A structural damage identification method based on global conservation map physical neural network

Abstract

The application discloses a structure damage identification method based on a global conservation diagram physical neural network, which collects structure dynamic response signals through a sensor array, converts the signals into high-dimensional time characteristics through Fourier characteristic mapping, combines a structure physical topology to construct a diagram structure, inputs the diagram structure into a pre-constructed physical information diagram neural network to perform spatial characteristic extraction and motion state prediction, and further realizes intelligent identification of a structure damage position and a damage degree. The application innovatively fuses the spatial topology modeling capability of the diagram neural network with a physical partial differential equation, introduces a macroscopic energy-momentum global conservation loss mechanism based on time integration and a two-step alternating decoupling optimization strategy of Huber truncation, effectively solves the problems of physical baseline drift under a strong noise environment, invisible high-frequency characteristics and pathological coupling of gradients between physical parameters of a traditional physical information model, and realizes accurate positioning and quantitative evaluation of structure damage under an extreme noise working condition.

Description

Technical Field

[0001] This invention relates to the field of civil engineering structural health monitoring technology, and in particular to a structural damage identification method based on a global conservation graph physical neural network. Background Technology

[0002] Structural health monitoring aims to assess the service status of a structure by analyzing its dynamic response data. In practical engineering applications, due to the harsh on-site environment, vibration data collected by sensors is often accompanied by high-intensity environmental noise. Traditional data-driven deep learning methods (MLP, CNN), while capable of fitting nonlinear mappings, have significant limitations: 1) they heavily rely on massive amounts of clean labeled data; 2) the output results lack physical interpretability and are prone to overfitting under high-noise conditions.

[0003] Currently, structural damage identification technologies are mainly divided into three categories. Each category has significant technical bottlenecks, making it difficult to meet the practical application requirements of small samples, high noise, and strong robustness in engineering testing scenarios. The specific shortcomings are as follows: The first category comprises traditional physical model-based structural damage identification methods. These methods mainly include damage identification based on modal analysis and damage identification based on finite element model correction. Modal analysis-based methods extract modal parameters such as natural frequencies, mode shapes, and damping ratios from structural vibration signals, and determine the presence of damage based on changes in these parameters. However, they have extremely low sensitivity to minor local damage, only achieving a qualitative judgment of "whether damage exists" without accurately locating the damage or quantitatively assessing its severity. Furthermore, they are easily affected by environmental noise and testing errors, resulting in a high false positive rate. Finite element model correction methods require constructing a high-precision structural finite element model and continuously correcting model parameters to match the simulated response with the measured response, thereby inverting the structural damage. However, these methods have extremely high computational costs, are prone to getting trapped in local optima during the inversion process, and experience a sharp decline in inversion accuracy when the structure is complex and has a large number of degrees of freedom, failing to meet the needs of real-time online monitoring.

[0004] The second category comprises purely data-driven deep learning-based damage identification methods. With the development of artificial intelligence, deep learning models such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Graph Convolutional Neural Networks (GCNs) are widely used in structural damage identification. These methods train the model using a large number of labeled damage samples, automatically extracting deep features of the damage signal to achieve end-to-end damage identification. However, these methods have inherent and insurmountable drawbacks: First, they require an extremely large number of labeled damage samples, while in actual engineering practice, the cost of obtaining structural damage samples, especially severe damage samples, is extremely high and their quantity is extremely scarce. The model is prone to overfitting in small sample scenarios, resulting in extremely poor generalization ability. Second, the pure black-box model architecture does not incorporate any structural dynamics or physical laws, and the model output lacks physical interpretability. When the input data is affected by environmental noise or load variations, it is very easy to output results that are mathematically well-fitted but physically completely wrong, making it impossible to guarantee the reliability of the recognition results. Third, most existing graph neural network-related damage recognition patents and technical solutions only use the graph structure as a post-processing module for feature fusion, without embedding structural physical constraints into the core of the network architecture. In essence, they are still pure data-driven models, which cannot solve the engineering pain points in small sample and high-noise scenarios.

[0005] The third type, Physics-Informed Neural Network (PINN), embeds physical partial differential equations as soft constraints into the loss function, providing a new paradigm for addressing data scarcity and the lack of physical meaning. However, for damage identification of complex multi-degree-of-freedom structures, existing traditional PINN techniques face the following insurmountable theoretical bottlenecks and engineering challenges: 1. Spectral bias and spatial blind zone. Traditional neural networks preferentially fit low-frequency smooth signals, making it difficult to capture the severe damage characteristics of structures under high-frequency excitation; moreover, traditional fully connected networks treat each degree of freedom node in isolation, failing to perceive the physical topology of the structure's force transmission path.

[0006] 2. Physical baseline drift and energy divergence. Most existing PINNs rely solely on the dynamic residuals at microscopic discrete time points (e.g., The MSE (mean squared error) network is prone to getting trapped in local optima, fitting spurious curves with high-frequency zero-mean oscillations, resulting in "spurious thrust" and energy leakage in the system macroscopically, causing severe baseline drift in long-term forecasts.

[0007] 3. Ill-conditioned coupling of parameter gradients. In noisy environments (such as up to 25% Gaussian white noise), the data gradients generated by outliers can overwhelm the physical residual gradients. At the same time, the network weights, which contain tens of thousands of parameters, and the very few structural physical parameters (such as stiffness) suffer from a severe imbalance in magnitude during joint optimization, causing the optimizer to fail to converge smoothly and parameter inversion to fail.

[0008] 4. Traditional PINN models based on Multilayer Perceptrons (MLPs) suffer from a severe "spatial topological blind spot" when dealing with multi-degree-of-freedom structures. MLPs treat the responses of each degree of freedom as independent or only implicitly related output nodes, ignoring the stiffness coupling and shear force transmission paths between adjacent floors or nodes in the real physical world. This lack of physical topological information makes MLP-PINN highly susceptible to misjudging localized damage on one floor or dispersing it to non-adjacent floors under strong noise interference, significantly reducing spatial positioning accuracy and failing to meet the health monitoring needs of complex engineering structures. Summary of the Invention

[0009] In view of this, the purpose of this invention is to overcome the core defects of existing structural damage identification methods, such as insufficient high-frequency fitting ability, poor robustness under high noise, weak physical consistency, low parameter inversion accuracy, and unstable training process. This invention provides a structural damage identification method and system based on a global conservation graph physical neural network. This solution deeply integrates the spatial topology modeling capability of graph neural networks, the high-frequency signal fitting capability of Fourier feature encoding, the physical constraint capability of global conservation laws, and the training stability of decoupling optimization strategies. Even in high-noise, small-sample engineering scenarios, it can still achieve accurate inversion of structural stiffness parameters and accurate damage identification, meeting the practical application needs of long-term online health monitoring of multi-degree-of-freedom shear structures.

