Method for identifying damage mode of carbon fiber reinforced composite material by using interpretable convolutional neural network

Abstract

The invention discloses a method for identifying a damage mode of a carbon fiber reinforced composite material by using an interpretable convolutional neural network. The method comprises the following steps: establishing a database; preprocessing the data; constructing a convolutional neural network model; and explaining a convolutional neural network model and generalization application of the convolutional neural network model, and after model training is completed, applying model generalization to identification of a multi-damage-mode image and a composite material impact damage mode in actual engineering so as to verify model robustness. According to the technical scheme provided by the invention, the damage mode of the carbon fiber reinforced composite material can be accurately and efficiently identified, and accurate quantification and efficient decoupling of the damage mode of the composite material under the service condition are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of composite material damage detection, and in particular to a method for identifying damage patterns of carbon fiber reinforced composite materials using an interpretable convolutional neural network. Background Art

[0002] Carbon fiber reinforced composite materials refer to composite materials that use carbon fiber or carbon fiber fabric as reinforcement and are formed by resin, ceramic, metal, cement, carbon or rubber matrix. Due to its advantages of high specific strength, high specific stiffness, corrosion resistance, strong designability, etc., it is widely used in aerospace, civil construction, chemical equipment, medical equipment and other fields. However, since carbon fiber reinforced composite materials usually serve in harsh environments and complex load conditions, they are inevitably subjected to impact loads, causing various types of damage. [1] .

[0003] The damage sensitivity and severity of composite materials can be affected by many factors, including composite material factors, manufacturing factors, load and environmental factors, design factors, etc. [2, 3] Carbon fiber reinforced composite materials have complex failure mechanisms. Common damage modes of carbon fiber reinforced composite materials include matrix cracking, fiber-matrix interface peeling, delamination, fiber extraction and fiber breakage, which can easily lead to material performance degradation and structural damage, affecting its service status throughout its life cycle. [4-6] .

[0004] Currently, the traditional damage detection methods for composite materials include ultrasonic nondestructive testing [7] , Acoustic Emission Testing [8] , Optical Nondestructive Testing [9] , infrared thermal imaging detection

[10] and thermal imaging technology

[11] However, traditional non-destructive testing technologies have their own scope of application and certain limitations.

[12] For example, ultrasonic testing technology is not suitable for complex service conditions.

[13] ; Acoustic emission technology requires good coupling, expensive equipment and certain technical requirements for testers

[14] Optical nondestructive testing technology requires materials to have good heat absorption rate and high depth requirements. The deeper the depth, the lower the detection sensitivity. It is suitable for relatively thin laminates, and the edge imaging is easily blurred during detection.

[15] In order to overcome the shortcomings of the existing technology, it is necessary to study a composite material damage detection method with wide applicability, high reliability and strong robustness, improve the efficiency and accuracy of composite material structure damage detection, and improve the applicability of damage detection. Summary of the invention

[0005] The present invention provides a method and device for predicting dynamic stress at structural defects based on a BP neural network, aiming to effectively solve the technical problems faced in the prior art of time-consuming calculation of multiple parameters, poor applicability when facing different structures, materials and defect forms, resulting in high time and labor costs.

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for identifying the damage mode of carbon fiber reinforced composite materials using an interpretable convolutional neural network, so as to solve the problems of high cost, low accuracy, and inability to characterize the degree and type of internal damage of materials under low-speed impact loads in current damage detection methods for composite materials. At the same time, in view of the "black box" problem of machine learning models, the present invention provides a method for explaining the internal working mechanism of convolutional neural network models, effectively identifying and preventing potential deviation problems.

[0007] According to a first aspect of the present invention, the present invention provides a method for identifying damage modes of carbon fiber reinforced composite materials using an interpretable convolutional neural network, comprising the following steps: Database establishment: collect and organize CT images of carbon fiber composite laminates damaged under low-speed impact loads, and establish CT damage pattern image dataset; Data preprocessing, resizing the image size to a preset size, adjusting the mean and standard deviation of the computed tomography damage pattern image dataset for standardization; Constructing a convolutional neural network model, wherein the convolutional neural network model includes a convolutional layer and a fully connected layer; Train and test the convolutional neural network model, use the computed tomography damage pattern image dataset to train and test the established convolutional neural network model, and perform classification, recognition and quantitative analysis on the carbon fiber composite material impact damage pattern image dataset; Explain the convolutional neural network model, introduce class activation maps and Shapley additive interpretation value interpretable tools to transparently explain the convolutional neural network model, reveal the potential variables and mechanisms that affect the model's prediction results, visualize the contribution of each type of damage pattern to the final prediction results, and perform rotational validation on the computed tomography damage image dataset; Generalization application of the convolutional neural network model,After the model training is completed, the model is generalized and applied to the identification of multiple damage pattern images and composite material impact damage patterns in actual engineering to verify the robustness of the model.

[0008] In addition to or as an alternative to one or more of the features disclosed above, the method further comprises: Composite material preparation steps: - Using carbon fiber or carbon fiber fabric as reinforcement and resin as matrix; -Preparing carbon fiber composite laminates by lamination process; Sample preparation steps: -Select carbon fiber composite laminate as the sample plate; -Manufacture multiple groups of samples according to predetermined specifications, where the shapes and widths of different groups of samples are consistent but the thicknesses are different; - Prepare two identical parallel samples for each set of specimens; Impact test steps: -Choose between two shock sources with different energy levels; - Two parallel samples of each group of specimens are placed under impact sources of different energies for impact tests; Data processing steps: - Use computer tomography to inspect the specimens after impact testing; - Obtain computer tomography damage image data of the damaged area under different impact energy loads.

