Proxy model and finite element fusion method and system for bridge safety early warning

Abstract

The invention belongs to the technical field of bridge engineering, and particularly discloses a proxy model and finite element fusion method and system for bridge safety early warning. Comprising the following steps: simulating a real load of a bridge by adopting a structural multi-scale finite element model to calculate a bridge structural response, constructing a finite element agent model, and training the finite element agent model by adopting a data set to extract a relationship between weight matrix analysis and output items of a neural network; and in the refined grid division part, considering the damage directly observed by the structure, dividing the model parameters extracted from the monitoring point by the model parameters directly corrected by the damage to obtain a final corrected model. According to the method, sensor signal processing, model correction and finite element calculation are closely combined together through neural network analysis, a method basis is provided for dynamic evolution of a digital object fusion model, and technical support is provided for subsequent bridge structure safety evaluation and health early warning.

Description

Technical Field

[0001] The present invention belongs to the technical field of bridge engineering, and more specifically, relates to a method and system for fusion of an agent model and finite element for bridge safety early warning. Background Art

[0002] The concept of digital twins is gaining more and more importance with the advancement of science and technology and the growing traffic construction. What follows is the improvement of the last link of the bridge health monitoring system using digital-physical fusion technology. In the construction of the concept of digital-physical fusion, dynamically updating and correcting the structural model based on the on-site sensor signals is an effective means in technical construction, and it can also reflect the real condition of the structure in a timely manner.

[0003] In the process of digital-physical integration, it is very important to reasonably design the information description method of the model for the twin object and improve the structural information included in the model building process. For the finite element model with perfect physical and mechanical relationships, the current structural analysis needs to update the current information based on sensors in a timely manner. The use of multi-scale finite element models can not only show the changes and predictions of structural behavior at the scale of the entire bridge, but also use refined simulations at the scale of components to carefully analyze the stress evolution trend. Therefore, the calculation and state prediction of the health status of the bridge structure by establishing a finite element model is a method with high computational efficiency and high calculation accuracy.

[0004] In the finite element model correction, there are three commonly used methods, mainly analyzing the sensitivity matrix of the structural correction parameters, deriving the likelihood function expression of the structure by Monte Carlo method sampling Bayesian probability method, and establishing a proxy model to directly calculate the appropriate adjustment parameters of the model for rapid correction; and the indicator parameter involved in the model correction process to judge whether the structure is damaged is the frequency response function. Through the existing model correction methods, it is difficult to use the frequency response function of the structure to judge the damage information of the structure in actual engineering applications, and the resulting model correction cannot verify whether its correction result is reasonable in long-term applications. This problem can be effectively solved by correcting the structural model based on the static response of the structure through a proxy model. Summary of the invention

[0005] In view of the above defects or improvement needs of the prior art, the present invention provides a method and system for fusion of proxy model and finite element for bridge safety early warning, which generates simulated loads at excitation monitoring points by establishing a multi-scale finite element model, obtains response monitoring point signals, integrates the excitation and response information of each node as input, extracts multi-scale finite element model calculation information as output, and serves as the input data set of the neural network. Establish and train a proxy model as an analysis of the relationship between [excitation signal-response signal] and [finite element model parameters], use the proxy model to analyze the on-site excitation monitoring point signals and response monitoring point signals, and obtain the finite element model parameters (material properties and geometric parameters of the unit) extracted by the monitoring point signals. In the refined mesh division part, the directly observed damage needs to be corrected, and the model parameters extracted from the monitoring point are divided by the model parameters directly corrected by the damage to obtain the final corrected model. By establishing a PINN proxy model to quickly process the on-site sensor excitation-response signals, directly output the calculation parameters of the finite element model, and quickly correct the multi-scale finite element model. Furthermore, real-time dynamic correction results of the finite element model can be obtained through limited time series signals. It provides a methodological basis for the dynamic evolution of the digital-physical fusion model and provides technical support for subsequent bridge structure safety assessment and health warning.

[0006] To achieve the above object, according to one aspect of the present invention, a method for fusion of agent model and finite element for bridge safety early warning is proposed, comprising the following steps: Step 1: Use a multi-scale finite element model to simulate the real load of the bridge to calculate the structural response of the bridge, and build a data set based on the real load and structural response; Step 2: construct a finite element proxy model and use the data set to train the finite element proxy model to extract the relationship between the weight matrix analysis of the neural network and the output items; Step 3: Use the structural damage information from on-site monitoring to directly modify the finite element model for damage correction; Step 4: Use the trained finite element proxy model to analyze the excitation monitoring point signal and the response monitoring point signal of the on-site monitoring to obtain the model correction parameters based on data analysis; Step 5, obtaining a modified multi-scale finite element model according to the model modification parameters obtained through damage modification and data analysis; Step six: using the modified multi-scale finite element model to generate a data set for training and updating the finite element proxy model.

