Large-span space steel structure damage identification method and system combining AI and numerical simulation

Abstract

The invention discloses an AI and numerical simulation combined large-span space steel structure damage identification method and system, and particularly relates to the technical field of building monitoring, and the method comprises the following steps: P1, according to a structural design drawing and material characteristics, constructing a large-span steel structure finite element simulation model, and simulating the response of the structure under various load and boundary conditions; p2, pre-training a damage identification model according to the obtained response data, and simulating multi-scale damage evolution in combination with numerical simulation data and experimental observation data; according to the method, the problem of no physical interpretation of a pure data driving model under complex working conditions is avoided, the credibility and engineering applicability of a prediction result are enhanced, the overall process characteristics of structural damage are effectively captured, early recognition and trend prediction are facilitated, the recognition performance under a small sample is improved, and the recognition efficiency is improved. Error accumulation and efficiency loss caused by multi-link splitting in a traditional method are avoided.

Description

Technical Field

[0001] This invention relates to the field of building monitoring technology, specifically to a method and system for identifying damage in large-span spatial steel structures that combines AI and numerical simulation. Background Technology

[0002] With the increasing scale and intelligence of urban infrastructure and public buildings, large-span steel structures are widely used in key facilities such as stadiums, convention centers, transportation hubs, and industrial plants due to their excellent load-bearing performance and spatial adaptability. These structures typically feature large spans, numerous components, complex nodes, and variable operating environments. They are highly susceptible to localized damage, loose connections, or member failures due to loads, corrosion, temperature and humidity changes, or fatigue. Failure to identify and address these issues promptly can lead to cascading damage or even structural collapse, causing significant safety accidents and economic losses. In the field of structural health monitoring, traditional damage identification methods mainly include frequency domain analysis, modal identification, strain / displacement threshold judgment, and numerical inversion techniques based on physical models. However, these methods rely on structural simplification assumptions, manual feature selection, and linear modeling, making it difficult to capture the complex nonlinear characteristics and multi-scale transmission effects during damage evolution. Furthermore, their accuracy is unstable under complex environmental interference.

[0003] Existing methods and systems for identifying damage in large-span spatial steel structures suffer from the problem of "no physical interpretation" in complex working conditions due to purely data-driven models. This reduces the reliability and engineering applicability of the prediction results, fails to capture the full-process characteristics of structural damage, hinders early identification and trend prediction, and reduces identification performance under small sample conditions. To address this, we propose a method and system for identifying damage in large-span spatial steel structures that combines AI and numerical simulation. Summary of the Invention

[0004] The purpose of this invention is to solve the problem, and to propose a method and system for damage identification of large-span spatial steel structures that combines AI and numerical simulation.

[0005] In a first aspect of this invention, a method and system for damage identification of large-span spatial steel structures combining AI and numerical simulation are proposed, the method comprising:

[0006] P1. Based on the structural design drawings and material properties, construct a finite element simulation model of a large-span steel structure to simulate the structure's response under various load and boundary conditions.

[0007] P2. Based on the acquired response data, a damage identification model is pre-trained, and multi-scale damage evolution is simulated by combining numerical simulation data and experimental observation data.

[0008] P3. Through various sensors, real-time response data of large-span steel structures under service environment is collected, and the monitored real-time response data is identified and decoupled.

[0009] P4. Simulate the propagation path of ultrasonic waves or stress waves in the structure, and extract and identify the feature information in each real-time response data after decoupling.

[0010] P5. Dynamically adjust the sensor network topology and identify the spatial distribution, severity, and development trend of structural damage based on the monitored real-time response data.

[0011] As a further aspect of the present invention, the specific steps for constructing the finite element simulation model of the large-span steel structure described in P1 are as follows:

[0012] S1.1: Obtain the geometric information of the current large-span steel structure based on the structural design drawings, establish a three-dimensional geometric model of the real structure based on the collected geometric information of the current large-span steel structure, then select the corresponding element type and mesh density, perform meshing on the established three-dimensional geometric model, define its corresponding constitutive relation according to the performance parameters of the steel, and analyze the stress of the current large-span steel material through a nonlinear material model.

[0013] S1.2: Based on the supports, connectors and constraint types in the actual large-span steel structure, set the corresponding support conditions for the three-dimensional geometric model after meshing, and set the load types and distribution methods of the current large-span steel structure under various working conditions;

[0014] S1.3: Using the basic governing equations of the finite element method, under preset load and boundary conditions, obtain the nodal displacements, component stresses and reactions in the current three-dimensional geometric model of the large-span steel structure to generate the corresponding finite element simulation model. At the same time, extract the displacement, stress, reaction and deformation data of the nodes as a reference baseline for the structural integrity state.

