Neural network-combined carbon fiber recovery path prediction method

Abstract

The invention provides a method for predicting a recycled carbon fiber path in combination with a neural network, and the method comprises the steps: carrying out the real-time monitoring of a whole cracking process through employing a continuous imaging method, obtaining a surface topography image sequence, carrying out the analysis of the surface topography image sequence through an image processing method, and carrying out the prediction of a recycled carbon fiber path. Judging the variation trend of the microstructure parameters of the depth and roughness of the groove on the surface of the carbon fiber; establishing a nonlinear mapping relation between the carbon fiber surface structure and the recovery quality according to the microstructure parameters of the depth and the roughness of the carbon fiber surface groove, and outputting a corresponding recovery quality predicted value when a parameter value representing the carbon fiber surface structure is input; for recycled carbon fiber samples obtained under different cracking process conditions, the quality and performance of recycled carbon fibers are predicted, prediction results are analyzed and compared, and cracking process parameters are optimized.

Description

Technical Field

[0001] The invention relates to the field of information technology, and in particular to a prediction method for recycling carbon fiber pathways combined with a neural network. Background Art

[0002] In the carbon fiber recycling process, the pyrolysis stage is one of the key steps. During this stage, the surface micromorphology of the carbon fiber changes significantly, such as the change in the depth of the surface grooves resulting in obvious differences in roughness. These changes in surface structure have an important impact on the performance of the final recycled carbon fiber, especially its mechanical properties and application potential. However, there is an urgent need to conduct in-depth research and establish a mapping relationship between surface structure and recycling quality to reveal the intrinsic mechanism of how structural changes affect the performance of carbon fiber. Specifically, due to the complex and diverse changes in the surface micromorphology of carbon fiber during the recycling process and its obvious anisotropic characteristics, traditional characterization methods such as scanning electron microscopy (SEM) and atomic force microscopy (AFM) are difficult to fully characterize these fine structures. At the same time, the mechanism of the influence of external factors such as temperature and pressure on the evolution of surface structure during the pyrolysis process is still unclear, resulting in insufficient accuracy of the correlation model between surface structure and recycling quality. In addition, existing quality assessment methods mainly rely on macroscopic mechanical property tests, ignoring the influence of surface structure and lack of quantitative correlation between surface micromorphology and mechanical properties. Therefore, it is urgent to develop new surface structure characterization technologies, combine them with advanced neural network algorithms, establish an accurate mapping model of surface micromorphology and recycling quality, clarify the evolution law of carbon fiber surface structure and its influence on mechanical properties, and provide theoretical guidance and technical support for high-quality recycled carbon fiber. Summary of the invention

[0003] The present invention provides a prediction method for recycling carbon fiber pathways combined with a neural network, which mainly includes:

[0004] Obtain recycled carbon fiber samples, observe the surface micromorphology of the carbon fiber using a scanning electron microscope to obtain surface morphology image data, scan the carbon fiber surface to obtain a three-dimensional topographic map, digitize the obtained surface morphology image, measure the groove depth on the carbon fiber surface by a line scanning analysis method, and obtain the numerical value of the carbon fiber surface micromorphology characterization parameter, including the groove depth and roughness;

[0005] The whole cracking process is monitored in real time by a continuous imaging method, and a surface morphology image sequence is obtained. The surface morphology image sequence is analyzed by an image processing method to determine the change trend of the microscopic morphology parameters of the groove depth and roughness of the carbon fiber surface;

[0006] A nonlinear mapping relationship between the carbon fiber surface structure and the recycling quality is established based on the microscopic morphological parameters of the carbon fiber surface groove depth and roughness. When the numerical value of the carbon fiber surface structure parameter is input, the corresponding recycling quality prediction value is output;

[0007] The internal three-dimensional structure data of carbon fiber is obtained by X-ray tomography, and the annular structure image of the carbon fiber cross section is obtained by image reconstruction. The integrity and uniformity of the annular structure are quantitatively analyzed based on the image to determine the quality evaluation index of the annular structure.

[0008] The carbon fiber surface micro-morphology characterization parameters and the ring structure quality evaluation index are integrated to evaluate the comprehensive performance of recycled carbon fiber, analyze the influence of surface structure changes and ring structure characteristics on the mechanical properties and conductive properties of recycled carbon fiber, and obtain the correlation law of carbon fiber structure performance;

[0009] For the recycled carbon fiber samples obtained under different pyrolysis process conditions, the quality and performance of the recycled carbon fiber are predicted, the prediction results are analyzed and compared, and the pyrolysis process parameters are optimized.

[0010] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0011] The present invention discloses a prediction method for recycling carbon fiber pathways combined with a neural network. The method obtains carbon fiber surface microscopic morphology data through scanning electron microscopy and nanometer-level resolution scanning, and uses line scanning analysis to measure surface groove depth and roughness. At the same time, X-ray tomography technology is used to obtain the internal three-dimensional structure data of the carbon fiber to analyze the integrity and uniformity of the annular structure. The present invention integrates the surface microscopic morphology parameters with the annular structure quality indicators, constructs a comprehensive performance evaluation model, and analyzes the impact of structural changes on mechanical properties and conductive properties. In addition, a nonlinear mapping model of surface structure and recycling quality is established through a neural network algorithm to achieve recycling quality prediction. Finally, the present invention uses these models to predict the quality and performance of recycled carbon fibers under different cracking process conditions, optimize the cracking process parameters, and thus improve the quality and performance of the recycled carbon fibers. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 The present invention is a flowchart of a method for predicting a recycling carbon fiber pathway in combination with a neural network. DETAILED DESCRIPTION

[0013] The technical solution of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0014] like Figure 1 In this embodiment, a prediction method for recycling carbon fiber pathways combined with a neural network may specifically include:

[0015] Step S101, obtain a recycled carbon fiber sample, use a scanning electron microscope to observe the surface micromorphology of the carbon fiber, obtain surface morphology image data, scan the carbon fiber surface to obtain a three-dimensional topographic map, digitize the obtained surface morphology image, measure the groove depth of the carbon fiber surface by a line scanning analysis method, and obtain the carbon fiber surface micromorphology characterization parameter value, including groove depth and roughness.

