A composite material damage diagnosis method integrating attention mechanism and deep learning

By combining acoustic emission technology with convolutional neural networks, integrating attention mechanisms and deep learning, the time-consuming and labor-intensive problems of traditional non-destructive testing are solved, and efficient and accurate diagnosis of composite material damage is achieved.

CN120445872BActive Publication Date: 2025-09-12CHINA STATE SHIPBUILDING CORP LTD RESEARCH INSTITUTE 719
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510947000.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-07-10
Publication Date
2025-09-12
Estimated Expiration
2045-07-10

AI Technical Summary

Technical Problem

Traditional nondestructive testing of composite material damage requires manual intervention, which is time-consuming and labor-intensive, and it is difficult to achieve high-precision real-time diagnosis.

Method used

Combining acoustic emission technology with convolutional neural networks, by integrating attention mechanism and deep learning, the dual-spectrum characteristics of acoustic emission signals are used to diagnose composite material damage, and a damage diagnosis model is constructed and trained.

Benefits of technology

It achieves efficient and accurate diagnosis of composite material damage, and improves detection efficiency and the accuracy of diagnostic results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120445872B_ABST
    Figure CN120445872B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of composite material damage diagnosis, and discloses a composite material damage diagnosis method that integrates an attention mechanism and deep learning, comprising: step 1: preparing a test sample; step 2: performing an impact test; step 3: obtaining a contour map of an acoustic emission signal; step 4: marking the damage of the composite material under different impact energies; step 5: constructing and training a damage diagnosis model; and step 6: evaluating the damage diagnosis model. By providing steps 3, 4, and 5, it is advantageous to combine the acoustic emission signal with a convolutional neural network, obtain composite material impact damage information according to the signal bispectral characteristics, improve detection efficiency, and make the diagnosis result more accurate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of composite material damage diagnosis, and more specifically to a composite material damage diagnosis method that integrates attention mechanism and deep learning. Background Art

[0002] Carbon fiber reinforced composites (CFRPs) are increasingly being used in aerospace, marine, and automotive industries due to their significant strength-to-weight ratio, corrosion resistance, and design flexibility. Impact loads are one of the most common external loads on composite structures during service. While they typically do not leave noticeable damage at the impact site, they may cause severe matrix cracking, delamination, and fiber failure within the structure. This can degrade the integrity of the composite structure and affect its normal service life. Therefore, developing a real-time, high-precision structural health monitoring (SHM) method is crucial for ensuring the proper operation of composite structures.

[0003] However, traditional nondestructive testing methods mostly require manual intervention, rely on experience, and are time-consuming and labor-intensive. Currently, artificial intelligence (AI) technology is rapidly developing, and deep learning, at its core, has demonstrated strong application potential across various fields. Convolutional neural networks (CNNs), a deep learning framework, have demonstrated remarkable capabilities in image processing and pattern recognition. Integrating this technology into composite material damage diagnosis, combined with acoustic emission technology, has shown great feasibility. Acoustic emission, a highly sensitive nondestructive testing method, captures the acoustic wave signals generated by composite materials during stress or damage. Convolutional neural networks (CNNs) automatically extract the features of these complex signals and, through trained models, accurately identify and classify the type and extent of damage in composite materials. Therefore, combining CNNs with acoustic emission technology can provide an efficient and accurate intelligent solution for composite material damage diagnosis. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a composite material damage diagnosis method that integrates attention mechanism and deep learning. By combining acoustic emission signals with convolutional neural networks, composite material impact damage information is obtained according to the signal bispectral characteristics, thereby improving detection efficiency and making the diagnosis results more accurate, so as to solve the problems existing in the above-mentioned background technology.

[0005] The present invention provides the following technical solution: a composite material damage diagnosis method integrating attention mechanism and deep learning, comprising the following steps:

[0006] Step 1: Prepare the test sample;

[0007] Step 2: Perform an impact test on the test sample and use the acoustic emission signal during the impact test;

[0008] Step 3: Perform bispectral analysis on the collected acoustic emission signal to obtain the contour map of the acoustic emission signal and use it as the input image of the damage diagnosis model;

[0009] Step 4: Combine and analyze the acoustic emission signal, the mechanical response curve during the impact test, and the impact surface damage of the test sample. Label the damage of the composite material under different impact energies and use it as the output label of the damage diagnosis model.

[0010] Step 5: Perform image data enhancement processing on the contour map of the acoustic emission signal in step 3; construct a damage diagnosis model and train it, and input the data into the constructed damage diagnosis model to obtain the output result;

[0011] Step 6: Use recall, precision, and F1 score to evaluate the performance of the damage diagnosis model for acoustic emission damage diagnosis of test samples.