[0015] As a further aspect of the present invention, the geometric information of the large-span steel structure in S1.1 specifically includes component length, cross-sectional dimensions, and node connection methods, etc.

[0016] The element types mentioned in S1.1 specifically include beam elements, shell elements, and solid elements, etc.

[0017] The load types mentioned in S1.2 specifically include dead load, live load, wind load, and seismic load.

[0018] As a further aspect of the present invention, the specific steps of the pre-trained damage recognition model described in P2 are as follows:

[0019] S2.1: Using the finite element simulation model of the corresponding large-span steel structure, simulate the response of the corresponding large-span steel structure under various loads and boundary conditions, and organize the simulated response data and historical real response data into a structural response dataset. Then, divide the structural response dataset into a training set and a validation set, and then establish a damage identification model based on the Transformer architecture.

[0020] S2.2: Initialize the parameters of the damage identification model. Based on the simulated response data and historical real response data, use the mean square error loss function as the data loss term for supervised learning of the damage identification model. Then, based on the basic equations in elasticity, construct the corresponding physical consistency loss function and use the physical consistency loss function as the physical constraint loss term for supervised learning of the damage identification model. Finally, linearly combine the data loss term and the physical constraint loss term to establish the total loss function of the damage identification model.

[0021] S2.3: Divide the data in the training set into time series of a preset length, perform linear mapping and position encoding on each time series, and then input the processed time series into the damage recognition model. The damage recognition model calculates the information correlation between each input time series based on the multi-head self-attention mechanism, extracts the dynamic dependency features between different time steps, and then processes each group of dynamic dependency features through the residual connection layer and the feedforward network in sequence, and generates the corresponding feedforward output.

[0022] S2.4: Transmit the feedforward output of each group to the regression or classification head and predict the damage information of the corresponding large-span steel structure. Then, calculate the loss value between the prediction result and the actual structural response through the total loss function, and use the backpropagation algorithm to pass the loss value from the output layer to the input layer of the damage identification model layer by layer. The gradient of each layer of the damage identification model is calculated by automatic differentiation, and then the network parameters of each layer are updated by the Adam optimizer.

[0023] S2.5: After each round of training, the validation set data is input into the damage identification model, the prediction result is output through forward propagation, and the total loss function is used to calculate the loss value between the prediction result and the real structural response data. If the loss value is higher than the preset threshold, the damage identification model is retrained and the training and update are repeated until the loss value of the damage identification model on the validation set converges to the preset range, and the weight parameters of the trained damage identification model are saved.

[0024] As a further aspect of the present invention, the specific calculation formula for the total loss function described in S2.2 is as follows:

[0025]

[0026] L total =λ1Ldata +λ2L phys

[0027] In the formula, L data N1 represents the number of data points in the current structural response; The predicted structural response value representing the i1th structural response data; L represents the i1th measured or simulated true structural response value; phys M1 represents the physical constraint loss term; M1 represents the current number of calculation points for the large-span steel structure. This represents the predicted equivalent flux density at the j1-th calculation point; This represents the coordinates of the j1th calculation point; This represents the actual applied body force at the j1st calculation point in the current large-span steel structure; L represents the predicted mass force or inertial response term at the j1-th calculation point; total λ represents the total loss value; λ1 and λ2 represent the weighting factors of the data loss term and the physical constraint loss term, respectively.

[0028] As a further aspect of the present invention, the specific steps for simulating multi-scale damage evolution described in P2 are as follows:

[0029] S3.1: During the numerical simulation of the finite element simulation model, damage variables at each micro-level, such as microcrack density, void ratio, and particle debonding rate, are extracted to form a microscale damage state tensor. Then, through statistical averaging, the damage variables at each micro-level are mapped to the mesoscale physical damage field.

[0030] S3.2: The mesoscale physical damage field is subjected to feature compression processing through principal component analysis or autoencoder, the changes in the mesoscale physical damage field during the simulation process are monitored in real time, and the macroscopic response of the overall structure is calculated.

[0031] S3.3: By mapping micro-scale variables, meso-scale variables and macro-scale responses end-to-end, a multi-scale mapping network structure corresponding to large-span steel structures is established to form a complete damage evolution chain. Then, the prediction results of the damage identification model are compared with simulated response data or historical real response data, and the multi-scale mapping parameters are corrected based on the comparison results.

