Composite board damage detection method based on physical-virtual heterogeneous graph neural network

Abstract

The application discloses a kind of based on physical-virtual heterogeneous graph neural network composite board damage detection method, the multiple regions obtained by dividing composite board are regarded as virtual damage node, the channel of ultrasonic transducer array deployed on the surface of composite board is regarded as physical measurement node, and physical-virtual heterogeneous graph is constructed;Composite board damage detection model is constructed, and the data feature extraction module of the data collected by ultrasonic transducer array carries out feature extraction, and the feature obtained is regarded as the initial feature vector of physical measurement node in heterogeneous graph, the initial feature vector of virtual damage node is set to zero vector, graph convolution is extracted by physical heterogeneous graph convolution extractor Feature, then by prediction module, the probability that corresponding area exists damage is obtained, and damage detection result is obtained.The application explicitly mines the physical geometry prior of sensor array by physical-virtual heterogeneous graph, realizes the feature aggregation of structure perception, to improve the damage detection performance of composite board.

Description

Technical Field

[0001] This invention belongs to the field of composite material technology, and more specifically, relates to a method for detecting damage in composite materials based on a physical-virtual heterogeneous graph neural network. Background Technology

[0002] Composite panels, with their advantages of being lightweight and having high rigidity, have become core load-bearing components in aerospace, transportation, and civil engineering. However, during long-term service, these materials are susceptible to impact, cyclic loading, and environmental erosion, resulting in highly concealed and cumulative internal micro-damage that severely weakens structural performance and can even lead to catastrophic failure. Therefore, developing real-time, precise structural health monitoring technologies is crucial for early damage warning, ensuring structural safety, and reducing maintenance costs.

[0003] Current monitoring methods largely rely on accurate modeling of material properties. However, the multimodal superposition, boundary reflections, and environmental noise within composite sandwich structures make the time-domain waveforms extremely complex, making it very difficult to accurately obtain the arrival time of damage-scattered waves. Furthermore, traditional methods lack the ability to adaptively learn from massive amounts of monitoring data, making it difficult to effectively process complex damage signals. In recent years, graph neural networks have been widely used for ultrasonic guided wave damage localization in plate-like composite materials. However, these networks often ignore the unstructured spatial topological relationships between sensor array data, resulting in significantly insufficient diagnostic performance in complex scenarios. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for damage detection of composite materials based on physical-virtual heterogeneous graph neural networks. This method uses physical-virtual heterogeneous graphs to explicitly mine the physical geometric priors of the sensor array, realizes feature aggregation for structural perception, and thus improves the damage detection performance of composite materials.

[0005] To achieve the above-mentioned objectives, the present invention provides a method for detecting damage in composite plates based on a physical-virtual heterogeneous graph neural network, comprising the following steps:

[0006] S1: Deploy an array of ultrasonic transducers on the surface of the composite material to divide the composite material into sections. The ultrasonic transducer array is divided into several regions, and each region is used as a virtual damage node. Each channel serves as a physical measurement node, thus constructing a physical-virtual heterogeneous graph. , where the set of nodes express A virtual damage node and A set consisting of physical measurement nodes, edge set ,in, Indicates the first One physical measurement node, Indicates the first A virtual damage node, Represents physical measurement nodes The corresponding main coverage area of ​​the sound beam , , ;

[0007] S2: For normal composite panels and several composite panels with damage in different areas, deploy ultrasonic transducer arrays using the same scheme as in step S1 and divide the area. Through the deployed... The channel ultrasonic transducer array acquires several array guided wave monitoring signal samples. , Represent the spatial dimension of the sample and label its damage category. ,in This is indicated as a normal composite board. Indicates the first Composite panels with damage in certain areas. Thus, the training sample set is obtained;

[0008] S3: Construct a damage detection model for composite plates based on a physically heterogeneous graph neural network, including a data feature extraction module, a physically heterogeneous graph convolutional extractor, and a prediction module, wherein:

[0009] The data feature extraction module is used to extract input array guided wave monitoring signal samples. Arrayed waveguide signal characteristics And send it to the physical heterogeneous graph convolution extractor;

[0010] Physical heterogeneous graph convolution extractor is used for physical-virtual heterogeneous graphs Perform graph convolution, where the initial feature vectors of the physical measurement nodes are based on the features of the array guided wave signals. The initial feature vector of the virtual damaged node is set to zero, and the feature matrix is ​​obtained through processing. And send it to the prediction module;

[0011] The prediction module is used to predict based on the feature matrix. generate A multidimensional prediction vector, where each element represents the probability that damage exists in the corresponding region;

[0012] S4: Use the training sample set obtained in step S2 to train the composite board damage detection model to obtain a trained composite board damage detection model.

[0013] S5: When it is necessary to perform damage detection on the composite board, the same scheme as in step S1 is used to deploy the ultrasonic transducer array and divide the area. The array guided wave monitoring signal samples are collected and input into the composite board damage detection model trained in step S4. The damage detection result is obtained according to the output prediction vector.

[0014] This invention presents a composite material damage detection method based on a physical-virtual heterogeneous graph neural network. An ultrasonic transducer array is deployed on the surface of the composite material, dividing the material into multiple regions. Each region is designated as a virtual damage node, and the channels of the ultrasonic transducer array are used as physical measurement nodes, thus constructing a physical-virtual heterogeneous graph. A composite material damage detection model is then constructed, comprising a data feature extraction module, a physical heterogeneous graph convolutional extractor, and a prediction module. The data feature extraction module extracts features from the data collected by the ultrasonic transducer array, using these features as the initial feature vectors of the physical measurement nodes in the heterogeneous graph. The initial feature vectors of the virtual damage nodes are set to zero. The physical heterogeneous graph convolutional extractor performs graph convolution to extract features, and the prediction module obtains the probability of damage in the corresponding region, thus yielding the damage detection result.

[0015] The present invention has the following beneficial effects:

[0016] 1) This invention achieves structure-aware feature aggregation by constructing a physical-virtual heterogeneous graph and combining it with graph convolution to explicitly mine the physical geometric priors of the sensor array. Experimental results show that the proposed model outperforms several mainstream diagnostic models in terms of localization accuracy, and its decision-making process exhibits excellent physical interpretability.

[0017] 2) Based on the frequency invariance of guided wave propagation, this invention combines a self-attention mechanism to perform joint attention modeling between and within channels to suppress noise signal interference, thereby further improving the accuracy of plate damage detection. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of ultrasonic guided wave propagation in a composite material.

[0019] Figure 2 This is a flowchart illustrating a specific implementation of the composite plate damage detection method based on a physical-virtual heterogeneous graph neural network according to the present invention.

[0020] Figure 3 This is a structural diagram of the composite plate damage detection model in this invention;

[0021] Figure 4 This is a schematic diagram of the CFRP board area division in this embodiment;

[0022] Figure 5 These are typical guided wave signals for different channels in this embodiment. Detailed Implementation

[0023] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0024] To better illustrate the technical solution of the present invention, the technical principles of the present invention will be briefly explained first.

[0025] During ultrasonic guided wave propagation, dispersion effects cause wave packet distortion. Although the frequencies of each frequency component in the excitation signal remain constant, the wavenumber varies with the material, making it difficult for the wave packet to maintain its original integrity during refraction. This phenomenon makes it difficult for traditional digital signal processing methods to effectively process the signal as a whole. However, frequency invariance provides crucial theoretical support for data-driven methods.

[0026] The propagation and damage interaction of ultrasonic guided waves in composite plates can be described by the wave equation. For an isotropic homogeneous elastic medium, the particle motion satisfies the Navier–Cauchy equation:

[0027] (1)

[0028] in, For time-varying displacement vector fields, and These are the Lamé constants of the materials, respectively. This represents the material density.