[0012] Preferably, the test sample is a carbon fiber high-performance composite material sample subjected to a standard impact test; the impact test is a drop hammer impact test; the impact test may use different impact energies to obtain different acoustic emission signals; and the output labels include three categories: slight damage, moderate damage, and severe damage.

[0013] Preferably, the image data enhancement processing includes adding noise, image translation, adjusting image brightness and scaling; the output result includes ; For minor injuries, For moderate damage, Corresponding to severe damage; the contour map of the acoustic emission signal used to train the damage diagnosis model and the label corresponding to the damage diagnosis result are used as training data of the damage diagnosis model for training.

[0014] Preferably, in step 3, the bispectral analysis of the collected acoustic emission signal is performed using bispectral transformation, and the bispectral transformation of the acoustic emission signal is performed by express:

[0015] ;

[0016] in, is the third-order cumulant of the three-dimensional vector X, is a parameter vector with the same dimensions as vector X, and The parameter introduced by Fourier transform, j is the imaginary unit (j 2 =-1), its high-order cumulants It is absolutely possible to find the sum.

[0017] Preferably, in step 5, the damage diagnosis model is a deep learning neural network, specifically a convolutional neural network model integrating an attention mechanism, which mainly includes:

[0018] Input layer, feature extraction module, attention mechanism module, prediction module and output layer;

[0019] The input layer is used to input the input image in step 3, that is, the contour map of the acoustic emission signal;

[0020] The output layer is used to output the results of acoustic emission damage diagnosis of the test sample.

[0021] Preferably, the feature extraction module is composed of six convolution modules; the convolution module is composed of a convolution layer, a normalization layer, a ReLu function activation layer and a maximum pooling layer; the formula is as follows:

[0022] ;

[0023] in, It is the first layer after the convolution operation. The output of the layer, It is The convolution kernel of the layer, It is The output of the layer, It is the convolutional layer Layer bias;

[0024] The normalization layer normalizes the input or output of each layer. The formula is as follows:

[0025] ;

[0026] in, is the normalized input, represents the input features of the i-th layer, and are the mean and variance of the training batch, A small constant is introduced to avoid the denominator being zero;

[0027] The ReLu function activation layer introduces nonlinearity, and the formula is as follows:

[0028] .

[0029] Preferably, the attention mechanism module is composed of a convolutional module attention mechanism layer; the input first passes through the channel attention module, which is expressed by the formula:

[0030] ;

[0031] ;

[0032] Among them, F represents the feature map input into the attention mechanism layer, represents the output of the channel attention mechanism, represents the ReLu activation function, Represents the calculation of element-by-element multiplication;

[0033] The calculation results are then input into the spatial attention module, and its formula is expressed as:

[0034] ;

[0035] ;

[0036] in, represents the spatial attention mechanism, Denotes the output of the spatial attention mechanism, and Conv denotes the convolution operation.

[0037] Preferably, the prediction module consists of a flattening layer, two fully connected layers and a softmax layer; the cross entropy function is used to describe the loss function during the training process, and the formula is expressed as:

[0038] ;

[0039] in, is the true label corresponding to the training data, Represents the label predicted by the neural network.

[0040] The technical effects and advantages of the present invention are as follows:

[0041] The present invention is provided with steps 3, 4 and 5, which is conducive to performing bispectral analysis on the collected acoustic emission signals to obtain a contour map of the acoustic emission signals, combining and analyzing the acoustic emission signals, the mechanical response curve during the impact test, and the impact surface damage of the test sample, marking the damage conditions of the composite material under different impact energies, and then constructing a damage diagnosis model and training it; by combining the acoustic emission signals with a convolutional neural network, the impact damage information of the composite material is obtained according to the bispectral characteristics of the signals, thereby improving the detection efficiency and making the diagnosis result more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 This is a flow chart of the composite material damage diagnosis method that integrates attention mechanism and deep learning.

[0043] Figure 2 This is an overall flow chart of an example of the damage diagnosis model of the present invention.

[0044] Figure 3Schematic diagram of the mechanical response curve of the impact process and the surface damage of the specimen under different impact energies.

[0045] Figure 4 Schematic diagram of image data enhancement processing method.

[0046] Figure 5 Schematic diagram of the damage diagnosis model.

[0047] Figure 6 Schematic diagram of the attention mechanism structure of the convolution module.

