Method for realizing carbon fiber extraction control by using deep learning

Abstract

The invention provides a method for realizing carbon fiber extraction control by using deep learning, and the method comprises the steps: obtaining a scattered light intensity matrix and an interference image of a carbon fiber component, carrying out the feature extraction of image data through combining the scattered light intensity matrix and the interference image, and obtaining a feature vector reflecting the damage state of the carbon fiber component; the extracted feature vectors are input into a pre-trained damage state evaluation model, the damage degree of the carbon fiber component is judged, and if the damage degree exceeds a preset threshold value, the carbon fiber component is marked as a to-be-recycled component; according to the damage degree of the carbon fiber component, a multi-target optimization model of pyrolysis temperature, carbon fiber decomposition rate and carbon fiber purity is constructed, and target pyrolysis furnace temperature field distribution is determined; and according to the current performance evaluation or future performance change interval of the carbon fibers, a grading recycling scheme of the carbon fibers is generated by combining the damage degree of the carbon fibers, and full-life-cycle closed-loop management of carbon fiber recycling is formed.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for realizing carbon fiber extraction control by utilizing deep learning. Background Art

[0002] Carbon fiber materials, due to their high strength, light weight, and excellent corrosion resistance, have key application value in fields such as aerospace and high-end manufacturing. In particular, in high-performance scenarios such as Formula One racing, the performance of carbon fiber components directly affects safety and efficiency. However, carbon fiber components are susceptible to complex stresses, fatigue, and environmental factors during use, leading to microscopic damage and even failure. The efficiency and quality of their recycling and reuse have become a major bottleneck restricting their application throughout their life cycle. Existing monitoring and recycling methods mostly rely on traditional point sensors or single physical field detection, which makes it difficult to fully and real-time capture internal component damage. In addition, the pyrolysis temperature control during the recycling process is imprecise, and the material decomposition rate is unstable, making it difficult to ensure the purity and performance consistency of the recycled carbon fiber. These limitations make it difficult to accurately assess the damage state of carbon fiber components, and the recycling process is prone to resource waste and performance loss. The core challenges are concentrated in the following aspects: First, the technical difficulty of obtaining multi-dimensional damage information inside the component in real time is high, and traditional methods have shortcomings in spatial resolution and dynamic response speed. Second, the problem of accurately matching the pyrolysis temperature field distribution with the material decomposition behavior during the recycling process has not yet been solved, resulting in the need to improve the purity and performance of the carbon fiber. The lack of breakthroughs in these technical factors has led to incomplete component status assessment and low recycling efficiency. Therefore, how to build a real-time monitoring system based on a distributed fiber optic sensor array to accurately obtain the damage characteristics of carbon fiber components and establish a reliable assessment model, while achieving precise control of the pyrolysis process and prediction of carbon fiber properties through multi-objective optimization, has become a key issue in promoting the efficient recycling and performance optimization of carbon fiber components. Summary of the Invention

[0003] The present invention provides a method for realizing carbon fiber extraction control using deep learning, which mainly includes:

[0004] Obtain the scattered light intensity matrix and interference image of the carbon fiber component, combine the scattered light intensity matrix and interference image to perform feature extraction on the image data, and obtain a feature vector reflecting the damage state of the carbon fiber component;

[0005] The extracted feature vectors are input into a pre-trained damage state assessment model to determine the damage degree of the carbon fiber component. If the damage degree exceeds a preset threshold, the component is marked as a component to be recycled.

[0006] According to the damage degree of carbon fiber components, a multi-objective optimization model of pyrolysis temperature, carbon fiber decomposition rate, and carbon fiber purity is constructed to determine the target pyrolysis furnace temperature field distribution;

[0007] The target temperature field distribution of the pyrolysis furnace is used as the control target to generate a temperature control method for the pyrolysis furnace. The temperature control method includes adjusting the power of the heating element, matching the temperature distribution in the pyrolysis furnace with the target temperature field distribution, and obtaining the spatial distribution of the carbon fiber dissociation rate.

[0008] Continuously collect temperature data in the pyrolysis furnace during the pyrolysis extraction of carbon fiber, calculate the deviation between the temperature data in the pyrolysis furnace and the target temperature field distribution, and combine it with the spatial distribution of the carbon fiber dissociation rate to evaluate the potential impact of the deviation on the directional extraction efficiency. Dynamically adjust the power of the heating element based on the potential impact and obtain the carbon fiber properties after pyrolysis extraction;

[0009] Match historical cases in the preset database according to the degree of damage, compare and analyze the carbon fiber performance results corresponding to the historical cases obtained after matching, establish a performance reference system based on historical data mapping, and output the current performance evaluation of the carbon fiber and the future performance change range obtained by reference comparison;

[0010] Based on the current performance evaluation of carbon fiber or the future performance change range, combined with the degree of damage to the carbon fiber, a graded recycling and utilization plan for carbon fiber is generated, forming a closed-loop management of the entire life cycle of carbon fiber recycling.

[0011] Furthermore, a distributed fiber optic sensor array is used to obtain the scattered light intensity matrix and interference image of the carbon fiber component, and feature extraction is performed on the image data to obtain a feature vector reflecting the damage state of the carbon fiber component. This method includes: setting a sensor layout grid based on the surface stress distribution of the carbon fiber component, placing a fiber Bragg grating sensor at each grid node, and exciting the fiber Bragg grating sensor with a preset wavelength light source. Scattered light signals are obtained from the fiber Bragg grating sensor, time-domain sampling of the scattered light signals is performed, and an initial scattered light intensity matrix is constructed according to the corresponding relationship between the sensor grid coordinates. The fiber coupling efficiency value at each grid node is calculated based on the initial scattered light intensity matrix. The fiber coupling efficiency value is corrected using a fiber transmission loss compensation function, and then converted to the frequency domain using a Fourier transform to extract phase information. The interference image is then calculated based on the phase difference. The interference image is then multi-scale decomposed using a wavelet transform to extract grayscale co-occurrence matrix features, local binary pattern features, and directional gradient histogram features from the interference image. The extracted features are standardized and reduced in dimension using principal component analysis. Clustering is then performed using a density clustering algorithm. If the Euclidean distance between cluster centers is greater than a preset damage determination threshold, the grid node is considered damaged. Based on the damage determination results and clustering characteristics, a damage feature vector is constructed. This vector contains the damage location coordinates, damage severity coefficient, and damage area range parameters.

[0012] Furthermore, the extracted feature vectors are input into a pre-trained damage assessment model. A pattern recognition algorithm is used to determine the damage extent of the carbon fiber component. If the damage extent exceeds a preset threshold, the component is marked for recycling. This process involves normalizing the input feature vectors, calculating feature vector normalization parameters using a maximum-minimum normalization method, and transforming the feature vectors based on the normalization parameters to obtain a normalized feature vector. A training sample set is constructed based on the normalized feature vectors and damage type labels. A stratified sampling method is used to divide the training sample set into a training subset and a validation subset to obtain basic damage assessment data. A support vector machine is used to train the training subset data. The optimal damage assessment model is obtained by optimizing the kernel function parameters and penalty factor. The validation accuracy is calculated using the validation subset. The normalized feature vectors are input into the damage assessment model. The damage type probability distribution is obtained based on the model output. The type with the highest probability is selected as the component damage type. The strain energy density value of the damaged area is calculated using a strain energy density calculation function. The damage extension trend is determined using a preset strain energy density threshold. A comprehensive damage severity index is calculated based on the strain energy density value and the damage type probability distribution. This comprehensive damage severity index is then graded using a preset damage severity threshold. If the comprehensive damage degree index exceeds the preset recovery threshold, a component recovery mark is generated, the damage assessment result and assessment time are recorded, and a recovery component data record is constructed.