[0032] In a second aspect of this invention, a damage identification system for large-span spatial steel structures combining AI and numerical simulation is proposed, comprising: a modeling and simulation module, a deployment and acquisition module, an identification and decoupling module, a dynamic modeling module, a special focusing module, a model training module, a collaborative analysis module, a learning and optimization module, a damage assessment module, an early warning and decision-making module, and an interactive display module.

[0033] The modeling and simulation module is used to establish a finite element simulation model of a large-span steel structure to simulate the structural response under various loads, working conditions and damage scenarios.

[0034] The deployment and acquisition module is used to set the sensor deployment and acquisition strategy and to acquire structural operation status data in real time.

[0035] The identification and decoupling module is used to identify and remove interference information in the structural operating status data;

[0036] The dynamic modeling module is used to model the sensor network topology as a graph structure and dynamically update the sensor network graph structure.

[0037] The special focusing module is used to extract damage-sensitive features from real-time monitoring data and analyze weak points in large-span steel structures.

[0038] The model training module is used to pre-train the constructed AI model based on the laws of solid mechanics;

[0039] The collaborative analysis module is used to construct a dual-channel digital twin for collaborative identification of damage to large-span steel structures.

[0040] The learning optimization module generates adversarial examples based on the comparison results of the collaborative analysis module, and optimizes and adjusts the AI ​​model in real time.

[0041] The damage assessment module is used to map the damage identification results to the corresponding large-span steel structure model and assess the severity of the damage.

[0042] The early warning decision module generates early warning information and formulates maintenance and reinforcement strategies based on the location of damage, development trend, and structural safety reserves.

[0043] The interactive display module is used to display the results of each stage in real time through a 3D visualization platform.

[0044] As a further aspect of the present invention, the specific steps of the identification and decoupling module in identifying and removing interference information from the structural operating status data are as follows:

[0045] S4.1: After normalizing the raw monitoring data collected in real time by various sensors, the data are input into the environmental feature channel and the damage feature channel respectively. The environmental feature channel and the damage feature channel use two convolutional encoders with different structures or weights to perform nonlinear mapping on each set of raw monitoring data to obtain the corresponding potential environmental feature vector and potential damage feature vector.

[0046] S4.2: The potential environmental feature vector and the potential damage feature vector are concatenated to obtain the corresponding joint input vector. Then, each joint input vector is transmitted to the shared generator to reconstruct the monitoring data. Finally, the discriminator is used to calculate the difference between the reconstructed monitoring data and the original monitoring data.

[0047] S4.3: Based on the calculated difference value, drive the encoder and shared generator to perform adversarial optimization with the discriminator, and establish an adversarial objective function based on the reconstruction error, the discriminator adversarial loss and the decoupling regularization term. Then, calculate the discriminator loss value through the adversarial objective function. If the loss value is higher than the preset threshold, update the parameters of the discriminator and shared generator based on the loss value, and re-perform feature mapping, feature concatenation, data reconstruction and data discrimination.

[0048] S4.4: Repeat multiple rounds of iterative updates until the discriminator loss value converges to the preset range, then stop the iteration. After that, extract features from the latest monitoring data through the environmental feature channel and the damage feature channel, generate the corresponding joint input vector through splicing, and then use the shared generator to reconstruct the monitoring data to remove environmental interference in the original monitoring data.

[0049] As a further aspect of the present invention, the specific steps of the collaborative analysis module for collaborative identification of damage in large-span steel structures are as follows:

[0050] S5.1: Process various sensors as nodes of a graph, and process the physical connectivity between adjacent sensors, the distance between measurement points, or the signal correlation as edges of the graph to form an initial graph. Then calculate the Euclidean distance between each sensor and initialize the connection weight of each node according to the distance attenuation coefficient.

[0051] S5.2: Real-time monitoring of sensor information deployed on the large-span steel structure. Whenever the sensor status, structural layout, or boundary conditions change, it enters the dynamic graph update state and triggers structural reconstruction. Then, based on the current physical position change or the signal correlation between sensors, a new adjacency matrix is ​​constructed.

[0052] As a further aspect of the present invention, the specific formula for determining the dynamic graph update state in S5.2 is as follows:

[0053]

[0054] In the formula, θ t Nt represents the average rate of change of the node state at time t; N2 represents the total number of sensor nodes. The measurement feature vector represents the i2th node at time t; Represents the measurement feature vector of the i2th node at time t-1; ε represents a small positive number to avoid division by zero; if θ tIf the value is greater than τ, then a graph structure update is triggered, where τ is a preset threshold.