[0029] Figure 1 This is a schematic diagram of ultrasonic guided wave propagation in a composite board. These represent the horizontal direction of wave propagation and the depth direction perpendicular to the surface, respectively. Indicates time. (Regarding...) Figure 1 The waveguide propagation structure shown is assumed to be... The surface boundary conditions are set. Since the analysis object is a surface wave, they can be ignored. The influence of direction. The expression for the excitation source is as follows:

[0030] (2)

[0031] in, Represents scalar potential. Representing vector potential Quantity, This indicates that the scalar potential only varies with depth. Changing complex amplitude, Represents vector potential The complex amplitude whose component varies only with depth y. Represents the imaginary unit. Indicates wave number, Indicates the excitation angular frequency. This represents the traveling wave phase factor.

[0032] Through Helmholtz decomposition, the time-varying displacement vector field It can be represented as a scalar potential With vector potential Combinations:

[0033] (3)

[0034] Substitute equation (3) into equation (1) to solve, and consider along The frequency-wavenumber expression for the displacement vector field can be obtained from the solution of the directional propagation harmonic wave:

[0035] (4)

[0036] in, , They represent the particles along... direction and Displacement components in the direction.

[0037] This formula shows that, in the matrix material, a single frequency Under excitation, the displacement field maintains a simple harmonic wave form, and its modal structure and frequency components remain unchanged. However, in actual detection, broadband pulse signals are usually used as the excitation source, in which case the frequency components... Refraction and reflection may occur at the interlayer interface of composite materials, thereby exciting or converting multiple propagation modes, resulting in the received signal exhibiting multimode aliasing and dispersion characteristics:

[0038] (5)

[0039] in, Represents the offset vector. Indicates the modal order. Indicates the first First mode at frequency The amplitude distribution function at that location, Indicates the first Wavenumber of the first mode.

[0040] In practical applications, damage and noise are distributed across unknown frequency components, and dispersion further complicates the signal structure in the time and frequency domains. Therefore, a targeted feature extraction mechanism can be designed to better characterize the array guided wave signal features of plate damage. .

[0041] Arrayed waveguide signal characteristics It can be regarded as the intensity of each equivalent scattering source under the Born approximation framework. According to the Born approximation, under weak scattering conditions, the total scattered wave field can be approximated as a linear superposition of the equivalent scattered wave fields at each damage point, and its mathematical expression is:

[0042] (6)

[0043] in, Green's function describes the propagation law of a unit impulse excitation in a medium. Represents the spatial position vector of the observation receiving point. No. Spatial location vectors of discrete damaged nodes Indicates the number of damaged nodes. Indicates the dummy time of integration. Represents the first Equivalent source strength at each virtual damage node.

[0044] From a priori physical perspective, the existence of damage can be equivalent to introducing a virtual bodily force into the wave equation. The corrected governing equations are:

[0045] (7)

[0046] Based on the principle of wave field superposition, the total wave field It can be decomposed into incident waves With scattered waves sum:

[0047] (8)

[0048] To achieve the inverse mapping of the physical forward process shown in the above equation, the equivalent source strength is inverted from the received signal. This invention divides the composite board into multiple regions, using each region as a virtual damage node, thereby constructing a physical-virtual heterogeneous graph. Among them, the node set From physical measurement nodes With virtual damage nodes Together constitute, that is . Corresponding to the actual "excitation-receiver" probe pair, its initial characteristics are determined by the array guided wave signal characteristics. Assign a value; This represents the potential damage location, with its initial features set as a zero vector. Further features are then extracted using graph convolution. (Edge set of heterogeneous graph) It is defined strictly according to the physical principles of sound beam coverage: only when the virtual damage node Located at physical node The corresponding main coverage area of ​​the sound beam Only then is a connection established between the two. Based on this, edge sets... This can be formally represented as:

[0049] (9)

[0050] If the nodes are sorted with physical measurement nodes first and virtual damage nodes second, then the overall adjacency matrix of the nodes in the physical-virtual heterogeneous graph is... It has a block sparse structure:

[0051] (10)

