A method and device for real-time damage detection of plate structures based on physical information spatiotemporal graph neural networks

Abstract

This invention discloses a method and apparatus for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network in the field of structural health monitoring technology. The method includes: constructing a three-dimensional physical mesh model of the plate structure based on the acquired dimensional information and material properties of the plate structure to be monitored; generating a sensor and excitation source layout in the physical mesh model based on a pre-determined sensor placement range; arranging corresponding sensors and excitation sources on the plate structure to be monitored based on the sensor and excitation source layout; acquiring feedback signals received by the sensors after pulse signals are emitted to the plate structure by the excitation sources; and sending the feedback signals to a pre-trained physical information spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle, and severity index of each damage. This invention achieves low-cost and highly robust structural health monitoring.

Description

Technical Field

[0001] This invention relates to a method and device for real-time detection of damage in plate structures based on a physical information spatiotemporal graph neural network, belonging to the field of structural health monitoring technology. Background Technology

[0002] Plate-like structures are widely used in bridges, aerospace, shipbuilding, and storage tanks. If internal cracks, delamination, and voids are not detected in time, they can lead to catastrophic damage. Traditional point-based detection methods such as ultrasound, X-ray, and eddy current testing are inefficient and costly; modal vibration detection is insensitive to early, minute damage; and while active monitoring based on Lamb waves is effective, existing methods generally suffer from the following challenges:

[0003] (1) The limited number of sensors (cost and wiring constraints) results in severely sparse observation data;

[0004] (2) Pure data-driven deep learning models lack physical constraints, have poor generalization ability, and require a large number of labeled and damaged samples.

[0005] (3) Existing graph neural network methods only treat sensor positions as nodes and do not make full use of structural geometry and wave physics priors, resulting in low reconstruction accuracy and large damage localization error;

[0006] (4) It lacks robustness against sensor failure and confidence assessment of prediction results, making it difficult to apply in engineering.

[0007] The relevant technologies mostly adopt ordinary ST-GCN (Spatial Temporal Graph Convolutional Network, a deep learning model based on spatiotemporal graph convolution) or PINN (Physics-Informed Neural Networks). One type treats wave velocity as a global scalar, while the other type attempts to predict local wave velocity but does not embed learnable physical parameters on the edges, resulting in a localization error of more than 0.3 m for internal closed damage, and the performance drops sharply when 30% of the sensors fail. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to overcome the defects of the prior art and provide a method and device for real-time detection of plate structure damage based on physical information spatiotemporal graph neural network.

[0009] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0010] In a first aspect, the present invention provides a method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network, comprising:

[0011] Based on the obtained dimensional information and material properties of the plate structure to be tested, a three-dimensional physical mesh model of the plate structure is constructed; the dimensional information includes the length, width and thickness of the plate structure.

[0012] Based on a predetermined sensor arrangement range, the sensor and excitation source location layout is generated in the physical mesh model of the plate structure;

[0013] Based on the aforementioned sensor and excitation source layout, corresponding sensors and excitation sources are arranged on the structure of the board to be tested. The health status of the arranged sensors is monitored. When a sensor failure is detected, the features of the node corresponding to the sensor are set to zero or masked, and missing information is filled in.

[0014] The sensor receives the feedback signal after the pulse signal is emitted to the side plate structure by the excitation source. The feedback signal is sent to a pre-trained spatiotemporal neural network model based on physical information to generate the location, area, major axis length, minor axis length, orientation angle and severity index of each damage.

[0015] Furthermore, the step of generating the sensor and excitation source location layout in the physical mesh model of the plate structure based on the predetermined sensor arrangement range includes:

[0016] Based on the pre-defined prohibited and key monitoring areas, as well as the number of sensors and excitation sources to be deployed, the random farthest point sampling algorithm is used to automatically optimize the deployment, resulting in the optimal sensor and excitation source location layout covering the three-dimensional spatial geometric features.

[0017] Furthermore, the training of the physical information-based spatiotemporal graph neural network model includes:

[0018] A training dataset is obtained. The training dataset is set in the physical mesh model of the plate structure according to the location layout of the sensor and the excitation source. After the setting is completed, a finite element simulation experiment is performed to generate finite element simulation sensor response samples for a target time period, including multiple closed or partially open cracks inside the plate structure, stereoscopic damage to the plate structure, and health benchmark samples of the plate structure.