[0009] In addition to or as an alternative to one or more of the features disclosed above, the damage mode of the carbon fiber composite laminate under load impacts of different energy magnitudes includes at least one of matrix cracking, delamination or fiber breakage.

[0010] In addition to or instead of one or more of the features disclosed above, the computed tomography injury image dataset is divided into a single injury pattern dataset and a multiple injury pattern dataset, wherein the image data in the single injury pattern dataset contains only one injury pattern, and the image data in the multiple injury pattern dataset contains two or more injury patterns.

[0011] In addition to or as an alternative to one or more of the features disclosed above, the step of adjusting the mean and standard deviation of the computed tomography lesion pattern image dataset for standardization further comprises: Each computed tomography damage pattern image data was transformed into a normal distribution with a mean of 1 and a variance of 0.

[0012] In addition to or as an alternative to one or more of the features disclosed above, the convolution layer is used to perform damage edge detection and feature extraction on the computed tomography damage image, and the fully connected layer is used to highly abstract and integrate the features extracted by the convolution layer to output the final laminate damage pattern recognition result; Among them, there are 5 convolutional layers and 3 fully connected layers; the network structure of the convolutional neural network model also includes 3 pooling layers and 1 normalized exponential function layer.

[0013] In addition to or instead of one or more of the features disclosed above, the number of channels of the first convolutional layer to the fifth convolutional layer of the convolutional neural network model are 16, 32, 64, 96 and 128 respectively.

[0014] In addition to one or more of the features disclosed above, or as an alternative, the learning rate of the convolutional neural network model adopts an adaptive adjustment strategy, with an initial learning rate of 1e-4 and a decay rate of 0.1, which changes every 10 epochs. The model adapts more quickly to the distribution characteristics of each damage mode in the target domain.

[0015] In addition to or as an alternative to one or more of the features disclosed above, the steps of training and testing the established convolutional neural network model using the computed tomography damage pattern image dataset include: randomly selecting 80 samples from each of the three types of damage patterns in the standardized computed tomography single damage pattern image dataset to form a test set, and randomly dividing the remaining computed tomography single damage pattern images into a training set and a validation set in a ratio of 4:1; using the training set and the validation set to train the convolutional neural network model, and adjusting the hyperparameter settings of the network model during the training process; after the model training is completed, using the test set to test the damage pattern recognition performance of the convolutional neural network model for the computed tomography image.

[0016] In addition to or instead of one or more of the features disclosed above, for the trained convolutional neural network model, a class activation map is introduced to reweight the weights of the last convolutional layer of the convolutional neural network model, visualize the importance of each input feature, visualize the decision process of the proposed convolutional neural network model, and explain the reasons behind the model decision.

[0017] In addition to or as an alternative to one or more of the features disclosed above, the class activation map includes at least one of Grad-CAM, Grad-CAM++, Score-CAM, XGrad-CAM, Ablation-CAM, EigenGrad-CAM, and LayerCAM.

[0018] In addition to one or more of the features disclosed above, or as an alternative, Shapley additive explanatory value analysis is introduced for the trained convolutional neural network model to calculate the contribution of different damage pattern features to the prediction results of the convolutional neural network model, provide an explanatory score for each damage pattern feature, and reveal the impact of different damage patterns on the model output.

[0019] In addition to or as an alternative to the above-disclosed one or more features, for the trained convolutional neural network model, the three damage pattern images of matrix cracking, delamination, and fiber breakage in the test set are randomly assigned to be rotated 90°, 180°, and 270° in a clockwise direction respectively, and the class activation map and Shapley additive explanatory value analysis visualization model are used to identify and predict the characteristics of the three damage patterns after rotation.

[0020] In addition to or instead of one or more of the features disclosed above, the hyperparameters and weight values ​​of each layer of the optimal convolutional neural network model are stored, and the network model is called to identify a computed tomography multiple damage pattern data set; the model is applied to actual engineering to identify different damage modes of carbon fiber reinforced composite laminates under low-speed impact.

[0021] In addition to or instead of one or more of the features disclosed above, multiple damage pattern images are efficiently decoupled and quantitatively analyzed, and interpretable tools such as class activation maps and Shapley additive explanatory values ​​are introduced for transparent interpretation to reveal potential variables and mechanisms that affect the model prediction results, and visualize the contribution of each type of damage pattern to the final prediction results; confidence values ​​are introduced to quantitatively analyze the prediction results.

[0022] In addition to or as an alternative to one or more features disclosed above, the verifying the initial dynamic stress prediction model using the test set comprises the following steps: Keeping all parameters of the initial dynamic stress prediction model unchanged, applying each test sample in the test set to the initial dynamic stress prediction model for verification; If the sum of the test errors of each sample in the test set reaches the preset error tolerance, it indicates that the initial dynamic stress prediction model has good generalization ability and the test passes; if the sum of the test errors is too large, underfitting or overfitting occurs, the test fails, and the training set is reused to perform initial training on the BP neural network prediction model.