[0016] The image is locally enhanced according to the grayscale value distribution characteristics of the carbon fiber surface microscopic image. After the edge of the carbon fiber surface groove contour is extracted, the image noise is filtered out by wavelet transform, the carbon fiber surface groove area is marked by morphological operation, and the contour data after the marked grayscale value distribution is obtained according to the grayscale value distribution in the marked area; a carbon fiber surface depth map is established according to the contour data after the marked grayscale value distribution, the groove area is segmented, and the local depth change is calculated by the grayscale value difference of adjacent pixels to obtain the groove depth value; for the carbon fiber surface depth map, the texture features are extracted by the grayscale co-occurrence matrix to obtain the surface roughness data.

[0017] Exemplarily, an image acquisition parameter library is established according to the gray value distribution characteristics of the microscopic image of the carbon fiber surface, the sub-region segmentation method is used to perform local enhancement processing on the image, the surface texture characteristics are highlighted by the contrast adaptive adjustment algorithm, and the edge of the groove contour on the carbon fiber surface is extracted to obtain the first contour data. For the first contour data, the image noise is filtered out by wavelet transform, the groove area on the carbon fiber surface is marked by morphological operation, and the spacing between adjacent contour lines is calculated according to the gray value distribution in the marked area to obtain the second contour data. A carbon fiber surface depth map is established according to the second contour data, the groove area is segmented by the regional growth algorithm, and the local depth change is calculated by the gray value difference of adjacent pixels to obtain the groove depth value. For the depth map, the gray co-occurrence matrix is ​​used to extract the surface texture characteristics, and the surface roughness data is obtained by calculating the statistical parameters such as the second-order moment of angle, correlation, and entropy value. A three-dimensional point cloud model is constructed according to the groove depth value and the surface roughness data, and the point cloud data is smoothed by the Gaussian kernel function, and the three-dimensional surface equation of the carbon fiber surface is fitted by the least squares method. According to the three-dimensional surface equation, the surface feature extraction algorithm is used to calculate the principal curvature and Gaussian curvature, and the carbon fiber surface concave-convex parameters are obtained according to the curvature distribution characteristics, and the quantitative results of the microscopic morphology characteristics of the carbon fiber surface are obtained. In the analysis of the microscopic morphology of the carbon fiber surface, the gray value distribution characteristics are an important data basis for characterizing the surface structure of the carbon fiber. Through the statistical analysis of the gray histogram, it is found that the gray value distribution of the groove area on the carbon fiber surface presents a bimodal feature, in which the brighter area corresponds to the convex part of the fiber surface, and the gray value range is between 180 and 220, and the darker area corresponds to the groove part, and the gray value range is between 90 and 130. Based on this feature, an image acquisition parameter library can be established. In the process of local image enhancement, the sub-region size is set to 16×16 pixels, and adaptive contrast adjustment is performed in each sub-region. When the difference between the maximum gray value and the minimum gray value in the region is less than 50, the histogram equalization method is used to enhance the regional contrast, so that the groove edge features are more prominent. The enhanced image is subjected to noise filtering by wavelet transform, and the db4 wavelet basis function is selected to perform soft threshold processing on the high-frequency coefficients, with the threshold set to 3 times the standard deviation. In the depth map construction stage, the growth criterion of the region growing algorithm is based on the grayscale value difference of adjacent pixels. The seed point is selected in the area with the lowest grayscale value. When the grayscale value difference of adjacent pixels is less than 20, they are included in the same area. This method can accurately divide the boundary of the groove area. For the marked groove area, the vertical distance between adjacent contour lines is calculated to obtain the depth value. For example, the spacing between adjacent contour lines in a groove area is 25 pixels, corresponding to an actual depth of 2.5 microns.The surface texture feature extraction adopts the gray level co-occurrence matrix method, setting the pixel displacement to 1, the direction angle to 0, 45, 90 and 135 degrees, and the calculated angular second-order moment reflects the uniformity of the image, the entropy value represents the complexity of the image, and the correlation describes the correlation of the local gray level. These parameters together constitute the quantitative index of surface roughness. In the three-dimensional surface reconstruction, the Gaussian kernel function is used for point cloud data smoothing. The standard deviation of the kernel function is set to 3, and the window size is 5×5. This parameter setting can effectively remove the measurement noise while maintaining the surface detail features. The principal curvature calculated by the surface feature extraction reflects the curvature of the surface in different directions, and the Gaussian curvature comprehensively represents the local geometric characteristics of the surface. These curvature parameters can accurately describe the microscopic morphology of the carbon fiber surface. In the quantitative results of the microscopic morphology, the root mean square value of the surface roughness is 0.8 microns, the maximum groove depth is 3.2 microns, and the average groove spacing is 12 microns.

[0018] Step S102, using a continuous imaging method to monitor the entire cracking process in real time, obtaining a surface morphology image sequence, analyzing the surface morphology image sequence through an image processing method, and determining the change trend of the carbon fiber surface groove depth and roughness microscopic morphology parameters.