[0019] Based on the training dataset and the joint loss function The physical information spatiotemporal graph neural network model to be trained is trained to obtain the trained physical information spatiotemporal graph neural network model; the joint loss function The formula is:

[0020] ;

[0021] in, The fitting loss for sensor measurement data; The physical equations contain the elastic wave equations, boundary conditions, and initial conditions that describe the wave propagation laws within the three-dimensional structure. The losses are L2 regularization and total variation regularization; , , These are the weighting coefficients for each type of loss.

[0022] Furthermore, the weight coefficients of each loss are dynamically adjusted using a relative loss balancing strategy. The weight update formula for the relative loss balancing strategy is as follows:

[0023] ;

[0024] in, Let k be the weight of the loss of the k-th item in the m-th update. Let k be the value of the loss at the current time. Let be the initial value of the k-th loss term, and n be the total number of loss terms. Let j be the value of the loss at the current time. Let j be the initial value of the loss term j, where j is the index of the loss term j; the relative loss balancing strategy is updated once after each preset training round.

[0025] Furthermore, the residuals of the elastic wave equation at each node Each compression time step The calculation formula is:

[0026] ;

[0027] in, The time second derivative of the node, This is the second spatial derivative, which is obtained through a weighted least-squares local fit of the K nearest neighbors in the graph. This is the wave field displacement or strain response. For nodes Local wave velocity at that location This represents the physical constraint residual value.

[0028] Further, the step of sending the feedback signal to a pre-trained physical information-based spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle, and severity index of each lesion includes,

[0029] The long-time-series feedback signal is compressed into a short-time-series feature sequence using a time window statistical feature compression method, and this short-time-series feature sequence is used as the initial feature of each node in the graph neural network.

[0030] A graph structure is constructed based on the 3D spatial coordinates of nodes and the Gaussian kernel distance function. The information transfer weights between nodes in the graph structure are determined by the adjacency matrix elements. The formula is as follows:

[0031] ;

[0032] in, For nodes The three-dimensional Euclidean distance between node j and node j The preset Gaussian bandwidth parameter, Here, is the preset cutoff radius threshold, and e is the natural constant;

[0033] Initialize the fixed geometric features and learnable physical features of each edge of the graph structure. The fixed geometric features include distance, direction cosine, direction sine, and boundary markers. The learnable physical features include wave velocity c, damping coefficient α, and impedance ratio. ;

[0034] The initial features of the nodes are input into 3-6 layers of spatiotemporal convolutional blocks for iterative updates. The structure of the spatiotemporal convolutional blocks consists of a graph convolutional network (GCN), a gated recurrent unit (GRU), and a layer normalization unit (LayerNorm) connected in sequence. Finally, the updated node features are output through residual connections.

[0035] After each spatiotemporal convolutional block completes the iterative update of node features, the learnable physical features are updated using a multilayer perceptron (MLP) network based on the features of neighboring nodes in the current layer. The edge feature update formula is as follows:

[0036] ;

[0037] in, and They are nodes and nodes The feature vector of the l-th layer, For the old edge features, The updated edge features;

[0038] Wave velocity channel values ​​are extracted from the edge features output by the last spatiotemporal convolutional block, and RBF interpolation is used to reconstruct the discrete wave velocity channel values ​​into a continuous three-dimensional wave velocity field. RBF interpolation is radial basis function interpolation.

[0039] Calculate the three-dimensional wave velocity field By comparing the relative differences with the healthy baseline samples of the plate structure, and combining gradient magnitude and adaptive threshold segmentation, morphological processing and connected component analysis, the location, area, major axis length, minor axis length, orientation angle and severity index of each damage are finally extracted.

[0040] Furthermore, the method of compressing the acquired long-time-series feedback signal into a short-time-series feature sequence using a time window statistical feature compression method includes:

[0041] Based on the feedback signal received by the sensor after the pulse signal is emitted to the side plate structure by the excitation source, the feedback signal is divided into overlapping time windows according to the time step. The mean and standard deviation, splicing position code and node type code are extracted for each time window to form a T′×6-dimensional temporal feature tensor for each node. The feature of nodes without sensors is set to zero. T′ represents the number of time windows.

