A Machine Learning-Based Method and Related Equipment for Predicting Crack Propagation Behavior of 3D Braided Composite Materials Under Multiple Impact Loads

By employing machine learning-based methods, adaptive threshold segmentation, and multi-feature fusion generative adversarial networks, the problems of prediction complexity and low efficiency in traditional methods are solved. This enables rapid and accurate prediction of crack propagation behavior in three-dimensional braided composite materials, meeting the needs of damage assessment and life prediction in aerospace and other fields.

CN123069552APending Publication Date: 2026-11-10XI'AN POLYTECHNIC UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611352984.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing technologies struggle to efficiently and accurately predict crack propagation behavior in three-dimensional braided composites under multiple impact loads. Traditional physical experiments are costly, numerical simulation modeling is complex and computationally inefficient, and machine learning predictions have a single dimension and cannot characterize crack spatial propagation features.

Method used

A machine learning-based approach was adopted to obtain a dataset of original crack images of three-dimensional braided composite materials under different impact conditions. The binarized crack images were extracted using an adaptive threshold segmentation algorithm. An AdaBoost regression prediction model and a multi-feature fusion conditional generative adversarial network were established to jointly predict the crack area and distribution morphology.

Benefits of technology

It enables rapid and high-precision prediction of crack propagation patterns in composite materials under multiple impact loads. It can simultaneously obtain the two-dimensional geometry and quantitative damage status of composite materials without preparing new samples or performing complex tests, meeting the needs of damage safety assessment and life prediction in aerospace and other fields.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123069552A_ABST
    Figure CN123069552A_ABST
Patent Text Reader

Abstract

This invention relates to the field of composite material damage assessment technology, and discloses a method and related equipment for predicting the crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning. The method includes: acquiring original crack images and operating parameters of the sample under different impact conditions; extracting binarized images using an adaptive threshold segmentation algorithm and statistically obtaining the true crack area labels; inputting the operating parameters and area labels into an AdaBoost regression prediction model for training, obtaining a crack area prediction model; constructing a multi-feature fusion conditional generative adversarial network, mapping parameters to conditional features, and alternatingly training based on a multi-loss joint function to obtain a crack distribution morphology prediction model; and jointly outputting the predicted crack area and distribution morphology images of the tested operating conditions using both models. This invention overcomes the constraints of mechanical modeling, achieving quantitative assessment of the cumulative damage area of ​​composite materials under multiple impacts and end-to-end visualization of high-precision crack morphology prediction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of composite material structure damage assessment and life prediction technology, and more specifically, to a method and related equipment for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning. Background Technology

[0002] Three-dimensional braided composite materials, with their interwoven three-dimensional braided structure as reinforcement, possess high specific strength, specific modulus, and excellent integrity, delamination resistance, and impact damage tolerance, making them important advanced materials for lightweight aircraft structural design. However, aircraft structural components are susceptible to low-velocity impact loads such as tool drops, gravel impacts, and maintenance collisions during service. Under repeated impact loads, the material will experience internal damage such as matrix cracking, interface debonding, and microcrack initiation and propagation, gradually forming complex cumulative damage that may ultimately lead to structural stiffness degradation or even sudden failure, seriously threatening equipment operational safety. Therefore, rapidly predicting the crack propagation morphology of three-dimensional braided composite materials after repeated impacts is of great significance for aircraft structural safety assessment and damage early warning.

[0003] Currently, crack propagation prediction in composite materials mainly employs numerical simulation methods such as physical experiments, finite element analysis, representative volume element methods, and damage mechanics models. Traditional multi-impact damage research heavily relies on repetitive physical impact tests, which suffers from problems such as high sample consumption, long test cycles, and severe equipment wear and tear, and it is difficult to capture the entire dynamic evolution process of crack accumulation and propagation. Furthermore, traditional numerical simulation methods such as the finite element method, due to the complex spatial interwoven structure and significant anisotropy of three-dimensional braided composite materials, require complex artificial construction of damage constitutive and evolution equations. This not only presents significant modeling difficulties and cumbersome parameter calibration, but also often necessitates iterative calculations when dealing with the cumulative coupling effects of multiple impact loads, resulting in significant time consumption and low efficiency. These methods struggle to accurately characterize the propagation law of nonlinear cumulative cracks and cannot meet the needs of rapid assessment and batch prediction in engineering applications.

[0004] In recent years, with the development of machine learning technology, data-driven methods have provided new approaches for damage prediction in composite materials. However, most existing machine learning methods are numerical predictions of damage indices from parameter data, resulting in overly simplistic evaluation dimensions. These methods can only map impact parameters to single scalar indices such as damage area, crack length, and residual strength, quantifying the degree of damage numerically but failing to intuitively reflect the spatial distribution, propagation direction, and penetration trend of cracks. Furthermore, existing research largely focuses on single impacts or single damage modes, insufficiently considering the cumulative damage effect under repeated impacts.

[0005] In summary, existing technologies suffer from numerous drawbacks, such as high physical testing costs and limited operating conditions, complex traditional numerical simulation modeling and low computational efficiency, and the single-dimensional prediction of existing machine learning which cannot characterize the spatial propagation characteristics of cracks. These limitations make it difficult to efficiently and accurately predict the cumulative crack propagation behavior of three-dimensional braided composite materials under multiple impact loads, severely restricting the refined assessment of the safety status and life prediction design of composite material structures. Summary of the Invention

[0006] To address the problems existing in the prior art, this invention provides a method and related equipment for predicting the crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning. It aims to break away from the constraints of complex mechanical modeling in traditional experimental testing, and achieve rapid, high-precision, and highly generalizable intelligent prediction of the crack propagation law of composite materials under multiple impact cumulative damage conditions, so as to meet the engineering needs of safety assessment and life prediction of composite material components in aerospace, protective equipment and other fields.

[0007] This invention is achieved through the following technical solution: A machine learning-based method for predicting crack propagation behavior of 3D braided composite materials under multiple impact loads includes: Obtain the original crack image dataset of three-dimensional braided composite materials under different impact conditions and the corresponding impact condition parameters; An adaptive threshold segmentation algorithm is used to perform image separation processing on the original crack image dataset, extracting the binarized crack images as the real label images, and obtaining the corresponding real crack area labels based on the pixel statistics of the binarized crack images. Using impact condition parameters as input variables and actual crack area labels as output variables, an AdaBoost regression prediction model is trained to establish the mapping relationship between parameters and area, thus obtaining a crack area prediction model. A multi-feature fusion conditional generative adversarial network is constructed, which includes a conditional coding module, a generator, and a discriminator. The impact condition parameters are mapped to conditional features through the conditional encoding module, and the conditional features are input into the generator to output the predicted crack image. The discriminator is used to distinguish between the predicted crack image and the corresponding real label image. The network parameters of the generator and the discriminator are alternately updated based on the multi-loss joint function until the model converges, and the crack distribution morphology prediction model is obtained. The impact parameters of the three-dimensional braided composite material to be tested are obtained and input into the crack area prediction model and crack distribution morphology prediction model, and the predicted crack area value and crack distribution morphology image are output.