[0019] According to the carbon fiber surface morphology image sequence, the continuous frame images are spatially aligned by using the image registration algorithm, and the displacement vector field between adjacent frames is calculated by the optical flow method to obtain the first sequence of image data; for the first sequence of image data, the image sequence background is extracted by using the temporal median filter, and the change area is extracted by the background difference method to obtain the second sequence of image data; for the second sequence of image data, the surface groove contour is extracted by using the morphological operator, and the three-dimensional coordinates of the contour feature points are calculated by the least squares method to obtain the groove depth time series data; for the second sequence of image data, the surface texture features are extracted by using the gray level co-occurrence matrix, and the local statistical parameters are calculated by the sliding window to obtain the surface roughness time series data.

[0020] Exemplarily, a time-series image library is established based on a sequence of carbon fiber surface morphology images, and continuous frame images are spatially aligned using an image registration algorithm. The displacement vector field between adjacent frames is calculated using an optical flow method to obtain a first sequence of image data. For the first sequence of image data, a temporal median filter is used to extract the image sequence background, and the change area in each frame of the image is extracted using a background difference method. The change area is marked according to an adaptive threshold segmentation algorithm to obtain a second sequence of image data. According to the second sequence of image data, a morphological operator is used to extract the surface groove contour, and the three-dimensional coordinates of the contour feature points are calculated using the least squares method to obtain the groove depth time series data. For the groove depth time series data, a Kalman filter is used for data smoothing, and the depth change rate is calculated using the recursive least squares method to obtain a depth trend curve. According to the second sequence of image data, a gray level co-occurrence matrix is ​​used to extract surface texture features, and local statistical parameters are calculated using a sliding window to obtain surface roughness time series data. For the depth trend curve and roughness time series data, the support vector regression method is used to construct the morphology parameter prediction function, and the parameter change inflection point is determined by local curvature analysis to obtain the evolution characteristics of the microscopic morphology parameters of the carbon fiber surface. In the real-time monitoring process of the carbon fiber surface morphology, image registration is the key link to ensure data continuity. Accurate alignment is achieved by calculating the displacement vector field between adjacent frames. When the image acquisition frequency is set to 10 frames per second, the maximum displacement between adjacent frames is usually no more than 5 pixels. At this time, the optical flow method can accurately capture the tiny deformation process. The time series image background extraction adopts the time median filtering method. For each pixel position, the median of the grayscale values ​​of the corresponding positions of all frames in the time window is taken as the background value. When the window size is set to 15 frames, it has good robustness, which can effectively suppress random noise and accurately reflect slow background changes. The change area obtained by background difference is segmented by adaptive threshold, and the threshold is taken as twice the standard deviation of the local area to achieve accurate detection of weak changes. In the extraction of groove contour features, the morphological operator uses 3×3 structural elements for opening and closing operations to eliminate small noise while maintaining contour integrity. The extracted contour points show a discrete distribution in space, and a continuous depth curve is obtained by least squares fitting. The time series of depth data shows obvious stage characteristics. In the early stage of cracking, the groove depth increases rapidly, increasing by an average of 0.5 microns per minute. As the cracking process progresses, the growth rate gradually decreases. The grayscale co-occurrence matrix method is used for surface texture feature analysis. The displacement is set to 1 pixel, and the directions include 0 degrees, 45 degrees, 90 degrees and 135 degrees. The calculated angular second-order moment, entropy value and other statistical parameters change with time. The curve can quantitatively characterize the evolution of surface roughness. Experimental data show that the surface roughness shows a trend of increasing first and then stabilizing during the cracking process. The root mean square roughness in the initial stage is 0.3 microns, which increases to 0.8 microns after 15 minutes and remains basically stable after 30 minutes.The morphology parameters are predicted by using the support vector regression method, and the nonlinear mapping relationship is established through the radial basis kernel function. The cross-validation method is selected for kernel parameter optimization, and the root mean square error on the validation set is controlled within 0.05 microns. Local curvature analysis found that the change curves of groove depth and surface roughness have an obvious inflection point when the cracking is about 40%. At this time, the microstructure of the carbon fiber surface begins to change significantly, the depth growth rate is reduced to less than 0.1 microns per minute, and the growth of roughness also tends to be flat.

[0021] Step S103, establishing a nonlinear mapping relationship between the carbon fiber surface structure and the recycling quality according to the microscopic morphological parameters of the carbon fiber surface groove depth and roughness, and when the numerical value representing the carbon fiber surface structure parameter is input, the corresponding recycling quality prediction value is output.

[0022] A feature data set is constructed based on the numerical values ​​of the groove depth and roughness of the carbon fiber surface, and the missing values ​​in the feature data set are processed by the median filling method. Outliers are eliminated by the triple standard deviation principle to obtain a data set after eliminating outliers. The numerical values ​​are normalized by a standardization method, and the key dimensions in the data set after eliminating outliers are extracted by principal component analysis to obtain a data set with extracted key dimensions. Based on the data set of the key dimensions, a nonlinear mapping structure is constructed by a residual neural network, and the output of the hidden layer neurons is nonlinearly transformed by a rectified linear unit function to obtain a mapping function. For the mapping function, a fixed-length sliding window is used to segment the input data, and the prediction value of each segment is obtained by forward calculation.