[0048] Figure 7 A diagram showing the accuracy, loss function, and confusion matrix of the model training process and the prediction results on the test set. DETAILED DESCRIPTION

[0049] The technical solutions of the present invention will be described clearly and completely below in conjunction with the drawings in the present invention. In addition, the forms of the various structures described in the following embodiments are merely examples. The composite material damage diagnosis method that integrates attention mechanism and deep learning involved in the present invention is not limited to the various structures described in the following embodiments. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0050] like Figure 1-7 As shown, the present invention provides a composite material damage diagnosis method integrating attention mechanism and deep learning, comprising the following steps:

[0051] Step 1: Prepare a test sample; the test sample is a carbon fiber high-performance composite material sample for standard impact testing, which can be prepared using a domestic fiber placement machine; the high-performance composite material sample uses T700 / 7901 high-performance prepreg, which is automatically laid by a domestic fiber placement machine to produce a 330mm×330mm orthogonal ply [0 / 90] 4s laminate, which is compacted and cured in an autoclave to produce a high-performance composite material board, which is cut into a standard size specimen sample of 150mm×100mm by water jet cutting; in this embodiment, the test sample size is 150mm long, 100mm wide, and 4mm thick, and the ply is orthogonal ply [0 / 90] 4s;

[0052] Step 2: Perform an impact test on the test sample and use the acoustic emission signal during the impact test; the impact test is a drop hammer impact test; the impact test may use different impact energies to obtain different acoustic emission signals, and the different impact energies are achieved by adjusting the initial height of the drop hammer; the impact energy range is 15J-45J, and the acoustic emission signal collector uses a 2 / 4 / 6 amplifier with a gain setting of 20dB; since the acoustic emission signal acquisition length during the impact process is basically in the range of 6-10ms, the acquisition frequency setting directly affects the number of discrete data points collected; if the acquisition frequency is set too high, the amount of collected data will be too large, increasing the computing time consumption; otherwise, the collected signal will be seriously distorted; therefore, in order to comprehensively consider computing efficiency and signal fidelity, the acquisition frequency is set to 10MHz in this embodiment; in this embodiment, the mass of the impact hammer punch of the impact test is 6.62kg, and the selected energies are 15J, 20J, 25J, 30J, 35J, 40J, and 45J. The acoustic emission signal collector uses a 2 / 4 / 6 amplifier with a gain setting of 20dB and an acquisition frequency of 10MHz;

[0053] Step 3: Perform bispectral analysis on the collected acoustic emission signal to obtain the contour map of the acoustic emission signal and use it as the input image of the damage diagnosis model;

[0054] Step 4: Combine and analyze the acoustic emission signal, the mechanical response curve during the impact test, and the impact surface damage of the test sample. Label the damage of the composite material under different impact energies and use it as the output label of the damage diagnosis model. The output label includes three categories: slight damage, moderate damage, and severe damage.

[0055] Step 5: Perform image data enhancement processing on the contour map of the acoustic emission signal in step 3; the image data enhancement processing includes but is not limited to adding noise, image translation, adjusting image brightness and scaling; construct a damage diagnosis model and train it to obtain output results ; For minor injuries, For moderate damage, Corresponding to severe injuries; using the cross entropy function to describe the loss function during model training;

[0056] Step 6: Use recall, precision, and F1 score to evaluate the performance of the damage diagnosis model for acoustic emission damage diagnosis of test samples.

[0057] In this embodiment, it should be specifically explained that in step 3, the bispectral analysis of the collected acoustic emission signal is performed using bispectral transformation, and the obtained three-dimensional bispectral image is reduced to a two-dimensional contour map. The bispectral transformation of the acoustic emission signal is performed by express:

[0058] ;

[0059] Its high-order cumulants It is absolutely possible to find the sum.

[0060] In this embodiment, it should be specifically noted that in step 4, when the impact energy is 15J and 20J, the mechanical response curve is relatively flat, with no sudden load drop. At this time, only a small amount of matrix cracking occurs on the specimen surface. At the impact energy of 25J and 30J, the mechanical response curve shows a significant sudden drop, and at this time, in addition to matrix cracking, fiber failure also occurs on the specimen surface. At the impact energy of 35J-45J, the mechanical response curve shows multiple sudden drops, corresponding to severe fiber breakage failure, and the fiber failure on the specimen surface also extends to a larger extent. Therefore, the specimens corresponding to the impact energy of 15J and 20J are classified as slightly damaged, the specimens corresponding to the impact energy of 25J and 30J are classified as moderately damaged, and the specimens corresponding to the impact energy of 35J-45J are classified as severely damaged.