[0055] The beneficial effects of this invention are:

[0056] This invention proposes a method and system for damage identification of large-span spatial steel structures that combines AI and numerical simulation. A structural response dataset is constructed using finite element simulation and historical monitoring data to establish a damage identification model. A total loss function is built by introducing physical constraints. Subsequently, based on multi-scale modeling technology, starting from microscopic damage variables, the model is mapped layer by layer to mesoscale physical fields and macroscopic responses, constructing an end-to-end multi-scale evolution network. Then, a dual-channel encoder and a generative adversarial mechanism are designed to decouple environmental features from damage features, removing environmental interference from real-time monitoring data. Finally, a graph modeling method is used to represent the sensor network structure as a dynamic graph, enabling adaptive sensing and updating of real-time topology changes. This avoids the problem of "no physical interpretation" in complex working conditions caused by purely data-driven models, enhancing the credibility and engineering applicability of prediction results. It effectively captures the full-process characteristics of structural damage, facilitating early identification and trend prediction, improving identification performance under small sample sizes, and avoiding the error accumulation and efficiency loss caused by the fragmentation of multiple stages in traditional methods. Attached Figure Description

[0057] The present invention will now be further described with reference to the accompanying drawings.

[0058] Figure 1 A flowchart illustrating a method for damage identification of large-span spatial steel structures that combines AI and numerical simulation, provided in an embodiment of the present invention.

[0059] Figure 2 A framework diagram of a damage identification system for large-span spatial steel structures that combines AI and numerical simulation, provided in an embodiment of the present invention. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0061] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0062] This invention provides a method and system for damage identification of large-span spatial steel structures that combines AI and numerical simulation. See also... Figure 1 , Figure 1A flowchart illustrating a damage identification method for large-span spatial steel structures combining AI and numerical simulation, provided as an embodiment of the present invention. The method includes the following steps:

[0063] Based on the structural design drawings and material properties, a finite element simulation model of a large-span steel structure was constructed to simulate the structure's response under various load and boundary conditions.

[0064] Specifically, the geometric information of the current large-span steel structure is obtained based on the structural design drawings. A three-dimensional geometric model of the actual structure is then established based on the collected geometric information. Subsequently, the corresponding element type and mesh density are selected to mesh the established three-dimensional geometric model. The corresponding constitutive relation is defined according to the performance parameters of the steel, and the stress of the current large-span steel material is analyzed through a nonlinear material model. Based on the support, connector, and constraint types in the actual large-span steel structure, corresponding support conditions are set for the meshed three-dimensional geometric model. The load types and distribution methods of the current large-span steel structure under various working conditions are set. Through the basic control equations of the finite element method, under preset load and boundary conditions, the nodal displacements, component stresses, and reactions in the three-dimensional geometric model of the current large-span steel structure are obtained to generate the corresponding finite element simulation model. At the same time, the displacement, stress, reaction, and deformation data of each node are extracted as a reference baseline for the structural integrity state.

[0065] It should be further explained that the geometric information of large-span steel structures specifically includes component length, cross-sectional dimensions, and node connection methods; the element types specifically include beam elements, shell elements, and solid elements; and the load types specifically include dead load, live load, wind load, and seismic load.

[0066] Based on the acquired response data, a damage identification model is pre-trained, and multi-scale damage evolution is simulated by combining numerical simulation data and experimental observation data.

[0067] Specifically, a finite element simulation model of a large-span steel structure is used to simulate its response under various loads and boundary conditions. The simulated response data and historical real response data are compiled into a structural response dataset, which is then divided into training and validation sets. A damage identification model is then built based on the Transformer architecture, and its parameters are initialized. Based on the simulated and historical real response data, the mean square error loss function is used as the data loss term for supervised learning of the damage identification model. Then, based on the fundamental equations of elasticity, a physical consistency loss function is constructed and used as the physical constraint loss term for supervised learning of the damage identification model. The data loss term and the physical constraint loss term are then linearly combined to establish the total loss function of the damage identification model. Each set of data in the training set is divided into time series of a preset length, and each time series is linearly mapped and its location encoded. The processed time series are then input into the damage identification model, which is based on a multi-head self-attention mechanism. The system calculates the information correlation between each input time series, extracts dynamic dependency features between different time steps, and then processes each group of dynamic dependency features sequentially through residual connection layers and feedforward networks to generate corresponding feedforward outputs. These feedforward outputs are then transmitted to regression or classification heads to predict damage information for the corresponding large-span steel structure. The total loss function is then used to calculate the loss value between the predicted result and the actual structural response. Backpropagation is used to propagate the loss value layer by layer from the output layer to the input layer of the damage identification model. Automatic differentiation is used to calculate the gradient of each layer of the damage identification model, and the Adam optimizer is used to update the network parameters of each layer. After each training round, validation set data is input into the damage identification model, and the prediction result is output through forward propagation. The total loss function is then used to calculate the loss value between the predicted result and the actual structural response data. If the loss value is higher than a preset threshold, the damage identification model is retrained, and the training and updating process is repeated until the loss value of the damage identification model on the validation set converges to a preset range. Finally, the weight parameters of the trained damage identification model are saved.