[0042] The formula for calculating the number of time windows T′ is:

[0043] ;

[0044] in, For window length, The step size is N, where N is the total number of sampling points in the original acquired signal, and the final node feature shape is the number of nodes. × × 6.

[0045] Furthermore, the method of compressing the feedback signal into a short time-series feature sequence using a time window statistical feature compression method further includes:

[0046] The sensor health status is determined based on three criteria: signal variance, abnormal amplitude, and neighbor correlation. If a sensor fails, the features of the corresponding node are automatically zeroed or masked, and the missing information is filled in by using the neighborhood aggregation capability of the graph neural network.

[0047] Furthermore, the formula for calculating graph convolution in the spatiotemporal convolution block is:

[0048] ;

[0049] in, For graph convolution filters, For node features, For the normalized Laplace matrix, , Laplace matrix The largest eigenvalue, It is a k-th order Chebyshev polynomial. Let K be the k-th order learnable weight matrix, where K is the order of the Chebyshev polynomial.

[0050] Secondly, the present invention also discloses a real-time damage detection device for plate structures based on a physical information spatiotemporal graph neural network, comprising:

[0051] The building module is used to construct a physical mesh model of the plate structure based on the acquired dimensional information and material properties of the plate structure to be tested;

[0052] The layout module is used to generate the sensor and excitation source position layout in the physical mesh model of the plate structure based on the predetermined sensor layout range; based on the sensor and excitation source position layout, the corresponding sensors and excitation sources are arranged on the plate structure to be tested, the health status of the arranged sensors is monitored, and when a sensor failure is detected, the features of the node corresponding to the sensor are set to zero or masked, and missing information is filled in.

[0053] The model processing module is used to collect the feedback signal received by the sensor after the pulse signal is emitted to the side plate structure by the excitation source, and send the feedback signal to the pre-trained physical information-based spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle and severity index of each damage.

[0054] The beneficial effects achieved by this invention are as follows:

[0055] This invention deeply integrates three-dimensional geometric constraints and wave physics laws, enabling high-precision reconstruction of the material parameter field inside and on the surface of a plate structure using sparse sensor acceleration response data in environments with partial sensor failure or high noise. It generates in real time the location, area, major axis length, minor axis length, orientation angle, and severity index of each damage, achieving low-cost and highly robust structural health monitoring. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of the process of a real-time damage detection method for plate structures based on a physical information spatiotemporal graph neural network provided by the present invention;

[0057] Figure 2 This is an architecture diagram of the spatiotemporal graph neural network model based on physical information provided by the present invention;

[0058] Figure 3 This is a schematic diagram of the working process of the three-dimensional plate structure physical mesh model provided by the present invention. Detailed Implementation

[0059] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0060] Example 1, as Figure 1 As shown in the figure, this embodiment introduces a real-time damage detection method for plate structures based on a physical information spatiotemporal graph neural network, including:

[0061] S101. Based on the obtained dimensional information and material properties of the plate structure to be detected, a three-dimensional physical mesh model of the plate structure is constructed; the dimensional information includes the length, width and thickness of the plate structure.

[0062] S102. Based on the predetermined sensor arrangement range, generate the sensor and excitation source position layout in the physical mesh model of the plate structure;

[0063] S103. Based on the aforementioned sensor and excitation source position layout, arrange corresponding sensors and excitation sources on the structure of the board to be tested.

[0064] S104. The sensor receives the feedback signal after the pulse signal is emitted to the side plate structure by the excitation source. The feedback signal is sent to the pre-trained spatiotemporal neural network model based on physical information to generate the location, area, major axis length, minor axis length, orientation angle and severity index of each damage.

[0065] The step of generating the sensor and excitation source location layout in the physical mesh model of the plate structure based on a predetermined sensor arrangement range includes:

[0066] Based on the pre-defined prohibited and key monitoring areas, as well as the number of sensors and excitation sources to be deployed, the random farthest point sampling algorithm is used to automatically optimize the deployment, resulting in the optimal sensor and excitation source location layout covering the three-dimensional spatial geometric features.