[0023] One of the above technical solutions has the following advantages or beneficial effects: the convolutional neural network model is trained and tested through a single damage data set of computed tomography (CT), and the hyperparameters and weight values ​​of each layer of the convolutional neural network model after training are stored. It can be applied to multiple damage pattern data sets and other damage pattern recognition tasks of carbon fiber reinforced composite materials under impact loads, and has strong robustness.

[0024] Another technical solution among the above technical solutions has the following advantages or beneficial effects: due to the introduction of class activation mapping (CAM) technology and Shapley additive interpretation (SHAP) value interpretable tools to transparently interpret the convolutional neural network, CAM locally interprets the convolutional neural network, which well illustrates the reliability of the convolutional neural network model in black-box prediction of regions of interest in different damage patterns. SHAP calculates the marginal contribution of the characteristics of different damage patterns to the model prediction results, reveals the potential variables and mechanisms that affect the model prediction results, and truly establishes trust between people and models.

[0025] Another technical solution among the above technical solutions has the following advantages or beneficial effects: the constructed convolutional neural network model comprehensively considers the importance of convolutional layers and channels, contains 5 convolutional layers, and has a simple and lightweight structure; the network model adopts a larger number of channels in the early stage, which can process more input data information at the same time, extract more features of different damage modes of composite materials, and improve the accuracy in damage pattern recognition tasks. The network model adopts a smaller number of channels in the later stage and introduces the dropout method in the fully connected layer to effectively prevent the convolutional neural network model from overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The technical solutions and other beneficial effects of the present invention will be made apparent by describing in detail the specific embodiments of the present invention in conjunction with the accompanying drawings.

[0027] Figure 1 is a schematic diagram of computer tomography provided according to an embodiment of the present invention; Figure 2 is a schematic diagram of a flow chart provided according to an embodiment of the present invention; Figure 3 is a schematic diagram of the structure of an interpretable convolutional neural network model provided according to an embodiment of the present invention; Figure 4 is a schematic diagram of an interpretable convolutional neural network damage pattern recognition result provided according to an embodiment of the present invention; Figure 5 It is a schematic diagram of explaining a convolutional neural network model according to the class activation map technology provided by an embodiment of the present invention; Figure 6 It is a schematic diagram of a convolutional neural network model for analyzing and explaining the Shapley additive interpretation (SHAP) value according to an embodiment of the present invention. DETAILED DESCRIPTION

[0028] The technical scheme in the embodiment of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0029] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the term "and / or" herein is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " herein, unless otherwise specified, generally indicates that the associated objects before and after are in an "or" relationship.

[0030] Reference Figure 1 In one embodiment of the present invention, a method for identifying damage modes of carbon fiber reinforced composite materials using an interpretable convolutional neural network is disclosed, which may include the following steps: T01, database establishment, collect and organize the CT images of carbon fiber composite laminates damaged under low-speed impact loads, and establish a CT damage pattern image dataset; T02, data preprocessing, resizing the image size to a preset size, adjusting the mean and standard deviation of the computed tomography damage pattern image dataset for standardization; T03, constructing a convolutional neural network model, wherein the convolutional neural network model includes a convolutional layer and a fully connected layer; T04, training and testing convolutional neural network models, using computed tomography damage pattern image datasets to train and test the established convolutional neural network models, and classifying, identifying and quantitatively analyzing carbon fiber composite material impact damage pattern image datasets; T05, Interpreting Convolutional Neural Network Models, introduces interpretable tools such as class activation maps and Shapley additive interpretation values ​​to transparently interpret the convolutional neural network model, reveal the potential variables and mechanisms that affect the model's prediction results, visualize the contribution of each type of damage pattern to the final prediction results, and perform rotation verification on the computed tomography damage image dataset; among them, the position and direction of the cracks change to different degrees through rotation transformation, further verifying the robustness of the model, improving the interpretability of the model, and enabling us to understand the decision-making process and important features of the model; T06, generalization application of convolutional neural network model. After the model training is completed, the model is generalized and applied to identify multi-damage pattern images and composite material impact damage patterns in actual engineering to verify the robustness of the model.

[0031] It is understandable that the order of step T01 and step T03 can be changed. Step T01 can be executed first and then step T03, or step T03 can be executed first and then step T01, or step T01 and step T03 can be executed at the same time. The specific execution order can be set according to actual conditions and is not limited here.

[0032] It is also understandable that the order of step T02 and step T03 can be changed. Step T02 can be executed first and then step T03, or step T03 can be executed first and then step T02, or step T02 and step T03 can be executed at the same time. The specific execution order can be set according to actual conditions and is not limited here.

[0033] Among them, computed tomography (CT) can provide 3D images of the internal structure of an object without loss. CT uses the penetrating power of X-rays to obtain a series of 2D radiographs of an object observed from multiple angles, and creates a series of cross-sectional images through calculations to display the internal structure of the object. Nondestructive CT has been used to image delicate samples that are not easily sliced ​​(e.g. frozen ice cream) or that should not be damaged (e.g. cultural relics), or to assess the integrity of engineering components (e.g. turbine blades). CT has also been used for discontinuous or continuous time-dependent studies monitoring the evolution of three-dimensional structures, such as the growth of malignant tumors, the metamorphosis of pupae, the flow of fluids in rocks, or the catastrophic thermal runaway failure of lithium batteries. CT was the first method capable of producing images in transverse planes (slices) through the human body without the bias caused by the superposition of different anatomical structures. Since the first clinical CT scanner, tremendous technological developments involving X-ray sources and new detectors have resulted in substantial improvements in transmitted dose and image quality. CT technology has evolved to higher resolution devices. X-ray micro-tomography (μ-CT) systems are capable of achieving micron-level spatial resolution for centimeter-sized samples and submicron-level spatial resolution for 1-2 mm-sized samples (nano-tomography). Today, this technology is used in a wide range of research fields, from medicine to biology, and from geology to archaeology and materials science.