[0068] Specifically, during the numerical simulation of the finite element model, damage variables at each micro-level, such as microcrack density, void ratio, and particle debonding rate, are extracted to form a microscale damage state tensor. Then, through statistical averaging, the damage variables at each micro-level are mapped to a mesoscale physical damage field. The mesoscale physical damage field is then subjected to feature compression processing using principal component analysis or an autoencoder. The changes in the mesoscale physical damage field during the simulation are monitored in real time, and the macroscopic response of the overall structure is calculated. By mapping the micro-variables, mesoscale variables, and macroscopic response end-to-end, a multi-scale mapping network structure corresponding to the large-span steel structure is established to form a complete damage evolution chain. Finally, the prediction results of the damage identification model are compared with the simulated response data or historical real response data, and the multi-scale mapping parameters are corrected based on the comparison results.

[0069] It should be further explained that the specific formula for calculating the total loss function is as follows:

[0070]

[0071] L total =λ1L data +λ2L phys

[0072] In the formula, L data N1 represents the number of data points in the current structural response; The predicted structural response value representing the i1th structural response data; L represents the i1th measured or simulated true structural response value; phys M1 represents the physical constraint loss term; M1 represents the current number of calculation points for the large-span steel structure. This represents the predicted equivalent flux density at the j1-th calculation point; This represents the coordinates of the j1th calculation point; This represents the actual applied body force at the j1st calculation point in the current large-span steel structure; L represents the predicted mass force or inertial response term at the j1-th calculation point; total λ represents the total loss value; λ1 and λ2 represent the weighting factors of the data loss term and the physical constraint loss term, respectively.

[0073] Various sensors are used to collect real-time response data of large-span steel structures under service conditions, and the monitored real-time response data is identified and decoupled.

[0074] Simulate the propagation path of ultrasonic waves or stress waves in the structure, and extract and identify the feature information in each real-time response data after decoupling.

[0075] The sensor network topology is dynamically adjusted, and the spatial distribution, severity, and development trend of structural damage are identified based on the monitored real-time response data.

[0076] Based on the same inventive concept, embodiments of the present invention also provide a damage identification system for large-span spatial steel structures that combines AI and numerical simulation. See also Figure 2 , Figure 2 A schematic diagram of the structure of the damage identification system for large-span spatial steel structures combining AI and numerical simulation provided in this embodiment of the invention includes:

[0077] The system includes a modeling and simulation module, a deployment and acquisition module, an identification and decoupling module, a dynamic modeling module, a special focusing module, a model training module, a collaborative analysis module, a learning and optimization module, a damage assessment module, an early warning and decision-making module, and an interactive display module.

[0078] The modeling and simulation module is used to establish finite element simulation models of large-span steel structures and simulate the structural response under various loads, working conditions, and damage scenarios; the deployment and acquisition module is used to set sensor deployment and acquisition strategies and acquire structural operating status data in real time; the identification and decoupling module is used to identify and remove interference information in the structural operating status data.