[0052] Based on the above analysis, this invention proposes a method for detecting damage in composite plates based on a physical-virtual heterogeneous graph neural network. Figure 2 This is a flowchart illustrating a specific implementation of the composite material damage detection method based on a physical-virtual heterogeneous graph neural network according to the present invention. Figure 2 As shown, the composite plate damage detection method based on physical-virtual heterogeneous graph neural network of the present invention includes the following steps:

[0053] S201: Constructing a Physical-Virtual Heterogeneous Graph:

[0054] An array of ultrasonic transducers is deployed on the surface of the composite material to divide the composite material into sections. The ultrasonic transducer array is divided into several regions, and each region is used as a virtual damage node. Each channel serves as a physical measurement node, thus constructing a physical-virtual heterogeneous graph. , where the set of nodes express A virtual damage node and A set consisting of physical measurement nodes, edge set ,in, Indicates the first One physical measurement node, Indicates the first A virtual damage node, Represents physical measurement nodes The corresponding main coverage area of ​​the sound beam , , .

[0055] In this embodiment, the main coverage area of ​​the sound beam The generation method is as follows: for physical measurement nodes For the corresponding ultrasonic transducer channel, a straight line segment between the transmitting and receiving probes is first generated based on the position of the corresponding probe pair as the main acoustic axis. Then, with the main axis The vertical distance is not greater than the preset beamwidth. The area is considered as the effective detection range of the channel. During the determination, if the virtual damaged node... The corresponding area and the main coverage area of ​​the sound beam Intersecting, then virtual damaged nodes Belongs to the main coverage area of ​​the sound beam .

[0056] S202: Obtain the training sample set:

[0057] For normal composite boards and For composite panels with damage in different areas, an ultrasonic transducer array is deployed using the same scheme as in step S1, and the areas are divided. The deployed... The channel ultrasonic transducer array acquires several array guided wave monitoring signal samples. , Represent the spatial dimension of the sample and label its damage category. ,in This is indicated as a normal composite board. Indicates the first Composite panels with damage in certain areas. Thus, the training sample set is obtained.

[0058] In practical applications, preprocessing operations such as outlier removal and synchronization alignment can be performed on each array waveguide monitoring signal sample to form a standardized signal sample that meets the input requirements of subsequent models.

[0059] S203: Constructing a damage detection model for composite panels:

[0060] In order to fully integrate the spatiotemporal characteristics of multi-source monitoring information and significantly improve the accuracy and generalization ability of ultrasonic guided wave plate damage localization, this invention constructs a composite plate damage detection model based on physical-virtual heterogeneous graph neural network. Figure 3 This is a structural diagram of the composite plate damage detection model in this invention. (See diagram below.) Figure 3 As shown, the composite plate damage detection model in this invention includes a data feature extraction module, a physical heterogeneous graph convolution extractor, and a prediction module. Each module will be described in detail below.

[0061] The data feature extraction module is used to extract input array guided wave monitoring signal samples. Arrayed waveguide signal characteristics And send it to the physical heterogeneous graph convolution extractor.

[0062] To better extract data features, the data feature extraction module in this embodiment employs a multi-granularity self-attention mechanism that focuses on physical information between and within waveguide channels. This mechanism models signal correlation patterns at different scales from a multi-granularity perspective. In modeling the importance of inter-channel signals, the relationship between channels and damage localization needs to be emphasized; therefore, a self-attention mechanism is used to directly analyze the time-domain array waveguide monitoring signal samples. Feature extraction is performed. Meanwhile, in terms of channel importance modeling, according to equation (5), frequency invariance can be used to highlight specific frequency bands. For this purpose, this embodiment uses discrete wavelet transform to map the signal to the time-frequency domain to extract local oscillation frequency domain features, and then uses a self-attention mechanism to extract frequency domain features.