[0034] Class Activation Map (CAM) is a visualization technique used primarily in convolutional neural networks in deep learning to identify and explain which regions in images or time series data are most critical to the model's decision-making process. It is particularly useful for helping understand how the network makes classification decisions based on visual or time series features. Shapley Additive Explanations (SHAP) is an interpretability method based on Shapley values ​​that aims to explain the output of machine learning models. SHAP provides global and local interpretability by calculating the contribution of each feature to the model's predictions.

[0035] The present invention trains and tests the convolutional neural network model through a single damage data set of computed tomography (CT), and stores the hyperparameters and weight values ​​of each layer of the convolutional neural network model after training. It can be applied to multiple damage pattern data sets and other damage pattern recognition tasks of carbon fiber reinforced composite materials under impact loads, and has strong robustness.

[0036] Furthermore, due to the introduction of the class activation mapping (CAM) technology and the interpretable tool of Shapley additive explanation (SHAP) value, the convolutional neural network is transparently explained. CAM performs local interpretation of the convolutional neural network, which well illustrates the reliability of the convolutional neural network model in the black-box prediction of the region of interest of different damage patterns. SHAP calculates the marginal contribution of the characteristics of different damage patterns to the model prediction results, reveals the potential variables and mechanisms that affect the model prediction results, and truly establishes trust between people and models.

[0037] In order to collect and organize the CT images of carbon fiber composite laminates damaged under low-speed impact loads, some pre-processing is usually required: S01, composite material preparation steps: - Using carbon fiber or carbon fiber fabric as reinforcement and resin as matrix; -Preparing carbon fiber composite laminates by lamination process; S02, sample preparation steps: -Select carbon fiber composite laminate as the sample plate; -Manufacture multiple groups of samples according to predetermined specifications, where the shapes and widths of different groups of samples are consistent but the thicknesses are different; - Prepare two identical parallel samples for each set of specimens; S03, impact test steps: -Choose between two shock sources with different energy levels; - Two parallel samples of each group of specimens are placed under impact sources of different energies for impact tests; S04, data processing steps: - Use computer tomography to inspect the specimens after impact testing; - Obtain computer tomography damage image data of the damaged area under different impact energy loads.

[0038] In one embodiment, 4 groups of carbon fiber composite laminate specimens with different thicknesses are manufactured according to the specification of 100mm×100mm, and the thicknesses of the 4 groups of carbon fiber composite laminate specimens are 2mm, 4mm, 6mm and 7mm, respectively, and two identical parallel specimens are prepared for each group of specimens; then, impact tests are carried out on the two parallel specimens of each group of specimens in turn using impact loads of 50J and 80J; finally, the specimens after the impact test are detected using computer tomography technology to obtain computer tomography damage image data of the damaged area under different impact energy loads.

[0039] In another embodiment, 4 groups of carbon fiber composite laminate specimens with different thicknesses are manufactured according to the specification of 120mm×120mm, and the thicknesses of the 4 groups of carbon fiber composite laminate specimens are 3mm, 6mm, 8mm and 9mm, respectively, and two identical parallel specimens are prepared for each group of specimens; then, impact tests are carried out on the two parallel specimens of each group of specimens in turn using impact loads of 60J and 90J; finally, the specimens after the impact test are detected by computer tomography technology to obtain computer tomography damage image data of the damaged area under different impact energy loads.

[0040] It is understandable that the specifications, shape, thickness and other dimensions of the sample, as well as the size of the impact load can be adaptively set according to actual conditions and are not limited here.

[0041] In some embodiments, the damage mode of the carbon fiber composite laminate under load impacts of different energy magnitudes includes at least one of matrix cracking, delamination, or fiber breakage.

[0042] In some embodiments, the computed tomography damage image dataset is divided into a single damage pattern dataset and a multiple damage pattern dataset, wherein the image data in the single damage pattern dataset contains only one damage pattern, and the image data in the multiple damage pattern dataset contains two or more damage patterns. According to the characteristics and distribution of the damage pattern, the computed tomography damage pattern image dataset is divided into two categories: a single damage pattern image dataset and a multiple damage pattern image dataset. The CNN model is trained and tested using a single damage pattern dataset, and the optimal hyperparameters of the model are determined by classifying, identifying and quantitatively analyzing the single damage pattern, thereby verifying the performance and accuracy of the model; based on the optimal hyperparameter combination of the model, the CNN model is tested using a multiple damage pattern dataset, and the robustness and generalization performance of the model are verified by classifying, identifying and quantitatively analyzing multiple damage patterns.

[0043] In some embodiments, the step of adjusting the mean and standard deviation of the computed tomography lesion pattern image dataset for standardization further comprises: T021, transform each computed tomography damage pattern image data into a normal distribution with a mean of 1 and a variance of 0.

[0044] For example, in one embodiment, the mean vector corresponding to the mean of the three RGB channels in the computed tomography image is (0.485, 0.456, 0.406), and the standard deviation vector corresponding to the standard deviation of the three RGB channels is (0.229, 0.224, 0.225), thereby converting each CT image sample into a normal distribution with a mean of 1 and a variance of 0.