[0079] Specifically, the raw monitoring data collected in real time by various sensors is normalized and then input into the environmental feature channel and the damage feature channel, respectively. The environmental feature channel and the damage feature channel each use two convolutional encoders with different structures or weights to perform nonlinear mapping on each set of raw monitoring data to obtain corresponding potential environmental feature vectors and potential damage feature vectors. These potential environmental feature vectors and potential damage feature vectors are then concatenated to obtain corresponding joint input vectors. These joint input vectors are then transmitted to a shared generator for monitoring data reconstruction. A discriminator is then used to calculate the difference between the reconstructed monitoring data and the original monitoring data. Based on the calculated difference, the encoder is driven. Adversarial optimization is performed between the shared generator and the discriminator. An adversarial objective function is established based on the reconstruction error, the discriminator's adversarial loss, and the decoupling regularization term. The discriminator loss value is then calculated using the adversarial objective function. If the loss value is higher than a preset threshold, the parameters of the discriminator and the shared generator are updated based on the loss value, and feature mapping, feature concatenation, data reconstruction, and data discrimination are performed again. This process is repeated for multiple rounds of iterative updates until the discriminator loss value converges to a preset range, at which point the iteration stops. Then, the latest monitoring data is processed by extracting features through environmental and damage feature channels, and the corresponding joint input vector is generated through concatenation. Finally, the shared generator is used to reconstruct the monitoring data to remove environmental interference from the original monitoring data.

[0080] The dynamic modeling module is used to model the sensor network topology as a graph structure and dynamically update the sensor network graph structure.

[0081] Specifically, various sensors are treated as nodes in a graph, and the physical connectivity between adjacent sensors, the distance between measuring points, or the signal correlation are treated as edges in the graph to form an initial graph. Then, the Euclidean distance between each sensor is calculated, and the connection weight of each node is initialized according to the distance attenuation coefficient. The information of each sensor deployed on the large-span steel structure is monitored in real time. Whenever the sensor status, structural layout, and boundary conditions change, the system enters a dynamic graph update state and triggers structural reconstruction. Then, a new adjacency matrix is ​​constructed based on the current physical position change or the signal correlation between sensors.

[0082] It should be further explained that the specific formula for determining the update status of the dynamic graph is as follows:

[0083]

[0084] In the formula, θ t Nt represents the average rate of change of the node state at time t; N2 represents the total number of sensor nodes. The measurement feature vector represents the i2th node at time t; Represents the measurement feature vector of the i2th node at time t-1; ε represents a small positive number to avoid division by zero; if θ t If the value is greater than τ, then a graph structure update is triggered, where τ is a preset threshold.

[0085] The special focusing module extracts damage-sensitive features from real-time monitoring data and analyzes weak points in large-span steel structures; the model training module pre-trains the constructed AI model based on solid mechanics principles; the collaborative analysis module constructs a dual-channel digital twin to collaboratively identify damage in large-span steel structures; the learning optimization module generates adversarial examples based on the comparison results from the collaborative analysis module, and optimizes and adjusts the AI ​​model in real time; the damage assessment module maps the damage identification results to the corresponding large-span steel structure model and assesses the severity of the damage; the early warning decision module generates early warning information and formulates maintenance and reinforcement strategies based on the damage location, development trend, and structural safety reserves; and the interactive display module displays the results of each stage in real time through a 3D visualization platform.

[0086] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A damage identification method for large-span spatial steel structures combining AI and numerical simulation, characterized in that, Includes the following steps: P1. Based on the structural design drawings and material properties, construct a finite element simulation model of a large-span steel structure to simulate the structure's response under various load and boundary conditions. P2. Based on the acquired response data, a damage identification model is pre-trained, and multi-scale damage evolution is simulated by combining numerical simulation data and experimental observation data. P3. Through various sensors, real-time response data of large-span steel structures under service environment is collected, and the monitored real-time response data is identified and decoupled. P4. Simulate the propagation path of ultrasonic waves or stress waves in the structure, and extract and identify the feature information in each real-time response data after decoupling. P5. Dynamically adjust the sensor network topology and identify the spatial distribution, severity, and development trend of structural damage based on the monitored real-time response data.

2. The damage identification method for large-span spatial steel structures combining AI and numerical simulation according to claim 1, characterized in that, The specific steps for constructing the finite element simulation model of the large-span steel structure described in P1 are as follows: S1.1: Obtain the geometric information of the current large-span steel structure based on the structural design drawings, establish a three-dimensional geometric model of the real structure based on the collected geometric information of the current large-span steel structure, then select the corresponding element type and mesh density, perform meshing on the established three-dimensional geometric model, define its corresponding constitutive relation according to the performance parameters of the steel, and analyze the stress of the current large-span steel material through a nonlinear material model. S1.2: Based on the supports, connectors and constraint types in the actual large-span steel structure, set the corresponding support conditions for the three-dimensional geometric model after meshing, and set the load types and distribution methods of the current large-span steel structure under various working conditions; S1.3: Using the basic governing equations of the finite element method, under preset load and boundary conditions, obtain the nodal displacements, component stresses and reactions in the current three-dimensional geometric model of the large-span steel structure to generate the corresponding finite element simulation model. At the same time, extract the displacement, stress, reaction and deformation data of the nodes as a reference baseline for the structural integrity state.