[0013] Furthermore, based on the degree of damage to carbon fiber components, a multi-objective optimization model was constructed for pyrolysis temperature, carbon fiber decomposition rate, and carbon fiber purity to determine the target pyrolysis furnace temperature distribution. This model involved obtaining pyrolysis temperature, carbon fiber decomposition rate, and fiber purity data from a historical pyrolysis database, arranging the data in time series according to the recorded timestamps, and using a sliding average method to eliminate data fluctuations. A three-dimensional feature space was constructed for pyrolysis temperature, carbon fiber decomposition rate, and fiber purity. A temperature field prediction function was established using multivariate Gaussian process regression, and the kernel function parameters were determined using maximum likelihood estimation. A multi-objective optimization function was constructed based on the temperature field prediction function, with lower limits for the carbon fiber decomposition rate and fiber purity set as constraints. A genetic algorithm was used to solve for the optimal temperature field distribution. Based on the optimal temperature field distribution, a matrix decomposition reaction kinetic equation was established, and the matrix decomposition rate at each temperature point was calculated. The matrix residue distribution was then obtained by integrating the pyrolysis reaction time. The matrix residue distribution data was used to construct a temperature field correction function, which was then locally corrected to obtain the corrected temperature field distribution data. A temperature control parameter sequence is generated based on the corrected temperature field distribution data. The temperature control region is then divided based on the pyrolysis furnace pressure and oxygen concentration constraints. The local temperature gradient is calculated within the divided temperature control region. The temperature field uniformity is determined using the temperature gradient threshold to determine the target distribution of the pyrolysis furnace temperature field.

[0014] Furthermore, a temperature control scheme for the pyrolysis furnace is generated, using the target pyrolysis furnace temperature field distribution as the control target. This temperature control scheme involves adjusting the power of the heating elements, matching the temperature distribution within the pyrolysis furnace with the target temperature field distribution, and determining the spatial distribution of the carbon fiber dissociation rate. The scheme involves: arranging an array of temperature sensors axially and circumferentially along the inner wall of the pyrolysis furnace, using preset spacing to determine the sensor placement, and constructing a real-time temperature field acquisition network for the pyrolysis furnace. A temperature control curve is established based on the target pyrolysis furnace temperature field distribution. A neural network is used to map the input power of the heating elements to the temperature field output parameters, generating initial heating power control parameters. Real-time temperature data within the pyrolysis furnace is acquired through the temperature sensor array. A Kalman filter is used to eliminate temperature data noise, and filtered temperature values are calculated based on the state estimation equation and observation equation. Based on the filtered temperature values, the deviation between the actual temperature and the target temperature at each measurement point is calculated. A heating power correction factor is generated based on the temperature deviation distribution, and a heating power compensation sequence is constructed. A separate power controller is set for each group of heating elements, and a proportional-integral controller is used to adjust the heating power in real time. The heating power output is optimized using a feedback compensation function. The relationship between the adjusted heating power and the temperature field was recorded. The thermal conductivity coefficient was calculated based on the temperature field time series data, and the temperature field distribution was solved using the heat transfer equation. The pyrolysis reaction rate equation was established based on the temperature field distribution data. The dissociation rate was calculated using the carbon fiber dissociation activation energy, and the dissociation rate distribution was generated using a spatial interpolation function.

[0015] Furthermore, the measured temperature field within the pyrolysis furnace is constructed and discretized to obtain a discrete temperature point set. A target temperature field is retrieved from a preset database and converted into a target discrete point set. If the Euclidean distance between the measured discrete point set and the target discrete point set exceeds a preset threshold, the measured point set is adjusted to obtain a matching temperature field. The dissociation rate is calculated, and rate distribution data is generated and three-dimensionally mapped to generate a spatial distribution. Based on this spatial distribution, carbon fibers are then directionally extracted. This involves: arranging an array of temperature sensors along the axial and radial directions of the pyrolysis furnace's inner wall, determining the sensor position coordinates at preset spatial intervals, and collecting real-time temperature data from the pyrolysis furnace. The real-time temperature data is subjected to noise filtering, using a median filter to eliminate temperature fluctuations, and outliers are removed by determining the validity of the temperature data to obtain basic data on the pyrolysis furnace temperature field. A three-dimensional grid coordinate system is established in the axial, radial, and circumferential directions. Grid node temperature values are calculated using cubic spline interpolation to generate a discrete point set of the measured temperature field. The target temperature field data is extracted from the preset database and converted to the grid coordinate system of the measured point set according to spatial coordinate mapping rules to generate a discrete point set of the target temperature field. The temperature deviation between the measured discrete point set and each corresponding point in the target discrete point set is calculated. The degree of temperature field matching is evaluated using weighted Euclidean distance, and the temperature field adjustment region is determined based on a preset temperature matching threshold. A temperature distribution correction function is established for the temperature field adjustment region, and the gradient descent method is used to calculate the temperature correction amount for the grid nodes. The matching temperature field is then iterated and updated through temperature correction. The carbon fiber pyrolysis reaction rate is calculated based on the matching temperature field, and a three-dimensional scattered point interpolation function is established to generate continuous dissociation rate distribution data. The dissociation rate distribution data is spatially sliced using the isosurface method, and a carbon fiber directional extraction trajectory is constructed based on the dissociation rate gradient direction.

[0016] Furthermore, temperature data within the pyrolysis furnace during carbon fiber extraction is continuously collected. The deviation between the temperature data and the target temperature field distribution is calculated. Combined with the spatial distribution of the carbon fiber dissociation rate, the potential impact of this deviation on the directional extraction efficiency is assessed. The power of the heating element is dynamically adjusted based on this potential impact, and the carbon fiber properties after pyrolysis extraction are determined. This process involves continuously acquiring temperature data from a temperature sensor array at preset sampling intervals, filtering the temperature data in real time using a Kalman filter, and calculating the temperature field distribution within the pyrolysis furnace using a three-dimensional polynomial interpolation function. Support vector regression is used to calculate the spatial deviation between the measured and target temperature fields. The spatial distribution of the dissociation rate is calculated based on the carbon fiber pyrolysis activation energy, and a mapping function for the impact of temperature deviation on dissociation uniformity is established. Based on the temperature deviation and dissociation rate distributions, localized areas of insufficient dissociation, uneven dissociation, and excessive fiber damage are calculated. A weighted scoring function is used to quantify the impact of temperature field deviation. A heating power correction function is constructed based on the temperature field deviation impact score, with differentiated power adjustment coefficients applied to different impact areas. The heating element output power is calculated using the power response characteristic curve. A feedback compensation function was used to eliminate heating response lag, and power regulation parameters were optimized based on the temperature dynamic response curve to achieve rapid temperature field convergence. Test samples were obtained from recycled carbon fiber, and the degree of graphitization was measured using a Raman spectrometer. The carbon fiber structural integrity index was calculated based on the characteristic peak intensity ratio. Electron microscopy was used to scan and capture carbon fiber surface morphology images. Surface defect features were extracted using an image recognition algorithm, and the fiber surface damage index was calculated.

[0017] Furthermore, the damage severity is matched against historical cases in a preset database. The carbon fiber performance results corresponding to the matched historical cases are compared and analyzed. A performance reference system based on historical data mapping is established, outputting the current carbon fiber performance assessment and future performance change range derived from the reference comparison. This involves collecting three-dimensional feature data on the carbon fiber's damage area ratio, damage depth, and damage type, constructing a damage severity feature vector through feature normalization, and extracting historical case data from the preset database. Cosine similarity is calculated based on the damage severity feature vector, and a similarity threshold is set to filter matching historical cases. A mapping relationship between the current carbon fiber damage severity and the historical cases is established. Carbon fiber tensile strength, elastic modulus, elongation at break, and surface defect density parameters are extracted from the historical cases. Support vector regression is used to construct a performance reference mapping function. A piecewise polynomial fit is performed on the performance parameter sequences of the filtered historical cases to establish a time-series variation function for different performance parameters and determine performance evolution trends. The current carbon fiber performance assessment is calculated using the performance reference mapping function, and the future performance change range is predicted using the time-series variation function. The performance change range is sampled using the Monte Carlo method to generate a performance change probability distribution and calculate a confidence interval for the prediction result. A performance reference system is constructed based on the performance evaluation value and the variation range, the performance prediction data and confidence index are recorded, and a carbon fiber performance evaluation report is generated.