[0023] Exemplarily, a feature data set is constructed according to the numerical values ​​of the groove depth and roughness of the carbon fiber surface, the median filling method is used to process the missing values, and the outliers are eliminated by the three-times standard deviation principle to obtain the first data set after preprocessing. For the first data set, the standardization method is used to normalize the values, and the key dimensions in the feature data are extracted by the principal component analysis method to obtain the second data set after dimensionality reduction. According to the second data set, a nonlinear mapping structure is constructed by a five-layer residual neural network, and the output of the hidden layer neurons is nonlinearly transformed by the rectified linear unit function to obtain the first mapping function. For the first mapping function, the training data and the validation data are divided by the five-fold cross validation method, the prediction deviation is calculated by the mean square error loss function, and the network parameters are updated according to the back propagation algorithm to obtain the second mapping function. According to the second mapping function, the input data is segmented by a fixed-length sliding window, and the prediction value of each segment is obtained by forward calculation. The multi-segment prediction results are fused based on the weighted average method to obtain the carbon fiber recycling quality value. For the recycling quality value, the confidence interval method is used to calculate the reliability range of the prediction result, and the abnormal prediction result is identified by the outlier detection algorithm to obtain the carbon fiber recycling quality evaluation index. The carbon fiber surface feature data set includes two core dimensions: groove depth and roughness. In the data preprocessing stage, missing values ​​are processed by filling the median. For example, a set of groove depth data sequences are 2.5, 3.1, missing values, 2.8, and 3.2 microns. The median 2.95 microns is used to fill the missing value position to maintain the continuity and rationality of the data. The three-times standard deviation criterion is used for outlier removal. When the measured value deviates from the mean value by more than three times the standard deviation, it is identified as an outlier and removed. The minimum-maximum normalization method is used for data standardization to map the eigenvalues ​​of different dimensions to the range of 0 to 1, so that the groove depth and roughness data are comparable. The results of principal component analysis show that the cumulative variance contribution rate of the first two principal components is more than 95%. The first principal component mainly reflects the groove depth characteristics with a weight coefficient of 0.82, and the second principal component mainly reflects the surface roughness characteristics with a weight coefficient of 0.76. The residual neural network adopts a five-layer structure. The input layer corresponds to the feature dimension after dimensionality reduction. The number of neurons in the three hidden layers in the middle is 64, 32, and 16 respectively. The output layer corresponds to the predicted value of recycling quality. The jump connection structure in the network effectively alleviates the gradient vanishing problem in the deep network training process. The introduction of the rectified linear unit function enhances the model's ability to express nonlinear features. During the cross-validation process, the data set is evenly divided into 5 parts, 4 of which are taken as training sets each time, and the remaining 1 is used as a validation set. The generalization performance of the model is ensured through 5 rounds of rotation validation. The training process adopts a mini-batch method with a batch size of 32. The initial value of the learning rate is set to 0.001. When the error of the validation set does not decrease significantly for 3 consecutive rounds, the learning rate is automatically reduced to 0.1 times the original.In the online prediction stage, a sliding window of length 10 is used to segment the input data, and the overlap rate of adjacent windows is 50% to ensure the smoothness and stability of the prediction results. The prediction results of each segment are fused by the exponential weighted average method, with the weight coefficient of the newer data being 0.7 and the weight coefficient of the older data being 0.3. The confidence interval of the prediction results uses a 95% confidence level. When the width of the confidence interval of a certain prediction value exceeds twice the average value, it is marked as a low-confidence prediction result. Outlier detection is based on the local anomaly factor algorithm, which calculates the density ratio of each prediction point to its neighboring points. When the density ratio is significantly lower than the average level, the prediction point is marked as an abnormal prediction. Experimental data show that the root mean square error of the model on the validation set is 0.15, the average width of the confidence interval of the prediction results is 0.3, and the proportion of abnormal prediction points is controlled below 5%.

[0024] Step S104, using an X-ray tomography method to obtain the internal three-dimensional structure data of the carbon fiber, obtaining the annular structure image of the carbon fiber cross section through an image reconstruction method, performing a quantitative analysis on the integrity and uniformity of the annular structure based on the image, and determining the quality evaluation index of the annular structure.

[0025] For carbon fiber tomography data, the scanning plane is spatially registered by the marker point matching method, and the original data is denoised by Gaussian filtering to obtain a denoised image sequence; according to the denoised image sequence, the scattering attenuation data is compensated by an iterative back-projection algorithm, and the edge features are extracted to obtain an image sequence after edge feature extraction; according to the image sequence after edge feature extraction, the cross-sectional contour is extracted, and the annular structure boundary is obtained by morphological closing operation to obtain an image sequence after the annular structure boundary is obtained; for the image sequence after the annular structure boundary is obtained, the circularity calculation method is used to quantify the integrity of the annular structure, the density distribution of the annular structure is calculated by the gray level co-occurrence matrix, and the internal pores are identified based on the regional connectivity analysis to obtain structural defect data; according to the structural defect data, a feature extraction structure is constructed to obtain an annular structure quality evaluation index.