[0067] The training of the physical information-based spatiotemporal graph neural network model includes:

[0068] A training dataset is obtained. The training dataset is set in the physical mesh model of the plate structure according to the location layout of the sensor and the excitation source. After the setting is completed, a finite element simulation experiment is performed to generate finite element simulation sensor response samples for a target time period, including multiple closed or partially open cracks inside the plate structure, stereoscopic damage to the plate structure, and health benchmark samples of the plate structure.

[0069] Based on the training dataset and the joint loss function The physical information spatiotemporal graph neural network model to be trained is then trained to obtain the trained physical information spatiotemporal graph neural network model, such as... Figure 2 As shown; the joint loss function The formula is:

[0070] ;

[0071] in, The fitting loss for sensor measurement data; The physical equations contain the elastic wave equations, boundary conditions, and initial conditions that describe the wave propagation laws within the three-dimensional structure. The losses are L2 regularization and total variation regularization; , , These are the weighting coefficients for each type of loss.

[0072] The weight coefficients of each loss are dynamically adjusted using a relative loss balancing strategy. The weight update formula for the relative loss balancing strategy is as follows:

[0073] ;

[0074] in, Let k be the weight of the loss of the k-th item in the m-th update. Let k be the value of the loss at the current time. Let be the initial value of the k-th loss term, n be the total number of loss terms, and j be the index of the loss term; the relative loss balancing strategy is updated once every 10 training rounds.

[0075] The wave equation residual at each node Each compression time step The calculation formula is:

[0076] ;

[0077] in, The time second derivative of the node, This is the second spatial derivative, which is obtained through a weighted least-squares local fit of the K nearest neighbors in the graph. This is the wave field displacement or strain response. For nodes Local wave velocity at that location This represents the physical constraint residual value.

[0078] The step involves sending the feedback signal to a pre-trained physics-based spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle, and severity index of each lesion, including...

[0079] The long-term feedback signal is compressed into a short-term feature sequence using the time window statistical feature compression method, and this sequence is used as the initial feature of each node in the graph neural network.

[0080] A graph structure is constructed based on the 3D spatial coordinates of nodes and the Gaussian kernel distance function. The information transfer weights between nodes in the graph structure are determined by the adjacency matrix elements. The formula is as follows:

[0081] ;

[0082] in, For nodes The three-dimensional Euclidean distance between node j and node j The preset Gaussian bandwidth parameter, This is the preset cutoff radius threshold;

[0083] Initialize the fixed geometric features and learnable physical features of each edge. The fixed geometric features include Euclidean distance (1D), three-dimensional direction vector (3D), and boundary marker (2D), for a total of 6 dimensions. The learnable physical features include wave velocity c, damping coefficient α, and impedance ratio. It has a total of 3-dimensional features;

[0084] The initial features of the nodes are input into 3-6 layers of spatiotemporal convolutional blocks for iterative updates. The structure of the spatiotemporal convolutional blocks is GCN→GRU→LayerNorm+residual connections.

[0085] After each layer of GRU message passing, the learnable physical edge features are updated using a multilayer perceptron (MLP) based on the features of neighboring nodes in the current layer. The edge feature update formula is as follows:

[0086] ;

[0087] in, and They are nodes and nodes The feature vector of the l-th layer, For the old edge features, For the updated edge features, MLP is a multilayer perceptron network that uses a residual connection mechanism;

[0088] Wave velocity channel values ​​are extracted from the updated edge features, and RBF interpolation is used to reconstruct the discrete edge wave velocities into a continuous three-dimensional wave velocity field. ;

[0089] Calculate the reconstructed wave velocity field By comparing the relative differences with the healthy baseline samples of the plate structure, and combining gradient magnitude and adaptive threshold segmentation, morphological processing and connected component analysis, the location, area, major axis length, minor axis length, orientation angle and severity index of each damage are finally extracted.