[0018] Furthermore, based on the current performance evaluation or future performance change range of carbon fiber and the degree of carbon fiber damage, a graded recycling and utilization plan for carbon fiber is generated, forming a closed-loop management system for the entire carbon fiber recycling lifecycle. This includes: constructing a performance index system based on carbon fiber tensile strength, elastic modulus, elongation at break, and surface integrity. A random forest algorithm is used to combine and analyze these performance indicators with damage characteristics to generate a graded evaluation score. Based on the graded evaluation scores, a grading system is established for superior, standard, and basic carbon fiber grades. Grading stability is determined by performance decay rate and fluctuation amplitude, and grade timestamps are recorded. A hierarchical analysis method is used to establish a performance parameter limit matrix for different application scenarios. An application scenario sequence is generated through performance matching calculations to construct a recycled carbon fiber application allocation plan. A unique carbon fiber identification code is generated based on recycling batch, grading results, and application scenario. Recycling process parameters, performance test data, and application scenario limits are recorded. A data archive for the entire carbon fiber lifecycle is established, recording original component information, damage assessment data, recycling process parameters, performance test results, and application scenario data. Performance warning thresholds are set based on performance prediction curves. Performance test data is used to determine carbon fiber performance decay trends and generate performance warning signals. A recurrent neural network is used to learn the evolution of carbon fiber performance, predict performance change trends in different application scenarios, and update digital traceability archives.

[0019] Furthermore, performance data is collected and analyzed from carbon fiber samples. Damage data is also extracted from the samples to determine the damage severity assessment result. If the damage severity assessment result is below a preset threshold, damage type classification is performed to obtain a damage type distribution. Based on the damage type distribution and performance analysis results, grading standards are developed and a graded recycling plan is determined. Recycling methods are derived from this, utilization efficiency is optimized, and recycling paths are obtained. This grading recycling plan is then formed. This includes: using a mechanical property tester to obtain carbon fiber tensile strength, elastic modulus, elongation at break, and surface integrity data; scanning electron microscopes are used to obtain surface topography images to establish a carbon fiber performance database. Characteristic parameters such as crack length, damage area, and defect density are extracted from the surface topography images. A deep convolutional neural network is used to classify the damage characteristics and generate damage type identification results. The damage characteristic parameters are compared with a preset threshold, and a hierarchical clustering algorithm is used to classify the damage types into a hierarchy, generating a spatial distribution map of the damage types. A multi-layer perceptron is used to map performance indicators to damage types. Based on the mapping results, the carbon fibers are graded and scored to generate graded assessment data. Recycling grades are determined based on graded assessment data. Performance parameter limits are set for different recycling grades, and a recycling standards database is established. Genetic algorithms are used to optimize recycling process parameters. Recycling efficiency evaluation functions are used to calculate recycling rates and identify the optimal recycling path. A process parameter sequence is generated based on the optimal recycling path, and a graded recycling data archive is established to record data from the entire recycling process.

[0020] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0021] The present invention discloses a method for realizing carbon fiber extraction control by using deep learning. The scattered light intensity matrix and interference image of the carbon fiber component are obtained through a distributed optical fiber sensor array, and the characteristic vector is extracted and input into a pre-trained model to evaluate the degree of damage. According to the degree of damage, a multi-objective optimization model of pyrolysis temperature, decomposition rate and purity is constructed to determine the temperature field distribution of the target pyrolysis furnace. During the pyrolysis process, temperature data is continuously collected and the impact on the directional extraction efficiency is evaluated, and the power of the heating element is dynamically adjusted. At the same time, a performance reference system is established based on historical data mapping to evaluate the current and future performance of carbon fiber. Finally, based on the performance evaluation and the degree of damage, a graded recycling and utilization plan is generated to realize closed-loop management of the entire life cycle of carbon fiber recycling. Through intelligent perception, multi-objective optimization and dynamic control, the present invention improves the accuracy and efficiency of carbon fiber recycling and realizes high-value utilization of carbon fiber. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 This is a flow chart of a method for realizing carbon fiber extraction control using deep learning according to the present invention.

[0023] Figure 2This is a schematic diagram of a method for realizing carbon fiber extraction control using deep learning according to the present invention. DETAILED DESCRIPTION

[0024] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0025] like Figure 1 In this embodiment, a method for implementing carbon fiber extraction control using deep learning may specifically include:

[0026] S101 collects scattered light signals and interference images of carbon fiber components through a distributed fiber optic sensor array, and extracts feature vectors that reflect the damage status of the components.

[0027] Damage monitoring of carbon fiber components is achieved through a distributed fiber optic sensor array. Based on the Bragg grating principle, this array accurately captures strain and damage information within the component. Sensor placement follows the stress distribution pattern on the component surface, prioritizing denser meshing in areas of stress concentration to improve monitoring accuracy.

[0028] like Figure 2 ,S1011 sets up a sensor grid according to the surface stress distribution law of carbon fiber components, arranges fiber Bragg grating sensors at the grid nodes, uses a preset wavelength light source to excite the sensor and collects scattered light signals.

[0029] Stress concentration areas on component surfaces, such as geometric discontinuities or load application points, require a denser sensor layout. The grid spacing is set at 5 mm in stress concentration areas and 10 mm elsewhere to balance monitoring accuracy and resource consumption. The fiber Bragg grating sensor uses a central wavelength of 1550 nm, a reflectivity of 90%, and a grating length of 10 mm. A laser light source is used to excite the sensor, collecting scattered light signals at a sampling rate of 100 kHz and a sampling duration of 1 second to generate an initial signal matrix containing spatial information of the grid nodes.

[0030] S1012 performs time-domain sampling and loss compensation on the scattered light signal, constructs a corrected scattered light intensity matrix, and generates an interference image through frequency-domain conversion.

[0031] The scattered light signal is sampled in the time domain using a high-speed data acquisition card, and an initial scattered light intensity matrix is constructed according to grid coordinates. A loss compensation function is applied to correct the signal intensity for the inherent loss of 0.2 dB / km in optical fiber transmission and bending loss, generating a corrected scattered light intensity matrix. A Fourier transform is used to convert the matrix into the frequency domain, extracting phase information. Interference patterns are then calculated based on the phase differences between adjacent nodes, reflecting the strain field changes caused by component deformation and damage.

[0032] S1013 performs multi-scale feature extraction and cluster analysis on the interference image to construct a damage feature vector.

[0033] The interference image was decomposed at multiple scales using wavelet transform to extract gray-level co-occurrence matrices, local binary patterns, and histograms of oriented gradients (HGGs). The gray-level co-occurrence matrix was calculated considering four directions: 0, 45, 90, and 135 degrees, extracting statistics such as energy, contrast, and correlation. The LBPs used a neighborhood window with a radius of 2 to extract uniform pattern features. The HGGs collected edge information in nine directions. After normalization, the extracted features were dimensionality reduced using principal component analysis. Eigenvectors with a cumulative contribution rate of 95% were selected to generate the reduced-dimensionality feature matrix. A density clustering algorithm was used with a neighborhood radius of 0.5 and a minimum sample size of 4 to calculate the Euclidean distance between cluster centers. If the distance was greater than the preset damage threshold (twice the average distance to a healthy component), the node was deemed damaged. The damage feature vector contains the damage location coordinates, severity coefficient, and regional extent, determined by the cluster density deviation and the convex hull of the boundary.

[0034] This method has demonstrated excellent performance in damage detection for carbon fiber laminates, achieving over 90% accuracy for 2mm impact damage and achieving positioning accuracy superior to grid spacing. In online monitoring of large components such as wind turbine blades, this method can provide early warning of damage, significantly improving component safety.

[0035] S102 inputs the extracted damage feature vector into a pre-trained damage state assessment model to determine the damage degree of the carbon fiber component and determines whether to mark it as a component to be recycled based on the damage degree classification.

[0036] The damage feature vector contains information about the location, extent, and scope of component damage. Through pattern recognition using a deep learning model, it can efficiently assess the damage state of carbon fiber components. The assessment results are used to determine whether components should be recycled, thereby optimizing resource utilization and component management.

[0037] S1021 normalizes the damage feature vector to generate a normalized feature vector, providing a unified input format for the damage assessment model.

[0038] The damage eigenvector contains multidimensional physical quantities such as displacement field, strain field, and stress field, with significant differences in numerical ranges. For example, displacement values may be at the micron level, while stress values may reach the megapascal level. To eliminate the influence of dimension, the maximum and minimum value normalization method is used to map the eigenvalue components to the range of 0 to 1. The specific operation involves calculating the maximum and minimum values of each eigenvalue component, generating standardized parameters, and then normalizing the eigenvalue vector through linear transformation. The normalized eigenvalue can effectively improve the stability and convergence speed of model training, providing consistent input for damage assessment.

[0039] S1022 constructs a training sample set based on the normalized feature vector, and trains a support vector machine model to generate a damage status assessment model.