[0007] As further preferred, step one comprises the following steps: Establish a refined component model based on the sensor layout and structural component division, divide the parts that need to be refined according to the sensor network and key monitoring points of the structure, analyze the degradation model of different parts and different materials of the model, and then establish a multi-scale finite element model of the structure based on the bridge design drawings; By applying an excitation signal or a generated simulated excitation to the multi-scale finite element model, a data set containing excitation monitoring point signals and response monitoring point signals for training a neural network is generated.

[0008] As further preferred, in step 2, the finite element proxy model is any one of a polynomial response surface, a BP neural network, a radial basis function, and a Kriging model.

[0009] As a further preferred embodiment, in step 2, in the finite element proxy model, the loss function part uses the virtual work principle equation as a physical constraint: , In the formula, represents the variation of the total virtual work, , are the stress tensor component and body density component in the i direction, is the surface force per unit area in the i direction, is the direction vector of the i-th end of the unit to be determined, is the strain component at end i, V is the volume of the unit for body force integration, and S is the area of ​​the unit for surface force integration; The loss function also includes: , In the formula, , , , It is the time series of acceleration, velocity, displacement and external force of the structure space nodes, and M, C, K are the mass, damping and stiffness distribution of the system.

[0010] As a further preferred embodiment, in step three, damage correction is performed on the finite element model according to the damage type of the structure, specifically including: If the structural damage is a crack in the concrete part, the finite element model is corrected by increasing the number of nodes in the finite element model and modifying the boundary conditions. The overall structural calculation parameter influence matrix a of the increased and decreased nodes needs to be recalculated. ij ; If the structural damage is the degradation of material properties, the model correction parameters are calculated by the equivalent property correction parameters obtained by analyzing the sensor signal. Then, it is used as the physical constraint of the neural network to directly perform damage correction on the finite element model; If the structural damage is a change in the structural geometric parameters, the relevant cross-sectional property parameters of the structure are recalculated to obtain the model correction parameters Then it is used as the physical constraint of the neural network to directly perform damage correction on the finite element model.

[0011] As a further preferred embodiment, in step 4, the trained finite element proxy model is used to analyze the excitation monitoring point signal and the response monitoring point signal of the field monitoring, and the model correction parameter β based on data analysis is obtained by comparing the neural network calculation parameters of each node with the calculation parameters of the initial model: , In the formula, represents the model parameters of the structure in the current state, represents the model parameters of the bridge structure in the initial state, are the model parameters of the structure calculated from the sensor data.

[0012] As a further preferred embodiment, step five specifically includes the following steps: If the model correction parameters If the correction target is consistent with the model correction parameter β but the correction time is inconsistent, the model correction parameter is directly used. as physical constraints to train neural networks; If the model completes health monitoring at this time, the model correction parameters directly obtained through monitoring Can explain the state of the structure, then the model correction parameters represents the refined component adjustment parameter γ; If the model correction parameters The overall model will be adjusted, or the monitoring bridge structure will generate new physical constraints, then the model correction parameters β and model correction parameters A refined component adjustment parameter γ is constructed, and the refined component adjustment parameter γ is used to complete the finite element model update to obtain a modified multi-scale finite element model.

[0013] As a further preferred embodiment, the calculation process of the refined component adjustment parameter γ is as follows: If the direct correction object can be solved by data analysis, but has not been obtained through experiments at this stage, then = ; If the object of correction has completed the test and obtained accurate data that can be directly used for correction, then = ; If the data obtained through field tests or direct monitoring cannot fully describe the structural state, then = ; in, , They are the model correction parameters of the model j node position directly obtained through on-site monitoring of damage and the model correction parameters obtained through proxy model calculation data. Adjust the parameters for the refined components used in the j-node finite element model, is the model correction parameter of the damage directly monitored on-site at the model node i and the model correction parameter obtained by calculating the data of the proxy model, a ij It is the influence matrix of the changes in the calculation parameters of all nodes after the calculation parameters of the model related to the structure i node change.

[0014] As a further preferred embodiment, it also includes using an updated multi-scale finite element model for bridge safety early warning.

[0015] According to another aspect of the present invention, a bridge safety early warning system based on the fusion of agent model and finite element is provided, comprising: The first main control module is used to simulate the real load of the bridge using a multi-scale finite element model of the structure to calculate the structural response of the bridge, and to construct a data set based on the real load and structural response; A second main control module is used to construct a finite element proxy model and train the finite element proxy model using a data set to extract the relationship between the weight matrix analysis of the neural network and the output items; The third main control module is used to directly modify the finite element model using the structural damage information monitored on site to perform damage correction; The fourth main control module is used to analyze the excitation monitoring point signal and the response monitoring point signal of the on-site monitoring by using the trained finite element proxy model to obtain the model correction parameters based on the data analysis; A fifth main control module, used for obtaining a corrected multi-scale finite element model according to model correction parameters of damage correction and data analysis; The sixth main control module is used to generate a data set for training and updating the finite element proxy model using the modified multi-scale finite element model.