[0063] like Figure 3 As shown, the data feature extraction module in this embodiment includes a Discrete Wavelet Transform (DWT) module, a first self-attention module, an Inverse Discrete Wavelet Transform (IDWT) module, a second self-attention module, and a feature fusion module, wherein:

[0064] The discrete wavelet transform module is used to process the input array guided wave monitoring signal samples. Perform a discrete wavelet transform to obtain the coefficient matrix. ,in Representing the discrete wavelet transform matrix, the coefficient matrix is... Send to the first self-attention module.

[0065] In this embodiment, the discrete wavelet transform uses the Haar wavelet basis, whose scaling function is... With wavelet function The definition is as follows:

[0066] (9)

[0067] (10)

[0068] For single-channel discrete signals Perform single-level decomposition to obtain approximate coefficients. With detail coefficient :

[0069] (11)

[0070] (12)

[0071] The constructed discrete wavelet transform matrix Each row of the coefficient matrix consists of discrete samples of the scaling function and the wavelet function. The former Listed as approximation coefficients, then Listed as detail coefficients.

[0072] The first self-attention module is used to apply a self-attention mechanism to the coefficient matrix. After processing, the enhancement coefficient matrix is ​​obtained. And send it to the inverse discrete wavelet transform module.

[0073] The inverse discrete wavelet transform module is used to determine the enhancement coefficient matrix. Perform inverse discrete wavelet transform to obtain the reconstructed time-domain features. And send it to the feature fusion module, This represents the inverse discrete wavelet transform matrix. Orthogonality must be satisfied to ensure complete signal reconstruction. A self-attention mechanism can adaptively enhance damage-related frequency components, thereby enhancing the temporal features.

[0074] The second self-attention module is used to sample the array waveguide monitoring signal using a self-attention mechanism. Processing is performed to obtain temporal self-attention features. And send it to the feature fusion module.

[0075] The feature fusion module is used to analyze arrayed waveguide monitoring signal samples. Reconstructing temporal features Temporal self-attention features The characteristics of the arrayed waveguide signal are obtained by fusion. :

[0076] (13)

[0077] in, , This represents the weighting parameter.

[0078] Through the above fusion, complementary information between and within channels can be integrated while preserving the underlying structure of the original signal. In practical applications, the weighting parameters can be adjusted. , By setting the parameters as learnable and optimizing them during backpropagation, the feature extraction module adaptively balances the importance of global dependencies and local oscillation features, thereby deeply capturing the dependencies and degradation trend features within the signal sequence.

[0079] In terms of specific implementation, both the first self-attention module and the second self-attention module in this embodiment adopt a block-based self-attention mechanism. The specific method is as follows: Let the input data be... First, its sample space is divided into There are 3 independent subspaces. For the 1st... Each subspace is used to obtain three sets of transformation features through linear projection. , , :

[0080] (14)

[0081] in, , For learnable projection matrices, This represents the dimension of each subspace.

[0082] Based on this, the correlation matrix between features is calculated and normalized to obtain the weighted aggregated features. :

[0083] (15)

[0084] Will Weighted aggregation features of individual subspaces By concatenating along the last dimension, the output feature is obtained. :

[0085] (16)

[0086] Physical heterogeneous graph convolution extractor is used for physical-virtual heterogeneous graphs Perform graph convolution, where the initial feature vectors of the physical measurement nodes are based on the features of the array guided wave signals. We obtain the feature matrix by setting the initial feature vector of the virtual damaged node to zero. And send it to the prediction module.

[0087] In the simulated diffusion process of graph convolution, the initial features of nodes propagate information under the constraints of a pre-constructed heterogeneous edge set. Information propagates directionally from physical measurement nodes to virtual damage nodes along a pre-defined physical connection path, gradually updating the latter's feature representation. This process, constrained by prior physical features, achieves reverse inference from the observed signal to the scattering source distribution. Its forward propagation formula can be expressed as follows:

[0088] (17)

[0089] in, To add the adjacency matrix after adding self-loops, Represents the identity matrix; For use in normalization The diagonal matrix. After multiple iterations and updates, the node features... The implicit representation of damage location and intensity information is gradually encoded in the code. Finally, this representation is input into subsequent modules to obtain the damage probability corresponding to each virtual node.