[0008] Preferably, the step of using an adaptive threshold segmentation algorithm to perform image separation processing on the original crack image dataset and extracting the binarized crack image as the real label image includes: Convert the images in the original crack image dataset into single-channel grayscale images; Using Otsu's maximum inter-class variance method, the inter-class variance of the single-channel grayscale image is calculated, and the grayscale value that maximizes the inter-class variance is taken as the optimal segmentation threshold. Pixels with gray values ​​lower than the optimal segmentation threshold in the single-channel grayscale image are assigned a value of 1 as crack damage pixels, and pixels with gray values ​​higher than the optimal segmentation threshold are assigned a value of 0 as matrix background pixels, generating a black-and-white binary classification of the crack image as the real label image. The total number of crack damage pixels in the binarized crack image is counted, and the total number of crack damage pixels is determined as the true crack area label under the corresponding impact condition parameters.

[0009] Preferably, the specific process for establishing the crack area prediction model includes: The impact condition parameters are used as input variables, and the actual crack area labels are used as output variables in the AdaBoost regression prediction model. The training samples in the AdaBoost regression prediction model are assigned initial sample weights; the training samples are iteratively learned using the AdaBoost algorithm, and the sample weights are updated according to the prediction error distribution of the previous weak learner in each iteration. The outputs of multiple weak learners after multiple iterations are weighted and summed to output the crack area prediction model that has been trained and converged. The output of the AdaBoost regression prediction model Represented as: ;in, This represents the area of ​​crack damage.

[0010] Preferably, the generator comprises a U-Net architecture consisting of a downsampling encoding module and an upsampling decoding module; The generator is embedded with a conditional batch normalization module and an attention mechanism module. The conditional batch normalization module is used to continuously regulate the intermediate features of the network by the impact condition parameters, and the attention mechanism is used to enhance the network's ability to express the features of the crack backbone region and the local propagation region. The discriminator adopts a conditional PatchGAN architecture. The input of the discriminator receives a first sample combination and a second sample combination. The first sample combination is composed of the real label image and the conditional features, and the second sample combination is composed of the predicted crack image and the conditional features. The discriminator divides the input first sample combination and the second sample combination into local image blocks for output discrimination.

[0011] Preferably, the impact parameters include weaving angle, impact energy, and number of impacts; Before using the impact condition parameters as input variables, the impact condition parameters are normalized to obtain a normalized input vector: ; In the formula, The normalized weaving angle, The normalized impact energy. This represents the normalized number of impacts.

[0012] Preferably, the acquisition of the original crack image dataset of the three-dimensional braided composite material under different impact conditions includes: acquiring three-dimensional braided composite material samples with different braiding angles prepared by a four-step three-dimensional braiding process and a vacuum-assisted resin transfer molding process; applying a low-speed impact load along the thickness direction of the three-dimensional braided composite material sample, and performing 1 to 5 repeated impacts under the same impact energy; acquiring crack images on the surface of the three-dimensional braided composite material sample after each impact, and generating the original crack image dataset.

[0013] Preferably, the calculation formula for the joint multi-loss function is as follows:

[0014] In the formula, For the total loss, To reduce the overall pixel deviation between the predicted image and the real image, Edge loss is used to enhance the consistency of crack profile and boundary region. To improve the perceptual loss for enhancing the similarity between predicted and real images in high-level structural features, To constrain the generated image to approximate the distribution of the real crack image, , , and These are the weighting coefficients for each loss term.

[0015] A machine learning-based system for predicting crack propagation behavior of 3D braided composite materials under multiple impact loads includes: The data acquisition module is used to acquire the original crack image dataset of three-dimensional braided composite materials under different impact conditions and the corresponding impact condition parameters. The image separation and label extraction module is configured to use an adaptive threshold segmentation algorithm to perform image separation processing on the original crack image dataset, extract the binarized crack image as the real label image, and obtain the corresponding real crack area label based on the pixel statistics of the binarized crack image. The area model training module is used to train an AdaBoost regression prediction model by taking the impact condition parameters as input variables and the actual crack area label as output variables, in order to establish the mapping relationship between parameters and area and obtain the crack area prediction model. The morphology model training module is used to construct a multi-feature fusion conditional generative adversarial network, which includes a conditional encoding module, a generator, and a discriminator. The impact condition parameters are mapped to conditional features through the conditional encoding module, and the conditional features are input into the generator to output a predicted crack image. The discriminator is used to distinguish between the predicted crack image and the corresponding real label image. The network parameters of the generator and the discriminator are alternately updated based on a multi-loss joint function until the model converges, thus obtaining a crack distribution morphology prediction model. The joint prediction execution module is used to acquire the impact condition parameters of the three-dimensional braided composite material to be tested, input them into the crack area prediction model and the crack distribution morphology prediction model, and output the predicted crack area value and crack distribution morphology image.

[0016] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the machine learning-based method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads.

[0017] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the machine learning-based method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads.

[0018] Compared with the prior art, the present invention has the following beneficial technical effects: This invention provides a machine learning-based method and system for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads. By acquiring original crack image datasets of three-dimensional braided composite materials under different impact conditions and corresponding impact parameters, it overcomes the shortcomings of traditional physical experiments in continuously capturing the dynamic process of cumulative damage evolution, providing a reliable physical benchmark for multi-condition mapping. By employing an adaptive threshold segmentation algorithm to separate the original crack image dataset and extracting the true crack area label based on pixel statistics, it avoids missed detections and distortions caused by manually fixed thresholds and automatically eliminates interference from illumination and substrate texture, generating high-quality standardized training and supervision data. By inputting the impact parameters and the true crack area label into an AdaBoost regression prediction model to establish a mapping relationship, it overcomes the shortcomings of traditional finite element simulation modeling, which is complex and time-consuming, accurately characterizing multiple impacts. The strong cumulative coupling effect of the load enables high-precision and rapid quantitative prediction of the nonlinear crack propagation scale under cumulative load. By constructing a multi-feature fusion conditional generative adversarial network containing a conditional encoding module, generator, and discriminator, and inputting the conditional features into the network for true / false discrimination and alternating updating and optimization of the multi-loss joint function, it can overcome the limitation of existing data-driven methods that can only output a single scalar value in the evaluation dimension, and directly reconstruct the spatial distribution, irregular propagation boundary, and connectivity trend of cracks end-to-end. By inputting the parameters of the test condition into the trained area and morphology dual model and jointly outputting the results, it achieves a technological leap in obtaining the two-dimensional geometric morphology and quantitative damage status of composite materials simultaneously without preparing new samples or performing complex impact tests. This effectively meets the needs of visualized damage safety fine assessment and rapid early warning of lightweight composite material components in aerospace, protective equipment, and other multi-impact service scenarios.