[0016] In general, the above technical solution conceived by the present invention has the following technical advantages compared with the prior art: 1. The model fusion method provided by the present invention uses a proxy model to quickly analyze the excitation-response signal, obtains the correction calculation parameters of the finite element model, and directly corrects the finite element model using the damage obtained by the sensor. It takes into account the defect information obtained by on-site sensor monitoring and the excitation-response monitoring point data information, realizes the dynamic evolution of the digital model, and improves the efficiency and accuracy of risk assessment.

[0017] 2. The model fusion method provided by the present invention can use the stiffness, geometric characteristics and other information of the structure on the basis of the updated finite element model to identify structural damage. Compared with the frequency response characteristic function often used in current model correction, it has the advantages of high response accuracy and can be directly applied to subsequent finite element calculations.

[0018] 3. The present invention uses vector finite element as the finite element model for model correction, which can not only avoid the integration of stiffness matrix in the calculation process and reduce the finite element analysis time of integrating the overall stiffness matrix due to changes in structural parameters, but also optimize and update the establishment of the model in the analysis process of multi-scale finite element model. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flow chart of a proxy model and finite element fusion method for bridge safety early warning involved in an embodiment of the present invention; Figure 2 It is a convergence diagram of the calculated MSE of the proxy model involved in the embodiment of the present invention according to the number of iteration steps. DETAILED DESCRIPTION

[0020] like Figure 1 As shown, an agent model and finite element fusion method for bridge safety early warning provided by an embodiment of the present invention analyzes the on-site sensor signals and uses a neural network instead of a multi-scale finite element model to analyze the dynamic evolution information of the bridge in real time, including the following steps: S1: Establish a multi-scale finite element model of the structure based on the bridge design drawings. The multi-scale here refers to the non-uniform model mesh size under the same finite element model, especially at the layout positions of the excitation sensors and response sensors. The finite element model is divided into fine grids to facilitate the comparison of calculation results at different monitoring positions in the same section. The multi-scale finite element model can use the excitation data collected in advance or stored in the database. The macroscopic damage of the structure that can be reflected in the finite element model is corrected by directly adding and reducing nodes, and other numerical aspects are achieved by modifying the model calculation parameters. After the initial multi-scale finite element model is established, different model parameters are used to generate bridge structure models corresponding to different safety states according to the needs of model training, and generate a pre-training data set of excitation-response pairs of neural networks.

[0021] In this step, the proxy model is established based on neural networks, especially PINN (Physics-Informed Neural Networks). Compared with the traditional neural network proxy model, the use of PINN to establish the proxy model can ensure the physical prior characteristics of the generated data, reduce data requirements and optimize the solution process of complex models.

[0022] The finite element model used is a multi-scale finite element model, especially a multi-scale vector finite element. In the vector finite element, the node behavior is used to describe the structural response imposed by external excitation instead of the unit analysis method of the traditional finite element. The unit stiffness analysis is used instead of the integrated stiffness matrix in the traditional finite element. This can effectively speed up the finite element analysis process and has good compatibility with the model correction process because there is no need to integrate the stiffness matrix.

[0023] S2: Establish a neural network proxy model based on the excitation monitoring point-response monitoring point signal to quickly solve the current structural model information. The input variables of the neural network are the bridge excitation monitoring point signal-response monitoring point signal, and the structural node connection information is determined by setting the weight function of the first few connection layers. The model calculation parameters required for structural calculation are obtained through proxy model analysis.

[0024] S3: Directly correct the structural model based on the structural damage obtained by on-site sensor monitoring, and calculate the model correction parameters for preliminary adjustment of the structure. When the damage obtained by on-site sensor monitoring requires the addition or reduction of finite element model nodes, such as crack extension, etc., it is necessary to adjust the structural material properties and geometric parameters, including cross-sectional area, cross-sectional bending moment of inertia, cross-sectional torsional moment of inertia, etc. When the damage information only involves adjusting the material properties of the structure, such as concrete carbonization, steel bar corrosion, etc., the cross-sectional area is corrected in this step. The specific structural parameters to be adjusted are related to the type of sensor monitoring data and the type of structure. The formula here only shows the adjustment calculation of stiffness. The adjustment parameters of the structural model stiffness are calculated as follows:

[0025] In the formula, represents the stiffness of the structure in the current state, represents the stiffness of the bridge structure in the initial state, The equivalent model property adjustment parameters are directly corrected based on the damage data obtained by direct monitoring of the structure by on-site sensors. Different types of parameters are distinguished by subscripts, such as Represents the correction parameter of Young's modulus.

