, forming a scientific quality characterization system.

[0070] S1051 used Pearson correlation analysis and principal component analysis to extract the significant correlation parameters and main component weights between peeling depth and mechanical properties.

[0071] In an embodiment of the present application, a parameter matrix is constructed based on the data sets of peeling depth, axial tensile strength and graphite lattice orientation, which contains about 120 sets of sample data. The correlation coefficients between the parameters are calculated using the Pearson correlation analysis method. The results show that the correlation coefficient between peeling depth and tensile strength is -0.91, and the correlation coefficient between peeling depth and orientation is -0.87, both of which are strongly negatively correlated at a significance level of 0.01. Parameters with an absolute value of the correlation coefficient greater than 0.8 are screened to construct a significantly correlated parameter set. For this parameter set, the principal component analysis method is used to calculate the eigenvalues and eigenvectors. The cumulative variance contribution rate of the first two principal components is 92%, of which the first principal component contribution rate is 78%, which mainly reflects the influence of peeling depth; the second principal component contribution rate is 14%, which is related to orientation and tensile strength. The main component weights are calculated based on the eigenvectors, for example, the peeling depth weight is 0.50 and the orientation is 0.30, ensuring that the main components cover the core change information of material properties. This method highlights the role of key parameters through dimensionality reduction processing and provides a data basis for subsequent indicator screening.

[0072] S1052 screens strongly correlated and independent quality indicators through grey relational analysis and variance inflation factor test.

[0073] In an embodiment of the present application, based on a set of significantly correlated parameters, a grey correlation algorithm is used to calculate the correlation coefficients between parameters to reveal potential correlation patterns. Peeling depth, orientation and tensile strength are used as reference sequences to calculate their correlation with material quality. The results show that the peeling depth correlation is 0.90, the orientation is 0.86, and the tensile strength is 0.82, all exceeding the threshold of 0.8, indicating that these parameters are closely related to quality. The variance inflation factor is further used to test parameter independence. The calculation results show that the maximum variance inflation factor is 3.5, which is much lower than the threshold of 5, proving that the multicollinearity between parameters is low and has good independence. Parameters with a correlation coefficient greater than 0.8 and a variance inflation factor less than 5 are screened to generate a set of strongly correlated indicators. This method ensures the high correlation and independence of the indicators by combining grey correlation and collinearity analysis, thereby improving the scientific nature of the screening results.

[0074] S1053 uses random forest regression and Bootstrap validation to construct parameter importance scores and screen key indicators.

[0075] In the embodiment of the present application, for a set of strongly correlated indicators, a random forest regression method is used to calculate the parameter importance score and quantify the contribution of each parameter to the quality of carbon fiber. A random forest model containing 500 decision trees is constructed, and the importance score is calculated by feature splitting gain. The results show that the stripping depth score is 0.48, the orientation is 0.34, and the tensile strength is 0.18, all higher than the mean of 0.25. The score stability is verified by 1000 Bootstrap resamplings. The coefficient of variation of the stripping depth score is 0.08, and the orientation is 0.09, indicating that the scoring results are highly reliable. Parameters with importance scores greater than the mean are screened, a key parameter set is generated, and the confidence interval of the regression coefficient is evaluated using the leave-one-out cross-validation method. For example, the 95% confidence interval of the stripping depth coefficient is [-0.68,-0.60], and the interval width is 0.08, showing high stability. Random forest regression accurately captures the complex relationship between parameters through nonlinear modeling, significantly improving the accuracy of indicator screening.

[0076] In the examples of this application, through multi-level analysis and verification, peel depth, orientation, and tensile strength were selected as key characterization indicators of carbon fiber quality. Compared with traditional methods, this application combines Pearson correlation analysis, gray correlation, and random forest regression to comprehensively quantify the importance of parameters and ensure the robustness of the results through bootstrap and cross-validation. These indicators not only reflect the degradation patterns of the material's microstructure and mechanical properties, but also provide high-quality input data for the construction of quality grading models.