[0090] The prediction module is used to predict based on the feature matrix. generate A dimensional prediction vector, where each element represents the probability that damage exists in the corresponding region.

[0091] S104: Training the composite plate damage detection model:

[0092] The composite material damage detection model is trained using the training sample set obtained in step S102 to obtain a trained composite material damage detection model.

[0093] To guide the model in accurately inferring damage categories, this embodiment designs a joint optimization loss for model training, including cross-entropy loss. and central loss loss function The calculation formula is as follows:

[0094] (18)

[0095] in, This represents the weights that balance the contributions of the two types of losses.

[0096] Traditional cross-entropy loss uses one-hot labels, which can easily lead to the model overconfidence in the training samples, reducing generalization ability. To alleviate this problem, this embodiment introduces label smoothing technology to construct a smoothed label distribution, thus reducing the label smoothing cross-entropy loss for each batch. The calculation formula is as follows:

[0097] (19)

[0098] in, Indicates the number of samples in the current batch. Indicates the first in the current batch The true category of each sample , Indicates the first in the current batch The predicted category for each sample The probability, , Represents the smoothing factor. Indicates an indicator function, when The value is 1 if the condition is met, and 0 otherwise. This loss function makes the model's output probability distribution smoother, avoids overfitting, and thus improves the generalization ability to unknown working conditions.

[0099] To further enhance the discriminability of features, this embodiment also introduces a center loss function, which minimizes the distance between sample features and their corresponding class centers, making features of the same damage class more clustered in the embedding space. Center Loss The calculation formula is as follows:

[0100] (20)

[0101] in, Indicates the number of samples in the current batch. Indicates the first in the current batch Each sample is an intermediate feature vector output from a certain layer in the composite plate damage detection model, typically the first layer counting backwards from the classification layer in the prediction model. This indicates all samples in the current batch that are related to the first... True category of each sample The mean of the intermediate feature vectors of the same samples, i.e., the class The feature center of the class.

[0102] In this embodiment, the composite plate damage detection model is trained end-to-end using the Adam optimizer, with an initial learning rate set to [value missing]. Central loss balance coefficient The learning rate is dynamically adjusted using a cosine annealing strategy to ensure stable convergence and improve generalization performance. This is achieved by minimizing the joint loss. The model can learn damage features with strong discriminative power and generalization performance under the joint guidance of prior physical features and monitoring data features, thereby achieving high-precision damage localization.

[0103] S105: Damage Detection of Composite Panels

[0104] When damage detection of composite panels is required, the same scheme as in step S101 is used to deploy an ultrasonic transducer array and divide the area. The array guided wave monitoring signal samples are collected and input into the composite panel damage detection model trained in step S104. The damage detection result is obtained based on the output prediction vector.

[0105] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the invention. In this embodiment, a Structural Health Monitoring (SHM) guided wave dataset is used. This dataset is acquired based on an ultrasonic guided wave system test bench, which consists of a CFRP board, a constant temperature chamber, an ultrasonic probe, and a PD200 voltage amplifier. Figure 4 This is a schematic diagram of the CFRP board area division in this embodiment. Figure 4As shown, in this embodiment, the CFRP board is divided into 20 regions. To simulate different service environments, the temperature in this embodiment is adjusted using a constant temperature chamber, with a range of 20°C to 60°C and a step size of 0.5°C. At each temperature point, tests are performed for five structural states, including the normal state (NC) and four typical damage regions, namely Region I damage (… ), Area VII damage ( ), regional IX damage ( ) and regional XV damage ( Six piezoelectric sensors are symmetrically arranged on each side of the CFRP, and each sensor can be used as an excitation source or a receiver. The excitation signal is a 5-cycle sinusoidal pulse modulated by a Hanning window, and the guided wave response is acquired using a pair of other methods. Figure 5 These are typical guided wave signals for different channels in this embodiment.