[0019] Furthermore, by employing the Otsu method to calculate the optimal segmentation threshold of a single-channel grayscale image for binary classification, the objective grayscale difference between the crack and the substrate can be highlighted to the maximum extent, ensuring the accuracy of area statistics. By assigning sample weights to the AdaBoost model and dynamically updating the weighted sum based on the error distribution of the previous weak classifier, the nonlinear fitting and correction capabilities for large-damage, high-error samples can be significantly enhanced.

[0020] Furthermore, by embedding conditional batch normalization and attention mechanisms within the U-Net generator and employing PatchGAN for local block discrimination, the network can be continuously and deeply tuned using physical parameters and focused on capturing fine crack branches, greatly improving the robustness of the model output in matching the response characteristics of different processes and loads.

[0021] Furthermore, by normalizing the weaving angle, impact energy, and number of impacts, the differences in underlying physical dimensions can be eliminated, ensuring the stability of the forward mapping. By constructing a joint function to constrain the process by fusing four loss terms—pixel reconstruction, edge detection, perception, and adversarial analysis—the consistency between the clarity of microscopic boundaries and the global topology can be taken into account, making the prediction results infinitely close to the real fracture morphology. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 The overall framework and prediction flowchart of the damage prediction model provided in the embodiments of the present invention; Figure 2 This is a comparison chart of crack area prediction results provided by different machine learning models in embodiments of the present invention; Figure 3 Radar chart and bar chart comparing the quantitative evaluation indexes of different models in crack distribution morphology prediction provided in the embodiments of the present invention; Figure 4 The images provided in this invention show a comparison of actual test crack morphology, model prediction results, and absolute error under different impact conditions. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0025] See Figure 1In one embodiment of the present invention, a method for predicting the crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning is provided. This method is used to achieve joint prediction of crack area values ​​and crack distribution morphology images from impact condition parameters, thereby solving the technical problems of high physical test costs, difficulty in numerical simulation modeling, and single prediction dimension of machine learning in the prior art.

[0026] S1: Obtain the original crack image dataset of three-dimensional braided composite materials under different impact conditions and the corresponding impact condition parameters.

[0027] Three-dimensional braided composite material samples with different braiding angles were obtained using a four-step three-dimensional braiding process and a vacuum-assisted resin transfer molding process. Specifically, domestically produced T700S-12K carbon fiber tow was selected as the reinforcing material, and a three-dimensional braided preform was prepared using a four-step three-dimensional braiding process. By adjusting the braiding process parameters, three-dimensional braided structures with braiding angles of 15°, 23°, and 30° were obtained, respectively. After the preform was formed, the composite material was prepared using a vacuum-assisted resin transfer molding process. Bisphenol A type epoxy resin (JC-05A) and modified acid anhydride curing agent (JC-05B) were mixed at a mass ratio of 100:85, and after vacuum degassing, the mixture was used as the infusion resin system. Under vacuum negative pressure, the resin was introduced into the mold and impregnated the three-dimensional braided preform. The mold was then placed in an oven for step curing treatment. The curing regimes were 30℃ / 10min, 90℃ / 120min, 110℃ / 60min and 135℃ / 360min respectively. After curing, the sample was cooled to room temperature in the oven. After demolding and cutting, a three-dimensional braided composite material sample was obtained.

[0028] A low-velocity impact load was applied along the thickness direction of the three-dimensional braided composite material specimen, and 1 to 5 repeated impacts were performed at the same impact energy. Specifically, the prepared three-dimensional braided composite material specimens were subjected to low-velocity repeated impact experiments. Different braiding angles, impact energies, and number of impacts were selected as influencing factors to explore the damage evolution characteristics of three-dimensional braided composite materials with different braiding structures under different impact conditions. During the experiment, the impact load was applied along the thickness direction of the specimen, and the same group of specimens underwent 1 to 5 repeated impacts at the same impact energy, ranging from 4J to 10J.

[0029] Crack images of the surface of the three-dimensional braided composite material sample are acquired after each impact, generating the original crack image dataset. Specifically, the crack damage morphology of the sample is acquired and recorded after each impact to form crack image datasets with different braiding angles, impact energies, and impact numbers. The impact condition parameters include braiding angle, impact energy, and impact number.

[0030] Interpretive analysis reveals that by preparing three-dimensional braided composite material specimens with different braiding angles and conducting repeated low-velocity impact tests, crack damage data covering various structural parameters and impact loading conditions can be obtained, providing a data foundation for training subsequent machine learning models. The braiding angle characterizes the differences in three-dimensional braided structures, the impact energy describes the external load intensity, and the number of impacts reflects the damage accumulation effect during repeated impacts; together, these three elements constitute a complete parameter space describing the impact conditions.

[0031] S2: An adaptive threshold segmentation algorithm is used to perform image separation processing on the original crack image dataset, extract the binarized crack image as the real label image, and obtain the corresponding real crack area label based on the pixel statistics of the binarized crack image.

[0032] In one possible implementation, the step of using an adaptive threshold segmentation algorithm to perform image separation processing on the original crack image dataset and extracting the binarized crack image as the real label image includes: The images in the original crack image dataset are converted into single-channel grayscale images. The original acquired images are color photographs of the sample surface. First, the images are converted into single-channel grayscale images to remove irrelevant features introduced by the color dimension, retaining only the grayscale information that can distinguish the cracks from the matrix.

[0033] The Otsu's maximum inter-class variance method (an adaptive image binarization algorithm that determines the optimal segmentation threshold by traversing gray levels and maximizing the inter-class variance between foreground and background pixel gray values) is used to calculate the inter-class variance of the single-channel grayscale image. The gray value that maximizes the inter-class variance is taken as the optimal segmentation threshold. Specifically, the Otsu's maximum inter-class variance method does not require manually setting a segmentation grayscale threshold; it relies on the image's own grayscale distribution characteristics to solve for the optimal segmentation threshold. The algorithm first counts the number of pixels corresponding to each grayscale level from 0 to 255 in the grayscale image, constructs an image grayscale distribution histogram, traverses all potential segmentation thresholds t, and divides the image into two categories based on pixel grayscale: a cracked foreground region and a intact substrate background region. Simultaneously, the pixel proportion, average grayscale, and inter-class variance between the two categories are calculated. The gray value that maximizes the global inter-class variance is determined as the optimal segmentation threshold t. At this point, the difference in grayscale between the crack and the substrate pixels reaches its maximum, resulting in the optimal segmentation and differentiation effect.

[0034] Pixels in the single-channel grayscale image with grayscale values ​​below the optimal segmentation threshold are assigned a value of 1 as crack damage pixels, and pixels with grayscale values ​​above the optimal segmentation threshold are assigned a value of 0 as matrix background pixels, generating a binary crack image for black-and-white classification as the true label image. The optimal threshold t is then used. The grayscale image is subjected to pixel-by-pixel binary classification. Crack damage pixels with grayscale values ​​below the threshold are assigned a value of 1, while matrix background pixels with grayscale values ​​above the threshold are assigned a value of 0. The output is a black and white binary crack image with clear boundaries and no extra noise. This image is used as the real label image for subsequent conditional generative adversarial network training.