[0026] Exemplarily, spatial registration is performed according to the carbon fiber tomography data, the scanning plane is aligned and corrected by the marker point matching method, and the original data is subjected to noise reduction processing by Gaussian filtering to obtain a first image sequence. For the first image sequence, the scattering attenuation data is compensated by the iterative back-projection algorithm, the edge features are extracted by four-layer wavelet decomposition, and the edge significance is calculated according to the gradient amplitude to obtain a second image sequence. According to the second image sequence, the cross-sectional contour is extracted by the region growing method, the edge gap is filled by the morphological closing operation, and the ring structure boundary is obtained based on the contour tracing algorithm to obtain a third image sequence. For the third image sequence, the integrity of the ring structure is quantified by the circularity calculation method, and the structural center offset is calculated by the radial symmetry evaluation algorithm to obtain the structural feature parameters. According to the structural feature parameters, the gray level co-occurrence matrix is ​​used to calculate the density distribution of the ring structure, the structural uniformity is evaluated by the entropy method, and the internal pores are identified based on the regional connectivity analysis to obtain the structural defect data. For the structural defect data and structural feature parameters, a six-layer feature extraction structure is constructed by a deep convolutional neural network, and the feature transfer process is optimized by residual connection to obtain the ring structure quality evaluation index. In the process of spatial registration of X-ray tomography images, the selection of markers has an important influence on the accuracy of the results. Usually, four high-contrast markers are set on the outside of the carbon fiber. The relative position deviation between the markers is controlled within 0.1 mm. After registration, the spatial correspondence error of adjacent layers is less than 0.05 mm. The Gaussian filter uses a 5×5 filter kernel and a standard deviation of 1.2, which effectively suppresses random noise while maintaining edge features. In scattering attenuation compensation, the iterative back-projection algorithm uses 10 rounds of iterative optimization, and the projection angle interval in each round of iteration is 1 degree. The signal-to-noise ratio of the compensated image is improved by about 40%. Wavelet decomposition uses two-dimensional discrete wavelet transform, and selects db4 wavelet basis function for four-layer decomposition to obtain 16 sub-band images, among which the high-frequency sub-band reflects the edge feature information, and the edge significance threshold is set to 2.5 times the standard deviation of the local area. In the extraction of ring structure, the region growing method uses the highest gray value point as the seed point, and the growth threshold is set to 85% of the gray value of the seed point. The morphological closing operation uses a circular structure element with a radius of 3 pixels to fill the tiny gaps in the edge contour. In the boundary point sequence obtained by contour tracing, the spacing between adjacent points is maintained between 1 and 2 pixels, ensuring the continuity of the boundary description. The roundness parameter is used to evaluate the structural integrity, and the deviation between the actual contour and the ideal circle is calculated. A roundness value between 0.95 and 1 indicates good structural integrity. In the radial symmetry assessment, the contour is divided into 36 sectors, and the standard deviation of the distance from the center of gravity of each sector to the center point is calculated. When the standard deviation is less than 0.2 mm, the structure is considered to have good symmetry.The density distribution of the ring structure adopts the gray-level co-occurrence matrix method, setting the displacement to 1 pixel, and the direction angles to 0, 45, 90 and 135 degrees. The calculated entropy value reflects the uniformity of the structure. An entropy value less than 3.5 indicates good structural uniformity. The internal pore identification adopts 8-neighborhood interconnection analysis. The connected area with an area less than 5 square pixels is marked as a tiny pore, and the area greater than 20 square pixels is marked as a significant defect. The deep convolutional neural network consists of six convolutional layers, with a convolution kernel size of 3×3 and the number of channels of 32, 64, 128, 256, 512, and 512, respectively. A residual connection is added between every two layers to avoid the loss of deep feature information. The last two layers use a fully connected structure to output the quality score of the carbon fiber ring structure. The score ranges from 0 to 100, where 90 points or more indicate excellent structural quality, 80 to 90 points indicate good quality, 70 to 80 points indicate qualified, and 70 points or less indicate obvious defects.

[0027] Step S105, integrating the carbon fiber surface micro-morphology characterization parameters with the annular structure quality evaluation index, evaluating the comprehensive performance of the recycled carbon fiber, analyzing the influence of the surface structure changes and the annular structure characteristics on the mechanical properties and the conductive properties of the recycled carbon fiber, and obtaining the correlation law of the carbon fiber structure performance.

[0028] After normalization processing is performed on the carbon fiber surface micromorphology characterization parameters and the ring structure quality indicators, outlier data points are eliminated; the principal component analysis method is used to reduce the dimension of the characteristic data, the number of principal components is determined by the variance contribution rate, and the coupling coefficient between parameters is calculated based on the correlation analysis to obtain the characteristic matrix; based on the characteristic matrix, an adaptive neural network is used to construct a performance prediction structure, the network parameters are optimized by the back propagation algorithm, and the prediction accuracy is calculated based on the mean square error function; the mechanical property prediction function is constructed by the multivariate regression method, the key influencing factors are screened by the significance test, and the contribution of each parameter is calculated based on the variance analysis to obtain the mechanical property mapping relationship; the conductivity performance prediction function is established by the sensitivity analysis method, the parameter action order is determined by the partial correlation analysis, and the conductivity performance mapping relationship is obtained; based on the mechanical performance mapping relationship and the conductivity performance mapping relationship, the carbon fiber structure performance correlation law is obtained.