[0106] Damage detection is performed on guided wave signals at a fixed temperature point. In this embodiment, the training set and test set are randomly divided in a 4:1 ratio. Table 1 shows the dataset partitioning information in this embodiment.

[0107] A 20 2600 B 30 6240 C 40 6240 D 50 6500

[0108] Table 1

[0109] Table 2 is a list of hyperparameters used for model training in this embodiment.

[0110] Learning rate 0.001 Optimizer AdamW Training rounds 300 Weight degradation 0.0001 Activation function ReLU Drop rate 0.4

[0111] Table 2

[0112] To comprehensively evaluate the diagnostic performance and generalization ability of this invention in nondestructive testing of composite panels, this embodiment uses three general evaluation indicators:

[0113] Accuracy (ACC, the higher the better) is used to evaluate diagnostic performance.

[0114] Detection rate (DR, the higher the better) is used to assess the rate at which damage is detected.

[0115] The false alarm rate (FAR, the lower the better) is used to evaluate the probability of a model making an incorrect impairment judgment.

[0116] Based on the single-temperature classification scenario constructed in Table 1, this embodiment compares the present invention with several typical methods under the same experimental settings to verify its effectiveness. The comparison methods include classic deep learning methods DD-BCNN, WDCNN, and MLE-LSTM, as well as graph neural network methods MCAGN and MA-STGNN. Table 3 is a comparison table of the classification capabilities of the present invention and the comparison methods in this embodiment.

[0117]

[0118] Table 3

[0119] As shown in Table 3, traditional machine learning methods exhibit poor stability, primarily due to their heavy reliance on manual feature extraction, which makes it difficult to fully capture deep-seated fault characteristics. For example, CNNs perform excellently at 30℃ and 50℃, but their performance drops significantly at 20℃ and 40℃, revealing their dependence on specific temperature conditions. While deep learning methods attempt to incorporate modern large-scale model techniques to enhance interpretability and convergence, their understanding of the physical nature of faults remains insufficient, leaving room for improvement in accuracy. Particularly at 20℃, the false positive rate of GAF reaches a high of 82.14%, indicating its near-failure. In contrast, this invention achieves the highest diagnostic accuracy across all metrics and demonstrates better stability. This not only demonstrates superior performance compared to traditional machine learning methods but also outperforms other compared deep learning methods.

[0120] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for damage detection of composite plates based on physical heterogeneous graph neural networks, characterized in that, Includes the following steps: S1: Deploy an array of ultrasonic transducers on the surface of the composite material to divide the composite material into sections. The ultrasonic transducer array is divided into several regions, and each region is used as a virtual damage node. Each channel serves as a physical measurement node, thus constructing a physical-virtual heterogeneous graph. , where the set of nodes express A virtual damage node and A set consisting of physical measurement nodes, edge set ,in, Indicates the first One physical measurement node, Indicates the first A virtual damage node, Represents physical measurement nodes The corresponding main coverage area of ​​the sound beam , , ; S2: For normal composite panels and several composite panels with damage in different areas, deploy ultrasonic transducer arrays using the same scheme as in step S1 and divide the area. Through the deployed... The channel ultrasonic transducer array acquires several array guided wave monitoring signal samples. , Represent the spatial dimension of the sample and label its damage category. ,in This is indicated as a normal composite board. Indicates the first Composite panels with damage in certain areas. Thus, the training sample set is obtained; S3: Construct a damage detection model for composite plates based on a physically heterogeneous graph neural network, including a data feature extraction module, a physically heterogeneous graph convolutional extractor, and a prediction module, wherein: The data feature extraction module is used to extract input array guided wave monitoring signal samples. Arrayed waveguide signal characteristics And send it to the physical heterogeneous graph convolution extractor; Physical heterogeneous graph convolution extractor is used for physical-virtual heterogeneous graphs Perform graph convolution, where the initial feature vectors of the physical measurement nodes are based on the features of the array guided wave signals. The initial feature vector of the virtual damaged node is set to zero, and the feature matrix is ​​obtained through processing. And send it to the prediction module; The prediction module is used to predict based on the feature matrix. generate A multidimensional prediction vector, where each element represents the probability that damage exists in the corresponding region; S4: Use the training sample set obtained in step S2 to train the composite board damage detection model to obtain a trained composite board damage detection model. S5: When it is necessary to perform damage detection on the composite board, the same scheme as in step S1 is used to deploy the ultrasonic transducer array and divide the area. The array guided wave monitoring signal samples are collected and input into the composite board damage detection model trained in step S4. The damage detection result is obtained according to the output prediction vector.