[0035] The total number of crack damage pixels in the binarized crack image is counted, and this total number of crack damage pixels is determined as the true crack area label under the corresponding impact condition parameters. If the pixel value of the crack area in the binary image is 1, and the pixel value of the background area is 0, then the crack area can be characterized by the number of crack pixels.

[0036] Interpretively, raw images are susceptible to interference from background texture, uneven lighting, and local grayscale fluctuations, which significantly increases the difficulty of feature learning when directly used for model training. By combining grayscale conversion with Otsu's adaptive threshold segmentation method, accurate extraction of crack regions can be achieved, resulting in a black-and-white binary image. The Otsu algorithm can adapt to sample images with significant differences in grayscale distribution under different weaving angles, impact energies, and impact cycles, avoiding problems such as missed detection of small cracks and segmentation distortion caused by manually fixed threshold segmentation. Based on the binary image obtained from Otsu segmentation, crack area quantification can be performed. The total number of crack pixels with a value of 1 in the image represents the crack damage area, providing a quantitative label for subsequent crack area regression models.

[0037] In one possible implementation, the binary crack image is normalized and pseudo-color mapped to generate a visualized crack distribution map, which facilitates a direct comparison between the actual damage morphology and the model prediction image.

[0038] S3: Using the impact condition parameters as input variables and the actual crack area label as output variables, train an AdaBoost regression prediction model to establish the mapping relationship between parameters and area, and obtain a crack area prediction model.

[0039] In one possible implementation, the specific process of establishing the crack area prediction model includes: The impact condition parameters are used as input variables, and the actual crack area label is used as output variable to input the AdaBoost regression prediction model (AdaBoost algorithm, also known as adaptive boosting algorithm, is an ensemble learning method that builds a strong regression model by iteratively adjusting sample weights and combining multiple weak learners).

[0040] The training samples in the AdaBoost regression prediction model are assigned initial sample weights, and the training samples are iteratively learned using the AdaBoost algorithm. In each iteration, the sample weights are updated according to the prediction error distribution of the previous weak learner.

[0041] The outputs of multiple weak learners after multiple iterations are weighted and summed to output the crack area prediction model that has been trained and converged.

[0042] The output of the AdaBoost regression prediction model Represented as: ; in, This represents the area of ​​crack damage. The crack pixel regions in the binary crack image are statistically analyzed and used to characterize the macroscopic damage degree of the sample after repeated impacts.

[0043] Specifically, the model uses weaving angle, impact energy, and number of impacts as input variables, and crack area as the output variable. The input vector is defined as follows:

[0044] Specifically, the weaving angle θ The spatial interlacing structural parameters of the three-dimensional braided composite material range from 15° to 30°; the impact energy E The kinetic energy applied to the specimen by the drop hammer in the low-velocity impact test ranges from 4J to 10J; the number of impacts... N This refers to the number of repeated impacts that the same sample can withstand under the same impact energy, ranging from 1 to 5 times.

[0045] Because different input parameters have different dimensions and value ranges, directly inputting them into the model may affect training stability. Therefore, this invention normalizes the input features before modeling. The normalized input vector is represented as: ; In the formula, The normalized weaving angle, The normalized impact energy. This represents the normalized number of impacts.

[0046] Interpretive normalization processes can bring input parameters with different dimensions within a similar order of magnitude, which is beneficial for improving the stability and convergence speed of model training. The weaving angle is used to characterize the differences in three-dimensional weaving structures, the impact energy is used to describe the intensity of the external load, and the number of impacts is used to reflect the cumulative damage effect during repeated impacts. Together, these three elements constitute a parameter vector describing the complete impact condition.

[0047] For interpretability, since the units and ranges of different input parameters vary, directly inputting them into the model may affect training stability. Therefore, the input features are normalized before modeling. The processed dataset is divided into training, validation, and test sets (in an 8:1:1 ratio) to ensure the reliability of the model training, validation, and testing processes.

[0048] In one possible implementation, to compare the predictive capabilities of different machine learning methods for crack area, six typical regression models (gradient boosting model, support vector regression model, K-nearest neighbor model, extreme random tree model, random forest model, and AdaBoost model) are selected for comparative analysis. All models use the same input variables, data partitioning methods, and evaluation metrics to ensure fairness in the model comparison. The crack area prediction process can be uniformly represented as:

[0049] in, f ( ) represents a regression prediction model. To predict the crack area.

[0050] To quantitatively evaluate the crack area prediction performance, the AdaBoost model was used for crack area prediction, and the coefficient of determination R was employed. 2 The mean squared error (MSE) and mean absolute error (MAE) are used as evaluation indicators, and their calculation formulas are as follows:

[0051]

[0052]

[0053] in, For the first i The actual crack area of ​​each sample For the first i Predicted crack area for each sample This represents the average area of ​​all actual cracks. n R represents the sample size. 2 The closer the value is to 1, the stronger the model's ability to explain the crack area variation; the smaller the MSE and MAE, the lower the model's prediction error.

[0054] Explained, the above modeling process enables rapid quantitative prediction of crack area based on impact parameters. After model training, only controllable process parameters such as weaving angle, impact energy, and number of impacts need to be input. Without preparing new samples or conducting a new round of physical impact tests, the crack area prediction results can be quickly output, providing a foundation for subsequent crack damage degree analysis and crack distribution morphology prediction.

[0055] S4: Construct a multi-feature fusion conditional generative adversarial network, which includes a conditional coding module, a generator, and a discriminator.

[0056] The generator comprises a U-Net architecture consisting of a downsampling encoding module and an upsampling decoding module. Specifically, the generator uses an encoder-decoder structure, where the encoding part is responsible for acquiring deep features of the working condition parameters, and the decoding part progressively enlarges the image scale and outputs a predicted image with the same size as the real crack label. To fully leverage the influence of working condition factors on image generation, the generator embeds a conditional batch normalization module and an attention mechanism module.

[0057] The Conditional Batch Normalization (CBN) module is a feature regulation module that integrates external conditional information into the normalization layer and adjusts the intermediate layer feature response through affine transformation parameters. This module continuously regulates the intermediate features of the network using impact condition parameters. The Attention Mechanism (A feature selection module that enhances the response of key regions and suppresses background interference by adaptively calculating feature weights) strengthens the network's ability to represent features in the main crack region and its local extension regions. Based on the morphological characteristics of cracks—long, discrete, and with complex boundaries—the Attention Mechanism enhances the network's representation of features in the main crack region and its local extension regions. The discriminator adopts a Conditional PatchGAN architecture (a conditional block adversarial discriminant network that divides the input image into multiple local image blocks and independently performs true / false discrimination on each block to constrain the consistency of the generated image's local morphology). The discriminator's input receives a first sample combination and a second sample combination. The first sample combination is constructed by concatenating the real label image and the conditional features, and the second sample combination is constructed by concatenating the predicted crack image and the conditional features. The discriminator divides the input first sample combination and second sample combination into local image blocks for output discrimination.