[0061] In this embodiment, it should be specifically explained that the image data enhancement processing in step 5 is to avoid overfitting of the damage diagnosis model, increase the diversity of training data, and perform random transformations on the training data to improve the generalization ability of the model; the training data in this embodiment is a picture, so methods such as adding noise, image translation, adjusting image brightness, and scaling are adopted to perform data enhancement, and ultimately 2205 data are generated. Among them, there are 735 data of categories I, II, and III, and this is to ensure the balance of data categories. The main reason for ensuring that there are the same amount of data in each category is to avoid the model's bias towards the majority category, improve the model's performance on the minority category, improve the model's generalization ability, and avoid deviations in evaluation indicators;

[0062] When training the damage diagnosis model, the data is divided into training set, validation set and test set in a ratio of 7:1:2; the resolution of the input image data is 224×224.

[0063] In this embodiment, it should be specifically noted that in step 5, the damage diagnosis model is a deep learning neural network, specifically a convolutional neural network model integrating an attention mechanism, which mainly includes:

[0064] Input layer, feature extraction module, attention mechanism module, prediction module and output layer;

[0065] The input layer is used to input the input image in step 3, that is, the contour map of the acoustic emission signal;

[0066] The output layer is used to output the results of acoustic emission damage diagnosis of the test sample;

[0067] The feature extraction module consists of six convolution modules; the convolution module consists of a convolution layer, a normalization layer, a ReLu function activation layer, and a maximum pooling layer; the convolution layer is the core part of the convolutional neural network, which processes the input data through a series of learnable kernels. The formula is as follows:

[0068] ;

[0069] in, It is the first layer after the convolution operation. The output of the layer, It is The convolution kernel of the layer, It is The output of the layer, It is the convolutional layer Layer bias;

[0070] The normalization layer accelerates the training process and improves the stability of the model by standardizing the input or output of each layer. The formula is as follows:

[0071] ;

[0072] in, and are the mean and variance of the training batch, A small constant is introduced to avoid the denominator being zero;

[0073] The ReLu function activation layer helps break the symmetry of the network by introducing nonlinearity and allows the model to learn sparse representation. Its formula is as follows:

[0074] ;

[0075] Max pooling is a downsampling technique that reduces the spatial size of the feature map by selecting the maximum value within a certain size window.

[0076] In this embodiment, it should be specifically noted that the attention mechanism module is composed of a convolutional module attention mechanism layer; the convolutional module attention mechanism layer combines the channel and spatial attention mechanisms. Specifically, the input first passes through the channel attention module, which is expressed by the formula:

[0077] ;

[0078] ;

[0079] in, Represents the calculation of element-by-element multiplication;

[0080] The calculation results are then input into the spatial attention module, and its formula is expressed as:

[0081] ;

[0082] ;

[0083] The prediction module consists of a flattening layer, two fully connected layers, and a softmax layer; the batch size, number of training rounds, and learning rate are 64, 40, and 0.01, respectively; the cross entropy function is used to describe the loss function during training, and the formula is expressed as:

[0084] ;

[0085] in, is the true label corresponding to the training data, Represents the label predicted by the neural network.

[0086] In this embodiment, it should be specifically noted that the accuracy and loss function are used in step 6 to characterize the training of the deep learning model's predictive performance. Accuracy and loss function are two important indicators in the neural network training process. Accuracy and loss function can indirectly indicate whether the model has learned useful features from the data. Normally, as training progresses, the accuracy will gradually increase and the loss function will gradually decrease.

[0087] The confusion matrix provides an intuitive way to view the performance of the classification model. By comparing the actual category with the predicted category through the confusion matrix, we can understand the performance of the model in different categories. In order to comprehensively evaluate the damage diagnosis performance of the damage model, the recall rate (Recall), precision rate (Precision) and F1 score ( ) Evaluate the accuracy of model diagnosis. Recall measures the model's ability to predict correct positive examples. A high recall means that the model is able to predict most true positive samples well. Precision measures the model's accuracy in predicting the positive class. A high precision means that when the model says a sample is positive, this judgment is usually correct. The F1 score provides a single value that balances the relationship between precision and recall;

[0088] The formula is:

[0089] ;

[0090] ;

[0091] ;

[0092] in, represents the correctly predicted positive example, represents the incorrectly predicted positive examples, represents a counterexample that was incorrectly predicted.