[0090] The step of compressing the feedback signal into a short time-series feature sequence using the time window statistical feature compression method includes:

[0091] Based on the feedback signal received by the sensor after the pulse signal is emitted to the side plate structure by the excitation source, the feedback signal is divided into overlapping time windows according to the time step. The mean and standard deviation, splicing position code and node type code are extracted for each time window to form a T′×6-dimensional temporal feature tensor for each node. The feature of nodes without sensors is set to zero. T′ represents the number of time windows.

[0092] The formula for calculating the time window T′ is:

[0093] ;

[0094] in For window length, The step size is N, where N is the total number of sampling points in the original acquired signal, and the final node feature shape is the number of nodes. × × 6.

[0095] The method of using time window statistical feature compression to compress the feedback signal into a short time-series feature sequence also includes: determining whether the sensor health status is faulty based on three criteria: signal variance, abnormal amplitude, and neighbor correlation. If the sensor is faulty, the features of the corresponding node of the sensor are automatically set to zero or masked. The missing information is filled in by using the neighborhood aggregation capability of graph neural networks to achieve sensor failure monitoring.

[0096] The formula for calculating graph convolution in the spatiotemporal convolution block is:

[0097] ;

[0098] in, For graph convolution filters, For node features, For the normalized Laplace matrix, , It is a k-th order Chebyshev polynomial. Let K be the k-th order learnable weight matrix, where K is the order of the Chebyshev polynomial.

[0099] The method for determining sensor health status based on three criteria—signal variance, abnormal amplitude, and neighbor correlation—specifically includes: calculating the sensor signal variance; if it falls below a preset silence threshold, it is considered an open-circuit failure; calculating the maximum signal amplitude; if it exceeds a range threshold, it is considered an abnormal failure; and calculating the average cross-correlation coefficient between the sensor and its K nearest neighbor sensors in the spatial neighborhood; if it falls below a consistency threshold, it is considered a degraded failure. If a failure is determined, the features of the corresponding node of the sensor are automatically zeroed or masked, and the neighborhood aggregation capability of the graph neural network is used to fill in the missing information.

[0100] Example 2, based on the same inventive concept as Example 1, introduces a real-time damage detection device for plate structures based on a physical information spatiotemporal graph neural network, comprising:

[0101] The building module is used to construct a physical mesh model of the plate structure based on the acquired dimensional information and material properties of the plate structure to be tested;

[0102] The layout module is used to generate the sensor and excitation source position layout in the physical mesh model of the plate structure based on a predetermined sensor arrangement range; and to arrange the corresponding sensors and excitation sources on the plate structure to be tested based on the sensor and excitation source position layout.

[0103] The model processing module is used to collect the feedback signal received by the sensor after the pulse signal is emitted to the side plate structure by the excitation source, and send the feedback signal to the pre-trained physical information-based spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle and severity index of each damage.

[0104] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0105] Example 3, based on the same inventive concept as other examples, introduces a real-time damage detection method for plate structures based on a physical information spatiotemporal graph neural network, including:

[0106] Divide the 2.7 m × 1.2 m concrete panel into a grid with a spacing of 0.2 m, as follows: Figure 3 As shown, the central area of ​​the marker plate is a densely arranged area. The program automatically arranges 23 piezoelectric sensors and 2 excitation sources, exciting 50 kHz, 5-cycle sine pulses, with the sampling rate set to 1 MHz, and collecting 4096 time steps.

[0107] 5000 samples were generated using a commercial finite element method (FEM) with explicit dynamics. Of these, 4000 samples contained 1-3 internal closed cracks, and 1000 contained delamination or pores. Each sample was compressed to T′=255 time steps with 50% overlap and a window length of 32 steps. The node features were 6-dimensional, the network consisted of 6 spatiotemporal layers, the Chebyshev polynomial order was K=3, and the edge features were 9-dimensional. Training on a single RTX 3090 GPU converged in approximately 4 hours, with a root mean square error (RMSE) and a wave velocity reconstruction error of 4.2%.

[0108] Real-time signal acquisition → compression → forward inference → RBF interpolation to obtain a 2 cm resolution wave velocity field → post-processing to output damage parameters, while the confidence level given by Monte Carlo Dropout is 0.93.

[0109] This method also verifies the robustness of sensor failure by randomly causing 10% of the sensors to fail and repeating the Monte Carlo experiment 100 times. The results show that the average positioning error is 0.11 m, proving that the system has extremely strong engineering robustness.