[0045] In another embodiment, the mean vector corresponding to the means of the three RGB channels in the computed tomography image is (0.5, 0.5, 0.5), and the standard deviation vector corresponding to the standard deviation of the three RGB channels is (0.5, 0.5, 0.5), thereby converting each CT image sample into a normal distribution with a mean of 1 and a variance of 0.

[0046] In some embodiments, the convolution layer is used to perform damage edge detection and feature extraction on the computer tomography damage image, and the fully connected layer is used to highly abstract and integrate the features extracted by the convolution layer to output the recognition result of the final laminate damage pattern; Among them, there are 5 convolutional layers and 3 fully connected layers; the network structure of the convolutional neural network model also includes 3 pooling layers and 1 normalized exponential function layer.

[0047] In some embodiments, the number of channels of the first convolutional layer to the fifth convolutional layer of the convolutional neural network model are 16, 32, 64, 96 and 128 respectively.

[0048] In some embodiments, the learning rate of the convolutional neural network model adopts an adaptive adjustment strategy, with an initial learning rate of 1e-4 and a decay rate of 0.1. It changes every 10 epochs, so that the model can adapt to the distribution characteristics of each damage mode in the target domain more quickly.

[0049] In some embodiments, the step of training and testing the established convolutional neural network model using the computed tomography damage pattern image dataset comprises: T041, 80 samples of each of the three types of damage patterns in the standardized computed tomography single damage pattern image dataset were randomly selected to form the test set, and the remaining computed tomography single damage pattern images were randomly divided into training set and validation set in a ratio of 4:1; T042, use the training set and validation set to train the convolutional neural network model, and adjust the hyperparameter settings of the network model during the training process; T043, after model training is completed, the test set is used to test the damage pattern recognition performance of the convolutional neural network model on computed tomography images.

[0050] Among them, the training set is used to estimate the model parameters, the validation set is used to adjust the model hyperparameters and preliminarily evaluate the model's ability to recognize different damage patterns, and the test set is used to finally evaluate the performance of the model to ensure that the model can generalize to unseen data.

[0051] In some embodiments, for the trained convolutional neural network model, a class activation map is introduced to reweight the weight of the last convolutional layer of the convolutional neural network model, visualize the importance of each input feature, visualize the decision process of the proposed convolutional neural network model, and explain the reasons behind the model decision. Visualizing the importance of each input feature These techniques improve the interpretability of the model, allowing us to understand the decision process and important features of the model.

[0052] In some embodiments, the class activation map includes at least one of Grad-CAM, Grad-CAM++, Score-CAM, XGrad-CAM, Ablation-CAM, EigenGrad-CAM, and LayerCAM.

[0053] In some embodiments, for the trained convolutional neural network model, Shapley additive explanatory value analysis is introduced to calculate the contribution of different damage pattern features to the prediction results of the convolutional neural network model, provide an explanatory score for each damage pattern feature, and reveal the impact of different damage patterns on the model output.

[0054] In some embodiments, for the trained convolutional neural network model, the three damage mode images of matrix cracking, delamination, and fiber breakage in the test set are randomly assigned to rotate 90°, 180°, and 270° in a clockwise direction, respectively, and the class activation map and Shapley additive interpretation value analysis are used to visualize the model's recognition and prediction results of the three damage mode features after rotation. The image is rotated (90°, 180°, 270°) to test whether the model can still accurately identify the damage mode in different directions. Combined with the class activation map and Shapley interpretation value analysis, the model's recognition ability of the rotated image is further visualized.

[0055] In some embodiments, the hyperparameters and weight values ​​of each layer of the optimal convolutional neural network model are stored, and the network model is called to identify a computed tomography multiple damage pattern data set; the model is applied to actual engineering to identify different damage modes of carbon fiber reinforced composite laminates under low-speed impact.

[0056] In some embodiments, multiple damage pattern images are efficiently decoupled and quantitatively analyzed, and interpretable tools such as class activation maps and Shapley additive explanatory values ​​are introduced for transparent interpretation, revealing the potential variables and mechanisms that affect the model prediction results, and visualizing the contribution of each type of damage pattern to the final prediction results; confidence values ​​are introduced to quantitatively analyze the prediction results.

[0057] The following describes in detail a method for identifying damage patterns of carbon fiber reinforced composite materials using an interpretable convolutional neural network through several specific embodiments.

[0058] Embodiment 1: This embodiment provides a method for identifying damage modes of carbon fiber reinforced composite materials using an interpretable convolutional neural network, comprising the following steps: P1, making carbon fiber reinforced composite laminate samples with the same process parameters: the polymer matrix used in the carbon fiber reinforced composite laminate is a thermosetting epoxy vinyl resin, the resin is cured with a curing agent at room temperature, and the mass ratio of the accelerator and the hardener is 100:2:0.2. The accelerator and the hardener are dimethylaniline and methyl ketone peroxide (MEKP), which are purchased from Dow Chemical Company, USA. The reinforcement material is a polyacrylic acid (PAN) derived carbon fiber plain weave cloth (surface mass density of 300g / m2), the single fiber diameter is 10μm, and the gap between the warp yarn and the filling fiber yarn in the cloth is 0.2mm. The carbon fiber fabric is purchased from Xi'an Xiangda Aviation Materials Co., Ltd., China. The conductive silver glue is purchased from Huiteng Company, China, which is mainly used to connect the electrode sheets and wires in the carbon fiber reinforced composite laminate. The colloid is cured after heating at 100℃ for 2h.