[0077] S106 uses machine learning training on key indicators of recycled carbon fiber to build a support vector machine classification model to predict axial tensile strength and establish quality grading standards based on strength grading.

[0078] In this application's examples, key indicators such as peel depth, orientation, and tensile strength were used to construct and standardize a feature matrix. A classification model was trained using a support vector machine algorithm to categorize carbon fiber strength grades. Clustering and validation methods were used to optimize the grading boundaries, ensuring the scientific and stable nature of the grading standards and providing a reliable reference for the screening of high-performance carbon fibers.

[0079] S1061 normalizes key indicator data and estimates probability distribution to construct a training data set for the support vector machine classifier.

[0080] In an embodiment of the present application, a feature matrix is constructed based on peeling depth, orientation and tensile strength, and the matrix contains about 1200 samples, each of which includes three index values. The peeling depth ranges from 0 to 90 microns, the orientation ranges from 0.2 to 1.0, and the tensile strength ranges from 800 to 4200 MPa. The maximum and minimum value normalization method is adopted to normalize the data to the interval of 0 to 1 to eliminate dimensional differences. The probability distribution of normalized data is calculated by the kernel density estimation method, and it is found that the samples form three clustering areas in the feature space, corresponding to high, medium and low intensity levels respectively. The stratified random sampling method is used to divide the training set and the validation set into a ratio of 8:2. The training set contains 960 samples and the validation set contains 240 samples to ensure that the proportion of samples of each intensity level is balanced. This data preprocessing method enhances the training efficiency and generalization ability of the classification model through standardization and distribution estimation, and provides high-quality input for subsequent support vector machine training.

[0081] S1062 constructs a support vector machine classifier and determines the optimal classification boundary through grid search and sequential minimum optimization algorithm.

[0082] In an embodiment of the present application, a three-category support vector machine classifier is constructed for the training set data, and the radial basis kernel function is selected for nonlinear mapping. The radial basis kernel function maps the data to a high-dimensional space through a Gaussian kernel, the kernel parameter γ controls the kernel function width, and the penalty factor C balances the model complexity and error. The parameters are optimized by the grid search method, and the search is carried out within the γ range of 0.01 to 50 and the C range of 0.1 to 500. The optimal combination is selected based on the five-fold cross-validation error. The results show that when γ is 0.08 and C is 120, the cross-validation accuracy reaches 93.2%. A one-to-many strategy is adopted to decompose the three-category problem into three two-category problems, and each classifier distinguishes one intensity level from other levels. The optimal separating hyperplane is solved by the sequential minimum optimization algorithm, and the convergence threshold is set to 0.0008 and the maximum number of iterations is 1200 to ensure the convergence of the algorithm. After the training is completed, the accuracy of the classifier on the validation set is 91%, the accuracy of the high intensity level is 94%, and the medium and low intensity levels are 89% and 87% respectively. Compared with linear classifiers, radial basis kernel functions significantly improve the classification performance of nonlinear data.

[0083] S1063 uses K-means clustering to classify tensile strength grades, and verifies the stability of the grading standard through Bootstrap.

[0084] In an embodiment of the present application, an axial tensile strength classification interval is established based on the prediction results of the support vector machine classifier. The tensile strength values are divided into three levels using the K-means clustering method, and the initial cluster centers are set to 3600, 2400, and 1400 MPa. By maximizing the inter-group variance criterion, the strength boundary values are determined to be 3100 and 1900 MPa, corresponding to high, medium, and low strength levels, respectively. The stability of the grading standard is further verified, and 1000 groups of validation data sets are generated using the Bootstrap resampling method, and 75% of the samples are randomly sampled for testing each time. The statistical results show that the 95% confidence interval of the classification accuracy is 87% to 94%, the precision is 86% to 93%, and the recall is 85% to 91%, indicating that the grading standard has high stability. The grading results show that the high-strength grade corresponds to a peel depth of less than 18 microns and an orientation greater than 0.85; the medium-strength grade corresponds to a peel depth of 20 to 45 microns and an orientation of 0.55 to 0.85; and the low-strength grade corresponds to a peel depth greater than 45 microns and an orientation less than 0.55. This multi-index collaborative grading method, combined with clustering and statistical verification, significantly improves the engineering applicability of the grading standard.