[0010] To achieve the above-mentioned technical objectives, the technical solution adopted by this invention is as follows: A structural damage identification method based on a global conservation graph physical neural network includes the following steps: S1. Graph Structure Construction and Dynamic Signal Acquisition Preprocessing: Dynamic response signals of the target structure (multi-degree-of-freedom shear-type target structure) are acquired by a sensor array deployed on the key degrees of freedom of the target structure. The multi-degree-of-freedom structure (multi-degree-of-freedom shear-type target structure) is then abstracted into a graph structure, defined as follows: In this context, node V represents the independent degrees of freedom of the structure, and the node attributes are the time-series characteristics of the dynamic response at the corresponding degree of freedom position. Based on the physical coupling relationship between adjacent degrees of freedom of the structure, the edge E and the standardized adjacency matrix of the graph structure are constructed, and the collected raw time-series signals are preprocessed by filtering and denoising, standardizing the dimensions, and dividing the dataset. S2. Construction of a dual-scale physical constraint system: Based on the dynamic characteristics of the target structure (multi-degree-of-freedom shear structure), a dual-scale physical regularization framework including point-state differential constraints and global integral constraints is constructed; the point-state differential constraints are the residual constraints of the dynamic control equations of the multi-degree-of-freedom shear structure, and the global integral constraints include at least the global momentum conservation constraints and the global energy conservation constraints of the system. S3. Construction of Physical Information Graph Neural Network Architecture: A physical information graph neural network model, abbreviated as PIGNN model, is constructed by integrating Fourier feature encoding layers and graph convolutional layers. The input of the PIGNN model is normalized time series and structural graph topological information, and the output is the displacement response prediction value of each degree of freedom of the structure. The PIGNN model includes a series of cascaded Fourier feature encoding layers, temporal embedding layers, multi-layer graph convolutional layers, and node output layers. The preprocessed time series signal and the topological adjacency matrix are input into the pre-constructed physical information graph neural network model to extract spatial topological features and temporal features, and output the response prediction value of each degree of freedom of the structure. S4. Robust Loss Function Design and Model Training: A data fitting loss is constructed based on the measured dynamic response signal and the predicted response value. A multi-dimensional composite loss function is constructed by combining the physical constraint loss output by the dual-scale physical regularization framework. An alternating decoupling optimization strategy is used to independently update the network parameters and the physical parameters of the structure to be inverted. Specifically, this can be achieved by constructing a multi-dimensional composite robust loss function and using a two-stage alternating decoupling optimization strategy to complete the end-to-end training of the PIGNN model. The multi-dimensional composite robust loss function is composed of a weighted fusion of Huber robust data fitting loss, point-state dynamic equation residual loss, global momentum conservation loss, and global energy conservation loss. The two-stage alternating decoupling optimization strategy includes a joint pre-training stage and a decoupling fine-tuning stage, which uses gradient switching control to achieve alternating independent optimization of the PIGNN network weights and the stiffness parameters to be inverted. S5. Intelligent Output of Structural Damage Information: Based on the structural physical parameters obtained by inversion, damage identification indicators are constructed to identify the spatial location and severity of structural damage. Specifically, this can be achieved by inputting the real-time dynamic response data of the structure to be tested into the trained PIGNN model, obtaining the stiffness parameters of each layer of the structure through model forward propagation and physical constraint inversion, identifying the spatial location and severity of structural damage based on the reduction ratio of the stiffness parameters relative to the undamaged initial state of the structure, and finally generating structural health status assessment results and damage warning information.

[0011] As a possible implementation, further, in step S1 of this scheme, the dynamic response signal includes displacement, velocity, and acceleration vibration signals of each degree of freedom of the structure, and the data sampling frequency is not less than 2.5 times the highest natural frequency of the structure; the preprocessing specifically includes: using Butterworth low-pass filtering and Savitzky-Golay smoothing filtering in a cascade to suppress high-frequency observation noise and environmental interference, using Z-score normalization to eliminate the influence of different dimensions of response data, and dividing the preprocessed dataset into training set, validation set and test set in a 6:2:2 ratio, maintaining the temporal continuity of the time series and not disrupting the order of time dimensions during the division process.

[0012] As one possible implementation, further, in step S1 of this solution, for multi-degree-of-freedom structures, the method for constructing the standardized adjacency matrix is ​​specifically as follows: Based on the unit nodes and the physical coupling relationships of the degrees of freedom in a multi-degree-of-freedom structure, a topological adjacency matrix A is constructed, where the matrix elements... Representing the i The and the first j There is a direct interlayer shear coupling relationship between each degree of freedom; otherwise... ; Construct a degree matrix D based on the topological adjacency matrix, where the diagonal elements of the degree matrix are... n is the total number of degrees of freedom of the structure, and all off-diagonal elements are 0; The standardized adjacency matrix is ​​constructed using a symmetric normalization method, and the calculation formula is as follows:

[0013] in, It is the inverse square root matrix of the degree matrix.

[0014] As one possible implementation, step S2 of this solution further includes: 1) Establish point-state dynamic control equation constraints. Based on the dynamic characteristics of the target structure, the control equation expression is constructed as follows:

[0015] Where M is the mass matrix of the structure, C is the modal damping matrix of the structure, and K is the stiffness matrix of the structure to be inverted. These are the displacement, velocity, and acceleration vectors for each degree of freedom of the structure. The external excitation vector acting on the structure; Based on the governing equations, the residual loss of the point-state dynamic equations in the entire time domain is constructed, and the expression is as follows:

[0016] in, For the first i One degree of freedom in t The residuals of the dynamic equations at time N are the total number of time sampling points and n is the total number of degrees of freedom of the structure. 3) Establish a global momentum conservation constraint, and construct the global momentum conservation loss based on the residuals of the dynamic equations across the entire time domain. The expression is as follows:

[0017] 4) Establish global energy conservation constraints. Based on the energy balance relationship between structural kinetic energy change, external force work, and damping dissipation, construct the global energy conservation loss, expressed as:

[0018] in, Let be the velocity vectors of each degree of freedom of the structure at time t. Let be the residual vector of the dynamic equations for each degree of freedom of the structure at time t.

[0019] As a possible implementation, further, in step S3 of this scheme, the physical information graph neural network model includes a Fourier feature encoding layer, a temporal embedding layer, a multi-layer graph convolutional layer, and a node output layer cascaded in sequence; the Fourier feature encoding layer is used to map a one-dimensional time variable into a high-dimensional periodic feature vector through a random frequency matrix.