[0040] The training sample set consists of normalized feature vectors and corresponding damage type labels, which include various types such as delamination damage, matrix cracking and fiber fracture. Due to the uneven number of damage samples of various types, for example, there may be fewer fiber fracture samples, a stratified sampling method is used to divide the training subset and the validation subset in a ratio of 7:3 to ensure that the proportion of each type of sample is consistent. The training subset contains no less than 600 damage samples of each type, and the validation subset contains no less than 250. The support vector machine model uses a radial basis kernel function, and the kernel function parameter γ is optimized through grid search in the range of 0.001 to 10, and the penalty factor C is in the range of 1 to 1000. The optimal parameter combination is determined when the accuracy of the validation set reaches 93%. During the training process, the model learns the mapping relationship between feature vectors and damage types through soft margin classification to generate a damage status assessment model.

[0041] S1023 inputs the normalized eigenvector into the damage state assessment model, obtains the probability distribution of damage types and determines the component damage type.

[0042] The normalized feature vector is input into a pre-trained support vector machine model, which outputs a probability distribution for each damage type, reflecting the confidence level in the damage type judgment. By selecting the type with the highest probability, the specific damage type of the component, such as delamination or fiber fracture, is determined. Confidence analysis of the probability distribution helps identify model uncertainties. For example, when the probabilities of multiple damage types are close, further verification can be triggered. Once the damage type is determined, the corresponding confidence value is recorded to provide a basis for damage severity calculation.

[0043] S1024 calculates the strain energy density of the damaged area, generates a comprehensive damage degree index based on the probability distribution of the damage type, and performs a graded judgment.

[0044] For each identified damage type, a strain energy density calculation function is used to assess the energy concentration in the damaged area. Taking into account the anisotropic properties of carbon fiber components, the ratio of the elastic modulus in the fiber direction to the perpendicular direction is approximately 15:1. The strain energy density is calculated using the displacement field and the material stiffness matrix, with a threshold of 0.5 joules per cubic millimeter in the fiber direction and 0.2 joules per cubic millimeter in the perpendicular direction. The calculated results reflect the damage extension trend; when the strain energy density exceeds the threshold, it indicates that the damage may worsen. A comprehensive damage severity index is developed by weighting the strain energy density value and the damage type probability, with weights of 0.65 and 0.35, respectively, to ensure that high-confidence damage types contribute more to the index. The comprehensive index is categorized as mild, moderate, and severe, with thresholds of 0.25, 0.55, and 0.85, respectively. If the index exceeds 0.85, the component is marked for recycling, and a recycling mark is automatically generated, along with the assessment timestamp and damage details.

[0045] The combination of feature vector preprocessing and model training significantly improves the accuracy of damage assessment. For example, in wind turbine blade inspection, this method identified delamination damage at the root of the blade at 75% of its span, with an area of approximately 180 square millimeters, a depth of 1.8 millimeters, a confidence level of 0.96, a strain energy density of 0.58 joules per cubic millimeter, and a comprehensive index of 0.87, triggering a recycling flag. In aerospace composite panel inspection, fiber fracture damage was successfully identified, providing a reliable basis for maintenance decisions. Through standardized processing and stratified sampling, the model's generalization capability was enhanced, and the validation set accuracy remained stable at over 93%, significantly reducing the false positive rate and optimizing the efficiency of the component recycling process.

[0046] S103 constructs a multi-objective optimization model based on the damage degree of the carbon fiber component to determine the target temperature field distribution of the pyrolysis furnace, and optimizes the pyrolysis temperature, carbon fiber decomposition rate and fiber purity.

[0047] The pyrolysis recycling of carbon fiber components requires comprehensive consideration of the impact of damage on the pyrolysis process. A multi-objective optimization model is used to efficiently control the temperature field, ensuring the high purity and performance stability of the recycled fiber. The optimization process, combined with historical data analysis and dynamic correction, significantly improves the accuracy and resource utilization of the pyrolysis process.

[0048] S1031 extracts relevant data from the historical pyrolysis database and applies filtering processing to generate a stable three-dimensional feature space.

[0049] The pyrolysis history database stores time series data of key parameters such as pyrolysis temperature, carbon fiber decomposition rate, and fiber purity. To eliminate random fluctuations in the data, the sliding average method is used to filter the data arranged by timestamp, and the sliding window is set to 60 minutes to generate a smooth filtered data set. The filtered data is used to construct a three-dimensional feature space, with pyrolysis temperature as the independent variable and decomposition rate and purity as dependent variables, reflecting the nonlinear relationship between parameters. For example, historical data shows that the decomposition of matrix resin is mainly concentrated in the range of 380 to 450 degrees Celsius, while carbon fiber maintains structural stability below 600 degrees Celsius. The construction of the feature space provides a reliable basis for subsequent regression analysis.

[0050] S1032 establishes a temperature field prediction function through multivariate Gaussian process regression and optimizes kernel function parameters.

[0051] Based on a three-dimensional feature space, a multivariate Gaussian process regression method was used to establish a mapping relationship between pyrolysis temperature, decomposition rate, and purity. A radial basis kernel function was selected, and kernel parameters were optimized using maximum likelihood estimation. The characteristic length scale was determined to be 15 degrees Celsius, indicating the high sensitivity of the pyrolysis process to temperature changes. The regression model predicts decomposition behavior at different temperatures by learning from the distribution characteristics of historical data, ensuring the accuracy of the temperature field design. During model training, the mean squared error on the validation set was kept within 0.5%, demonstrating high prediction accuracy.

[0052] S1033 constructs a multi-objective optimization function and applies a genetic algorithm to solve the optimal temperature field distribution.

[0053] The multi-objective optimization function aims to maximize the carbon fiber decomposition rate and fiber purity while minimizing the pyrolysis energy consumption. The constraints set include a decomposition rate lower limit of 95%, a purity lower limit of 98%, and a temperature range of 350 to 750 degrees Celsius. A genetic algorithm was used to solve the problem, with a population size of 120 and 600 generations of evolution. After convergence, the optimal temperature field distribution was generated. The optimization results show that the temperature field maintains a long residence time at 400 degrees Celsius, the key point of resin decomposition, forming a step-type heating curve, which effectively prolongs the stable stage of matrix decomposition and improves decomposition efficiency.

[0054] S1034 calculates the residual distribution based on the matrix decomposition kinetic equation, corrects the temperature field distribution and generates target control parameters.

[0055] Based on the optimal temperature field distribution, a matrix decomposition reaction kinetic equation is constructed to describe the decomposition rate of the resin at different temperatures. The equation takes into account the coupling effect of temperature and oxygen concentration. The decomposition rate peak is 0.18% per minute at 400 degrees Celsius. The matrix residue is calculated by time integration, which drops from the initial 35% to below 2%. Using the residue distribution data, a temperature field correction function is constructed. Local correction is performed on areas within the 200 mm horizontal range of the pyrolysis furnace where the temperature gradient is greater than 5 degrees Celsius per meter to generate a corrected temperature field distribution. The corrected temperature field is divided into three temperature control areas. An independent temperature control parameter sequence is generated for each area. Combined with the constraint of an oxygen concentration below 2%, temperature uniformity is ensured. The control parameter sequence guides the dynamic adjustment of the heating power of the pyrolysis furnace.

[0056] This method has demonstrated excellent performance in the pyrolysis recycling of wind turbine blades. For example, using a 50 kg blade fragment as an example, the initial temperature was 380°C, then raised at 2.5°C / minute to 420°C, held for 100 minutes, and then raised to 480°C for 70 minutes. The final fiber purity reached 98.7%, with a residual resin content of less than 1.3%. The tensile strength of the recycled fiber was retained at 92%, meeting high-end recycling requirements. In the pyrolysis of aviation composite materials, the segmented temperature control strategy effectively minimized fiber damage, resulting in stable recycled fiber performance and providing a highly efficient solution for material recycling.

[0057] S104 takes the target pyrolysis furnace temperature field distribution as the control target, generates the temperature control strategy of the pyrolysis furnace, achieves accurate matching between the measured temperature field and the target temperature field by adjusting the power of the heating element, and calculates the spatial distribution of the carbon fiber dissociation rate to guide the directional extraction.