[0085] In this application's examples, a quality grading model for recycled carbon fiber was successfully constructed using support vector machine classification and K-means clustering. Compared to traditional grading methods, this application improved the model's classification accuracy and robustness through kernel density estimation and grid search optimization. Bootstrap validation further ensured the reliability of the grading criteria, providing an efficient and scientific tool for performance prediction and material screening of recycled carbon fiber.

[0086] S107 aims to improve the accuracy of recycled carbon fiber quality grading. It uses genetic algorithms to optimize the parameters of the support vector machine classifier, obtains the optimal combination of kernel function parameters and penalty factors, and constructs a performance-optimized carbon fiber quality grading model.

[0087] In this application example, a genetic algorithm optimization process was designed to optimize the kernel function parameters and penalty factors of a support vector machine classifier. Through binary encoding, population iteration, and fitness evaluation, the parameter space was searched to obtain the optimal parameter combination. Cross-validation and performance indicator analysis were used to verify the classification performance and stability of the optimized model, ensuring the efficient application of the grading model in recycled carbon fiber quality assessment.

[0088] S1071 binary encodes the support vector machine parameters and initializes the population through the genetic algorithm to start the optimization process.

[0089] In an embodiment of the present application, the kernel function parameter γ and penalty factor C of the support vector machine classifier are selected as optimization objects, the range of γ is set to 0.01 to 80, and the range of C is set to 0.1 to 800. A 32-bit binary code is used to generate a parameter sequence, where γ and C each occupy 16 bits, which are mapped to a specified range to ensure search accuracy. For example, γ = 0.1 is encoded as a 16-bit binary string, and the corresponding decimal value is linearly interpolated in the range of 0.01 to 80. An initial population is constructed, and the scale is set to 120 individuals, each of which represents a set of parameter combinations. Taking the five-fold cross-validation accuracy as the fitness function, the fitness value of the best individual in the initial population is approximately 0.87, corresponding to a classification accuracy of 87%. Through encoding and random initialization, the genetic algorithm provides a diverse starting point for parameter optimization, balances the search range and computational efficiency, and lays the foundation for subsequent iterative optimization.

[0090] S1072 performs genetic operations to update the population and improve fitness through roulette wheel selection, single-point crossover, and bit reversal mutation.

[0091] In the embodiments of the present application, a roulette wheel selection method is used to assign selection probabilities based on individual fitness values. The higher the fitness, the greater the probability of selection, ensuring the transmission of high-quality genes. The top 25% of elite individuals, approximately 30 in total, are retained to form an elite individual sequence. For this elite sequence, a single-point crossover operation is performed with a crossover probability of 0.75, splitting the two parent codes at random positions and recombining to generate offspring. For example, after the 14th crossover, the offspring inherits part of the γ and C code segments of the parent. Bit reversal mutation is performed with a mutation probability of 0.12, randomly flipping the code bits to increase population diversity, for example, changing a bit from 0 to 1 to fine-tune parameter values. Through an elite retention strategy, the individuals with the highest fitness are directly passed on to the next generation to generate an updated population. Genetic operations gradually optimize the population parameter combination through selection, crossover, and mutation, significantly improving fitness. For example, in the first 30 iterations, the maximum fitness increased from 0.87 to 0.92. This operation method effectively balances the exploratory and convergent properties of the population.

[0092] S1073 decodes the optimal parameters based on the fitness convergence criterion and evaluates the classification performance of the optimized model through the validation set.