[0020] In some embodiments, in step S3, the specific construction method of each layer of the PIGNN model is as follows: 1) The Fourier feature coding layer is used to map the one-dimensional time input to a high-dimensional periodic feature space, solving the spectral bias problem of traditional PINN. Its mapping formula is as follows:

[0021] in, The matrix is ​​a random frequency matrix, and its elements follow a normal distribution. The output is a high-dimensional Fourier periodic feature; 2) Temporal embedding layer, using a fully connected neural network structure, is used to map high-dimensional Fourier features to node features of the adaptation graph convolutional layer, and the output feature dimension matches the structural degrees of freedom and node feature dimension; 3) Multi-layer graph convolutional layers, employing a two-layer symmetric normalized graph convolutional structure. The forward propagation formula for a single-layer graph convolution is:

[0022] in, Let l be the node feature matrix of the l-th layer. , For the first The learnable weight matrix and bias vector of the layer, where σ is the Tanh nonlinear activation function. For a standardized adjacency matrix; 4) Node output layer: A linear fully connected layer is used to map the node features output by the graph convolutional layer into the displacement response prediction values ​​of each degree of freedom. The output dimension is consistent with the number of structural degrees of freedom.

[0023] As one possible implementation, further, in step S4 of this solution, the calculation expression of the multi-dimensional composite loss function is:

[0024] in, Huber robust data fitting loss covers the fitting deviation between the measured values ​​and model predictions of three types of responses: displacement, velocity, and acceleration. The formula for calculating Huber loss is:

[0025] In the formula, The threshold parameter for Huber loss. This is the measured response value. Predict response values ​​for the model; These are the weighting factors for each loss term. The balance coefficient for global energy conservation loss is denoted by , and all weighting factors are positive real numbers.

[0026] As one possible implementation, further, in step S4 of this solution, the specific execution steps of the alternating decoupling optimization strategy are as follows: 1) The joint pre-training phase is divided into two sub-steps: A01. The Adam first-order optimizer is used to jointly pre-train the network parameters of the PIGNN model and the stiffness parameters to be inverted. During the training process, a course learning strategy is adopted. In the first half of the iteration cycle, only the displacement data fitting loss is used for optimization. In the second half of the iteration cycle, the velocity and acceleration data fitting loss are added. A02. The LBFGS second-order optimizer is used to jointly fine-tune the network parameters and stiffness parameters to be inverted in the PIGNN model to improve the model convergence accuracy. Among them, the joint pre-training adopts an early stopping mechanism throughout the process, and monitors the validation set loss in real time. When the validation set loss does not decrease for a continuous preset iteration period, the training is terminated in advance and the optimal parameter state is saved. 2) In the decoupling and fine-tuning stage, two alternating sub-steps are executed cyclically to remove the gradient coupling interference between network fitting and stiffness parameter inversion. This includes: B01: Fix all network parameters of the PIGNN model, turn off gradient backpropagation of network parameters, and use only the LBFGS optimizer to iteratively fine-tune the stiffness parameters to be inverted; B02: Fix the stiffness parameters to be inverted, turn off the gradient backpropagation of the stiffness parameters, and use only the LBFGS optimizer to iteratively optimize all network parameters of the PIGNN model; Sub-steps B1 and B2 are executed cyclically for a preset number of rounds. In each iteration, an early stopping mechanism is used to monitor the loss of the validation set and save the globally optimal parameters.

[0027] As a possible implementation, further, in step S4 of this scheme, the physical parameters of the structure to be inverted are physically constrained using the Softplus activation function, and its mapping formula is:

[0028] in, The structural stiffness parameters to be inverted are: For an unconstrained, learnable parameter vector, Let be the initial stiffness vector of the structure in its undamaged state. To prevent numerically singular local minima, the Softplus function is used to ensure the physical non-negativity of the stiffness parameter, and its expression is: , This is the smoothing coefficient.

[0029] Preferably, during the model initialization phase, to ensure that the stiffness of the initial network output is as close as possible to the prior lossless stiffness of the structure, Unconstrained parameters Initialize to satisfy The value, that is This ensures physical stability during the initial training phase and accelerates model convergence.

[0030] As one possible implementation, step S5 of this scheme further includes: based on the physical parameters of each layer of the structure obtained by inversion, calculating the reduction ratio of the physical parameters at each location relative to the undamaged initial state; evaluating the stiffness reduction ratio in a graded manner according to structural design specifications, historical health benchmarks, or set thresholds; setting early damage thresholds and functional damage thresholds; and generating an assessment report containing specific damage locations, damage severity, and structural health status levels by comprehensively considering the spatial distribution of stiffness reduction, the confidence level of inversion results, and structural design safety threshold information, and providing graded early warning for functional damage.

[0031] In some embodiments, the multi-degree-of-freedom target structure includes any one of a multi-story building frame structure, an industrial warehouse rack structure, a bridge pier structure, and a tower structure. By adjusting the topological adjacency matrix and the dynamic control equations, it can be adapted to structures with different degrees of freedom and different topological forms.

[0032] As one possible implementation, the method described in this solution further possesses universal adaptability to different structural forms, including: 1) Physical topology adaptation: By adjusting the topology adjacency matrix The connectivity relationships are used to match the actual physical topology of the target structure, which includes, but is not limited to, multi-story building frames, industrial warehouse racks, bridge piers, towers, and truss structures. 2) Force Mechanism Adaptation: Based on the force characteristics of the target structure, the mass matrix is ​​changed by adjusting the dynamic control equations in the dual-scale physical constraint system. Damping matrix With the stiffness matrix to be inverted The mathematical structure is designed to adapt to structural systems with different degrees of freedom, different boundary conditions, and different dominant force mechanisms.

[0033] Based on the above, this solution also proposes a structural damage identification system based on a global conservation graph physical neural network, which is used to execute the structural damage identification method based on a global conservation graph physical neural network described above. The system includes: The data acquisition and preprocessing module is used to perform the core functions of constructing the graph structure of multi-degree-of-freedom shear-type structures, acquiring dynamic response signals, filtering and denoising, dimensional standardization, and data set partitioning. The physical constraint construction module is used to construct a dual-scale physical regularization framework that combines point-state differential constraints and global integral constraints, and outputs the loss calculation results for each physical constraint term. The network construction module is used to perform the construction, initialization, and forward propagation functions of a physical information graph neural network model that integrates Fourier feature encoding and graph convolution. The model training module is used to construct multi-dimensional composite robust loss functions, execute two-stage alternating decoupling optimization strategies, monitor the model training process, and save optimal parameters. The damage identification output module is used to perform functions such as structural stiffness parameter inversion, damage location and severity identification, structural health status assessment report generation, and damage early warning information output.