[0093] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

[0094] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A composite material damage diagnosis method integrating attention mechanism and deep learning, characterized by: The following steps are involved: Step 1: Prepare the test sample; Step 2: Perform an impact test on the test sample and use the acoustic emission signal during the impact test; Step 3: Perform bispectral analysis on the collected acoustic emission signal to obtain the contour map of the acoustic emission signal and use it as the input image of the damage diagnosis model; Step 4: Combine and analyze the acoustic emission signal, the mechanical response curve during the impact test, and the impact surface damage of the test sample. Label the damage of the composite material under different impact energies and use it as the output label of the damage diagnosis model. Step 5: Perform image data enhancement processing on the contour map of the acoustic emission signal in step 3; construct a damage diagnosis model and train it, and input the data into the constructed damage diagnosis model to obtain the output result; Step 6: Use recall, precision, and F1 score to evaluate the performance of the damage diagnosis model for acoustic emission damage diagnosis of test samples; The damage diagnosis model is a deep learning neural network, specifically a convolutional neural network model that integrates an attention mechanism; The attention mechanism module in the damage diagnosis model is composed of a convolutional module attention mechanism layer; the input first passes through the channel attention module, which is expressed by the formula: ; ; Among them, F represents the feature map input into the attention mechanism layer, represents the output of the channel attention mechanism, represents the ReLu activation function, Represents the calculation of element-by-element multiplication; The calculation results are then input into the spatial attention module, and its formula is expressed as: ; ; in, represents the spatial attention mechanism, Denotes the output of the spatial attention mechanism, and Conv denotes the convolution operation.

2. The composite material damage diagnosis method integrating attention mechanism and deep learning according to claim 1 is characterized in that: The test sample is a carbon fiber high-performance composite material sample subjected to a standard impact test; the impact test is a drop hammer impact test; the impact test may use different impact energies to obtain different acoustic emission signals; and the output labels include three categories: slight damage, moderate damage, and severe damage.

3. The composite material damage diagnosis method integrating attention mechanism and deep learning according to claim 2 is characterized in that: The image data enhancement processing includes adding noise, image translation, adjusting image brightness and scaling; the output results include ; For minor injuries, For moderate damage, Corresponding to severe damage; the contour map of the acoustic emission signal used to train the damage diagnosis model and the label corresponding to the damage diagnosis result are used as training data of the damage diagnosis model for training.

4. The composite material damage diagnosis method integrating attention mechanism and deep learning according to claim 3 is characterized by: In step 3, the bispectral analysis of the collected acoustic emission signal is performed by bispectral transformation, and the bispectral transformation of the acoustic emission signal is performed by express: ; in, is the third-order cumulant of the three-dimensional vector X, is a parameter vector with the same dimensions as vector X, and The parameter introduced for Fourier transform, j is the imaginary unit, j 2 =-1, its high-order cumulant It is absolutely possible to find the sum.

5. The composite material damage diagnosis method integrating attention mechanism and deep learning according to claim 4 is characterized in that: In step 5, the damage diagnosis model mainly includes: Input layer, feature extraction module, attention mechanism module, prediction module and output layer; The input layer is used to input the input image in step 3, that is, the contour map of the acoustic emission signal; The output layer is used to output the results of acoustic emission damage diagnosis of the test sample.

6. The composite material damage diagnosis method integrating attention mechanism and deep learning according to claim 5 is characterized by: The feature extraction module consists of six convolution modules; the convolution module consists of a convolution layer, a normalization layer, a ReLu function activation layer and a maximum pooling layer; the formula is as follows: ; in, It is the first layer after the convolution operation. The output of the layer, It is The convolution kernel of the layer, It is The output of the layer, It is the convolutional layer Layer bias; The normalization layer normalizes the input or output of each layer. The formula is as follows: ; in, is the normalized input, represents the input features of the i-th layer, and are the mean and variance of the training batch, A small constant is introduced to avoid the denominator being zero; The ReLu function activation layer introduces nonlinearity, and the formula is as follows: , where X represents the input features.

7. The composite material damage diagnosis method integrating attention mechanism and deep learning according to claim 6, characterized in that: The prediction module consists of a flattening layer, two fully connected layers and a softmax layer. The cross entropy function is used to describe the loss function during the training process, and the formula is expressed as: ; in, is the true label corresponding to the training data, Represents the label predicted by the neural network.

Citation Information

Patent Citations

  • Bearing fault diagnosis method based on parallel multi-fusion deep learning network model

    CN119128690A

  • Acoustic-optical-electric synergistic welding material fatigue performance testing method and system

    CN119246289A