[0093] In an embodiment of the present application, the fitness value of each individual is calculated for the updated population, and the algorithm converges when the maximum fitness change rate for 60 consecutive generations is less than 0.0005. The optimal individual is decoded and a parameter combination is obtained, for example, γ = 0.09, C = 110, and the corresponding cross-validation accuracy is 94.5%. The support vector machine classifier is retrained using the optimized parameters, and the performance is evaluated using a validation set (about 250 samples). Confusion matrix analysis shows that the true positive rate for high intensity levels is 96%, for medium intensity levels is 93%, for low intensity levels is 90%, and the overall accuracy is 92%. The receiver operating characteristic curve is further calculated, and the area is 0.96, indicating that the classifier has excellent distinguishing ability. Compared with before optimization, the classification accuracy of boundary samples increased from 72% to 89%, showing stronger robustness. The parameter stability was evaluated by 10-fold cross validation, and the standard deviation of the accuracy was reduced to 0.015, the 95% confidence interval of γ was 0.07 to 0.11, and C was 90 to 130, with high parameter reliability. This optimization process ensures the high accuracy and generalization ability of the model through convergence judgment and performance verification.

[0094] In this application's examples, a genetic algorithm was used to optimize support vector machine parameters, significantly improving the classification performance of a carbon fiber quality grading model. Compared to traditional grid search, this application uses binary encoding and genetic manipulation to efficiently search the parameter space, reducing computational costs. The optimized model demonstrated excellent accuracy, precision, and recall, providing highly reliable technical support for the quality assessment of recycled carbon fiber.

[0095] Although the present invention has been described in detail above using general descriptions and specific embodiments, it will be apparent to those skilled in the art that modifications and improvements may be made thereto. Therefore, such modifications and improvements, without departing from the spirit of the present invention, are intended to be within the scope of protection claimed herein.

Claims

1. A carbon fiber quality analysis method based on surface morphology, characterized in that: The method comprises: Scanning electron microscopy was used to obtain the surface morphology of the recycled carbon fiber, identify the microstructure images of the carbon fiber surface at different recycling stages, and determine the degree of carbon fiber surface peeling based on the layered peeling characteristics of the carbon fiber surface in the image; Determine the orientation of the graphite lattice on the surface of the recycled carbon fiber, obtain the orientation diffraction patterns of the graphite lattice on the surface of the carbon fiber at different recycling stages, perform background subtraction and normalization on the orientation diffraction patterns, obtain the change pattern of the orientation diffraction peak parameters in different recycling stages, and determine the change pattern of the orientation of the graphite lattice on the surface of the carbon fiber; According to the variation law of carbon fiber surface peeling degree and graphite lattice orientation, an exponential function model between carbon fiber surface peeling depth and graphite lattice orientation was established. By fitting the exponential function model, the mapping relationship between peeling depth and graphite lattice orientation was simulated. The axial tensile strength of carbon fibers at different peeling depths was identified through tensile testing, and the correlation between the peeling depth of the carbon fiber surface and the axial tensile strength was analyzed. The key indicators characterizing carbon fiber quality were screened out through the correlation between carbon fiber surface peeling depth and axial tensile strength, as well as the mapping relationship between peeling depth and graphite lattice orientation. The key indicators of recycled carbon fiber are trained using a support vector machine algorithm. The axial tensile strength of the recycled carbon fiber is predicted using a support vector machine classifier. The quality of the recycled carbon fiber is graded according to the axial tensile strength, and a recycled carbon fiber quality grading standard is obtained. Taking the accuracy of recycled carbon fiber quality classification as the optimization goal, the genetic algorithm was used to optimize the parameters of the support vector machine classifier to obtain the optimized support vector machine classifier.