[0029] Exemplarily, an original data set is established based on the micro-morphology characterization parameters of the carbon fiber surface and the quality index of the annular structure, the numerical values ​​are normalized by a standardized method, and outlier data points are removed by outlier detection to obtain a first characteristic matrix. For the first characteristic matrix, the principal component analysis method is used to reduce the dimension of the characteristic data, the number of principal components is determined by the variance contribution rate, and the coupling coefficient between parameters is calculated based on the correlation analysis to obtain a second characteristic matrix. According to the second characteristic matrix, a five-layer adaptive neural network is used to construct a performance prediction structure, the network parameters are optimized by the back propagation algorithm, and the prediction accuracy is calculated based on the mean square error function to obtain a third characteristic matrix. For the third characteristic matrix, a multivariate regression method is used to construct a mechanical property prediction function, the key influencing factors are screened by a significance test, and the contribution of each parameter is calculated based on variance analysis to obtain a mechanical property mapping relationship. According to the mechanical property mapping relationship, a sensitivity analysis method is used to establish a conductive property prediction function, the parameter action order is determined by partial correlation analysis, and the prediction accuracy is verified based on the orthogonal test method to obtain a conductive property mapping relationship. According to the mapping relationship between the mechanical properties and the conductive properties, a weighted fusion algorithm is used to construct a comprehensive evaluation function, and the weight coefficient is determined by the cross-validation method to obtain the correlation law of carbon fiber structure performance. The original data set consisting of the microscopic morphology characteristics of the carbon fiber surface and the ring structure quality index contains multiple dimensions, among which the groove depth ranges from 0.5 to 5 microns, the surface roughness ranges from 0.2 to 2 microns, the ring structure roundness ranges from 0.85 to 1, and the structural uniformity index ranges from 0.6 to 0.95. The minimum-maximum normalization method is used for data standardization, and the outlier detection is based on the three-times standard deviation criterion to eliminate data points that deviate too much from the mean. The results of principal component analysis show that the cumulative variance contribution rate of the first three principal components reaches 92%, among which the first principal component mainly reflects the surface morphology characteristics, explaining 48% of the data variance, the second principal component reflects the ring structure characteristics, explaining 32% of the variance, and the third principal component is related to structural defects, explaining 12% of the variance. The coupling coefficient analysis between features found that the surface groove depth and the uniformity of the annular structure showed a significant negative correlation, with a correlation coefficient of -0.75. The adaptive neural network adopts a five-layer structure. The input layer corresponds to the feature dimension after dimensionality reduction. The number of neurons in the three hidden layers is 64, 32, and 16 respectively. The output layer corresponds to the predicted values ​​of mechanical properties and conductive properties. The network training adopts the random gradient descent method with a batch size of 32. The initial value of the learning rate is 0.001. When the error of the validation set does not decrease significantly for 5 consecutive rounds, the learning rate is automatically reduced to 0.1 times the original. In the prediction of mechanical properties, the significance test adopts a confidence level of 0.05. The results of variance analysis show that the contribution of the surface groove depth is the highest, reaching 45%, followed by the circularity of the annular structure, with a contribution of 30%, and the contribution of the structural uniformity index is 15%. The relative error of the prediction model on the validation set is controlled within 8%.In terms of conductive performance prediction, sensitivity analysis shows that the integrity of the annular structure is the most critical factor, with a sensitivity coefficient of 0.68, followed by surface defect density, with a sensitivity coefficient of 0.42. In the comprehensive evaluation function, the weight coefficients of mechanical properties and conductive properties were determined by five-fold cross-validation, which were 0.6 and 0.4, respectively. The evaluation results show that when the surface groove depth is below 2.5 microns, the roundness of the annular structure is greater than 0.95, and the structural uniformity index is higher than 0.85, the comprehensive performance of carbon fiber is optimal. Correlation analysis found that the effect of surface micromorphology on mechanical properties is more significant, explaining 75% of the performance variation, while the annular structure characteristics have a greater impact on conductive properties, explaining 82% of the conductive properties variation.

[0030] Step S106, predicting the quality and performance of the recycled carbon fiber samples obtained under different pyrolysis process conditions, analyzing and comparing the prediction results, and optimizing the pyrolysis process parameters.

[0031] The surface structure data of recycled carbon fiber under different pyrolysis process conditions were obtained, and the correlation coefficients between process parameters and structural characteristics were calculated through correlation analysis. A nonlinear mapping function was used to calculate the recovery mass prediction value of each group of samples, and the accuracy of the recovery mass prediction value was evaluated through cross-validation. A multi-objective optimization equation was constructed using a comprehensive evaluation function, and the weight coefficients of performance indicators were determined by the entropy weight method. The process parameters were grouped, and the optimal solution of the performance indicators was found by the gradient descent method to obtain the process parameter optimization plan.