[0026] S4: Use the proxy model processor to analyze the excitation monitoring point information-response monitoring point information, output the model information obtained according to the [excitation signal-response signal] data analysis, compare the model calculation parameters of each finite element model node (such as Young's modulus E, cross-sectional calculation area A, cross-sectional inertia moment I, etc.) with the information of the initial finite element model, and obtain the model correction parameter β based on data analysis. Derive the inversion function of the weight matrix of the neural network for the model calculation information, improve the physical relationship between the excitation and response revealed by the established proxy model, and improve the rationality of the model. For example, analyze the relevant relationship between the information of a certain neuron node and the final output, and whether its calculation process can be reflected in the specific calculation formula of the finite element. Or by analyzing the connection between multiple neuron nodes, calculate how to derive the relationship between the weight matrix and the final output through the inter-layer connection of these neurons.

[0027]

[0028] In the formula, represents the model parameters of the structure in the current state, represents the model parameters of the bridge structure in the initial state, are the model parameters of the structure calculated from the sensor data.

[0029] S5: The final update of the finite element model includes not only direct damage correction, but also model correction information at the data level obtained from the analysis of on-site sensor signals. If the corresponding calculation parameters (such as Young's modulus and other properties that characterize concrete deterioration) can be directly obtained by monitoring the damage through on-site sensors, it means that the direct correction parameter α should be consistent with the data-driven correction parameter β, and the direct correction parameter α can be used as a physical constraint for neural network training. If the direct correction parameter α will produce adjustments to the overall model (such as local cracks in the concrete component causing changes in the cross-sectional area, etc.), or the bridge structure state is monitored at this time (generating new physical constraints), then only the direct correction parameter α or the data-driven correction parameter β cannot be used to obtain the model adjustment parameters that accurately describe the current state of the model. At this time, the data-driven model correction parameter β is used to subtract the influence of the structural model parameter change α obtained through on-site sensor analysis in S3 on the relevant nodes to obtain a more accurate correction result. The refined component adjustment parameter in this step is This process needs to be iterated for several time steps until the calculated response of the structure and the on-site monitoring response of the sensor are less than the preset range; if the calculation result error is greater than the preset range, it is necessary to retrain the neural network proxy model and return to S4 to recalculate the adjustment parameters of the structure.

[0030] The calculation process of the refined component adjustment parameter γ is as follows: If the direct correction object can be solved by data analysis, but has not been obtained through experiments at this stage, then = ; If the object of correction has completed the test and obtained accurate data that can be directly used for correction, then = ; If the data obtained through field tests or direct monitoring cannot fully describe the structural state, then = ; in, , They are the model correction parameters of the model j node position directly obtained through on-site monitoring of damage and the model correction parameters obtained through proxy model calculation data. Adjust the parameters for the refined components used in the j-node finite element model, is the model correction parameter of the damage directly monitored on-site at the model node i and the model correction parameter obtained by calculating the data of the proxy model, a ij It is the influence matrix of the changes in the calculation parameters of all nodes after the calculation parameters of the model related to the structure i node change.

[0031] The method further comprises the following steps: S6: Use the modified finite element model in S5 to generate [excitation signal-response signal] and [finite element model calculation parameter] data sets for updating the neural network proxy model.

[0032] Example 2 In this embodiment, a neural network proxy model is used to quickly update the model based on the field sensor signal. Finite element model information and field sensor monitoring stimulus-response information are input into the proxy model, and the calculation information of the structural finite element model is directly output.

[0033] The proxy model is established using PINN. Since PINN needs to impose constraints during the calculation process, the constraints are not only the structural mechanics equations as hard constraints, but also the value range of the signal adjustment parameters and the boundary conditions of the finite element model as soft constraints.

[0034] Structural model adjustment parameters Represents the ratio of the current structural parameters to the initial state, and the structural model adjustment parameters The value range is between 0 and 1, because the performance of the structure is always in a degraded state during use. The structural model adjustment parameters are adjusted under regular structural maintenance. It may exceed 1, in which case the neural network needs to be redesigned.

[0035] The data input to PINN varies according to the model adjustment parameters that need to be modified. For example, if the cross-sectional adjustment parameters of the structure need to be adjusted, the tensile response information of the structure under excitation needs to be input; if the torsional section inertia moment of the structure needs to be adjusted, the monitoring information of the torque sensor of the structure needs to be input.

[0036] A multi-scale finite element model is established, and different finite element parts are established using units with different meshing accuracies. For parts that require refined simulation of component behavior, the modeling scale is smaller than other units. Through refined modeling, not only can detailed damage be simulated on the entire finite element model, but also damage input by field sensors can be isolated and analyzed locally in the refined model.

[0037] The types of bridge structure models established include, but are not limited to, beam bridges, arch bridges, suspension bridges, and cable-stayed bridges.

[0038] According to the established finite element structural model, the sensor monitoring points are arranged reasonably and the refined simulation parts are divided. According to the structural monitoring requirements, the detailed component model settings are designed. The refined components include but are not limited to the sensor monitoring points, and also involve some parts that are prone to performance degradation.

[0039] The established and modified finite element model can not only be used for structural calculations, but also generate neural network training sets.