2. The method for detecting damage to composite panels according to claim 1, characterized in that, The main coverage area of ​​the sound beam The generation method is as follows: For physical measurement nodes For the corresponding ultrasonic transducer channel, a straight line segment between the transmitting and receiving probes is first generated based on the position of the corresponding probe pair. As the main axis Then, with the main axis The vertical distance is not greater than the preset beamwidth. The area is considered the effective detection range of this channel. ; During the determination, if the virtual damaged node The corresponding area and the main coverage area of ​​the sound beam Intersecting, then virtual damaged nodes Belongs to the main coverage area of ​​the sound beam .

3. The method for detecting damage to composite panels according to claim 1, characterized in that, The feature extraction module includes a discrete wavelet transform module, a first self-attention module, an inverse discrete wavelet transform module, a second self-attention module, and a feature fusion module, wherein: The discrete wavelet transform module is used to process the input array guided wave monitoring signal samples. Perform a discrete wavelet transform to obtain the coefficient matrix. ,in Representing the discrete wavelet transform matrix, the coefficient matrix is... Send to the first self-attention module; The first self-attention module is used to apply a self-attention mechanism to the coefficient matrix. After processing, the enhancement coefficient matrix is ​​obtained. And send it to the inverse discrete wavelet transform module; The inverse discrete wavelet transform module is used to determine the enhancement coefficient matrix. Perform inverse discrete wavelet transform to obtain the reconstructed time-domain features. And send it to the feature fusion module, Represents the inverse discrete wavelet transform matrix; The second self-attention module is used to sample the array waveguide monitoring signal using a self-attention mechanism. Processing is performed to obtain temporal self-attention features. And send it to the feature fusion module; The feature fusion module is used to analyze arrayed waveguide monitoring signal samples. Reconstructing temporal features Temporal self-attention features The characteristics of the arrayed waveguide signal are obtained by fusion. : ， in, , This represents the weighting parameter.

4. The method for detecting damage to composite panels according to claim 3, characterized in that, The first and second self-attention modules employ a block-based attention mechanism, specifically as follows: Let the input data be... First, its sample space is divided into There are 3 independent subspaces. For the 1st... Each subspace is used to obtain three sets of transformation features through linear projection. , , : ， in, , For learnable projection matrices, Indicates the dimension of each subspace; The weighted aggregation features are obtained using the following formula. : ， Will Weighted aggregation features of individual subspaces By concatenating along the last dimension, the output feature is obtained. : 。 5. The method for detecting damage to composite panels according to claim 1, characterized in that, loss function in step S4 The calculation formula is as follows: ， in, Indicates weight, Represents cross-entropy loss, The central loss is expressed by the following formula: ， in, Indicates the number of samples in the current batch. Indicates the first in the current batch Each sample is an intermediate feature vector output from a certain layer in the composite plate damage detection model. This indicates all samples in the current batch that are related to the first... True category of each sample The mean of the intermediate feature vectors of the same samples.

6. The method for detecting damage to composite panels according to claim 5, characterized in that, The cross-entropy loss is the label-smoothed cross-entropy loss, calculated using the following formula: ， in, Indicates the number of samples in the current batch. Indicates the first in the current batch The true category of each sample , Indicates the first in the current batch The predicted category for each sample The probability, , Represents the smoothing factor. Indicates an indicator function, when It is set to 1 if it is true, otherwise it is set to 0.

Images