[0032] Exemplarily, a feature database is established according to the surface structure data of recycled carbon fiber under different cracking process conditions, and the correlation analysis method is used to calculate the correlation coefficient between the process parameters and the structural characteristics. The dimension effect is eliminated by regularization processing to obtain the first data group. For the first data group, a nonlinear mapping function is used to calculate the recovery quality prediction value of each group of samples, and a validation sample set is generated by the Monte Carlo method. The prediction accuracy is evaluated based on cross-validation to obtain the second data group. According to the second data group, a multi-objective optimization equation is constructed using a comprehensive evaluation function, and the weight coefficient of each performance indicator is determined by the entropy weight method. The prediction results are calibrated based on the Bayesian optimizer to obtain the third data group. For the third data group, a clustering algorithm is used to group the sample features, and the significance of the influence of the process parameters is calculated by variance analysis. The parameter response relationship is constructed based on the decision tree algorithm to obtain the fourth data group. According to the fourth data group, the cracking process parameters are traversed by the grid search method, the parameter change boundary is determined by the constraint propagation algorithm, and the optimal solution of the performance indicator is found based on the gradient descent method to obtain the fifth data group. For the fifth data group, the process parameters are locally tuned using an adaptive optimization algorithm, the parameter adjustment direction is determined by sensitivity analysis, the reliability of the optimization results is evaluated based on the confidence interval, and the optimization scheme for the cracking process parameters is obtained. The influence of the cracking process conditions on the surface structure of carbon fiber is manifested in many aspects. Through correlation analysis, it is found that the cracking temperature is significantly positively correlated with the surface groove depth, and the correlation coefficient reaches 0.82. When the temperature increases from 350 degrees Celsius to 450 degrees Celsius, the groove depth increases from 0.8 microns to 2.5 microns, and the correlation coefficient between the cracking time and the surface roughness is 0.65. For every 30 minutes of extension of the cracking time, the surface roughness increases by 0.3 microns. The Monte Carlo method is used to generate 1000 groups of validation samples for prediction accuracy evaluation. In the cross-validation process, the root mean square error of the recovery quality prediction is 0.12, and the correlation coefficient between the predicted value and the measured value reaches 0.89. 95% of the prediction errors in the validation samples are controlled within the range of ±15%, indicating that the prediction model has good generalization performance. In the comprehensive performance evaluation, the entropy weight method was used to determine the weight of each performance index, among which the weight of tensile strength was 0.35, the weight of elastic modulus was 0.25, the weight of electrical conductivity was 0.25, and the weight of surface structural integrity was 0.15. The Bayesian optimization process established the parameter response surface through Gaussian process regression, and the expected improvement criterion was selected for the acquisition function. Ten sampling points were selected for evaluation in each round of optimization. The cluster analysis results showed that the samples could be divided into three characteristic groups. The first group was characterized by high strength and low conductivity, the second group was characterized by medium strength and high conductivity, and the third group was characterized by balanced performance. Variance analysis showed that the pyrolysis temperature had the most significant effect on tensile strength, with an F statistic of 28.6, and the pyrolysis time had the greatest effect on surface structural integrity, with an F statistic of 22.4.During the parameter optimization process, the temperature step of the grid search was set to 10 degrees Celsius, the time step was 5 minutes, and the constraints included an upper temperature limit of 500 degrees Celsius and an upper time limit of 120 minutes. The gradient descent uses an adaptive learning rate with an initial value of 0.01. When there is no significant improvement after three consecutive rounds of optimization, the learning rate is reduced to 0.5 times the original. The local optimization uses an adaptive algorithm to dynamically adjust the parameter search range according to the degree of improvement of the performance indicators. The temperature adjustment accuracy is ±5 degrees Celsius and the time adjustment accuracy is ±2 minutes. Sensitivity analysis shows that near the optimal parameter point, a temperature change of 1 degree Celsius leads to a 0.5% change in strength, and a time change of 1 minute leads to a 0.3% change in structural integrity. The 95% confidence interval analysis of the optimization results shows that under the selected optimal process parameters, the expected value of tensile strength is 85% ± 3% of the original strength, and the expected value of electrical conductivity is 90% ± 2% of the original conductivity.

[0033] 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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for predicting the recycling of carbon fiber pathways in combination with a neural network, characterized in that: The method comprises: Obtain recycled carbon fiber samples, observe the surface micromorphology of the carbon fiber using a scanning electron microscope to obtain surface morphology image data, scan the carbon fiber surface to obtain a three-dimensional topographic map, digitize the obtained surface morphology image, measure the groove depth on the carbon fiber surface by a line scanning analysis method, and obtain the numerical value of the carbon fiber surface micromorphology characterization parameter, including the groove depth and roughness; The whole cracking process is monitored in real time by a continuous imaging method, and a surface morphology image sequence is obtained. The surface morphology image sequence is analyzed by an image processing method to determine the change trend of the microscopic morphology parameters of the groove depth and roughness of the carbon fiber surface; A nonlinear mapping relationship between the carbon fiber surface structure and the recycling quality is established based on the microscopic morphological parameters of the carbon fiber surface groove depth and roughness. When the numerical value of the carbon fiber surface structure parameter is input, the corresponding recycling quality prediction value is output; The internal three-dimensional structure data of carbon fiber is obtained by X-ray tomography, and the annular structure image of the carbon fiber cross section is obtained by image reconstruction. The integrity and uniformity of the annular structure are quantitatively analyzed based on the image to determine the quality evaluation index of the annular structure. The carbon fiber surface micro-morphology characterization parameters and the ring structure quality evaluation index are integrated to evaluate the comprehensive performance of recycled carbon fiber, analyze the influence of surface structure changes and ring structure characteristics on the mechanical properties and conductive properties of recycled carbon fiber, and obtain the correlation law of carbon fiber structure performance; For the recycled carbon fiber samples obtained under different pyrolysis process conditions, the quality and performance of the recycled carbon fiber are predicted, the prediction results are analyzed and compared, and the pyrolysis process parameters are optimized.

2. The method according to claim 1, characterized in that The method comprises obtaining a recycled carbon fiber sample, observing the surface micromorphology of the carbon fiber using a scanning electron microscope to obtain surface morphology image data, scanning the surface of the carbon fiber to obtain a three-dimensional topographic map, digitally processing the obtained surface morphology image, measuring the groove depth on the surface of the carbon fiber by a line scanning analysis method, and obtaining the numerical value of the micromorphology characterization parameter of the carbon fiber surface, including the groove depth and roughness, including: According to the gray value distribution characteristics of the carbon fiber surface microscopic image, the image is locally enhanced, the edge of the carbon fiber surface groove contour is extracted, and the image noise is filtered out by wavelet transform. The carbon fiber surface groove area is marked by morphological operation, and the contour data after the marked gray value distribution is obtained according to the gray value distribution in the marked area. Establishing a carbon fiber surface depth map based on the contour data after the grayscale value distribution of the markers, segmenting the groove area, calculating the local depth change through the grayscale value difference of adjacent pixels, and obtaining the groove depth value; For the carbon fiber surface depth map, the surface roughness data is obtained by extracting texture features through a gray level co-occurrence matrix.