3. The damage identification method for large-span spatial steel structures combining AI and numerical simulation according to claim 2, characterized in that, The specific steps of the pre-trained damage recognition model described in P2 are as follows: S2.1: Using the finite element simulation model of the corresponding large-span steel structure, simulate the response of the corresponding large-span steel structure under various loads and boundary conditions, and organize the simulated response data and historical real response data into a structural response dataset. Then, divide the structural response dataset into a training set and a validation set, and then establish a damage identification model based on the Transformer architecture. S2.2: Initialize the parameters of the damage identification model. Based on the simulated response data and historical real response data, use the mean square error loss function as the data loss term for supervised learning of the damage identification model. Then, based on the basic equations in elasticity, construct the corresponding physical consistency loss function and use the physical consistency loss function as the physical constraint loss term for supervised learning of the damage identification model. Finally, linearly combine the data loss term and the physical constraint loss term to establish the total loss function of the damage identification model. S2.3: Divide the data in the training set into time series of a preset length, perform linear mapping and position encoding on each time series, and then input the processed time series into the damage recognition model. The damage recognition model calculates the information correlation between each input time series based on the multi-head self-attention mechanism, extracts the dynamic dependency features between different time steps, and then processes each group of dynamic dependency features through the residual connection layer and the feedforward network in sequence, and generates the corresponding feedforward output. S2.4: Transmit the feedforward output of each group to the regression or classification head and predict the damage information of the corresponding large-span steel structure. Then, calculate the loss value between the prediction result and the actual structural response through the total loss function, and use the backpropagation algorithm to pass the loss value from the output layer to the input layer of the damage identification model layer by layer. The gradient of each layer of the damage identification model is calculated by automatic differentiation, and then the network parameters of each layer are updated by the Adam optimizer. S2.5: After each round of training, the validation set data is input into the damage identification model, the prediction result is output through forward propagation, and the total loss function is used to calculate the loss value between the prediction result and the real structural response data. If the loss value is higher than the preset threshold, the damage identification model is retrained and the training and update are repeated until the loss value of the damage identification model on the validation set converges to the preset range, and the weight parameters of the trained damage identification model are saved.

4. The damage identification method for large-span spatial steel structures combining AI and numerical simulation as described in claim 3, characterized in that, The specific calculation formula for the total loss function described in S2.2 is as follows: L total =λ1L data +λ2L phys In the formula, L data N1 represents the number of data points in the current structural response; The predicted structural response value representing the i1th structural response data; L represents the i1th measured or simulated true structural response value; phys M1 represents the physical constraint loss term; M1 represents the number of calculation points for the current large-span steel structure; ζ j1 This represents the predicted equivalent flux density at the j1-th calculation point; This represents the coordinates of the j1th calculation point; This represents the actual applied body force at the j1st calculation point in the current large-span steel structure; L represents the predicted mass force or inertial response term at the j1-th calculation point; total λ represents the total loss value; λ1 and λ2 represent the weighting factors of the data loss term and the physical constraint loss term, respectively.

5. The damage identification method for large-span spatial steel structures combining AI and numerical simulation according to claim 3, characterized in that, The specific steps for simulating multi-scale damage evolution described in P2 are as follows: S3.1: During the numerical simulation of the finite element simulation model, damage variables at each micro-level, such as microcrack density, void ratio, and particle debonding rate, are extracted to form a microscale damage state tensor. Then, through statistical averaging, the damage variables at each micro-level are mapped to the mesoscale physical damage field. S3.2: The mesoscale physical damage field is subjected to feature compression processing through principal component analysis or autoencoder, the changes in the mesoscale physical damage field during the simulation process are monitored in real time, and the macroscopic response of the overall structure is calculated. S3.3: By mapping micro-scale variables, meso-scale variables and macro-scale responses end-to-end, a multi-scale mapping network structure corresponding to large-span steel structures is established to form a complete damage evolution chain. Then, the prediction results of the damage identification model are compared with simulated response data or historical real response data, and the multi-scale mapping parameters are corrected based on the comparison results.