[0040] Based on the above content, the method in this embodiment includes: S1: Establish a multi-scale finite element model of the structure based on the bridge design drawings, and establish a refined component model based on the sensor layout and the structural component division. Divide the parts that need to be refined according to the key monitoring points of the sensor network and the structure. Analyze the degradation model of different parts and different materials of the model. Generate a pre-training data set of the neural network's excitation monitoring point-response monitoring point signal pair by applying the excitation signal stored in the cloud server or the generated simulated excitation to the multi-scale finite element model.

[0041] S2: Establish a PINN proxy model processor that quickly solves the current structural model information correction parameters based on the excitation monitoring point-response monitoring point signal. The input variable of the neural network is the bridge excitation monitoring point-response monitoring point signal pair of the previous time step, and the calculation parameters of the structural finite element model are obtained. The structural parameter adjustment information is extracted and the weight matrix between the fully connected layers is back-calculated for verification and then used for subsequent structural correction. In this step, only the neural network is pre-trained, and the generated adjustment parameters Subsequent verification of the structural correction results is required.

[0042] Among them, PINN is a network model that adds physical relationship equations as soft constraints or hard constraints in the calculation process of the neural network loss function, taking into account the physical relationship between the output variables to satisfy the prior conditions. In the establishment of the PINN model of structural mechanics, the loss function part usually uses the virtual work principle equation as a physical constraint. The loss function of the model can be calculated by the above formula to participate in the training process of the neural network. As shown in the following formula: .

[0043] This formula describes the structural state variables in spatial coordinates, where represents the variation of the total virtual work of the system, , are the variables on the volume of the system in the i direction, which are the stress tensor components and the body density components, is the surface force per unit area of ​​the system in the i direction.

[0044] Similarly, the loss function part also uses the motion equation as a physical constraint. This physical constraint relationship is well compatible with the calculation parameter verification of the vector finite element to be used in the present invention.

[0045] .

[0046] This formula describes the time state variable in spatial coordinates, where: , , , It is the time series of acceleration, velocity, displacement and external force of the structure space nodes, and M, C, K are the mass, damping and stiffness distribution of the system.

[0047] S3: Directly modify the structural model based on the structural damage obtained by sensor monitoring. According to the main damage type of the structure, design different damage correction methods for the finite element model. If the structural damage is a crack in the concrete part, the finite element model is corrected by increasing the number of nodes in the finite element model and modifying the boundary conditions. It is also necessary to recalculate the overall structural calculation parameter influence matrix a of the increased and decreased nodes. ij ; If the structural damage is the degradation of material properties, the model correction parameters are calculated by the equivalent property correction parameters obtained by analyzing the sensor signal Then, it is used as the physical constraint of the neural network to directly perform damage correction on the finite element model; if the structural damage is the change of the structural geometric parameters, the relevant cross-sectional property parameters of the structure are recalculated to obtain the model correction parameters Then, it is used as the physical constraint of the neural network to directly perform damage correction on the finite element model. The correction parameters of the remaining models that have not been corrected for damage are All of them are 1, which means that the ratio of the relevant parameters before and after the model correction is 1.

[0048] S4: Use the proxy model processor to analyze the signal pairs of the excitation monitoring point and the response monitoring point obtained by the field sensor monitoring, use the PINN proxy model to directly obtain the calculation parameters of the finite element, and use the significance of the weight matrix between the fully connected layers proposed in S2 to analyze the rationality of the neural network analysis results. Finally, by comparing the neural network calculation parameters of each node with the calculation parameters of the initial model, the model correction parameter β based on data analysis is obtained.

[0049]

[0050] In the formula, represents the model parameters of the structure in the current state, represents the model parameters of the bridge structure in the initial state, are the model parameters of the structure calculated from the sensor data.

[0051] S5: The final update of the finite element model includes not only direct damage correction, but also model correction information at the data level obtained from the analysis of field sensor signals. If the corresponding calculation parameters (such as Young's modulus and other properties that characterize concrete deterioration) can be directly obtained by monitoring the damage through field sensors, it means that the direct correction parameter α should be consistent with the data-driven correction parameter β, and the direct correction parameter α can be used as a physical constraint for neural network training. If the direct correction parameter α will cause adjustments to the overall model (such as local cracks in concrete components causing changes in cross-sectional area, etc.), the model adjustment parameters that accurately describe the current state of the model cannot be obtained only through the direct correction parameter α or the data-driven correction parameter β. At this time, the data-driven model correction parameter β is used to subtract the influence of the structural model parameter change α obtained through field sensor analysis in S3 on the relevant nodes to obtain a more accurate correction result. This process requires several iterations of time steps until the structural calculation response and the field monitoring response of the sensor are less than the predetermined range; if the calculation result error is greater than the preset range, it is necessary to retrain the neural network proxy model and return to S4 to recalculate the adjustment parameters of the structure. Finally, use Multiplying the calculated parameters of the corresponding nodes of the finite element model, the finite element model update based on the sensor signal using the proxy model is completed. The calculation process is as follows: If the direct correction object can be solved by data analysis, but has not been obtained through experiments at this stage, then = ; If the object of correction has completed the test and obtained accurate data that can be directly used for correction, then = ; If the data obtained through field tests or direct monitoring cannot fully describe the structural state, then = ; in, , They are the model correction parameters of the model j node position directly obtained through on-site monitoring of damage and the model correction parameters obtained through proxy model calculation data. Adjust the parameters for the refined components used in the j-node finite element model, a ij It is the influence matrix of the changes in the calculation parameters of all nodes after the calculation parameters of the model related to the structure i node change.