2. The method according to claim 1, characterized in that The scanning electron microscope is used to obtain the surface morphology of the recycled carbon fiber, identify the microstructure images of the carbon fiber surface at different recycling stages, and judge the degree of peeling of the carbon fiber surface based on the layered peeling characteristics of the carbon fiber surface in the image, including: Obtaining a grayscale histogram matrix based on the surface morphology image of the carbon fiber sample, and performing edge enhancement processing on the image using the grayscale histogram matrix to obtain a set of feature point coordinates; The minimum bounding rectangle algorithm is used to calculate the direction angle and area of each feature point region for the feature point coordinate set to obtain the surface topography feature vector matrix; Obtaining a feature classification result by a density clustering method according to the surface morphology feature vector matrix, and calculating the grayscale gradient value between adjacent layers in the feature classification result by using a Sobel operator; A gradient threshold interval is set for the grayscale gradient value to obtain a wrinkle contour line coordinate sequence, and a curvature change rate is obtained by performing cubic spline curve fitting on the wrinkle contour line coordinate sequence to judge the degree of peeling of the carbon fiber surface.

3. The method according to claim 1, characterized in that The method comprises: determining the orientation of the graphite lattice on the surface of the recycled carbon fiber, obtaining the orientation diffraction patterns of the graphite lattice on the surface of the carbon fiber at different recycling stages, performing background subtraction and normalization on the orientation diffraction patterns, obtaining the variation pattern of the orientation diffraction peak parameters in different recycling stages, and determining the variation pattern of the orientation of the graphite lattice on the surface of the carbon fiber, including: Obtaining an original diffraction intensity matrix based on the graphite lattice diffraction pattern on the surface of the recycled carbon fiber, and performing a Fourier transform operation on the original diffraction intensity matrix to obtain a frequency domain signal matrix; Calculating a power spectral density distribution value for the frequency domain signal matrix, comparing the power spectral density distribution value with a preset density threshold, and applying Gaussian filtering to obtain a diffraction peak signal matrix after noise reduction; Constructing an intensity distribution curve for the noise-reduced diffraction peak signal matrix, and fitting the intensity distribution curve using the least squares method to obtain a standardized diffraction peak shape characteristic value; The intensity ratio sequence between adjacent diffraction peaks is extracted according to the standardized diffraction peak shape characteristic value, and the intensity ratio sequence is fitted to obtain the graphite lattice orientation angle distribution function, and determine the change law of the graphite lattice orientation degree on the carbon fiber surface.

4. The method according to claim 1, wherein According to the variation law of carbon fiber surface peeling degree and graphite lattice orientation, an exponential function model between carbon fiber surface peeling depth and graphite lattice orientation is established, and the mapping relationship between peeling depth and graphite lattice orientation is simulated by fitting the exponential function model, including: A peeling depth distribution map is obtained by calculating the grayscale gradient matrix in the carbon fiber surface topography image, a multi-scale wavelet transform is used to perform a three-layer decomposition on the peeling depth distribution map, and a peeling depth feature sequence is obtained by reconstructing the wavelet coefficients; An orientation distribution diagram is obtained by calculating the graphite lattice diffraction peak intensity matrix in the diffraction pattern, singular value decomposition is used to extract features of the orientation distribution diagram, and an orientation feature sequence is constructed according to the main eigenvalues; Establishing a corresponding point set according to the peeling depth feature sequence and the orientation feature sequence, numerically interpolating the corresponding point set using a cubic spline function, and obtaining a dense sampling data set using a local weighted average method; An exponential function model is constructed for the densely sampled data set, the least squares method is used to determine the initial values of the model coefficients, the initial values of the model coefficients are iteratively optimized by the gradient descent method, and the optimization termination condition is determined according to the convergence threshold to obtain the optimized exponential function model parameters.