[0058] The temperature field control of the pyrolysis furnace is the core link of the carbon fiber recycling process. Through real-time monitoring, dynamic adjustment and data-driven optimization, the uniformity of the temperature field distribution and the stability of the dissociation process are ensured, thereby improving the purity and mechanical properties of the recycled fiber.

[0059] S1041 arranges a temperature sensor array to collect real-time temperature data, and uses Kalman filtering to generate smooth temperature field basic data.

[0060] An array of high-precision temperature sensors is arranged axially and circumferentially within the pyrolysis furnace, forming a multi-point temperature acquisition network. In the cylindrical pyrolysis furnace, which measures 2 meters in diameter and 4 meters in length, a temperature measurement ring is installed every 200 mm along the axial direction. Each ring contains eight K-type thermocouple sensors spaced 45 degrees apart circumferentially, for a total of 32 measurement points. The temperature range covers 0 to 1000 degrees Celsius, with a sampling frequency of 1 Hz. Real-time temperature data is subject to random fluctuations due to thermal noise and electromagnetic interference. To improve data reliability, a Kalman filter is used for processing. This combines a state estimation equation to describe the temperature change rate, and the observation equation incorporates sensor noise characteristics. After filtering, the temperature fluctuation amplitude is reduced from the original ±2 degrees Celsius to 0.4 degrees Celsius. This filtered temperature data provides a high-precision foundation for deviation analysis and power regulation.

[0061] S1042 calculates the deviation based on the filtered temperature data, generates a heating power correction value, and adjusts the heating element through a proportional-integral control algorithm.

[0062] The temperature deviation of each temperature measurement point is calculated based on the filtered temperature data and the target temperature field. If the deviation exceeds 5 degrees Celsius, the corresponding heating power correction is generated. The pyrolysis furnace is equipped with 6 groups of independently controlled heating elements, each with a rated power of 20 kilowatts. The power controller adopts a proportional-integral control algorithm. The proportional coefficient is dynamically adjusted according to the deviation range, and the integral time constant is set to 50 seconds to ensure both fast response and stability. The control signal is optimized through the feedback compensation function and output to the heating element to achieve real-time power regulation. For example, near the critical temperature point of 420 degrees Celsius, the area with large deviation is temperature compensated by rapid power increase, and the deviation converges to within plus or minus 3 degrees Celsius. The correspondence between the power regulation process record and the temperature field provides data support for heat transfer analysis.

[0063] S1043 constructs the heat transfer equation and the pyrolysis reaction rate equation, calculates the dissociation rate and generates the spatial distribution.

[0064] According to the adjusted heating power and real-time temperature field data, a heat transfer equation is established to describe the heat conduction, convection and radiation heat transfer characteristics in the pyrolysis furnace. The comprehensive heat transfer coefficient is about 28 watts per square meter Kelvin at 400 degrees Celsius. It is solved by the finite difference method, with a grid spacing of 50 mm, a time step of 1 second, and the temperature field uniformity controlled at plus or minus 3 degrees Celsius. Based on the temperature field distribution and combined with the carbon fiber dissociation activation energy of 150 kilojoules per mole, the Arrhenius equation is applied to calculate the dissociation rate. At 420 degrees Celsius, the dissociation rate peak is 0.16% per minute, the central area has a higher rate, and the peripheral area is about 80% of the center. A continuous dissociation rate distribution is generated by the spatial interpolation function, showing a circular ring feature, which guides the directional extraction path planning of carbon fiber.

[0065] S1044 discretizes the measured temperature field, optimizes the matching with the target temperature field, and generates a matching temperature field.

[0066] The measured temperature field is discretized using a three-dimensional grid coordinate system. The grid size is increased to 20 mm in areas with larger temperature gradients and increased to 50 mm in edge areas. Cubic spline interpolation is used to calculate the temperature values of the grid nodes to generate a discrete point set of the measured temperature field. The target temperature field is extracted from the preset database and mapped to the same coordinate system to form a target discrete point set. The weighted Euclidean distance between the two point sets is calculated, with the key temperature area having a higher weight. If the distance exceeds 5 degrees Celsius, adjustment is initiated. The temperature distribution correction function is optimized using the gradient descent method with a learning rate of 0.12. After 60 iterations, a matching temperature field is generated, and the mapping error is controlled within 1.5 degrees Celsius. The matching temperature field is used to verify the uniformity of the dissociation rate distribution.

[0067] For example, for a 50 kg blade fragment, the target temperature field is maintained at 420 degrees Celsius at the center and decreases at the edges. The measured temperature field is precisely matched through power regulation, with a dissociation rate distribution coefficient of less than 8%, a recycled fiber purity of 98.6%, and a mechanical property retention rate of 92%. In the recycling of automotive composite parts, the temperature field uniformity is controlled within ±2.5 degrees Celsius, the dissociation rate exhibits a spiral distribution, and fiber damage is reduced by 15% after directional extraction trajectory optimization. In the recycling of aviation composite materials, precise temperature field control ensures the uniformity of dissociation of large-scale components, and the recycled fibers meet the requirements of high-performance reuse.

[0068] S105 continuously monitors the temperature data in the pyrolysis furnace, analyzes the impact of temperature field deviation on carbon fiber dissociation efficiency, dynamically adjusts the power of the heating element to optimize the extraction process, and evaluates the performance indicators of the recycled carbon fiber.

[0069] Real-time control of the temperature field during pyrolysis is crucial to the uniformity of carbon fiber dissociation and the quality of recycled materials. Through continuous monitoring, deviation analysis, and dynamic adjustment, we ensure that the temperature field closely matches the target distribution. Performance testing is also combined with the structural integrity of the recycled fibers to ensure high-quality reuse.

[0070] S1051 uses a temperature sensor array to collect real-time temperature data from the pyrolysis furnace and reconstructs the temperature field distribution through Kalman filtering and three-dimensional interpolation.

[0071] The pyrolysis furnace is equipped with 128 platinum-rhodium thermocouple sensors, which are arranged every 150 mm along the axial direction and every 45 degrees circumferentially, forming a high-density temperature acquisition network with a sampling interval of 1 second and a temperature measurement accuracy of plus or minus 0.5 degrees Celsius. The real-time temperature data is affected by thermal noise and is processed using a Kalman filter. The state noise covariance is set to 0.08 and the observation noise covariance is set to 0.04. After filtering, the temperature fluctuation amplitude is reduced to 0.25 degrees Celsius. Based on the filtered data, a fifth-order three-dimensional polynomial interpolation function is applied to reconstruct the temperature field distribution on a grid with a spacing of 50 mm. The interpolation error is controlled within 0.8 degrees Celsius. The reconstructed temperature field reflects the spatial continuity of the temperature in the pyrolysis furnace, providing accurate input for deviation analysis.

[0072] The S1052 uses support vector regression to analyze temperature field deviations, quantify the impact on dissociation rates, and identify problem areas.

[0073] The measured temperature field is compared with the target temperature field through the support vector regression algorithm to generate a spatial deviation distribution. The center of the target temperature field is set to 420 degrees Celsius and the edge to 380 degrees Celsius. The deviation calculation shows that the maximum deviation in the end area is 7.5 degrees Celsius. Combined with the carbon fiber dissociation activation energy of 150 kilojoules per mole, the dissociation rate is calculated based on the Arrhenius equation, and the rate fluctuation in the deviation area can reach 25%. The deviation distribution is used to identify insufficient dissociation areas below 380 degrees Celsius, uneven dissociation areas with deviations exceeding 5 degrees Celsius, and excessive fiber damage areas above 450 degrees Celsius. A weighted scoring function is used to quantify the deviation impact, and the weight is assigned according to the sensitivity of the region to the dissociation efficiency. For example, the weight of the insufficient dissociation area is 0.5, and the excessive damage area is 0.3, to generate deviation impact scoring data.

[0074] S1053 dynamically adjusts heating power based on the deviation impact score to optimize temperature field convergence.

[0075] A heating power correction function is constructed based on the deviation impact score, with differentiated adjustment coefficients set for each zone: a coefficient of 1.25 for areas with insufficient dissociation and a coefficient of 0.75 for areas with excessive damage. The pyrolysis furnace is equipped with eight heating elements, each with a power of 25 kilowatts. A proportional-integral control algorithm is used, with a response time of 8 seconds and an integral time constant of 45 seconds. A feedback compensation function eliminates power regulation lag, and parameters are optimized through a dynamic response curve, ensuring that temperature deviations converge to within ±2.8 degrees Celsius within 2.5 control cycles. Power adjustment records correspond to the deviation distribution, ensuring rapid stabilization of the temperature field and minimizing uneven dissociation.