[0052] The method further comprises the following steps: S6: Generate excitation-response data pairs using the modified finite element model in S5 for updating the PINN proxy model.

[0053] Furthermore, for simple geometric boundary conditions, the loss function constraint of the displacement at the boundary condition can be calculated; for load boundary conditions, it can be achieved by establishing a calculation relationship of mechanical parameters in the node (for example, the equivalent shear spring stiffness can be obtained by calculating the ratio of the node shear force to the displacement, and then comparing them). This further reflects the use of physical constraints in the neural network established by the present invention.

[0054] Furthermore, compared with the traditional finite element model, the finite element model can effectively reduce the storage space of the model, avoid the integrated calculation of the stiffness matrix in the calculation process, and improve the calculation efficiency of the finite element structural model in the time series. Improve the efficiency and convergence speed of the finite element calculation process.

[0055] Furthermore, based on the established finite element model, the linear analysis of the structure is carried out in combination with the traditional finite element simulation of the structural elastic state.

[0056] Example 3 In this embodiment, a method for fusing a PINN agent model and a finite element structural model for bridge safety early warning is described in detail.

[0057] It should be noted that the finite element proxy model is used in the present invention to find model calculation parameters that can be used to calculate the mapping relationship between sensor excitation and response signals. The finite element proxy model can be a polynomial response surface (P-RSM), a BP neural network, a radial basis function (RBF), a Kriging model, etc. This embodiment does not impose any specific restrictions on this.

[0058] It should be noted that, except for the nonlinear behavior or the part with stiffness change, the finite element model needs to be established by the finite element method, the parts where the structural calculation parameters do not change can be established by traditional finite elements.

[0059] The embodiments described herein are specific embodiments of the present invention, which are used to illustrate the concept of the present invention, are illustrative and exemplary, and should not be construed as limiting the embodiments of the present invention and the scope of the present invention. In addition to the embodiments described herein, those skilled in the art can also adopt other obvious technical solutions based on the claims and the contents disclosed in the specification of this application, including technical solutions that adopt any obvious replacement and modification of the embodiments described herein.

[0060] This embodiment proposes a fusion method of PINN agent model and finite element structure model for bridge safety early warning. Figure 2 Taking the model fusion step diagram shown as an example, the following steps are included: S1: A multi-scale finite element model of the structure is established based on the bridge design drawings, and a refined component model is established according to the sensor layout and the structural component division. This case uses an eight-node simply supported beam model for calculation, and only shows the model correction of the full bridge scale, assuming that the sensors are arranged on nodes 4 and 5. The external excitation is a 1000kN downward impact load applied at the monitoring point.

[0061] Furthermore, under simple loads, the main information of the structural model is: the elastic modulus, cross-sectional area and bending moment of inertia of the unit, the length of the unit, and the main information provided by the node is the mass of the node. In the simple model listed, only the movement of the plane is involved. If the solution of a complex model involves the calculation of three-dimensional space, it is necessary to expand the calculation space, mainly the structural information of the model and the external excitation information to three-dimensional space. The response information of the structural sensor signal also needs to be expanded to three-dimensional space.

[0062] Furthermore, since the data dimensions of node information and unit information are different during the finite element calculation process, but the descriptive features of the structural information are provided, it is necessary to preprocess the information (such as filling 0 at the end or in the middle of the data, using 1-dimensional CNN to process the model data) to solve the problem of different numbers of neural network sample features.

[0063] S2: Establish a PINN proxy model processor that can quickly solve the current structural model information correction parameters based on the sensor point excitation-response signal. The input of the neural network is the [excitation signal-response signal] at each node at a certain time. The main information of the structural model at this time as the output of the neural network is: the elastic modulus, cross-sectional area and bending moment of inertia of the unit, the length of the unit, and the main information provided by the node is the mass of the node. Extract the relationship between the weight matrix analysis of the neural network and the output items.

[0064] Furthermore, the neural network input and output proposed in this case are classified according to the nodes of the finite element model. In other words, the excitation-response monitoring point data must first be sorted to the corresponding node positions as input. Similarly, the input data types can be classified, with the excitation monitoring point signals classified into one category and the response monitoring point signals classified into another category.