[0058] Specifically, the discriminator is used to determine whether the input image is a real image under given working conditions. The network input includes two types of sample combinations: real sample combinations ( Y , c ) and generated sample combination This structure uses local image blocks as discrimination units to constrain the local morphology of the generated image, thereby improving the matching between crack details and operating conditions.

[0059] S5: The impact condition parameters are mapped into conditional features through the conditional encoding module, and the conditional features are input into the generator to output the predicted crack image; the discriminator is used to distinguish between the predicted crack image and the corresponding real label image, and the network parameters of the generator and discriminator are alternately updated based on the multi-loss joint function until the model converges, thus obtaining the crack distribution morphology prediction model.

[0060] In one possible implementation, the input condition vector is defined as: ; in, θ For the weaving angle, E To impact energy, N Let be the number of impacts. The predicted image output by the model can be represented as:

[0061] Among them, G( ) represents an image generation network. Crack images generated for the model, Y The images are real crack images obtained for the experiment.

[0062] During the training phase, a strategy of alternating optimization between the generator and discriminator is employed. To balance pixel accuracy, edge contours, high-level structural features, and image realism, a total loss function jointly optimized by multiple loss mechanisms is constructed:

[0063] In the formula, For the total loss, This is the pixel reconstruction loss (i.e., L1 loss, used to constrain the consistency of the generated image with the real image in low-level visual features such as brightness and color on a pixel-by-pixel basis) to reduce the overall pixel deviation between the predicted image and the real image. The edge loss is used to enhance the consistency of crack profile and boundary region (i.e., edge loss, used to constrain the prediction accuracy of crack boundary region and improve profile blur and edge offset problems). This is a perceptual loss used to improve the similarity between predicted and real images in high-level structural features (i.e., perceptual loss, which extracts high-level semantic features and calculates feature map distance through a pre-trained neural network to maintain the consistency between predicted and real images in structural connectivity and overall morphology). The adversarial loss is used to constrain the generated image to approximate the distribution of the real crack image (i.e., the adversarial loss, which uses the discriminator's true / false judgment result to inversely constrain the generator, so that the data distribution of the generated image approximates the data distribution of the real image). , , and These are the weighting coefficients for each loss term, which are determined by optimizing the IoU metrics of different combinations on the validation set.

[0064] Interpretive, unlike the aforementioned crack area prediction task, crack distribution morphology prediction is an image-level prediction task. Besides the degree of damage, the predicted results must approximate the real sample in terms of crack location, contour morphology, and overall structure. Unlike traditional methods that only output scalar indicators such as crack area, length, or damage level, this invention achieves prediction from operating parameters to crack images, more intuitively reflecting the spatial evolution characteristics of crack damage under repeated impacts.

[0065] After the model training converges, inputting any set of weaving angle, impact energy, and number of impacts will quickly yield the crack morphology prediction results for the corresponding working conditions.

[0066] To comprehensively evaluate the crack morphology prediction performance, several metrics were used: Intersection over Union (IoU, measuring the degree of overlap between predicted and actual crack regions), Dice coefficient (measuring the similarity between two sets), precision (the proportion of actual crack pixels in the predicted crack region), recall (the proportion of actual crack pixels correctly predicted), mean squared error (MSE), mean absolute error (MAE), and structural similarity index (SSIM, measuring the similarity between the predicted and actual images in terms of brightness, contrast, and structure). These metrics comprehensively evaluate the crack distribution morphology prediction performance from multiple perspectives, including region overlap, pixel error, and structural similarity.

[0067] S6: Obtain the impact condition parameters of the three-dimensional braided composite material to be tested, input them into the crack area prediction model and crack distribution morphology prediction model, and output the predicted crack area value and crack distribution morphology image.

[0068] Specifically, for any three-dimensional braided composite material component to be tested, its braiding angle, the expected impact energy and number of impacts during service are obtained. These three parameters are then input into the trained and converged AdaBoost crack area prediction model and the multi-feature fusion conditional generative adversarial network crack distribution morphology prediction model, respectively. This allows for the simultaneous acquisition of the predicted crack area and crack distribution morphology image under the working condition, thus achieving a comprehensive characterization of the crack propagation behavior of three-dimensional braided composite materials after repeated impacts.

[0069] Explainingly, through the above process, this invention establishes an end-to-end joint prediction framework from impact condition parameters to crack area and crack morphology images. The area prediction model quantitatively assesses the damage degree from a numerical perspective, while the morphology prediction model visually presents the location, outline, propagation direction, and penetration trend of the crack from a spatial distribution perspective. These two models complement each other, jointly constituting a complete prediction scheme for repeated impact damage of three-dimensional woven composite materials. This scheme can predict crack propagation behavior under one to five repeated impact conditions, providing technical support for the visual assessment of composite material structural damage, rapid prediction of impact damage, and optimized design of lightweight aircraft structures.

[0070] In one possible implementation, the generator and discriminator of the crack distribution morphology prediction model employ an alternating optimization strategy.

[0071] Specifically, during each training iteration, the generator parameters are first fixed. A batch of real crack images and their corresponding impact condition parameters are sampled from the dataset, and a batch of random noise vectors are sampled from a random noise distribution. The impact condition parameters and random noise vectors are concatenated and input into the generator to obtain a batch of predicted crack images. The real crack images and impact condition parameters are concatenated to form real sample combinations, and the predicted crack images and impact condition parameters are concatenated to form generated sample combinations. These are then input into the discriminator to calculate the discrimination loss. The discriminator's network parameters are updated using gradient descent, enhancing the discriminator's ability to classify real sample combinations as true and generated sample combinations as false. Then, the discriminator parameters are fixed again, and the impact condition parameters and random noise vectors are input into the generator to obtain new predicted crack images. The predicted crack images and impact condition parameters are concatenated and input into the discriminator to calculate adversarial loss, pixel reconstruction loss, edge loss, and perception loss. The generator's network parameters are updated using gradient descent, making the discriminator's judgment result of the predicted crack images output by the generator closer to the real images.

[0072] Interpretably, the generator and discriminator engage in an adversarial game during their alternating optimization process: the generator continuously learns to produce more realistic crack images to deceive the discriminator, while the discriminator continuously learns to more accurately distinguish between real and generated images. As the number of iterations increases, the generator's generation ability and the discriminator's discrimination ability improve simultaneously, eventually reaching a Nash equilibrium state, at which point the generator can output a highly realistic crack morphology prediction image.

[0073] In one possible implementation, to further reveal the influence of impact condition parameters on crack area evolution, the SHAP method is used to perform interpretability analysis on the AdaBoost model.

[0074] SHAP (Shapley Additive exPlanations, a feature importance analysis method based on cooperative game theory, assesses feature importance by calculating the marginal contribution of each feature to the model's prediction) analysis results show that the number of impacts has the most significant impact on the crack area prediction results, with the largest SHAP value distribution range, indicating that the damage accumulation caused by repeated impacts is the main reason affecting crack area growth; impact energy is the second most significant factor, while the weaving angle has a relatively small impact. Crack area is not controlled by a single factor, but is the result of the combined effects of impact energy, the number of impacts, and the weaving angle. Overall, with the increase of the number of impacts, the crack area shows a gradual increasing trend, indicating that multiple impacts produce cumulative damage; when the impact energy is low, the crack area changes slowly, while at higher impact energies, the crack area expands rapidly with the increase of the number of impacts, indicating that impact energy accelerates the damage propagation process.