[0034] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described structural damage identification method based on a global conservation graph physical neural network.

[0035] The present invention also provides a computer device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described structural damage identification method based on a global conservation graph physical neural network.

[0036] By adopting the above technical solution, the present invention has the following beneficial effects compared with the prior art: 1) Significantly improved accuracy in fitting structural dynamic response and high-frequency capture capability: This invention innovatively constructs a PIGNN architecture that integrates Fourier feature encoding and graph convolution. By explicitly modeling the interlayer physical coupling characteristics of multi-degree-of-freedom sheared structures through graph convolution layers, it effectively alleviates the deficiency of traditional MLP-PINN in capturing the spatial topological relationships of structures. By mapping the one-dimensional time input to a high-dimensional periodic feature space through Fourier feature encoding, it effectively alleviates the inherent "spectral bias" problem of traditional PINN. The fitting capability of high-frequency vibration signals of structures is significantly improved compared with traditional MLP-PINN. Even under 25% extreme high noise environment, the determination coefficient R² of displacement response fitting can still remain high, and the sensitivity to early minor damage of structures is significantly improved.

[0037] 2) Extremely robust in high-noise environments, effectively addressing the convergence pain point of non-physical solutions: This invention constructs a dual-scale physical regularization framework that combines point-state differential constraints and global integral constraints. Based on traditional point-state PDE constraints, it introduces integral constraints of global momentum conservation and global energy conservation. The integral form of global constraints has natural noise suppression characteristics. Random noise at a single time point will cancel each other out after summing in the entire time domain, avoiding the residual calculation distortion problem caused by single-point noise. Combined with the suppression effect of Huber robust loss function on outliers, the relative error of structural stiffness parameter inversion is always controlled at an extremely low level in a wide noise range of 0-25%. The noise interference resistance is significantly improved compared with the traditional single-constraint PINN, solving the industry pain point that the model is prone to baseline drift and convergence to non-physical solutions in high-noise environments.

[0038] 3) Stable convergence during training and significantly improved parameter inversion accuracy: This invention designs a two-stage alternating decoupling optimization strategy, which realizes the alternating independent optimization of network weights and stiffness parameters through gradient switch control. This effectively solves the ill-conditioned problem of gradient coupling between large network weights and small stiffness parameters in the traditional joint optimization mode, avoids the problem of physical parameters deviating due to network fitting noise, and improves the training convergence speed compared with the traditional joint optimization. The stability and accuracy of the inversion results are higher than those of the existing technology.

[0039] 4) Strong physical consistency and physical interpretability of the identification results: Starting from the conservation laws at the bottom level of structural dynamics, this invention constructs a physical constraint system of "local accuracy guarantee + global physical compliance", which reduces the solution space of the inverse dynamic problem. At the same time, the Softplus function applies physical non-negative constraints to the stiffness parameters, which effectively alleviates the inversion results such as negative stiffness that violate physical common sense from the training mechanism. Every result output by the model strictly follows the physical laws of structural dynamics, which effectively alleviates the defects of pure data-driven model black box output and lack of physical interpretability. The reliability of the identification results meets the requirements of engineering safety assessment.

[0040] 5) High engineering practicality and extremely low dependence on labeled samples: This invention supplements the missing information in the data through physical constraints, eliminating the need for massive amounts of labeled damage samples. Only a small amount of structural vibration response data is required to complete the high-precision stiffness parameter inversion, making it suitable for real-world scenarios with small samples and high noise in engineering measurements. At the same time, the technical solution of this invention has strong versatility. By adjusting the topological adjacency matrix and dynamic control equations, it can be adapted to multi-degree-of-freedom shear structures with different numbers of layers and different topological forms without reconstructing the core network architecture. It can be applied to long-term online health monitoring of various structures such as multi-story buildings, industrial shelves, bridge piers, and towers, demonstrating its engineering application value and promising prospects for promotion. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0042] Figure 1 This is an overall flowchart of the structural damage identification method based on a global conservation graph physical neural network as described in this invention; Figure 2 is a schematic diagram of the architecture of the PIGNN physical information graph neural network model described in this invention; Figure 3 is a schematic diagram of the principle of the dual-scale physical constraint system described in this invention.

[0043] Figure 4 is a diagram showing the connection of the unit modules described in this invention. Detailed Implementation

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Combination Figures 1 to 3 As shown in one embodiment, this scheme proposes a structural damage identification method based on a global conservation graph physical neural network, which includes the following steps: S1. Graph Structure Construction and Dynamic Signal Acquisition Preprocessing By collecting dynamic response signals of the target structure (a multi-degree-of-freedom shear-type target structure) through a sensor array deployed on the key degrees of freedom of the target structure, the multi-degree-of-freedom structure (a multi-degree-of-freedom shear-type target structure) is abstracted into a graph structure, defined as follows: In this context, node V represents the independent degrees of freedom of the structure, and the node attributes are the time-series characteristics of the dynamic response at the corresponding degree of freedom position. Based on the physical coupling relationship between adjacent degrees of freedom of the structure, the edge E and the standardized adjacency matrix of the graph structure are constructed, and the collected raw time-series signals are preprocessed by filtering and denoising, standardizing the dimensions, and dividing the dataset. As an example, in step S1 of this scheme, the dynamic response signal includes displacement, velocity, and acceleration vibration signals of each degree of freedom of the structure, and the data sampling frequency is not less than 2.5 times the highest natural frequency of the structure. The preprocessing specifically includes: using Butterworth low-pass filtering and Savitzky-Golay smoothing filtering in a cascade to suppress high-frequency observation noise and environmental interference; using Z-score normalization to eliminate the influence of different dimensions of response data; and dividing the preprocessed dataset into training set, validation set and test set in a 6:2:2 ratio, maintaining the temporal continuity of the time series and not disrupting the order of the time dimensions.

[0046] Furthermore, in step S1 of this scheme, for multi-degree-of-freedom structures, the specific method for constructing the standardized adjacency matrix is ​​as follows: Based on the unit nodes and the physical coupling relationships of the degrees of freedom in a multi-degree-of-freedom structure, a topological adjacency matrix A is constructed, where the matrix elements... Representing the i The and the first j There is a direct interlayer shear coupling relationship between each degree of freedom; otherwise... ; Construct a degree matrix D based on the topological adjacency matrix, where the diagonal elements of the degree matrix are... n is the total number of degrees of freedom of the structure, and all off-diagonal elements are 0; The standardized adjacency matrix is ​​constructed using a symmetric normalization method, and the calculation formula is as follows:

[0047] in, It is the inverse square root matrix of the degree matrix.