[0065] Furthermore, in the pre-processing of data, for the input and output of complex data types, it is necessary to independently design custom pre-processing methods and integrate the processing neural network code according to task requirements. For example, if the model data of a multi-scale finite element model needs to be input into the proxy model for other analysis purposes, it is necessary to write a pre-processing program or set the inter-layer connections of the initial few layers of neurons.

[0066] Similarly, in the output of data, the relationship between the output values ​​and the rationality of the output data can be determined by analyzing the units of each parameter. The units of the geometric dimensions of the output parameters, cross-sectional area and cross-sectional bending moment of inertia, are m respectively. 2 With m 4 , the rationality of the output parameters can be verified by analyzing the derivatives of the two with respect to length.

[0067] S3: Directly modify the structural model based on the structural damage obtained by on-site sensor monitoring. This model is a simple simply supported beam model, and no refined component model is established, so only the material adjustment parameters are calculated and updated. The model correction parameters obtained through sensor data are obtained by setting random numbers through the code to assume model damage If the damage is a change in the geometric parameters of the structure, then the cross-sectional properties and other parameters of the structure are recalculated to directly correct the corresponding parameters.

[0068] As a further supplement to the above solution, the modification of the calculation parameters involved in this case will not affect the total number of nodes. If it is necessary to add finite element model nodes during the model modification process, it is necessary to return to S1, re-establish a model with added nodes, and re-train the proxy model.

[0069] S4: Use the proxy model processor to analyze the excitation monitoring point-response monitoring point signals obtained by field sensor monitoring, output the calculation parameters of the multi-scale finite element model, and extract the weight matrix of the neural network to analyze the rationality of the calculation results. Use the collected sensor excitation monitoring point data to input the finite element model established by the neural network calculation parameters, and compare the rationality of the revised model with the response monitoring point data analysis.

[0070] S5: The final finite element model correction needs to exclude the influence of the direct correction damage. However, in this case, the structure does not have model damage that requires direct correction, so the calculation results in S4 are used as the calculation parameters of the updated finite element model. This process requires several time steps of iteration until the structural calculation response is consistent with the sensor field monitoring response.

[0071] The method further comprises the following steps: S6: Generate excitation-response data pairs using the finite element model modified in S5 for updating the PINN proxy model. Because the finite element model has been judged to be damaged after calculation, the model adjustment parameters calculated by the PINN proxy model in the next time step are based on the damage model modified in S5.

[0072] Furthermore, the cumulative damage information of the model can be obtained by dividing the relevant parameters of the node by the calculation parameters of the original model. The cumulative damage parameters obtained at different times are plotted into a graph to obtain the degradation model of the relevant parameters. The rationality of the correction process can be determined by comparing the degradation model proposed through the experiment.

[0073] Specifically, the bridge to be evaluated is a physical bridge to be evaluated for building safety. This embodiment does not impose any restrictions on the structure and material properties of the bridge to be evaluated. The bridge to be evaluated includes but is not limited to bridges of various architectural forms and uses such as ancient building bridges and cross-sea bridges. The original design drawing of the bridge is a drawing reflecting various building data of the bridge drawn based on the built bridge or obtained based on the historical building information of the bridge. The original design drawing of the bridge is annotated with building data and building material information. The bridge connection point is the splicing site of multiple bridge body components for the purpose of constructing the bridge body and balancing the load-bearing force.

[0074] Of course, based on any of the above embodiments or a combination of multiple embodiments, the present invention also provides a system for performing bridge safety early warning using the above method.

[0075] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A proxy model and finite element fusion method for bridge safety early warning, characterized in that: The following steps are involved: Step 1: Use a multi-scale finite element model to simulate the real load of the bridge to calculate the structural response of the bridge, and build a data set based on the real load and structural response; Step 2: construct a finite element proxy model and use the data set to train the finite element proxy model to extract the relationship between the weight matrix analysis of the neural network and the output items; Step 3: Use the structural damage information from on-site monitoring to directly modify the finite element model for damage correction; Step 4: Use the trained finite element proxy model to analyze the excitation monitoring point signal and the response monitoring point signal of the on-site monitoring to obtain the model correction parameters based on data analysis; Step 5, obtaining a modified multi-scale finite element model according to the model modification parameters obtained through damage modification and data analysis; Step six: using the modified multi-scale finite element model to generate a data set for training and updating the finite element proxy model.

2. The agent model and finite element fusion method for bridge safety early warning according to claim 1 is characterized in that: Step one includes the following steps: Establish a refined component model based on the sensor layout and structural component division, divide the parts that need to be refined according to the sensor network and key monitoring points of the structure, analyze the degradation model of different parts and different materials of the model, and then establish a multi-scale finite element model of the structure based on the bridge design drawings; By applying an excitation signal or a generated simulated excitation to the multi-scale finite element model, a data set containing excitation monitoring point signals and response monitoring point signals for training a neural network is generated.