[0075] Under different weaving angles, the evolution of crack area with impact energy and number of impacts varies. The weaving angle affects crack propagation behavior and distribution morphology by altering fiber orientation and load transfer path. A comprehensive comparison of models, thermograms, and SHAP results reveals that the crack area evolution of three-dimensional braided composites after repeated impacts exhibits significant nonlinearity and parametric coupling characteristics. Specifically, the number of impacts determines the degree of damage accumulation, impact energy represents the magnitude of the external load, and the weaving angle reflects the material's intrinsic response characteristics.

[0076] In a preferred implementation, the key implementation parameters of the crack area prediction model (AdaBoost regression model) are as follows: the number of weak learners (i.e., the number of iterations) is set to 50, the learning rate is set to 1.0, the loss function is of the linear type, the base learner is a decision tree regressor, and the maximum depth of the decision tree is set to 6 layers.

[0077] The key implementation parameters of the crack distribution morphology prediction model (multi-feature fusion conditional generative adversarial network) are as follows: the input image resolution is set to 256×256 pixels; the generator adopts the U-Net architecture, the encoder contains 4 downsampling modules, each module contains a convolutional layer (4×4 kernel size, stride 2), a batch normalization layer and a LeakyReLU activation function (negative slope set to 0.2), and the decoder contains 4 upsampling modules, each module contains a transposed convolutional layer (4×4 kernel size, stride 2), a batch normalization layer, a Dropout layer (dropout rate set to 0.5) and a ReLU activation function; the conditional batch normalization module is embedded before each upsampling module of the decoder, and normalizes the working condition parameters (weaving angle) θ Impact energyE Number of impacts N ′) is mapped to affine transformation parameters (scaling factor) through a fully connected layer. γ and offset factor β The attention mechanism module is embedded in the bottleneck layer between the encoder and decoder. It extracts the global context information of the feature map through global average pooling and global max pooling, respectively. The attention weight map is generated by convolutional layers and sigmoid activation function, and then multiplied element-wise with the original feature map to achieve feature recalibration. The discriminator adopts the conditional PatchGAN architecture, which contains 5 convolutional modules. Each module contains a convolutional layer (4×4 kernel size, stride of 2), a batch normalization layer, and a LeakyReLU activation function (negative slope set to 0.2), and finally outputs a 30×30 discriminant matrix.

[0078] The weighting coefficients of the multi-loss joint function are set as follows: , , and

[0079] The optimizer uses the Adam optimizer (Adaptive Moment Estimation, a stochastic optimization algorithm that dynamically adjusts the learning rate of each parameter by calculating the first and second moment estimates of the gradient), with a learning rate of 0.0002, momentum hyperparameters β1=0.5, and β2=0.999. The batch size is set to 8, and the total number of training epochs is set to 200. An alternating optimization strategy is used between the generator and discriminator; the discriminator is updated twice and the generator is updated once in each iteration.

[0080] Figure 1 The overall framework of the damage prediction model constructed in this invention comprises two sub-parts: (a) an AdaBoost crack area prediction model and (b) a cGAN crack distribution morphology prediction model. Both models use weaving angle, impact energy, and number of impacts as input conditions, respectively predicting low-velocity impact damage of three-dimensional woven composite materials from two dimensions: scalar numerical values ​​and image morphology. Figure 1 As shown in (a), the AdaBoost regression model performs quantitative prediction of crack area. The model first assigns initial sample weights to the training samples, then iteratively learns using the AdaBoost algorithm, continuously adjusting the sample weights to enhance the learning effect on high-error samples. Next, it performs a weighted set prediction of the results from multiple weak learners, finally outputting a scalar predicted value of the crack area. This model can quickly obtain quantitative indicators of crack damage, but relying solely on numerical results cannot reflect the spatial location, contour morphology, and propagation path information of the crack.

[0081] like Figure 1 (b) A conditional generative adversarial network (cGAN) is used to achieve image-level prediction of crack distribution morphology. The operating condition data is input into the generator G (U-Net structure) to generate predicted crack images under the corresponding operating conditions. The discriminator D uses a PatchGAN structure, simultaneously receiving crack images generated by the model and real crack samples processed by Otsu to perform image authenticity judgment. During network training, supervised loss calculation is performed, and the generator and discriminator parameters are alternately updated through backpropagation using a multi-loss joint function. The results of the discrimination output further guide the iterative optimization of the network. After the model training converges, inputting any set of impact condition parameters will output a complete crack distribution image, compensating for the lack of spatial morphology information in scalar regression models and enabling a direct reproduction of the crack location, outline, and propagation morphology.

[0082] In one possible implementation, during the training process of the crack area prediction model, six regression models are selected for comparative analysis.

[0083] The six regression models selected were: Gradient Boost, Support Vector Regression (SVR), K-Nearest Neighbors (KNN), ExtraTree, Random Forest, and AdaBoost. All models used the same input variables (normalized weaving angle, impact energy, and number of impacts), the same data partitioning method (training set:validation set:test set = 8:1:1), and the same evaluation metric. R 2 MAE and MSE were used to ensure fairness in the model comparison. Each model made predictions on the test set, and the predicted values ​​were compared with the actual experimental values. The results are as follows: Figure 2 As shown, all models can establish a mapping relationship between impact parameters and crack area, but the prediction accuracy of each model varies significantly.

[0084] Depend on Figure 2 (a) It is evident that the predictive power of the Gradient Boost model is relatively limited, and the coefficient of determination... R 2 =0.746, Mean Absolute Error (MAE) = 13.227, Mean Squared Error (MSE) = 346.035. The relatively discrete distribution of the predicted sample points indicates that the model has difficulty accurately capturing the complex nonlinear damage evolution characteristics of crack propagation.

[0085] Figure 2 (b) shows improved predictive performance of the Support Vector Regression (SVR) model. R 2=0.839, MAE=10.015, MSE=218.941; however, for samples with larger crack areas, the prediction bias gradually increases, indicating that the model's ability to characterize nonlinear damage evolution under severe damage conditions is insufficient. The K-Nearest Neighbor (KNN) model further improves prediction accuracy ( Figure 2 (c) R 2 =0.862, MAE=8.831, MSE=188.127, but its predictive stability is still affected by the sample distribution, and the prediction error for some highly damaged samples is still relatively large. The ExtraTree model has stronger nonlinear fitting ability ( Figure 2 (d) R²=0.878, MAE=8.458, MSE=165.859, but there is still a certain prediction bias for samples with large crack areas.