[0048] S2, Construction of a dual-scale physical constraint system Based on the dynamic characteristics of the target structure (multi-degree-of-freedom shear structure), a dual-scale physical regularization framework including point-state differential constraints and global integral constraints is constructed; the point-state differential constraints are the residual constraints of the dynamic control equations of the multi-degree-of-freedom shear structure, and the global integral constraints include at least the global momentum conservation constraint and the global energy conservation constraint of the system. As an example, step S2 of this solution includes: 1) Establish point-state dynamic control equation constraints. Based on the dynamic characteristics of the target structure, the control equation expression is constructed as follows:

[0049] Where M is the mass matrix of the structure, C is the modal damping matrix of the structure, and K is the stiffness matrix of the structure to be inverted. These are the displacement, velocity, and acceleration vectors for each degree of freedom of the structure. The external excitation vector acting on the structure; Based on the governing equations, the residual loss of the point-state dynamic equations in the entire time domain is constructed, and the expression is as follows:

[0050] in, For the first i One degree of freedom in t The residuals of the dynamic equations at time N are the total number of time sampling points and n is the total number of degrees of freedom of the structure. 4) Establish a global momentum conservation constraint, and construct the global momentum conservation loss based on the residuals of the dynamic equations in the entire time domain. The expression is as follows:

[0051] 5) Establish global energy conservation constraints. Based on the energy balance relationship between structural kinetic energy change, external force work, and damping dissipation, construct the global energy conservation loss, expressed as:

[0052] in, Let be the velocity vectors of each degree of freedom of the structure at time t. Let be the residual vector of the dynamic equations for each degree of freedom of the structure at time t.

[0053] S3, Physical Information Graph Neural Network Architecture Construction A Physical Information Graph Neural Network (PIGNN) model is constructed, which integrates Fourier feature encoding layers and graph convolutional layers. The input of the PIGNN model is normalized time series and structural graph topological information, and the output is the predicted displacement response value of each degree of freedom of the structure. The PIGNN model includes a series of cascaded Fourier feature encoding layers, temporal embedding layers, multi-layer graph convolutional layers, and node output layers. The preprocessed time series signal and the topological adjacency matrix are input into the pre-constructed physical information graph neural network model to extract spatial topological features and temporal features, and output the predicted response value of each degree of freedom of the structure. In step S3 of this scheme, the physical information graph neural network model includes a cascaded Fourier feature encoding layer, a temporal embedding layer, a multi-layer graph convolutional layer, and a node output layer; the Fourier feature encoding layer is used to map a one-dimensional time variable into a high-dimensional periodic feature vector through a random frequency matrix.

[0054] As an example, in step S3, the specific construction method of each layer of the PIGNN model is as follows: 1) The Fourier feature coding layer is used to map the one-dimensional time input to a high-dimensional periodic feature space, solving the spectral bias problem of traditional PINN. Its mapping formula is as follows:

[0055] in, The matrix is ​​a random frequency matrix, and its elements follow a normal distribution. The output is a high-dimensional Fourier periodic feature; 2) Temporal embedding layer, using a fully connected neural network structure, is used to map high-dimensional Fourier features to node features of the adaptation graph convolutional layer, and the output feature dimension matches the structural degrees of freedom and node feature dimension; 3) Multi-layer graph convolutional layers, employing a two-layer symmetric normalized graph convolutional structure. The forward propagation formula for a single-layer graph convolution is:

[0056] in, Let l be the node feature matrix of the l-th layer. , For the first The learnable weight matrix and bias vector of the layer, where σ is the Tanh nonlinear activation function. For a standardized adjacency matrix; 4) Node output layer: A linear fully connected layer is used to map the node features output by the graph convolutional layer into the displacement response prediction values ​​of each degree of freedom. The output dimension is consistent with the number of structural degrees of freedom.

[0057] S4. Robust Loss Function Design and Model Training A data fitting loss is constructed based on the measured dynamic response signal and the predicted response value. A multi-dimensional composite loss function is constructed by combining the physical constraint loss output by the dual-scale physical regularization framework. An alternating decoupling optimization strategy is used to independently update the network parameters and the physical parameters of the structure to be inverted. Specifically, this can be achieved by constructing a multi-dimensional composite robust loss function and using a two-stage alternating decoupling optimization strategy to complete the end-to-end training of the PIGNN model. The multi-dimensional composite robust loss function is composed of a weighted fusion of Huber robust data fitting loss, point-state dynamic equation residual loss, global momentum conservation loss, and global energy conservation loss. The two-stage alternating decoupling optimization strategy includes a joint pre-training stage and a decoupling fine-tuning stage, which uses gradient switching control to achieve alternating independent optimization of the PIGNN network weights and the stiffness parameters to be inverted. In step S4 of this scheme, the calculation expression for the multi-dimensional composite loss function is as follows:

[0058] in, Huber robust data fitting loss covers the fitting deviation between the measured values ​​and model predictions of three types of responses: displacement, velocity, and acceleration. The formula for calculating Huber loss is:

[0059] In the formula, The threshold parameter for Huber loss. This is the measured response value. Predict response values ​​for the model; These are the weighting factors for each loss term. The balance coefficient for global energy conservation loss is denoted by , and all weighting factors are positive real numbers.

[0060] Furthermore, in step S4 of this scheme, the specific execution steps of the alternating decoupling optimization strategy are as follows: 1) The joint pre-training phase is divided into two sub-steps: A01. The Adam first-order optimizer is used to jointly pre-train the network parameters of the PIGNN model and the stiffness parameters to be inverted. During the training process, a course learning strategy is adopted. In the first half of the iteration cycle, only the displacement data fitting loss is used for optimization. In the second half of the iteration cycle, the velocity and acceleration data fitting loss are added. A02. The LBFGS second-order optimizer is used to jointly fine-tune the network parameters and stiffness parameters to be inverted in the PIGNN model to improve the model convergence accuracy. Among them, the joint pre-training adopts an early stopping mechanism throughout the process, and monitors the validation set loss in real time. When the validation set loss does not decrease for a continuous preset iteration period, the training is terminated in advance and the optimal parameter state is saved. 2) In the decoupling and fine-tuning stage, two alternating sub-steps are executed cyclically to remove the gradient coupling interference between network fitting and stiffness parameter inversion. This includes: B01: Fix all network parameters of the PIGNN model, turn off gradient backpropagation of network parameters, and use only the LBFGS optimizer to iteratively fine-tune the stiffness parameters to be inverted; B02: Fix the stiffness parameters to be inverted, turn off the gradient backpropagation of the stiffness parameters, and use only the LBFGS optimizer to iteratively optimize all network parameters of the PIGNN model; Sub-steps B1 and B2 are executed cyclically for a preset number of rounds. In each iteration, an early stopping mechanism is used to monitor the loss of the validation set and save the globally optimal parameters.