3. The agent model and finite element fusion method for bridge safety early warning according to claim 1 is characterized in that: In step 2, the finite element proxy model is any one of a polynomial response surface, a BP neural network, a radial basis function, and a Kriging model.

4. The agent model and finite element fusion method for bridge safety early warning according to claim 3 is characterized in that: In step 2, in the finite element proxy model, the loss function part uses the virtual work principle equation as a physical constraint: ， In the formula, represents the variation of the total virtual work, , are the stress tensor component and body density component in the i direction, is the surface force per unit area in the i direction, is the direction vector of the i-th end of the unit to be determined, is the strain component at end i, V is the volume of the unit for body force integration, and S is the area of ​​the unit for surface force integration; The loss function also includes: ， In the formula, , , , It is the time series of acceleration, velocity, displacement and external force of the structure space nodes, and M, C, K are the mass, damping and stiffness distribution of the system.

5. The agent model and finite element fusion method for bridge safety early warning according to claim 1 is characterized in that: In step 3, the finite element model is corrected according to the damage type of the structure, including: If the structural damage is a crack in the concrete part, the finite element model is corrected by increasing the number of nodes in the finite element model and modifying the boundary conditions. The overall structural calculation parameter influence matrix a of the increased and decreased nodes needs to be recalculated. ij ; If the structural damage is the degradation of material properties, the model correction parameters are calculated by the equivalent property correction parameters obtained by analyzing the sensor signal. Then, it is used as the physical constraint of the neural network to directly perform damage correction on the finite element model; If the structural damage is a change in the structural geometric parameters, the relevant cross-sectional property parameters of the structure are recalculated to obtain the model correction parameters Then it is used as the physical constraint of the neural network to directly perform damage correction on the finite element model.

6. The agent model and finite element fusion method for bridge safety early warning according to claim 1 is characterized in that: In step 4, the trained finite element proxy model is used to analyze the excitation monitoring point signals and response monitoring point signals of the field monitoring. By comparing the neural network calculation parameters of each node with the calculation parameters of the initial model, the model correction parameter β based on data analysis is obtained: ， In the formula, represents the model parameters of the structure in the current state, represents the model parameters of the bridge structure in the initial state, are the model parameters of the structure calculated from the sensor data.

7. The agent model and finite element fusion method for bridge safety early warning according to claim 1 is characterized in that: Step 5 specifically includes the following steps: If the model correction parameters If the correction target is consistent with the model correction parameter β but the correction time is inconsistent, the model correction parameter is directly used. as physical constraints to train neural networks; If the model completes health monitoring at this time, the model correction parameters directly obtained through monitoring Can explain the state of the structure, then the model correction parameters represents the refined component adjustment parameter γ; If the model correction parameters The overall model will be adjusted, or the monitoring bridge structure will generate new physical constraints, then the model correction parameters β and model correction parameters A refined component adjustment parameter γ is constructed, and the refined component adjustment parameter γ is used to complete the finite element model update to obtain a modified multi-scale finite element model.

8. The agent model and finite element fusion method for bridge safety early warning according to claim 7 is characterized in that: The calculation process of the refined component adjustment parameter γ is as follows: If the direct correction object can be solved by data analysis, but has not been obtained through experiments at this stage, then = ; If the object of correction has completed the test and obtained accurate data that can be directly used for correction, then = ; If the data obtained through field tests or direct monitoring cannot fully describe the structural state, then = ; in, , They are the model correction parameters of the model j node position directly obtained through on-site monitoring of damage and the model correction parameters obtained through proxy model calculation data. Adjust the parameters for the refined components used in the j-node finite element model, is the model correction parameter of the damage directly monitored on-site at the model node i and the model correction parameter obtained by calculating the data of the proxy model, a ij It is the influence matrix of the changes in the calculation parameters of all nodes after the calculation parameters of the model related to the structure i node change.

9. The proxy model and finite element fusion method for bridge safety early warning according to claim 7 further includes using an updated multi-scale finite element model for bridge safety early warning.

10. A bridge safety early warning system based on the fusion of agent model and finite element, characterized in that: include: The first main control module is used to simulate the real load of the bridge using a multi-scale finite element model of the structure to calculate the structural response of the bridge, and to construct a data set based on the real load and structural response; A second main control module is used to construct a finite element proxy model and train the finite element proxy model using a data set to extract the relationship between the weight matrix analysis of the neural network and the output items; The third main control module is used to directly modify the finite element model using the structural damage information monitored on site to perform damage correction; The fourth main control module is used to analyze the excitation monitoring point signal and the response monitoring point signal of the on-site monitoring by using the trained finite element proxy model to obtain the model correction parameters based on the data analysis; A fifth main control module, used for obtaining a corrected multi-scale finite element model according to model correction parameters of damage correction and data analysis; The sixth main control module is used to generate a data set for training and updating the finite element proxy model using the modified multi-scale finite element model.

Images