6. A damage identification system for large-span spatial steel structures combining AI and numerical simulation, used to implement the damage identification method for large-span spatial steel structures combining AI and numerical simulation as described in any one of claims 1-5, characterized in that, include: The system includes a modeling and simulation module, a deployment and acquisition module, an identification and decoupling module, a dynamic modeling module, a special focusing module, a model training module, a collaborative analysis module, a learning and optimization module, a damage assessment module, an early warning and decision-making module, and an interactive display module. The modeling and simulation module is used to establish a finite element simulation model of a large-span steel structure to simulate the structural response under various loads, working conditions and damage scenarios. The deployment and acquisition module is used to set the sensor deployment and acquisition strategy and to acquire structural operation status data in real time. The identification and decoupling module is used to identify and remove interference information in the structural operating status data; The dynamic modeling module is used to model the sensor network topology as a graph structure and dynamically update the sensor network graph structure. The special focusing module is used to extract damage-sensitive features from real-time monitoring data and analyze weak points in large-span steel structures. The model training module is used to pre-train the constructed AI model based on the laws of solid mechanics; The collaborative analysis module is used to construct a dual-channel digital twin for collaborative identification of damage to large-span steel structures. The learning optimization module generates adversarial examples based on the comparison results of the collaborative analysis module, and optimizes and adjusts the AI ​​model in real time. The damage assessment module is used to map the damage identification results to the corresponding large-span steel structure model and assess the severity of the damage. The early warning decision module generates early warning information and formulates maintenance and reinforcement strategies based on the location of damage, development trend, and structural safety reserves. The interactive display module is used to display the results of each stage in real time through a 3D visualization platform.

7. The damage identification system for large-span spatial steel structures combining AI and numerical simulation as described in claim 6, characterized in that, The specific steps of the identification and decoupling module in identifying and removing interference information from the structural operating status data are as follows: S4.1: After normalizing the raw monitoring data collected in real time by various sensors, the data are input into the environmental feature channel and the damage feature channel respectively. The environmental feature channel and the damage feature channel use two convolutional encoders with different structures or weights to perform nonlinear mapping on each set of raw monitoring data to obtain the corresponding potential environmental feature vector and potential damage feature vector. S4.2: The potential environmental feature vector and the potential damage feature vector are concatenated to obtain the corresponding joint input vector. Then, each joint input vector is transmitted to the shared generator to reconstruct the monitoring data. Finally, the discriminator is used to calculate the difference between the reconstructed monitoring data and the original monitoring data. S4.3: Based on the calculated difference value, drive the encoder and shared generator to perform adversarial optimization with the discriminator, and establish an adversarial objective function based on the reconstruction error, the discriminator adversarial loss and the decoupling regularization term. Then, calculate the discriminator loss value through the adversarial objective function. If the loss value is higher than the preset threshold, update the parameters of the discriminator and shared generator based on the loss value, and re-perform feature mapping, feature concatenation, data reconstruction and data discrimination. S4.4: Repeat multiple rounds of iterative updates until the discriminator loss value converges to the preset range, then stop the iteration. After that, extract features from the latest monitoring data through the environmental feature channel and the damage feature channel, generate the corresponding joint input vector through splicing, and then use the shared generator to reconstruct the monitoring data to remove environmental interference in the original monitoring data.

8. The damage identification system for large-span spatial steel structures combining AI and numerical simulation as described in claim 6, characterized in that, The specific steps of the collaborative analysis module for collaborative identification of damage in large-span steel structures are as follows: S5.1: Process various sensors as nodes of a graph, and process the physical connectivity between adjacent sensors, the distance between measurement points, or the signal correlation as edges of the graph to form an initial graph. Then calculate the Euclidean distance between each sensor and initialize the connection weight of each node according to the distance attenuation coefficient. S5.2: Real-time monitoring of sensor information deployed on the large-span steel structure. Whenever the sensor status, structural layout, or boundary conditions change, it enters the dynamic graph update state and triggers structural reconstruction. Then, based on the current physical position change or the signal correlation between sensors, a new adjacency matrix is ​​constructed.

9. The damage identification system for large-span spatial steel structures combining AI and numerical simulation as described in claim 8, characterized in that, The specific formula for determining the update status of the dynamic graph described in S5.2 is as follows: In the formula, θ t Nt represents the average rate of change of the node state at time t; N2 represents the total number of sensor nodes. The measurement feature vector represents the i2th node at time t; The measurement feature vector represents the i2th node at time t-1; ε represents a small positive number that avoids division by zero; if θ t If the value is greater than τ, then a graph structure update is triggered, where τ is a preset threshold.

Images