[0076] The S1054 tests the properties of recycled carbon fiber, evaluating structural integrity and surface defects.

[0077] Samples of recycled carbon fiber were collected and the degree of graphitization was measured using a Raman spectrometer. The structural integrity index was calculated by the intensity ratio of the D-peak to the G-peak. An intensity ratio below 0.18 indicated good graphitization of the fiber. Surface defects were identified using an electron microscope at 6000x magnification and a resolution of 3.5 nanometers per pixel. Defect density was calculated, and defects below 0.08 per square micron were considered high quality. These test results guide the tiered utilization of recycled fiber to meet the needs of different application scenarios.

[0078] When processing 100 kg of blade fragments, the temperature field deviation was controlled within ±3°C, the coefficient of variation of the dissociation rate was less than 7%, the tensile strength retention of the recycled fiber was 96%, and the residual resin content was 0.4%. In aviation composite recycling, dynamic power regulation eliminates end overheating, resulting in clear grooves on the fiber surface and a single filament diameter fluctuation of 0.15 microns. In automotive parts recycling, precise temperature field control ensures the integrity of the fiber bundle, allowing the recycled fiber to meet the requirements of high-performance composite materials and reduce defect density by 20%.

[0079] S106 establishes a performance reference system based on the damage degree of carbon fiber components and matches historical cases to evaluate current carbon fiber performance and predict future performance change ranges.

[0080] By analyzing the similarity between carbon fiber damage characteristics and historical data, a performance reference system is constructed to provide a scientific basis for the performance evaluation and long-term stability prediction of recycled carbon fiber, ensuring the reliability of recycled fiber in high-end applications.

[0081] S1061 extracts the three-dimensional damage characteristics of carbon fiber and performs normalization processing to generate a damage degree feature vector.

[0082] The damage characteristics of carbon fiber components include damage area ratio, damage depth and damage type, which are obtained through high-resolution image analysis and sensor data acquisition. The damage area ratio reflects the proportion of the damaged area to the surface of the component, the damage depth describes the extent of damage extension inside the material, and the damage type includes interlayer delamination, fiber breakage and other forms. Feature normalization uses maximum and minimum value normalization to map the damage area ratio and depth to the range of 0 to 1. The damage type is converted into a numerical vector through unique hot encoding to form a unified damage degree feature vector. Normalization eliminates dimensional differences and improves the accuracy of subsequent similarity calculations. For example, the damage area ratio of wind turbine blades is about 12% and the depth is 0.7 mm. The normalized feature vector is convenient for comparison with historical data.

[0083] S1062 calculates the similarity between the damage feature vector and historical cases, and selects matching cases to build a performance mapping basis.

[0084] The preset database stores over 1,200 historical case studies, including damage signatures and corresponding performance parameters. A cosine similarity algorithm is used to calculate the similarity between the current damage feature vector and historical cases, with a threshold of 0.88 set to screen for highly matching cases. The similarity calculation considers the directional consistency of the feature vectors to ensure that the screening results are highly correlated with the current component damage pattern. After screening, 50 to 80 matching cases are typically obtained, covering similar damage distributions and application scenarios, such as interlaminar delamination in aviation components or fiber fracture in pressure vessels, providing diverse data support for performance mapping.

[0085] S1063 extracts performance parameters based on matching cases, uses support vector regression to construct a performance reference mapping function and predict performance trends.

[0086] Key performance parameters of carbon fiber, such as tensile strength, elastic modulus, elongation at break and surface defect density, were extracted from matching historical cases. Using the support vector regression algorithm and the radial basis kernel function, the kernel function parameter γ was optimized in the range of 0.01 to 5 through 5-fold cross validation, and the penalty factor C was in the range of 1 to 500, with the prediction error controlled within 4.5%. The regression model learns the nonlinear mapping relationship between damage characteristics and performance parameters, and outputs the performance evaluation value of the current component. At the same time, a third-order polynomial fitting was performed on the performance parameter sequence of historical cases, with a goodness of fit of 0.96, to extract the time series change trend and predict the performance attenuation curve for the next 6 to 12 months. For example, the monthly attenuation rate of tensile strength was 1.8% at the initial stage and dropped to 0.4% after 3 months.

[0087] S1064 applies Monte Carlo sampling to analyze the probability distribution of performance changes, generate confidence intervals and build a performance reference system.

[0088] The output of the performance reference mapping function is combined with the time series trend to generate a performance change interval. A Monte Carlo method is used to perform 12,000 random samplings to simulate the uncertainty of damage characteristics and environmental factors, generating a probability distribution of performance changes. A 90% confidence interval is calculated. For example, the predicted range for tensile strength is 3600 to 3900 MPa, and the elastic modulus retention rate is 91% to 96%. Based on the prediction results, a performance reference system is constructed, divided into three levels: high-quality, qualified, and observation. The high-quality level requires a performance retention rate greater than 92% and a six-month decay rate less than 2.5%. The system records the evaluation value, prediction interval, and confidence level, and generates a detailed performance evaluation report.

[0089] For example, a wing component with a 10% damage area and a depth of 0.9 mm was matched against 58 historical case studies, resulting in a predicted tensile strength retention rate of 93% and a 12-month attenuation rate of 2.1%, with a confidence level of 0.94. In wind turbine blade recycling, 80% of cases achieved high-quality grades, with a surface defect density below 0.04 per square micron and an elastic modulus retention rate of 96%. In automotive parts recycling, the correlation coefficient between performance prediction and actual measurement was 0.93, and the attenuation of fiber elongation at break was kept within 12%, validating the accuracy and applicability of the system.

[0090] S107 generates a graded recycling plan based on the current performance evaluation and damage level of carbon fiber, and realizes the efficient reuse of carbon fiber through performance prediction and full life cycle management.

[0091] The graded recycling and utilization plan for carbon fiber optimizes resource allocation through comprehensive performance testing, damage analysis, and application scenario matching, while establishing a digital traceability system to ensure the performance stability and economic value of recycled fibers in different applications.

[0092] S1071 collects carbon fiber performance and damage data, uses a random forest algorithm to generate grading evaluation scores and determine recycling grades.

[0093] Mechanical property testing instruments measure the tensile strength, elastic modulus, and elongation at break of carbon fibers, combined with surface integrity data from electron microscopy to construct a comprehensive performance index system. Tensile strength reflects the fiber's load-bearing capacity, elastic modulus measures stiffness, elongation at break characterizes toughness, and surface integrity assesses the extent of microdefects. A random forest algorithm was employed, employing 600 decision trees trained on 6,000 sets of performance and damage signature data, achieving a classification accuracy of 93%. The algorithm integrates multiple decision trees to analyze the nonlinear relationship between performance indicators and damage signatures, generating a grading score. The score combines performance decay rate and fluctuation amplitude to categorize excellent, standard, and basic grades. The excellent grade requires a tensile strength retention rate exceeding 92%, a monthly decay rate below 0.25%, and a fluctuation amplitude of less than 2.5%. The standard grade requires a tensile strength retention rate of 80% to 92%, and the basic grade requires a retention rate of 70% to 80%. Grading results are timestamped to ensure traceability. For example, the tensile strength of recycled fibers from wind turbine blades is 3600 MPa, the elastic modulus is 235 GPa, the elongation at break is 1.9%, the surface integrity is 0.93, and the score is 88, which is classified as excellent.

[0094] S1072 uses the analytic hierarchy process to construct a performance parameter limit matrix, match the application scenario, and set up a performance warning mechanism.

[0095] Based on the grading results, a performance parameter limit matrix was constructed using the Analytic Hierarchy Process (AHP), comprehensively considering strength requirements, cost constraints, and usage environment. This matrix sets performance thresholds for different application scenarios: for example, automotive parts are weighted at 0.5 for strength, 0.3 for cost, and 0.2 for environment; and building reinforcements are weighted at 0.4 for strength, 0.4 for durability, and 0.2 for cost. Performance matching is calculated to generate a sequence of application scenarios, such as using premium-grade fiber for high-load structural parts and standard-grade fiber for non-load-bearing components. Performance warning thresholds were set based on 4,000 sets of historical data. Warning signals were triggered when the decay rate exceeded 0.4% three times in a row or when performance fell below 85% of the limit. The warning mechanism uses a recurrent neural network to predict performance trends. The network uses a long-short-term memory (LSTM) architecture, inputs 16 performance features, has three hidden layers, and is trained on two years of data, achieving a prediction accuracy of 91%. Predictions indicate that the strength of building reinforcement fibers drops to 84% after 14 months in a hot and humid environment, prompting adjustments to application scenarios.