3. The method according to claim 1, characterized in that The method adopts a continuous imaging method to monitor the whole cracking process in real time, obtains a surface morphology image sequence, and analyzes the surface morphology image sequence by an image processing method to determine the change trend of the carbon fiber surface groove depth and roughness microscopic morphology parameters, including: According to the carbon fiber surface morphology image sequence, the continuous frame images are spatially aligned using an image registration algorithm, and the displacement vector field between adjacent frames is calculated using an optical flow method to obtain the first sequence of image data; For the first sequence of image data, a temporal median filter is used to extract the image sequence background, and a background difference method is used to extract the change area to obtain a second sequence of image data; For the second sequence of image data, a morphological operator is used to extract the surface groove contour, and the three-dimensional coordinates of the contour feature points are calculated by the least square method to obtain the groove depth time series data; For the second sequence of image data, the gray level co-occurrence matrix is ​​used to extract surface texture features, and the local statistical parameters are calculated through a sliding window to obtain surface roughness time series data.

4. The method according to claim 1, characterized in that: The nonlinear mapping relationship between the carbon fiber surface structure and the recycling quality is established according to the microscopic morphology parameters of the carbon fiber surface groove depth and roughness, and when the numerical value of the carbon fiber surface structure parameter is input, the corresponding recycling quality prediction value is output, including: A feature data set is constructed based on the groove depth and roughness values ​​of the carbon fiber surface. The missing values ​​in the feature data set are processed using the median filling method. Outliers are eliminated using the triple standard deviation principle to obtain a data set after outliers are eliminated. The numerical values ​​are normalized by using a standardization method, and the key dimensions in the data set after the outliers are removed are extracted by using a principal component analysis method to obtain a data set with extracted key dimensions; According to the data set of the key dimension, a nonlinear mapping structure is constructed by using a residual neural network, and a nonlinear transformation is performed on the output of hidden layer neurons by using a rectified linear unit function to obtain a mapping function; For the mapping function, a fixed-length sliding window is used to process the input data in segments, and the prediction value of each segment is obtained through forward calculation.

5. The method according to claim 1, characterized in that The method adopts an X-ray tomography method to obtain the internal three-dimensional structure data of the carbon fiber, obtains the annular structure image of the carbon fiber cross section by an image reconstruction method, and quantitatively analyzes the integrity and uniformity of the annular structure according to the image to determine the quality evaluation index of the annular structure, including: For carbon fiber tomography data, the scanning plane is spatially registered by using the marker point matching method, and the original data is denoised by Gaussian filtering to obtain the image sequence after denoising. According to the image sequence after the noise reduction processing, after compensating the scattering attenuation data using an iterative back-projection algorithm, edge features are extracted to obtain an image sequence after edge feature extraction; Extracting the cross-sectional contour according to the image sequence after edge feature extraction, obtaining the annular structure boundary through morphological closing operation, and obtaining the image sequence after the annular structure boundary is obtained; For the image sequence after obtaining the boundary of the annular structure, the circularity calculation method is used to quantify the integrity of the annular structure, the density distribution of the annular structure is calculated by the gray level co-occurrence matrix, and the internal pores are identified based on the regional connectivity analysis to obtain the structural defect data; According to the structural defect data, a feature extraction structure is constructed to obtain a ring structure quality evaluation index.

6. The method according to claim 1, characterized in that The carbon fiber surface micro-morphology characterization parameters and the ring structure quality evaluation index are integrated to evaluate the comprehensive performance of the recycled carbon fiber, analyze the influence of the surface structure change and the ring structure characteristics on the mechanical properties and conductive properties of the recycled carbon fiber, and obtain the correlation law of the carbon fiber structure performance, including: After normalization according to the carbon fiber surface micro-morphology characterization parameters and the ring structure quality index, outlier data points are eliminated; The principal component analysis method is used to reduce the dimension of the characteristic data, the number of principal components is determined by the variance contribution rate, and the coupling coefficient between parameters is calculated based on the correlation analysis to obtain the characteristic matrix; According to the characteristic matrix, an adaptive neural network is used to construct a performance prediction structure, network parameters are optimized by a back propagation algorithm, and prediction accuracy is calculated based on a mean square error function; The multivariate regression method was used to construct the mechanical properties prediction function, the key influencing factors were screened through significance test, and the contribution of each parameter was calculated based on variance analysis to obtain the mechanical properties mapping relationship. The conductivity prediction function is established by using sensitivity analysis method, and the order of parameter action is determined by partial correlation analysis to obtain the conductivity mapping relationship. Based on the mechanical property mapping relationship and the conductive property mapping relationship, the correlation law of carbon fiber structure properties is obtained.

7. The method according to claim 1, characterized in that The method predicts the quality and performance of the recycled carbon fiber samples obtained under different pyrolysis process conditions, analyzes and compares the prediction results, and optimizes the pyrolysis process parameters, including: The surface structure data of recycled carbon fibers under different pyrolysis process conditions were obtained, and the correlation coefficients between process parameters and structural characteristics were calculated through correlation analysis; A nonlinear mapping function is used to calculate the recovery mass prediction value of each group of samples, and the accuracy of the recovery mass prediction value is evaluated by cross-validation; A comprehensive evaluation function is used to construct a multi-objective optimization equation, and the weight coefficients of performance indicators are determined by the entropy weight method; The process parameters are grouped, and the optimal solution of the performance indicators is found through the gradient descent method to obtain the process parameter optimization plan.

Images