5. The method according to claim 1, wherein The method of identifying the axial tensile strength of carbon fibers corresponding to different peeling depths through tensile testing and analyzing the correlation between the peeling depth of the carbon fibers and the axial tensile strength includes: Receiving test data collected by a stress and strain sensor, obtaining a frequency domain decomposition signal using Fourier transform according to the test data, and obtaining a smoothed stress and strain curve through inverse Fourier transform; The five-point difference method is used to calculate the slope of the smoothed stress-strain curve. If the slope of the curve reaches a maximum value, the yield strength and axial tensile strength of the carbon fiber are determined. Constructing a grayscale gradient matrix according to the carbon fiber surface topography image, obtaining a boundary contour line of the peeling area, and calculating a peeling depth value according to a vertical distance of the boundary contour line; A sample data matrix is established for the axial tensile strength and peeling depth values, a support vector regression method is used to obtain a mapping function, and a principal component analysis method is used to determine the quantitative relationship between the peeling depth and the axial tensile strength.

6. The method according to claim 1, characterized in that The key indicators characterizing the quality of carbon fiber are screened out through the correlation between the surface peeling depth of carbon fiber and the axial tensile strength, as well as the mapping relationship between the peeling depth and the graphite lattice orientation, including: A correlation coefficient matrix is obtained based on the carbon fiber surface peeling depth and axial tensile strength data, and the correlation coefficient is calculated using the Pearson correlation analysis method. A significant correlation parameter set is obtained when the correlation coefficient is greater than the significance level; Calculating eigenvalues and eigenvectors for the significantly correlated parameter set, obtaining the eigenvalues and eigenvectors using a principal component analysis method, and obtaining the principal component weights by determining that the cumulative variance contribution rate corresponding to the eigenvector is greater than a preset threshold; Calculating a parameter correlation coefficient based on the main component weights, and obtaining a strongly correlated indicator set when the parameter correlation coefficient is greater than a preset threshold; For the strongly correlated indicator set, a random forest regression method is used to obtain parameter importance scores, and key indicators characterizing carbon fiber quality are screened out through the parameter importance scores.

7. The method according to claim 1, characterized in that The key indicators of the recycled carbon fiber are trained using a support vector machine algorithm, the axial tensile strength of the recycled carbon fiber is predicted by the support vector machine classifier, and the quality of the recycled carbon fiber is graded according to the axial tensile strength to obtain a recycled carbon fiber quality grading standard, including: A characteristic matrix is constructed according to the peeling depth, orientation and tensile strength, the characteristic matrix is normalized by using the maximum and minimum value normalization method, and the normalized sample probability distribution is obtained by kernel density estimation; Constructing a support vector machine classifier for the sample probability distribution, mapping the sample probability distribution using a radial basis kernel function, and obtaining the optimal combination of kernel parameters and penalty factors through a grid search method; A support vector machine classifier is trained according to the optimal combination, a one-to-many strategy is used to perform classification training on the sample probability distribution, and an optimal separating hyperplane is obtained by a sequential minimum optimization algorithm; An axial tensile strength grading interval is established for the optimal separation hyperplane, and the axial tensile strength values are divided into three levels using the K-means clustering method. The strength boundary value is obtained by maximizing the inter-group variance criterion, and a recycled carbon fiber quality grading standard is established based on the boundary value.

8. The method according to claim 1, characterized in that The accuracy of the recycled carbon fiber quality classification is used as the optimization goal, and the parameters of the support vector machine classifier are optimized using a genetic algorithm to obtain an optimized support vector machine classifier, including: Obtain kernel function parameters and penalty factors of a support vector machine classifier, and generate a parameter encoding sequence using a thirty-two-bit binary code according to the kernel function parameters and the penalty factor; Selecting high-quality individuals by a roulette wheel selection method for the parameter coding sequence, and sorting the high-quality individuals according to their fitness values to obtain an elite individual sequence; A single-point crossover method is used to perform a crossover operation on the elite individual sequence, a bit reversal method is used to perform a mutation operation on the individuals after the crossover, and an updated population is obtained according to the mutation operation; The individual fitness value is calculated for the updated population. If the maximum change rate of the individual fitness value for fifty consecutive generations is less than a set threshold, the optimal individual is decoded to obtain the optimal parameter combination of the support vector machine, and an optimized support vector machine classifier is obtained.

Images