[0096] S1073 analyzes surface morphology and damage type, and uses deep convolutional neural networks and hierarchical clustering to generate damage distribution maps.

[0097] An electron microscope scanned the carbon fiber surface morphology at 6000x magnification and a resolution of 4 nanometers per pixel, extracting crack length, damage area, and defect density. Crack lengths ranged from 50 to 180 microns, with a damage area accounting for 10% and a defect density of 0.07 per square micron. A deep convolutional neural network, configured with a six-layer convolutional architecture and trained on 9000 samples, achieved 95% accuracy in identifying damage types, including surface microcracks, localized debonding, and fiber breakage. A hierarchical clustering algorithm, using Euclidean distance as a metric, generated a three-layer spatial distribution map of damage types, showing concentrated microcracks in edge regions and significant debonding in load-bearing areas. This map intuitively reflects the spatial characteristics of damage, guiding hierarchical assessments.

[0098] S1074 optimizes recycling process parameters, establishes a full life cycle data archive and generates digital identification.

[0099] A genetic algorithm was used to optimize the pyrolysis process parameters, with a population size of 120 and 600 generations of evolution, with the goal of maximizing recycling efficiency. The optimization results showed that the pyrolysis temperature was controlled at 405 degrees Celsius, the heating rate was 2.2 degrees Celsius per minute, the pyrolysis time was 85 minutes, and the recovery rate reached 93%. After the process parameter sequence was generated, a 28-digit unique identification code was generated based on the recycling batch, grading results, and application scenario, including a 5-digit batch number, an 8-digit timestamp, a 5-digit grading code, a 6-digit performance index, and a 4-digit application number. A full life cycle data archive was established to record original component information, damage assessment, recycling process, performance testing, and application data to ensure traceability throughout the process. The archive supports dynamic updates. For example, in the recycling of ship composites, the application is adjusted to low load after the early warning signal is triggered, and the archive records the changes in real time.

[0100] In aviation composite recycling, premium-grade fibers have a 94% tensile strength retention rate and are used in wing secondary components. They maintain stable performance for six months, with a degradation rate of 0.2%. Standard-grade recycled pressure vessel fibers, with a matching degree of 0.87, are used in non-critical components. Early warning mechanisms predict a 12-month strength drop to 82%, enabling proactive optimization of recycling paths. In automotive parts recycling, tiered management has increased overall utilization by 18%. Premium-grade fibers are used for body reinforcement, good-grade fibers for interior trim, and qualified-grade fibers for filling materials, maximizing resource value.

[0101] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for realizing carbon fiber extraction control using deep learning, characterized in that: The method comprises: Obtain the scattered light intensity matrix and interference image of the carbon fiber component, combine the scattered light intensity matrix and interference image to perform feature extraction on the image data, and obtain a feature vector reflecting the damage state of the carbon fiber component; The extracted feature vectors are input into a pre-trained damage state assessment model to determine the damage degree of the carbon fiber component. If the damage degree exceeds a preset threshold, the component is marked as a component to be recycled. According to the damage degree of carbon fiber components, a multi-objective optimization model of pyrolysis temperature, carbon fiber decomposition rate, and carbon fiber purity is constructed to determine the target pyrolysis furnace temperature field distribution; The target temperature field distribution of the pyrolysis furnace is used as the control target to generate a temperature control method for the pyrolysis furnace. The temperature control method includes adjusting the power of the heating element, matching the temperature distribution in the pyrolysis furnace with the target temperature field distribution, and obtaining the spatial distribution of the carbon fiber dissociation rate. Continuously collect temperature data in the pyrolysis furnace during the pyrolysis extraction of carbon fiber, calculate the deviation between the temperature data in the pyrolysis furnace and the target temperature field distribution, and combine it with the spatial distribution of the carbon fiber dissociation rate to evaluate the potential impact of the deviation on the directional extraction efficiency. Dynamically adjust the power of the heating element based on the potential impact and obtain the carbon fiber properties after pyrolysis extraction; Match historical cases in the preset database according to the degree of damage, compare and analyze the carbon fiber performance results corresponding to the historical cases obtained after matching, establish a performance reference system based on historical data mapping, and output the current performance evaluation of the carbon fiber and the future performance change range obtained by reference comparison; Based on the current performance evaluation of carbon fiber or the future performance change range, combined with the degree of damage to the carbon fiber, a graded recycling and utilization plan for carbon fiber is generated, forming a closed-loop management of the entire life cycle of carbon fiber recycling.

2. The method according to claim 1, characterized in that The method of obtaining a scattered light intensity matrix and an interference image of the carbon fiber component, and extracting features from the image data in combination with the scattered light intensity matrix and the interference image to obtain a feature vector reflecting the damage state of the carbon fiber component includes: Acquire scattered light signals from the surface of the carbon fiber component, perform time domain sampling on the scattered light signals to construct an initial scattered light intensity matrix, and correct the initial scattered light intensity matrix using an optical fiber transmission loss compensation function to obtain a corrected scattered light intensity matrix; Performing frequency domain conversion on the corrected scattered light intensity matrix using Fourier transform, and obtaining an interference image based on phase difference calculation; The interference image is decomposed into multiple scales by wavelet transform, and the gray-level co-occurrence matrix features are extracted. Clustering calculation is performed on the gray-level co-occurrence matrix features. If the Euclidean distance between cluster centers is greater than a preset damage judgment threshold, it is determined that the grid node is damaged, and a damage feature vector is constructed.

3. The method according to claim 1, characterized in that The extracted feature vector is input into a pre-trained damage state assessment model to determine the damage degree of the carbon fiber component. If the damage degree exceeds a preset threshold, the component is marked as a component to be recycled, including: Normalizing the eigenvector, calculating a eigenvector normalization parameter using a maximum and minimum value normalization method, and transforming the eigenvector according to the normalization parameter to obtain a normalized eigenvector; A training sample set is constructed based on the normalized eigenvector and the damage type label, the training sample set is divided into a training subset and a validation subset using a stratified sampling method, the training subset data is trained using a support vector machine, a damage assessment model is obtained by optimizing kernel function parameters and penalty factors, the normalized eigenvector is input into the damage assessment model, a probability distribution of damage types is obtained based on the model output results, and the type corresponding to the maximum probability is selected as the component damage type; For the determined damage type, the strain energy density value of the damaged area is calculated, and the damage extension trend is judged by the preset strain energy density threshold; The comprehensive damage degree index is calculated based on the strain energy density value and the probability distribution of damage type, and the comprehensive damage degree index is graded using the preset damage degree threshold; If the comprehensive damage degree index exceeds the preset recycling threshold, a component recycling mark is generated, the damage assessment result and the assessment time are recorded, and a data record of the component to be recycled is constructed.

4. The method according to claim 1, wherein According to the damage degree of the carbon fiber component, a multi-objective optimization model of pyrolysis temperature, carbon fiber decomposition rate, and carbon fiber purity is constructed to determine the target pyrolysis furnace temperature field distribution, including: A three-dimensional feature space was constructed for pyrolysis temperature, carbon fiber decomposition rate, and fiber purity. A temperature field prediction function was established through multivariate Gaussian process regression, and the kernel function parameters were determined using the maximum likelihood estimation method. A multi-objective optimization function was constructed based on the temperature field prediction function. The lower limit of carbon fiber decomposition rate and fiber purity was set as constraints, and the genetic algorithm was used to solve the optimal temperature field distribution. A matrix decomposition reaction kinetic equation is established based on the optimal temperature field distribution, the matrix decomposition rate at each temperature point is calculated, and the matrix residue distribution is obtained by integrating the pyrolysis reaction time; A temperature field correction function is constructed using the matrix residue distribution data, a temperature control parameter sequence is generated after local correction of the optimal temperature field distribution, and the temperature control area is divided in combination with the pyrolysis furnace pressure and oxygen concentration constraints; The local temperature gradient is calculated within the divided temperature control area, the temperature field uniformity is judged by the temperature gradient threshold, and the target distribution of the pyrolysis furnace temperature field is determined.