[0110] In summary, this invention deeply integrates three-dimensional geometric constraints and wave physics laws, enabling high-precision reconstruction of the material parameter field inside and on the surface of a plate structure using sparse data in environments with partial sensor failure or high noise. It generates the location, area, major axis length, minor axis length, orientation angle, and severity index of each damage in real time, achieving low-cost and highly robust structural health monitoring.

[0111] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0112] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0113] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0114] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0115] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A real-time damage detection method for plate structures based on a physical information spatiotemporal graph neural network, characterized in that, include: Based on the obtained dimensional information and material properties of the plate structure to be tested, a three-dimensional physical mesh model of the plate structure is constructed. The dimensional information includes the length, width, and thickness of the plate structure; Based on a predetermined sensor arrangement range, the sensor and excitation source location layout is generated in the physical mesh model of the plate structure; Based on the aforementioned sensor and excitation source layout, corresponding sensors and excitation sources are arranged on the structure of the board to be tested. The health status of the arranged sensors is monitored. When a sensor failure is detected, the features of the node corresponding to the sensor are set to zero or masked, and missing information is filled in. The sensor receives the feedback signal after the pulse signal is emitted to the side plate structure by the excitation source. The feedback signal is sent to a pre-trained spatiotemporal neural network model based on physical information to generate the location, area, major axis length, minor axis length, orientation angle and severity index of each damage.

2. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 1, characterized in that, The step of generating the sensor and excitation source location layout in the physical mesh model of the plate structure based on a predetermined sensor arrangement range includes: Based on the pre-defined prohibited and key monitoring areas, as well as the number of sensors and excitation sources to be deployed, the random farthest point sampling algorithm is used to automatically optimize the deployment, resulting in the optimal sensor and excitation source location layout covering the three-dimensional spatial geometric features.

3. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 1, characterized in that, The training of the physical information-based spatiotemporal graph neural network model includes: A training dataset is obtained. The training dataset is set in the physical mesh model of the plate structure according to the location layout of the sensor and the excitation source. After the setting is completed, a finite element simulation experiment is performed to generate finite element simulation sensor response samples for a target time period, including multiple closed or partially open cracks inside the plate structure, stereoscopic damage to the plate structure, and health benchmark samples of the plate structure. Based on the training dataset and the joint loss function The physical information spatiotemporal graph neural network model to be trained is trained to obtain the trained physical information spatiotemporal graph neural network model; the joint loss function The formula is: ； in, The fitting loss for sensor measurement data; The physical equation residual loss is the loss of the physical equation, which includes the elastic wave equation, boundary conditions, and initial conditions that can describe the wave propagation law inside the three-dimensional structure. The losses are L2 regularization and total variation regularization; , , These are the weighting coefficients for each type of loss.

4. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 3, characterized in that, The weight coefficients of each loss are dynamically adjusted using a relative loss balancing strategy. The weight update formula for the relative loss balancing strategy is as follows: ； in, Let k be the weight of the loss of the k-th item in the m-th update. Let k be the value of the loss at the current time. Let be the initial value of the k-th loss term, and n be the total number of loss terms. Let j be the value of the loss at the current time. Let j be the initial value of the loss term j, where j is the index of the loss term j; the relative loss balancing strategy is updated once after each preset training round.

5. The real-time damage detection method for plate structures based on a physical information spatiotemporal graph neural network according to claim 3, characterized in that, The residual of the elastic wave equation at each node Each compression time step The calculation formula is: ； in, The time second derivative of the node, This is the second spatial derivative, which is obtained through a weighted least-squares local fit of the K nearest neighbors in the graph. This is the wave field displacement or strain response. For nodes Local wave velocity at that location This represents the physical constraint residual value.

6. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 1, characterized in that, The feedback signal is then sent to a pre-trained physical information-based spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle, and severity index of each lesion. include, The long-time-series feedback signal is compressed into a short-time-series feature sequence using a time window statistical feature compression method, and this short-time-series feature sequence is used as the initial feature of each node in the graph neural network. A graph structure is constructed based on the 3D spatial coordinates of nodes and the Gaussian kernel distance function. The information transfer weights between nodes in the graph structure are determined by the adjacency matrix elements. The formula is as follows: ； in, For nodes The three-dimensional Euclidean distance between node j and node j The preset Gaussian bandwidth parameter, Here, is the preset cutoff radius threshold, and e is the natural constant; Initialize the fixed geometric features and learnable physical features of each edge of the graph structure. The fixed geometric features include distance, direction cosine, direction sine, and boundary markers. The learnable physical features include wave velocity c, damping coefficient α, and impedance ratio. ; The initial features of the nodes are input into 3-6 layers of spatiotemporal convolutional blocks for iterative updates. The structure of the spatiotemporal convolutional blocks consists of a graph convolutional network (GCN), a gated recurrent unit (GRU), and a layer normalization unit (LayerNorm) connected in sequence. Finally, the updated node features are output through residual connections. After each spatiotemporal convolutional block completes the iterative update of node features, the learnable physical features are updated using a multilayer perceptron (MLP) network based on the features of neighboring nodes in the current layer. The edge feature update formula is as follows: ； in, and They are nodes and nodes The feature vector of the l-th layer, For the old edge features, The updated edge features; Wave velocity channel values ​​are extracted from the edge features output by the last spatiotemporal convolutional block, and RBF interpolation is used to reconstruct the discrete wave velocity channel values ​​into a continuous three-dimensional wave velocity field. RBF interpolation is radial basis function interpolation. Calculate the three-dimensional wave velocity field By comparing the relative differences with the healthy baseline samples of the plate structure, and combining gradient magnitude and adaptive threshold segmentation, morphological processing and connected component analysis, the location, area, major axis length, minor axis length, orientation angle and severity index of each damage are finally extracted.

7. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 6, characterized in that, The method of compressing the acquired long-time-series feedback signal into a short-time-series feature sequence using a time window statistical feature compression method includes: Based on the feedback signal received by the sensor after the pulse signal is emitted to the side plate structure by the excitation source, the feedback signal is divided into overlapping time windows according to the time step. The mean and standard deviation, splicing position code and node type code are extracted for each time window to form a T′×6-dimensional temporal feature tensor for each node. The feature of nodes without sensors is set to zero. T′ represents the number of time windows. The formula for calculating the number of time windows T′ is: ； in, For window length, The step size is N, where N is the total number of sampling points in the original acquired signal, and the final node feature shape is the number of nodes. × × 6.

8. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 7, characterized in that, The step of compressing the feedback signal into a short time-series feature sequence using the time window statistical feature compression method further includes: The sensor health status is determined based on three criteria: signal variance, abnormal amplitude, and neighbor correlation. If a sensor fails, the features of the corresponding node are automatically zeroed or masked, and the missing information is filled in by using the neighborhood aggregation capability of the graph neural network.

9. The method for real-time damage detection of plate structures based on a physical information spatiotemporal graph neural network according to claim 6, characterized in that, The formula for calculating graph convolution in the spatiotemporal convolution block is: ； in, For graph convolution filters, For node features, For the normalized Laplace matrix, , Laplace matrix The largest eigenvalue, It is a k-th order Chebyshev polynomial. Let K be the k-th order learnable weight matrix, where K is the order of the Chebyshev polynomial.

10. A real-time damage detection device for plate structures based on a physical information spatiotemporal graph neural network, characterized in that, include: The building module is used to construct a physical mesh model of the plate structure based on the acquired dimensional information and material properties of the plate structure to be tested; The layout module is used to generate the sensor and excitation source position layout in the physical mesh model of the plate structure based on a predetermined sensor arrangement range; Based on the aforementioned sensor and excitation source layout, corresponding sensors and excitation sources are arranged on the structure of the board to be tested. The health status of the arranged sensors is monitored. When a sensor failure is detected, the features of the node corresponding to the sensor are set to zero or masked, and missing information is filled in. The model processing module is used to collect the feedback signal received by the sensor after the pulse signal is emitted to the side plate structure by the excitation source, and send the feedback signal to the pre-trained physical information-based spatiotemporal graph neural network model to generate the location, area, major axis length, minor axis length, orientation angle and severity index of each damage.

Images