[0086] Figure 2 (e) The predictive performance of the RandomForest model has been further improved. R 2 =0.901, MAE=8.641, MSE=134.777, proving that the model can effectively learn the nonlinear mapping relationship between impact conditions and crack area. Among all the comparative models, the AdaBoost model has the best overall prediction performance. Figure 2 (f) The predicted values ​​have the highest degree of agreement with the actual experimental values, and the predicted sample points are closer to the regression trend line; the model achieves the highest R²=0.930, while having the lowest MAE=8.133 and MSE=95.724.

[0087] Comprehensive R 2 As can be seen from the three evaluation indicators MAE, MSE, etc., the AdaBoost model can stably describe the nonlinear relationship between weaving angle, impact energy, number of impacts, and crack area. Therefore, this invention selects this model to carry out subsequent crack area prediction and analysis of the influence of each parameter.

[0088] To systematically evaluate the predictive performance of the proposed model for crack distribution morphology, the multi-feature fusion conditional generative adversarial network (Ours-cGAN, hereinafter referred to as Ours) of this invention was compared with four other models: DNN, CVAE, CVAE+Ours, and basic cGAN. The quantitative evaluation results of each model are as follows: Figure 3 As shown in the figure. This evaluation uses Intersection over Union (IoU), Dice coefficient, precision, recall, and structural similarity (SSIM) as evaluation metrics.

[0089] Depend on Figure 3As shown in (a)-(h), different models exhibit significant differences in crack distribution morphology reconstruction. The DNN model has the lowest IoU (Intersection over U) at only 0.229, indicating that the single feedforward neural network has limited ability to capture the complex nonlinear relationship between impact parameters and crack distribution morphology. Although the CVAE model possesses basic generation capabilities, its ability to reconstruct crack boundaries and fine structural details is insufficient, with IoU (0.466), Dice (0.506), and SSIM (0.268) all at low levels. After introducing the module proposed in this invention into the CVAE framework, all evaluation metrics of the CVAE+Ours model are simultaneously improved compared to the original CVAE. Figure 3 As shown in (a)-(e), the IoU improved from 0.392 to 0.503, the Dice coefficient improved from 0.506 to 0.593, and the SSIM improved from 0.268 to 0.712, but the overall performance of this model was still inferior to the Ours model. The prediction performance of the basic cGAN was better than that of DNN and CVAE, with an IoU of 0.466 and a Dice coefficient of 0.562, proving that adversarial training helps generate crack morphologies that are closer to the real distribution. In contrast, the Ours model achieved the best results on almost all evaluation metrics: IoU reached 0.799, Dice was 0.887, precision was 0.951, recall was 0.836, and SSIM was 0.875, all of which were significantly higher than the other comparative models.

[0090] also, Figure 3 The radar charts (i)-(j) intuitively demonstrate the comprehensive advantages of the Ours model. Figure 3 (i) Each coordinate axis corresponds to positive metrics such as IoU, Dice, precision, recall, F1 score, and SSIM. The larger the value on the coordinate axis, the better the model performance. The polygonal outline of the Ours model (red curve) extends the furthest outward in all dimensions, which is significantly better than cGAN, CVAE, CVAE+Ours, and DNN; while the curve of DNN is concentrated at the innermost part of the radar chart, which confirms that its crack reconstruction ability is the weakest. Figure 3 (j) The coordinate axes represent error metrics (MAE, MSE, RMSE, NMAE, NRMSE, Area_RelErr). The smaller the metric value, the higher the prediction accuracy. Ours model has the closest contour to the center of the radar chart across all error metrics, meaning it has the smallest prediction error; the polygon boundaries of the other baseline models expand outward, resulting in larger reconstruction errors.

[0091] In summary, the two radar charts visually demonstrate that the model of this invention achieves optimal performance in both similarity and error metrics, without any significant performance weaknesses in any single metric, thus verifying the model's excellent comprehensive crack morphology reconstruction capability. This indicates that the Ours model can effectively establish a nonlinear mapping relationship between impact conditions and crack distribution morphology, achieving accurate reconstruction of crack spatial distribution characteristics and boundary structures. Compared to DNN, CVAE, CVAE+Ours, and basic cGAN, the model proposed in this invention exhibits superior performance in crack morphology prediction accuracy and structural reconstruction capability.

[0092] To further explore the reliability of the model (Ours) in predicting crack distribution morphology under different loading conditions, a visual comparison and statistical analysis were conducted for different impact numbers, weaving angles and impact energies. Figure 4 The crack morphology prediction results and their corresponding error distribution are given. Figure 4 The actual crack morphology, model prediction results, and absolute error diagrams under different impact conditions were intuitively compared.

[0093] Under different impact numbers ( Figure 4 (a1)-(a5)), as the number of impacts increases, the crack morphology gradually evolves from localized damage to a large-scale crack network. Despite the continuous accumulation of damage and the emergence of secondary crack branches, the model of this invention can still accurately capture the main crack propagation path and spatial development trend. Figure 4 Sections (b1)-(b3) investigated the influence of the weaving angle on the crack morphology prediction results. For different weaving structures, the model prediction results showed good agreement with experimental observations, accurately reproducing the location of the main crack, the dispersed damage area, and the interconnected crack boundaries. These results indicate that the model has strong adaptability to crack morphology differences caused by variations in the reinforcement structure. Figure 4 (c1)-(c6) demonstrate the model's predictive performance under different impact energies. As the impact energy increases, the crack gradually develops from a scattered damage area into a larger and more interconnected crack network. The model of this invention can accurately reproduce this damage evolution process, indicating that the model can effectively establish a mapping relationship between the intensity of the impact load and the crack morphology evolution.

[0094] The following are embodiments of the apparatus of the present invention, which can be used to execute embodiments of the method of the present invention. For details not disclosed in the apparatus embodiments, please refer to the embodiments of the method of the present invention.

[0095] See Figure 1 In another embodiment of the present invention, a machine learning-based system for predicting the crack propagation behavior of three-dimensional braided composite materials under multiple impact loads is provided, comprising: The data acquisition module is used to acquire the original crack image dataset of three-dimensional braided composite materials under different impact conditions and the corresponding impact condition parameters. The image separation and label extraction module is configured to use an adaptive threshold segmentation algorithm to perform image separation processing on the original crack image dataset, extract the binarized crack image as the real label image, and obtain the corresponding real crack area label based on the pixel statistics of the binarized crack image. The area model training module is used to train an AdaBoost regression prediction model by taking the impact condition parameters as input variables and the actual crack area label as output variables, in order to establish the mapping relationship between parameters and area and obtain the crack area prediction model. The morphology model training module is used to construct a multi-feature fusion conditional generative adversarial network, which includes a conditional encoding module, a generator, and a discriminator. The impact condition parameters are mapped to conditional features through the conditional encoding module, and the conditional features are input into the generator to output a predicted crack image. The discriminator is used to distinguish between the predicted crack image and the corresponding real label image. The network parameters of the generator and the discriminator are alternately updated based on a multi-loss joint function until the model converges, thus obtaining a crack distribution morphology prediction model. The joint prediction execution module is used to acquire the impact condition parameters of the three-dimensional braided composite material to be tested, input them into the crack area prediction model and the crack distribution morphology prediction model, and output the predicted crack area value and crack distribution morphology image.