[0059] P2, low-speed impact test on the prepared carbon fiber reinforced composite laminate: The carbon fiber reinforced composite laminate is prepared by vacuum assisted resin infiltration (VARI), the length and width of the laminate are 500×500mm, the thickness is 2mm, 4mm, 6mm, 7mm respectively, and the carbon fiber reinforced composite laminate is cut into 100(W)×100(L)mm by water jet cutting process. The prepared carbon fiber composite laminate is subjected to low-speed impact test. The impact test is carried out on a computer-driven Instron / 9250HV low-speed drop hammer tester. The hammer mass is 2~70kg, the impact energy is 0.59~1800J, the position measurement accuracy is ±0.02mm, and the speed measurement accuracy is ±0.25%. The incident kinetic energy in the impact test is set to 50J and 80J according to Newton's law.

[0060] P3, obtain and organize the low-speed impact damage dataset of carbon fiber reinforced composite laminate specimens: all laminate specimens damaged after the impact test were observed by computer tomography (CT) scanning, and the scanning direction was along the thickness and width of the specimens. The equipment used was the YXLON FF35 equipment. The laminate impact damage image dataset was established and organized by computer tomography (CT) scanning, and the image data size was adjusted to 227×227. The mean and standard deviation of the image data were adjusted to achieve standardization. The mean vector corresponding to the mean of the three RGB channels in the computer tomography (CT) scan image was (0.485, 0.456, 0.406), and the standard deviation vector corresponding to the standard deviation of the three RGB channels was (0.229, 0.224, 0.225). The damage modes of carbon fiber reinforced composite laminate specimens are mainly three types: matrix cracking, delamination and fiber fracture. According to the distribution of damage modes on the image data, the computed tomography (CT) scanning image dataset is divided into a single damage mode dataset and a multiple damage mode dataset. The image data in the single damage mode dataset contains only one single damage mode (matrix cracking, delamination and fiber fracture), and the image data in the multiple damage mode dataset contains two or more damage modes.

[0061] P4, define the interpretable convolutional neural network model: construct a convolutional neural network model for damage pattern recognition of carbon fiber reinforced composites (CFRPs) laminates. The model consists of two parts: convolutional block layer and fully connected layer. The convolutional block layer performs damage edge detection and feature extraction on the computed tomography (CT) scan image. The fully connected layer highly abstracts and integrates the features extracted by the convolutional layer and outputs the final laminate damage pattern recognition result. The network structure of the convolutional neural network model includes 5 convolutional layers, 3 pooling layers, 3 fully connected layers and 1 Softmax layer. The activation function uses the ReLU activation function, and the dropout method is used to control the model complexity of the fully connected layer. The number of channels from the first convolutional layer to the fifth convolutional layer is 16, 32, 64, 96 and 128 respectively. The learning rate of the convolutional neural network model adopts an adaptive adjustment strategy. The initial learning rate is 1e-4, and the learning rate is reduced to 10% every 10 epochs.

[0062] P5, training and testing of interpretable convolutional neural network model: The single damage pattern dataset of CT scan images of CFRPs damaged specimens is divided into training set, validation set and test set. The test set is composed of 80 CT scan images of each of the three types of damage (matrix cracking, delamination and fiber fracture) in the single damage pattern dataset. The training set and validation set are composed of the remaining image data randomly divided in 8:2. The multi-damage pattern dataset of CT scan images of CFRPs damaged specimens is used to verify the robustness of the convolutional neural network model. Labels are added to the image data of different damage patterns, where matrix cracking damage is marked as 0, delamination damage is marked as 1, and fiber fracture damage is marked as 2. The convolutional neural network model is trained and validated using the single damage pattern training set and validation set. The hyperparameter settings of the network model are adjusted during the training process. After the model training is completed, the damage pattern recognition accuracy of the model for CT scan images is tested using the test set.

[0063] P6, introduces class activation map (CAM) technology and Shapley additive explanation (SHAP) value interpretable tools to transparently explain the convolutional neural network model: For the trained convolutional neural network model, the class activation map (CAM) technology is introduced to reweight the weight of the last convolutional layer of the convolutional neural network model, visualize the importance of each input feature, and visualize the decision process of the proposed convolutional neural network model. The introduced class activation map (CAM) technologies include Grad-CAM, Grad-CAM++, Score-CAM, XGrad-CAM, Ablation-CAM, EigenGrad-CAM, and LayerCAM. For the trained convolutional neural network model, the Shapley additive explanation (SHAP) value analysis is introduced to calculate the contribution of different damage pattern features to the prediction results of the convolutional neural network model, provide an explanatory score for each damage pattern feature, and reveal the impact of different damage patterns on the model output. For the trained convolutional neural network model, three damage mode images of matrix cracking, delamination and fiber fracture were randomly selected from the test set and rotated 90°, 180° and 270° clockwise respectively. The class activation map (CAM) technology and Shapley additive interpretation (SHAP) value were used to analyze the recognition and prediction results of the three damage mode features after rotation by the visualization model.