[0061] As a possible implementation, further, in step S4 of this scheme, the physical parameters of the structure to be inverted are physically constrained using the Softplus activation function, and its mapping formula is:

[0062] in, The structural stiffness parameters to be inverted are: For an unconstrained, learnable parameter vector, Let be the initial stiffness vector of the structure in its undamaged state. To prevent numerically singular local minima, the Softplus function is used to ensure the physical non-negativity of the stiffness parameter, and its expression is: , This is the smoothing coefficient.

[0063] Preferably, during the model initialization phase, to ensure that the stiffness of the initial network output is as close as possible to the prior lossless stiffness of the structure, Unconstrained parameters Initialize to satisfy The value, that is This ensures physical stability during the initial training phase and accelerates model convergence.

[0064] S5, Intelligent Output of Structural Damage Information Based on the structural physical parameters obtained by inversion, damage identification indicators are constructed to identify the spatial location and severity of structural damage. Specifically, this can be done by inputting the real-time dynamic response data of the structure to be tested into the trained PIGNN model, obtaining the stiffness parameters of each layer of the structure through model forward propagation and physical constraint inversion, identifying the spatial location and severity of structural damage based on the reduction ratio of the stiffness parameters relative to the undamaged initial state of the structure, and finally generating structural health status assessment results and damage warning information.

[0065] Step S5 of this scheme specifically includes: based on the physical parameters of each layer of the structure obtained by inversion, calculating the reduction ratio of the physical parameters at each location relative to the undamaged initial state; evaluating the stiffness reduction ratio in a graded manner according to structural design specifications, historical health benchmarks, or set thresholds; setting early damage thresholds and functional damage thresholds; and generating an assessment report containing specific damage locations, damage severity, and structural health status levels by comprehensively considering the spatial distribution of stiffness reduction, the confidence level of inversion results, and structural design safety threshold information, and providing graded early warnings for functional damage.

[0066] In some embodiments, the multi-degree-of-freedom target structure includes any one of a multi-story building frame structure, an industrial warehouse rack structure, a bridge pier structure, and a tower structure. By adjusting the topological adjacency matrix and the dynamic control equations, it can be adapted to structures with different degrees of freedom and different topological forms.

[0067] Furthermore, the method described in this solution possesses universal adaptability to different structural forms, including: 1) Physical topology adaptation: By adjusting the topology adjacency matrix The connectivity relationships are used to match the actual physical topology of the target structure, which includes, but is not limited to, multi-story building frames, industrial warehouse racks, bridge piers, towers, and truss structures. 2) Force Mechanism Adaptation: Based on the force characteristics of the target structure, the mass matrix is ​​changed by adjusting the dynamic control equations in the dual-scale physical constraint system. Damping matrix With the stiffness matrix to be inverted The mathematical structure is designed to adapt to structural systems with different degrees of freedom, different boundary conditions, and different dominant force mechanisms.

[0068] As an example of an implementation, the present solution provides the following simulation example to further illustrate the aforementioned solution.

[0069] The core inventive concept of this example solution is: to address the need for damage recognition of multiple structures, a PIGNN architecture that integrates Fourier feature encoding and graph convolution is constructed to adapt to the spatial topological coupling characteristics of the structure and solve the spectral bias problem of traditional PINN. In addition, a dual-scale physical regularization framework is established that combines point-state PDE constraints with global momentum-energy conservation constraints to ensure the physical consistency and high noise robustness of the inversion results. Furthermore, by designing a two-stage alternating decoupling optimization strategy, the ill-conditioned problem of gradient coupling is solved, training stability and parameter inversion accuracy are improved, and finally, accurate structural damage identification is achieved in high-noise, small-sample engineering scenarios.

[0070] This example uses a 4-DOF shear-type bookshelf structure as the experimental object, which includes the following: S1. Experimental Subjects and Parameter Settings The experimental subject in this embodiment is a 4-layer shear-type bookshelf structure, and its basic structural parameters are as follows: Mass parameters: The masses of the 4 layers of mass blocks are respectively , , , Structural mass matrix

[0071] Stiffness parameter: The initial inter-story stiffness of the structure in a lossless state is , , , The damage condition is set so that the stiffness of the first and second layers decreases by 50%, meaning the stiffness after damage is... , , , ; Damping parameters: Rayleigh damping is adopted, and the damping ratio is set to ξ=0.02. The modal damping matrix C is calculated based on the mass matrix and stiffness matrix of the structure in a lossless state. External excitation: Simple harmonic hybrid excitation is adopted, and the excitation form is as follows: The excitation is applied to the mass block of the first layer of the structure.

[0072] S2, Sensor Deployment and Data Acquisition Sensor deployment: One laser displacement sensor and one accelerometer are deployed on each side of the four-layer mass block. The sensor sensing axis is aligned with the direction of structural vibration to ensure measurement accuracy. The laser displacement sensor has a resolution of 0.01 mm and a sampling rate of 1000 Hz. The accelerometer has a range of ±5 g and a sampling rate of 1000 Hz. Data Acquisition: Vibration signals from four displacement channels and four acceleration channels were simultaneously acquired using an 8-channel synchronous sampling data acquisition card. The acquisition duration was set to 3 seconds, consistent with the simulation parameters. Vibration data were acquired at six noise levels: 0%, 1%, 5%, 10%, 15%, and 25%. Five sets of valid data were repeatedly acquired for each noise level. The noise level was achieved by introducing controllable environmental interference into the experimental environment. Data preprocessing: The acquired raw signals were preprocessed. First, a cascaded 4th-order Butterworth low-pass filter (cutoff frequency 15Hz) and Savitzky-Golay smoothing filter (window length 15, polynomial order 3) were used to suppress high-frequency noise. Then, the Z-score normalization method was used to eliminate the influence of data dimensions. Finally, the preprocessed dataset was divided into training set, validation set and test set in a 6:2:2 ratio, while maintaining the continuity of the time series during the partitioning process.