[0096] All relevant content of each step involved in the aforementioned embodiments of the machine learning-based method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads can be referenced to the functional description of the corresponding functional module of the machine learning-based system for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads in the embodiments of the present invention, and will not be repeated here.

[0097] In another embodiment of the present invention, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the above-described method for predicting the crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning.

[0098] In another embodiment of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-described method for predicting the crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning.

[0099] Unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0100] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Those skilled in the art can readily implement the present invention based on the accompanying drawings and the above description. However, any modifications, alterations, or variations made by those skilled in the art without departing from the scope of the present invention, utilizing the disclosed technical content, are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.

Claims

1. A method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning, characterized in that, include: Obtain the original crack image dataset of three-dimensional braided composite materials under different impact conditions and the corresponding impact condition parameters; An adaptive threshold segmentation algorithm is used to perform image separation processing on the original crack image dataset, extracting the binarized crack images as the real label images, and obtaining the corresponding real crack area labels based on the pixel statistics of the binarized crack images. Using impact condition parameters as input variables and actual crack area labels as output variables, an AdaBoost regression prediction model is trained to establish the mapping relationship between parameters and area, thus obtaining a crack area prediction model. A multi-feature fusion conditional generative adversarial network is constructed, which includes a conditional coding module, a generator, and a discriminator. The impact condition parameters are mapped to conditional features through the conditional encoding module, and the conditional features are input into the generator to output the predicted crack image. The discriminator is used to distinguish between the predicted crack image and the corresponding real label image. The network parameters of the generator and the discriminator are alternately updated based on the multi-loss joint function until the model converges, and the crack distribution morphology prediction model is obtained. The impact parameters of the three-dimensional braided composite material to be tested are obtained and input into the crack area prediction model and crack distribution morphology prediction model, and the predicted crack area value and crack distribution morphology image are output.

2. The method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning according to claim 1, characterized in that, The process of using an adaptive threshold segmentation algorithm to perform image separation processing on the original crack image dataset and extracting binarized crack images as the real label images includes: Convert the images in the original crack image dataset into single-channel grayscale images; Using Otsu's maximum inter-class variance method, the inter-class variance of the single-channel grayscale image is calculated, and the grayscale value that maximizes the inter-class variance is taken as the optimal segmentation threshold. Pixels with gray values ​​lower than the optimal segmentation threshold in the single-channel grayscale image are assigned a value of 1 as crack damage pixels, and pixels with gray values ​​higher than the optimal segmentation threshold are assigned a value of 0 as matrix background pixels, generating a black-and-white binary classification of the crack image as the real label image. The total number of crack damage pixels in the binarized crack image is counted, and the total number of crack damage pixels is determined as the true crack area label under the corresponding impact condition parameters.

3. The method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning according to claim 1, characterized in that, The specific process of establishing a crack area prediction model includes: The impact condition parameters are used as input variables, and the actual crack area labels are used as output variables in the AdaBoost regression prediction model. The training samples in the AdaBoost regression prediction model are assigned initial sample weights; the training samples are iteratively learned using the AdaBoost algorithm, and the sample weights are updated according to the prediction error distribution of the previous weak learner in each iteration. The outputs of multiple weak learners after multiple iterations are weighted and summed to output the crack area prediction model that has been trained and converged. The output of the AdaBoost regression prediction model Represented as: ;in, This represents the area of ​​crack damage.

4. The method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning according to claim 1, characterized in that, The generator includes a U-Net architecture comprising a downsampling encoding module and an upsampling decoding module; The generator is embedded with a conditional batch normalization module and an attention mechanism module. The conditional batch normalization module is used to continuously regulate the intermediate features of the network by the impact condition parameters, and the attention mechanism is used to enhance the network's ability to express the features of the crack backbone region and the local propagation region. The discriminator adopts a conditional PatchGAN architecture. The input of the discriminator receives a first sample combination and a second sample combination. The first sample combination is composed of the real label image and the conditional features, and the second sample combination is composed of the predicted crack image and the conditional features. The discriminator divides the input first sample combination and the second sample combination into local image blocks for output discrimination.

5. The method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning according to claim 1, characterized in that, The impact conditions parameters include weaving angle, impact energy, and number of impacts. Before using the impact condition parameters as input variables, the impact condition parameters are normalized to obtain a normalized input vector: ; In the formula, The normalized weaving angle, The normalized impact energy. This represents the normalized number of impacts.

6. The method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning according to claim 1, characterized in that, The acquisition of the original crack image dataset of the three-dimensional braided composite material under different impact conditions includes: acquiring three-dimensional braided composite material samples with different braiding angles prepared by a four-step three-dimensional braiding process and a vacuum-assisted resin transfer molding process; applying a low-speed impact load along the thickness direction of the three-dimensional braided composite material sample and performing 1 to 5 repeated impacts under the same impact energy; acquiring crack images on the surface of the three-dimensional braided composite material sample after each impact to generate the original crack image dataset.

7. The method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads based on machine learning according to claim 1, characterized in that, The calculation formula for the joint function of multiple losses is as follows: In the formula, For the total loss, To reduce the overall pixel deviation between the predicted image and the real image, Edge loss is used to enhance the consistency of crack profile and boundary region. To improve the perceptual loss for enhancing the similarity between predicted and real images in high-level structural features, To constrain the generated image to approximate the distribution of the real crack image, , , and These are the weighting coefficients for each loss term.

8. A machine learning-based system for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads, characterized in that... include: The data acquisition module is used to acquire the original crack image dataset of three-dimensional braided composite materials under different impact conditions and the corresponding impact condition parameters. The image separation and label extraction module is configured to use an adaptive threshold segmentation algorithm to perform image separation processing on the original crack image dataset, extract the binarized crack image as the real label image, and obtain the corresponding real crack area label based on the pixel statistics of the binarized crack image. The area model training module is used to train an AdaBoost regression prediction model by taking the impact condition parameters as input variables and the actual crack area label as output variables, in order to establish the mapping relationship between parameters and area and obtain the crack area prediction model. The morphology model training module is used to construct a multi-feature fusion conditional generative adversarial network, which includes a conditional encoding module, a generator, and a discriminator. The impact condition parameters are mapped to conditional features through the conditional encoding module, and the conditional features are input into the generator to output a predicted crack image. The discriminator is used to distinguish between the predicted crack image and the corresponding real label image. The network parameters of the generator and the discriminator are alternately updated based on a multi-loss joint function until the model converges, thus obtaining a crack distribution morphology prediction model. The joint prediction execution module is used to acquire the impact condition parameters of the three-dimensional braided composite material to be tested, input them into the crack area prediction model and the crack distribution morphology prediction model, and output the predicted crack area value and crack distribution morphology image.

9. A computer device, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the machine learning-based method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the machine learning-based method for predicting crack propagation behavior of three-dimensional braided composite materials under multiple impact loads as described in any one of claims 1-7.