5. The method according to claim 1, wherein The target pyrolysis furnace temperature field distribution is used as a control target to generate a temperature control method for the pyrolysis furnace. The temperature control method includes adjusting the power of the heating element, matching the temperature distribution in the pyrolysis furnace with the target temperature field distribution, and obtaining the spatial distribution of the carbon fiber dissociation rate, including: A temperature control curve is established based on the target pyrolysis furnace temperature field distribution. A neural network is used to construct a mapping relationship between the heating element input power and the temperature field output parameters, generate initial heating power control parameters, and calculate the filtered temperature value based on the state estimation equation and observation equation. Calculate the deviation between the actual temperature and the target temperature at each temperature measurement point based on the filtered temperature value, generate a heating power correction value according to the temperature deviation distribution, set an independent power controller for each group of heating elements, and use a proportional integral controller to adjust the heating power in real time; The corresponding relationship between the adjusted heating power and the temperature field was recorded, the thermal conductivity coefficient was calculated based on the temperature field time series data, the temperature field distribution was solved using the heat transfer equation, the pyrolysis reaction rate equation was established, the dissociation rate was calculated, and the spatial distribution of the carbon fiber dissociation rate was generated using the spatial interpolation function.

6. The method according to claim 5, characterized in that Also includes: Construct the measured temperature field in the pyrolysis furnace, discretize the measured temperature field to obtain a discrete temperature point set, obtain the target temperature field from the preset database, and convert it into a target discrete point set. If the Euclidean distance between the measured discrete point set and the target discrete point set exceeds the preset threshold, adjust the measured point set to obtain a matching temperature field, calculate the dissociation rate, generate rate distribution data and perform three-dimensional mapping to generate a spatial distribution, and perform directional extraction of carbon fibers according to the spatial distribution. Specifically, the following steps are involved: Based on the real-time temperature collected by the temperature sensor array arranged along the axial and radial directions of the inner wall of the pyrolysis furnace, a three-dimensional grid coordinate system is established according to the axial, radial and circumferential directions. The temperature values of the grid nodes are calculated using cubic spline interpolation to generate a discrete point set of the measured temperature field. Extract target temperature field data from a preset database, transform the target temperature field into a measured point set grid coordinate system according to spatial coordinate mapping rules, and generate a target temperature field discrete point set; Calculating the temperature deviation of each corresponding point between the measured discrete point set and the target discrete point set, determining the temperature field adjustment area according to a preset temperature matching threshold, calculating the grid node temperature correction amount using a gradient descent method, and obtaining a matching temperature field through iterative updating of the temperature correction; The carbon fiber pyrolysis reaction rate is calculated according to the matching temperature field to generate continuous dissociation rate distribution data. The dissociation rate distribution data is spatially sliced using the isosurface method, and a carbon fiber directional extraction trajectory is constructed according to the dissociation rate gradient direction.

7. The method according to claim 1, characterized in that The method continuously collects temperature data in the pyrolysis furnace during the pyrolysis extraction of carbon fibers, calculates the deviation between the temperature data in the pyrolysis furnace and the target temperature field distribution, and evaluates the potential impact of the deviation on the directional extraction efficiency in combination with the spatial distribution of the carbon fiber dissociation rate. The power of the heating element is dynamically adjusted according to the potential impact, and the carbon fiber properties after pyrolysis extraction are obtained, including: Continuously acquiring temperature data in the pyrolysis furnace from the temperature sensor array according to a preset sampling interval and calculating the measured temperature field in the pyrolysis furnace; Support vector regression was used to calculate the spatial deviation distribution between the measured temperature field and the target temperature field. The dissociation rate distribution was calculated based on the carbon fiber pyrolysis activation energy. Based on the temperature deviation distribution and the dissociation rate distribution, the local areas of insufficient dissociation, uneven dissociation, and excessive fiber damage were calculated. A weighted scoring method was used to quantify the impact of the temperature field deviation. Based on the degree of influence of temperature field deviation, differentiated power adjustment coefficients are adopted for different impact types, and the output power of the heating element is calculated through the power response characteristic curve; The heating response hysteresis was eliminated and the power adjustment parameters were optimized according to the temperature dynamic response curve. The graphitization degree of the carbon fiber was measured, and the carbon fiber structural integrity index was calculated by the characteristic peak intensity ratio. At the same time, the surface defect characteristics were extracted to calculate the degree of fiber surface damage.

8. The method according to claim 1, characterized in that The method matches historical cases in a preset database according to the degree of damage, compares and analyzes the carbon fiber performance results corresponding to the historical cases obtained after matching, establishes a performance reference system based on historical data mapping, and outputs the current performance evaluation of the carbon fiber and the future performance change range obtained by the reference comparison, including: Obtain test data on carbon fiber tensile strength, elastic modulus, elongation at break, and surface integrity, and use the random forest algorithm to perform combined analysis on the test data to obtain a graded evaluation score; The carbon fibers are graded according to the grading evaluation scores, and a grading stability determination result is obtained by calculating the performance attenuation rate and the fluctuation amplitude, so as to determine whether the carbon fibers belong to the superior grade, the standard grade, or the basic grade; Based on the carbon fiber grading results, the hierarchical analysis method is used to establish a performance parameter limit matrix. The performance matching degree is calculated through the performance parameter limit matrix to obtain the application scenario sequence and construct the recycled carbon fiber application allocation plan; Generate a unique carbon fiber identification code based on recycling batches, grading results, and application scenarios, and record recycling process parameters, performance test data, and application scenario limits; Establish a data archive for the entire life cycle of carbon fiber, recording original component information, damage assessment data, recycling process parameters, performance test results and application scenario data.

9. The method according to claim 1, characterized in that Based on the current performance evaluation or future performance change range of carbon fiber and the degree of damage to the carbon fiber, a graded recycling plan for carbon fiber is generated, forming a closed-loop management of the entire life cycle of carbon fiber recycling, including: A performance index system was established based on carbon fiber tensile strength, elastic modulus, elongation at break, and surface integrity. A random forest algorithm was used to combine performance indicators and damage characteristics to generate a graded evaluation score. Based on the grading evaluation scores, the grading standards for superior, standard and basic carbon fibers are set. The grading stability is judged by the performance attenuation rate and fluctuation amplitude, and the grading timestamp is recorded. The analytic hierarchy process was used to establish a performance parameter limit matrix for different application scenarios. The application scenario sequence was generated through performance matching calculation, and a recycled carbon fiber application allocation plan was constructed. Generate a unique carbon fiber identification code based on recycling batches, grading results, and application scenarios, and record recycling process parameters, performance test data, and application scenario limits; Establish a data archive for the entire carbon fiber life cycle, recording original component information, damage assessment data, recycling process parameters, performance test results, and application scenario data; Set the performance warning threshold according to the performance prediction curve, judge the carbon fiber performance attenuation trend through performance test data, and generate a performance warning signal; A recurrent neural network is used to learn the evolution of carbon fiber performance, predict performance change trends in different application scenarios, and update digital traceability archives.

10. The method according to claim 9, characterized in that Also includes: The performance status data of carbon fiber samples is obtained and analyzed. At the same time, damage data is extracted from the samples to determine the damage degree assessment result. If the damage degree assessment result is lower than the preset threshold, the damage type is classified to obtain the damage type distribution. Based on the damage type distribution and the performance status analysis results, a grading standard is formulated to determine a graded recycling plan, from which recycling methods are obtained, utilization efficiency is optimized, recycling paths are obtained, and a graded recycling plan is formed, which specifically includes: Obtain carbon fiber tensile strength data and elastic modulus data, and obtain carbon fiber surface morphology images through electron microscope scanning; Extracting crack length parameters and damage area parameters according to the surface topography image, and classifying the crack length parameters and damage area parameters using a deep convolutional neural network to obtain a damage type recognition result; Based on the damage type identification results, a hierarchical clustering algorithm is used to hierarchically divide the damage types to obtain a spatial distribution map of the damage types; A mapping relationship between carbon fiber performance indicators and the spatial distribution map of the damage type is constructed through a multi-layer perceptron. The carbon fiber is graded and evaluated based on the mapping relationship. Performance parameter limits are set for different recycling levels, and a graded recycling and utilization plan is constructed.

Images