[0064] P7, generalization application of convolutional neural network model: store the hyperparameters and weight values ​​of each layer of the trained interpretable convolutional neural network model, call the generalization of the network model to identify multi-damage pattern data sets of computed tomography (CT) scans, and efficiently decouple and quantitatively analyze multi-damage pattern images. The recognition accuracy of multi-damage pattern data sets can reach 92.99%. Introduce class activation map (CAM) technology and interpretable tools of Shapley additive interpretation (SHAP) values ​​for transparent interpretation, reveal the potential variables and mechanisms that affect the model prediction results, and visualize the contribution of each type of damage pattern to the final prediction results. At the same time, introduce confidence values ​​to quantitatively analyze the prediction results.

[0065] Experimental Results 1. Figure 4 The present invention proposes an example of an interpretable convolutional neural network model for the output results of a single injury pattern and a multi-injury pattern data set of a computed tomography (CT) scan. Figure 4 (a) is the loss function and accuracy function of the model for single damage pattern recognition during the training process. It can be seen that the damage function decreases exponentially with the training process of the network, and finally decreases to 0, and the accuracy increases exponentially with the training process of the network, and finally increases to 100%. Figure 4 (b) is the confusion matrix of the damage pattern recognition output results of the convolutional neural network model for the test set of single damage and multiple damage patterns. It can explain that the convolutional neural network has a recognition accuracy of 100% for single damage patterns and 92.99% for multiple damage patterns, realizing efficient decoupling and precise quantification of multiple damage patterns under service conditions.

[0066] 2. Figure 5 The schematic diagram of the convolutional neural network model is used to explain the Class Activation Map (CAM) technology of the present invention. The CAM technology provides a useful local evaluation index for CFRPs damage pattern recognition. Figure 5 The visualization results of different CAM techniques under single damage pattern image data are shown. It can be seen that different damage patterns have different visualization effects, but the highlights remain consistent and focus on the damaged area. The highlighted area is the focus area of ​​the network model, which explains the decision-making process of the convolutional neural network model. Due to the rectified linear unit activation function, the CAM heat map can only produce positive values.

[0067] 3. Figure 6This is a schematic diagram of the Shapley additive interpretation (SHAP) value analysis and interpretation of the convolutional neural network model for single damage pattern recognition results of the present invention example. From left to right, it shows the original damage pattern image, the SHAP value for predicting the matrix cracking damage pattern, the SHAP value for predicting the delamination damage pattern, and the SHAP value for predicting the fiber rupture damage pattern. For a given label and a single damage pattern test sample, a positive SHAP value in red indicates that the feature contributes positively to the output label, and a negative SHAP value in blue indicates that the feature contributes negatively to the output label. Figure 6 It can be observed that the red (positive SHAP value) in the diagonal image is more concentrated. The red area in the prediction results of the matrix cracking damage mode is concentrated in the first image, the red area in the prediction results of the delamination damage mode is concentrated in the second image, and the red area in the prediction results of the fiber rupture damage mode is concentrated in the third image. That is, the classification result of the convolutional neural network is consistent with the true label. The SHAP value analysis provides an explanatory score for each damage pattern feature. The pattern recognition result is correct, revealing the impact of different damage modes on the model output.

[0068] In summary, the beneficial effects of the present invention are: 1. The constructed convolutional neural network model comprehensively considers the importance of convolutional layers and channels. It contains 5 convolutional layers and has a simple and lightweight structure. The network model uses a larger number of channels in the early stage, which can process more input data information at the same time, extract more features of different damage modes of composite materials, and improve the accuracy in damage pattern recognition tasks. The network model uses a smaller number of channels in the later stage and introduces the dropout method in the fully connected layer to effectively prevent the convolutional neural network model from overfitting.

[0069] 2. The class activation map (CAM) technology and the interpretable tool Shapley additive interpretation (SHAP) value are introduced to provide a transparent interpretation of the convolutional neural network. CAM provides a local interpretation of the convolutional neural network, which well illustrates the reliability of the convolutional neural network model in black-box prediction of the region of interest of different damage patterns. SHAP calculates the marginal contribution of the characteristics of different damage patterns to the model prediction results, reveals the potential variables and mechanisms that affect the model prediction results, and truly establishes trust between people and models.

[0070] 3. The convolutional neural network model is trained and tested using a single damage dataset from computed tomography (CT) scans. The hyperparameters and weight values ​​of each layer of the convolutional neural network model after training are stored. This model can be applied to multiple damage pattern datasets and other damage pattern recognition tasks of carbon fiber reinforced composite materials under impact loads, and has strong robustness.

[0071] The above description is only a partial implementation of the embodiments of the present invention and does not impose any form of limitation on the application. The protection scope of the embodiments of the present invention is not limited to this. Any simple modifications, equivalent changes and modifications that can be easily thought of by any technician familiar with the technical field within the technical scope disclosed in the embodiments of the present invention should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A method for identifying damage patterns of carbon fiber reinforced composite materials using an interpretable convolutional neural network, characterized in that: The following steps are involved: Database establishment: collect and organize CT images of carbon fiber composite laminates damaged under low-speed impact loads, and establish CT damage pattern image data set; Data preprocessing, resizing the image size to a preset size, adjusting the mean and standard deviation of the computed tomography damage pattern image dataset for standardization; Constructing a convolutional neural network model, wherein the convolutional neural network model includes a convolutional layer and a fully connected layer; Train and test the convolutional neural network model, use the computed tomography damage pattern image dataset to train and test the established convolutional neural network model, and perform classification, recognition and quantitative analysis on the carbon fiber composite material impact damage pattern image dataset; Explain the convolutional neural network model, introduce class activation maps and Shapley additive interpretation value interpretable tools to transparently explain the convolutional neural network model, reveal the potential variables and mechanisms that affect the model's prediction results, visualize the contribution of each type of damage pattern to the final prediction results, and perform rotational validation on the computed tomography damage image dataset; Generalization application of the convolutional neural network model,After the model training is completed, the model is generalized and applied to the identification of multiple damage pattern images and composite material impact damage patterns in actual engineering to verify the robustness of the model.