[0073] S3, Structure Construction and Normalized Adjacency Matrix Calculation To address the interlayer coupling characteristics of a 4-DOF shear structure, a topological adjacency matrix A is constructed:

[0074] Calculate the degree matrix based on the topological adjacency matrix. The final normalized adjacency matrix is ​​calculated as follows:

[0075] S4. Building the PIGNN model The specific structure of the PIGNN model built in this embodiment is as follows: Fourier feature coding layer: The random frequency matrix B has a dimension of 1×32, and its elements follow a certain order. The normal distribution maps a one-dimensional time input to a 64-dimensional high-dimensional Fourier feature. Temporal embedding layer: A two-layer fully connected network is used. The first layer has an input dimension of 64 and an output dimension of 128, with the activation function being Tanh. The second layer has an input dimension of 128 and an output dimension of 4×64=256, with the activation function being Tanh. The output features are reshaped into a 4×64 node feature matrix. Graph Convolutional Layers: There are 2 graph convolutional layers in total. The first layer has an input feature dimension of 64 and an output feature dimension of 64, with the activation function being Tanh. The second layer has an input feature dimension of 64 and an output feature dimension of 32, with the activation function being Tanh. Node output layer: Linear fully connected layer, with an input feature dimension of 32 and an output dimension of 1, ultimately outputting the displacement response prediction value with 4 degrees of freedom.

[0076] S5, Construction of a Dual-Scale Physical Constraint System Based on the dynamic characteristics of a 4-DOF shear structure, a dual-scale physical constraint system is constructed: Point-state dynamics governing equation constraints: Construct the dynamics governing equations for a 4-DOF shear structure, calculate the dynamics residuals over the entire time domain, and obtain the point-state PDE loss. ; Global momentum conservation constraint: Calculate the global momentum conservation loss based on the residuals of the dynamic equations across the entire time domain. ; Global energy conservation constraints: Calculate global energy conservation losses based on structural energy balance relationships. .

[0077] S6. Loss Function Setting and Model Training Multi-dimensional composite robust loss function settings: weight factors are set to... , , β=0.1, Huber loss threshold δ=0.1; Two-stage alternating decoupling optimization strategy settings: Joint pre-training phase: Adam optimizer iterations 8000 times, learning rate 1e-3, course learning ratio 0.5; LBFGS optimizer iterations 5000 times, learning rate 0.1; early stopping mechanism patience value set to 500, minimum change 1e-6. Decoupling and fine-tuning phase: 5 iterations in the outer loop, 200 iterations of LBFGS for stiffness parameter optimization, and a learning rate of 0.05; 50 iterations of LBFGS for network parameter optimization, and a learning rate of 0.1. Non-negativity constraint for stiffness parameters: The Softplus function is used to implement the non-negativity constraint for stiffness parameters, with a smoothing coefficient β=5.0 to prevent singular minima. ; Training environment: The model was built and trained using the Python programming language and the PyTorch deep learning framework. The hardware environment consisted of an NVIDIA RTX 5070 Laptop GPU and an AMD R9-7845HX CPU.

[0078] S7, Damage Information Output The real-time dynamic response data of the actual structure to be tested, along with graph topology information, is input into a trained Physical Information Graph Neural Network (PIGNN) model. The model outputs predicted dynamic response values ​​through forward propagation, and iterative inversion is performed using a dual-scale physical constraint system to obtain the current true stiffness parameters at each level of the structure. Then, based on the reduction ratio of the inverted stiffness parameters relative to the undamaged initial state of the structure, the spatial location of structural damage is accurately identified, and the severity of the damage is quantitatively assessed.

[0079] After the model is trained, the stiffness inversion results under different noise levels are statistically analyzed, and the damage location, degree and overall confidence level are finally output to generate an evaluation report.

[0080] Based on the above, this solution also proposes a structural damage identification system based on a global conservation graph physical neural network, which is used to execute the structural damage identification method based on a global conservation graph physical neural network described above. The system includes: The data acquisition and preprocessing module is used to perform the core functions of constructing the graph structure of multi-degree-of-freedom shear-type structures, acquiring dynamic response signals, filtering and denoising, dimensional standardization, and data set partitioning. The physical constraint construction module is used to construct a dual-scale physical regularization framework that combines point-state differential constraints and global integral constraints, and outputs the loss calculation results for each physical constraint term. The network construction module is used to perform the construction, initialization, and forward propagation functions of a physical information graph neural network model that integrates Fourier feature encoding and graph convolution. The model training module is used to construct multi-dimensional composite robust loss functions, execute two-stage alternating decoupling optimization strategies, monitor the model training process, and save optimal parameters. The damage identification output module is used to perform functions such as structural stiffness parameter inversion, damage location and severity identification, structural health status assessment report generation, and damage early warning information output.

[0081] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A structural damage identification method based on a global conservation graph physical neural network, characterized in that, It includes the following steps: S1. By collecting dynamic response signals of the structure through a sensor array deployed on the key degrees of freedom of the target structure, the multi-degree-of-freedom structure is abstracted into a graph structure, defined as follows: In this context, node V represents the independent degrees of freedom of the structure, and the node attributes are the time-series characteristics of the dynamic response at the corresponding degree of freedom position. Based on the physical coupling relationship between adjacent degrees of freedom of the structure, the edge E and the standardized adjacency matrix of the graph structure are constructed, and the collected raw time-series signals are preprocessed by filtering and denoising, standardizing the dimensions, and dividing the dataset. S2. Based on the dynamic characteristics of the target structure, a dual-scale physical regularization framework including point-state differential constraints and global integral constraints is constructed; the point-state differential constraints are residual constraints of the dynamic control equations, and the global integral constraints include at least the system's global momentum conservation constraints and global energy conservation constraints. S3. Input the preprocessed time-series signal and the topological adjacency matrix into the pre-constructed physical information graph neural network model, extract spatial topological features and temporal features, and output the response prediction values ​​of each degree of freedom of the structure. S4. Construct a data fitting loss based on the measured dynamic response signal and the predicted response value, and construct a multi-dimensional composite loss function by combining the physical constraint loss output by the dual-scale physical regularization framework; use an alternating decoupling optimization strategy to independently update the network parameters and the physical parameters of the structure to be inverted. S5. Based on the structural physical parameters obtained from the inversion, construct damage identification indicators to identify the spatial location and severity of structural damage.

2. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, In step S1, for multi-degree-of-freedom structures, the method for constructing the standardized adjacency matrix is ​​as follows: Based on the unit nodes and the physical coupling relationships of the degrees of freedom in a multi-degree-of-freedom structure, a topological adjacency matrix A is constructed, where the matrix elements... Representing the i The and the first j There is a direct interlayer shear coupling relationship between each degree of freedom; otherwise... ; Construct a degree matrix D based on the topological adjacency matrix, where the diagonal elements of the degree matrix are... n is the total number of degrees of freedom of the structure, and all off-diagonal elements are 0; The standardized adjacency matrix is ​​constructed using a symmetric normalization method, and the calculation formula is as follows: in, It is the inverse square root matrix of the degree matrix.

3. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, Step S2 includes: 1) Establish point-state dynamic control equation constraints. Based on the dynamic characteristics of the target structure, the control equation expression is constructed as follows: Where M is the mass matrix of the structure, C is the modal damping matrix of the structure, and K is the stiffness matrix of the structure to be inverted. These are the displacement, velocity, and acceleration vectors for each degree of freedom of the structure. The external excitation vector acting on the structure; Based on the governing equations, the residual loss of the point-state dynamic equations in the entire time domain is constructed, and the expression is as follows: in, For the first i One degree of freedom in t The residuals of the dynamic equations at time N are the total number of time sampling points and n is the total number of degrees of freedom of the structure. 2) Establish a global momentum conservation constraint, and construct the global momentum conservation loss based on the residuals of the dynamic equations in the entire time domain. The expression is as follows: 3) Establish global energy conservation constraints. Based on the energy balance relationship between structural kinetic energy changes, external force work, and damping dissipation, construct the global energy conservation loss, expressed as: in, Let be the velocity vectors of each degree of freedom of the structure at time t. Let be the residual vector of the dynamic equations for each degree of freedom of the structure at time t.

4. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, In step S3, the physical information graph neural network model includes a cascaded Fourier feature encoding layer, a temporal embedding layer, a multi-layer graph convolutional layer, and a node output layer; the Fourier feature encoding layer is used to map a one-dimensional time variable into a high-dimensional periodic feature vector through a random frequency matrix.

5. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, In step S4, the calculation expression for the multi-dimensional composite loss function is as follows: in, Huber robust data fitting loss covers the fitting deviation between the measured values ​​and model predictions of three types of responses: displacement, velocity, and acceleration. The formula for calculating Huber loss is: In the formula, The threshold parameter for Huber loss. This is the measured response value. Predict response values ​​for the model; These are the weighting factors for each loss term. The balance coefficient for global energy conservation loss is denoted by , and all weighting factors are positive real numbers.

6. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, In step S4, the specific execution steps of the alternating decoupling optimization strategy are as follows: 1) The joint pre-training phase is divided into two sub-steps: A01. The Adam first-order optimizer is used to jointly pre-train the network parameters of the PIGNN model and the stiffness parameters to be inverted. During the training process, a course learning strategy is adopted. In the first half of the iteration cycle, only the displacement data fitting loss is used for optimization. In the second half of the iteration cycle, the velocity and acceleration data fitting loss are added. A02. The LBFGS second-order optimizer is used to jointly fine-tune the network parameters and stiffness parameters to be inverted in the PIGNN model to improve the model convergence accuracy. Among them, the joint pre-training adopts an early stopping mechanism throughout the process, monitors the validation set loss in real time, and terminates the training early and saves the optimal parameter state when the validation set loss has not decreased for a continuous preset iteration period. 2) In the decoupling and fine-tuning stage, two alternating sub-steps are executed cyclically to remove the gradient coupling interference between network fitting and stiffness parameter inversion. This includes: B01: Fix all network parameters of the PIGNN model, turn off gradient backpropagation of network parameters, and use only the LBFGS optimizer to iteratively fine-tune the stiffness parameters to be inverted; B02: Fix the stiffness parameters to be inverted, turn off the gradient backpropagation of the stiffness parameters, and use only the LBFGS optimizer to iteratively optimize all network parameters of the PIGNN model; Sub-steps B1 and B2 are executed cyclically for a preset number of rounds. In each iteration, an early stopping mechanism is used to monitor the loss of the validation set and save the globally optimal parameters.

7. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, In step S4, the physical parameters of the structure to be inverted are physically constrained using the Softplus activation function, and the mapping formula is as follows: in, The structural stiffness parameters to be inverted are: For an unconstrained, learnable parameter vector, Let be the initial stiffness vector of the structure in a lossless state. To prevent numerically singular minimum constants, the Softplus function is used to ensure the physical non-negativity of the stiffness parameter, and its expression is: , This is the smoothing coefficient.

8. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, Step S5 specifically includes: based on the physical parameters of each layer of the structure obtained by inversion, calculating the reduction ratio of the physical parameters at each location relative to the undamaged initial state; evaluating the stiffness reduction ratio in a graded manner according to structural design specifications, historical health benchmarks, or set thresholds; setting early damage thresholds and functional damage thresholds; and generating an assessment report containing specific damage locations, damage severity, and structural health status levels by comprehensively considering the spatial distribution of stiffness reduction, the confidence level of the inversion results, and structural design safety threshold information, and providing graded early warnings for functional damage.

9. The structural damage identification method based on a global conservation graph physical neural network according to claim 1, characterized in that, The method possesses universal adaptability to different structural forms, including: 1) Physical topology adaptation: By adjusting the topology adjacency matrix The connectivity relationships are used to match the actual physical topology of the target structure, which includes, but is not limited to, multi-story building frames, industrial warehouse racks, bridge piers, towers, and truss structures. 2) Force Mechanism Adaptation: Based on the force characteristics of the target structure, the mass matrix is ​​changed by adjusting the dynamic control equations in the dual-scale physical constraint system. Damping matrix With the stiffness matrix to be inverted The mathematical structure is designed to adapt to structural systems with different degrees of freedom, different boundary conditions, and different dominant force mechanisms.

10. A structural damage identification system based on a global conservation graph physical neural network, characterized in that, The system is used to perform the structural damage identification method based on a global conservation graph physical neural network according to any one of claims 1-9, the system comprising: The data acquisition and preprocessing module is used to perform the core functions of constructing the graph structure of multi-degree-of-freedom shear-type structures, acquiring dynamic response signals, filtering and denoising, dimensional standardization, and data set partitioning. The physical constraint construction module is used to construct a dual-scale physical regularization framework that combines point-state differential constraints and global integral constraints, and outputs the loss calculation results for each physical constraint term. The network construction module is used to perform the construction, initialization, and forward propagation functions of a physical information graph neural network model that integrates Fourier feature encoding and graph convolution. The model training module is used to construct multi-dimensional composite robust loss functions, execute two-stage alternating decoupling optimization strategies, monitor the model training process, and save optimal parameters. The damage identification output module is used to perform functions such as structural stiffness parameter inversion, damage location and severity identification, structural health status assessment report generation, and damage early warning information output.

Images