2. The method according to claim 1, characterized in that The method further comprises: Composite material preparation steps: - Using carbon fiber or carbon fiber fabric as reinforcement and resin as matrix; -Preparing carbon fiber composite laminates by lamination process; Sample preparation steps: -Select carbon fiber composite laminate as the sample plate; -Manufacture multiple groups of samples according to predetermined specifications, where the shapes and widths of different groups of samples are consistent but the thicknesses are different; - Prepare two identical parallel samples for each set of specimens; Impact test steps: -Choose between two shock sources with different energy levels; - Two parallel samples of each group of specimens are placed under impact sources of different energies for impact tests; Data processing steps: - Use computer tomography to inspect the specimens after impact testing; - Obtain computer tomography damage image data of the damaged area under different impact energy loads.

3. The method according to claim 2, characterized in that The damage modes of carbon fiber composite laminates under load impacts of different energy levels include at least one of matrix cracking, delamination or fiber breakage.

4. The method according to claim 3, characterized in that The computed tomography damage image dataset is divided into a single damage pattern dataset and a multiple damage pattern dataset. The image data in the single damage pattern dataset contains only one damage pattern, while the image data in the multiple damage pattern dataset contains two or more damage patterns.

5. The method according to claim 1, characterized in that The step of adjusting the mean and standard deviation of the computed tomography damage pattern image data set for standardization also includes: Each computed tomography damage pattern image data was transformed into a normal distribution with a mean of 1 and a variance of 0.

6. The method according to claim 1, characterized in that The convolution layer is used to perform damage edge detection and feature extraction on the computer tomography damage image, and the fully connected layer is used to highly abstract and integrate the features extracted by the convolution layer to output the final recognition result of the laminate damage pattern; Among them, there are 5 convolutional layers and 3 fully connected layers; the network structure of the convolutional neural network model also includes 3 pooling layers and 1 normalized exponential function layer.

7. The method according to claim 6, characterized in that The number of channels from the first convolutional layer to the fifth convolutional layer of the convolutional neural network model are 16, 32, 64, 96, and 128, respectively.

8. The method according to claim 6, characterized in that The learning rate of the convolutional neural network model adopts an adaptive adjustment strategy. The initial learning rate is 1e-4, the decay rate is 0.1, and it changes every 10 epochs. The model adapts to the distribution characteristics of various damage modes in the target domain more quickly.

9. The method according to claim 4, characterized in that The steps of training and testing the established convolutional neural network model using the computed tomography damage pattern image data set include: randomly selecting 80 samples from each of the three types of damage patterns in the standardized computed tomography single damage pattern image data set to form a test set, and randomly dividing the remaining computed tomography single damage pattern images into a training set and a validation set at a ratio of 4:1; using the training set and the validation set to train the convolutional neural network model, and adjusting the hyperparameter settings of the network model during the training process; after the model training is completed, using the test set to test the damage pattern recognition performance of the convolutional neural network model on the computed tomography image.

10. The method according to claim 1, characterized in that For the trained convolutional neural network model, the class activation map is introduced to reweight the weight of the last convolutional layer of the convolutional neural network model, visualize the importance of each input feature, visualize the decision process of the proposed convolutional neural network model, and explain the reasons behind the model decision.

11. The method according to claim 10, characterized in that The class activation map includes at least one of Grad-CAM, Grad-CAM++, Score-CAM, XGrad-CAM, Ablation-CAM, EigenGrad-CAM, and LayerCAM.

12. The method according to claim 1, characterized in that For the trained convolutional neural network model, Shapley additive explanatory value analysis is introduced to calculate the contribution of different damage pattern characteristics to the prediction results of the convolutional neural network model, provide an explanatory score for each damage pattern feature, and reveal the impact of different damage patterns on the model output.

13. The method according to claim 1, characterized in that For the trained convolutional neural network model, the three damage mode images of matrix cracking, delamination and fiber fracture in the test set were randomly assigned to rotate 90°, 180° and 270° clockwise respectively. The class activation map and Shapley additive explanatory value were used to analyze the recognition and prediction results of the three damage mode features after rotation by the visualization model.

14. The method according to claim 1, wherein: The hyperparameters and weight values ​​of each layer of the optimal convolutional neural network model are stored, and the network model is called to identify the multi-damage pattern data set of computed tomography; the model is applied to actual engineering to identify different damage modes of carbon fiber reinforced composite laminates under low-speed impact.

15. The method according to claim 14, characterized in that Efficiently decouple and quantitatively analyze images with multiple damage patterns, introduce interpretable tools such as class activation maps and Shapley additive explanatory values ​​for transparent interpretation, reveal the potential variables and mechanisms that affect the model prediction results, and visualize the contribution of each type of damage pattern to the final prediction results; introduce confidence values ​​for quantitative analysis of the prediction results.

Images