A bridge structure deformation real-time prediction method based on finite element analysis

Abstract

The application discloses a bridge structure deformation real-time prediction method based on finite element analysis, which comprises the following steps: a coupling model containing a bridge structure and a rock-soil finite element model is established; a proxy model of a graph neural network is introduced to reproduce the mechanical response of the rock-soil model; in the real-time prediction stage, an asynchronous time step framework is adopted, the bridge structure model is iterated at a small step, and the proxy model provides interface reaction force at a large step, so that the calculation efficiency is improved; by receiving the actual monitoring deformation data of the bridge, the key parameters of the rock-soil finite element model are reversely calculated and updated based on the Bayesian inference theory, the initial parameter deviation and the model time-varying effect are eliminated, the retraining of the proxy model is triggered after the parameter updating, and the prediction model is ensured to be long-term synchronized with the real state of the bridge; the application solves the problem that the traditional finite element analysis is difficult to consider real-time performance and long-term accuracy, and provides reliable prediction for the whole life cycle health monitoring and prospective maintenance of the bridge structure.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bridge engineering and computational mechanics, and particularly relates to a bridge structure deformation real-time prediction method based on finite element analysis. BACKGROUND

[0002] With the rapid development of transportation infrastructure, the service safety and long-term performance guarantee of large and super-large bridges have become the focus of attention in recent years. Among them, the numerical simulation method, especially the technology based on finite element analysis (FEA), has become an important means to realize the whole life cycle health monitoring and forward-looking maintenance of the bridge structure by predicting the deformation behavior of the bridge structure during application. Currently, high-fidelity bridge deformation prediction research usually needs to establish a refined finite element model including the superstructure, substructure and foundation, and focuses on the soil-structure interaction (Soil-Structure Interaction, SSI). By simulating traffic load, environmental changes and other factors, the mechanical response existing in SSI can be analyzed, which can provide a theoretical basis for the safety evaluation and maintenance decision of the bridge.

[0003] However, the existing technology still has significant technical bottlenecks in realizing high-precision and high-efficiency real-time prediction of bridge deformation. First of all, for the full-coupling SSI finite element model of the refined geotechnical model, the calculation degree of freedom usually reaches millions or even higher, so that the calculation cost of single model full analysis is high and the time consumption is huge, which conflicts with the real-time requirement of bridge application period deformation prediction. Secondly, since the accuracy of the traditional finite element model is highly dependent on the initial parameters provided by the geological exploration, there is inherent deviation between these parameters and the actual state of the project, and this deviation will continue to accumulate over time and environmental evolution. Therefore, the existing method generally lacks an effective feedback mechanism to dynamically and automatically use the actual monitoring data collected by the sensors arranged on the bridge to perform inversion correction and continuous calibration of the key parameters of the finite element model, so that the prediction model gradually deviates from the real state of the bridge, and the reliability of long-term prediction is greatly reduced. SUMMARY

[0004] This section is intended to summarize some aspects of the embodiments of the present application and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section and the abstract and title of the specification to avoid obscuring the purpose of this section, the abstract and the title, and such simplifications or omissions cannot be used to limit the scope of the present application.

[0005] In view of the above existing problems, the present application is proposed. Therefore, the present application provides a bridge structure deformation real-time prediction method based on finite element analysis to solve the problems proposed in the background art.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for real-time prediction of bridge structure deformation based on finite element analysis, comprising:

[0007] In computer-aided software, a coupled model containing a bridge structure finite element model and a geotechnical finite element model is generated, and coupling constraint relationships are established on this coupled model.

[0008] A surrogate model is used for prediction. The surrogate model is used to reproduce the mechanical response of the soil and rock finite element model. At the same time, when the surrogate model makes real-time predictions, it calls the surrogate model to calculate the interface reaction force based on the interface state of the bridge structure finite element model, and obtains the deformation state of the bridge structure.

[0009] The system receives actual monitored deformation data of the bridge structure and, based on the difference between the actual monitored deformation data and the model predicted deformation data, performs reverse calculations to update the parameters in the geotechnical finite element model and the surrogate model.

[0010] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, the generated coupled model includes:

[0011] Analyze the architectural information model data and geological exploration data of the bridge;

[0012] Furthermore, the soil and rock finite element model is divided into a near-field soil and rock model surrounding the substructure of the bridge and a far-field boundary model outside of it.

[0013] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, the method includes: parsing the geological exploration data, including:

[0014] Using natural language processing technology, soil layer information and basic physical and mechanical parameters are automatically identified from unstructured geological exploration report text, and constitutive model parameters for nonlinear analysis of the near-field soil and rock model are inferred from the basic physical and mechanical parameters.

[0015] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, it further includes:

[0016] The geogrid described in the exploration data is established as an embedded element, and the nodal degrees of freedom of the embedded element are constrained within the shape function interpolation field of the host soil element.

[0017] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, establishing the coupling constraint relationship includes:

[0018] On the non-coordinated mesh interface between the bridge structure finite element model and the soil finite element model, a Lagrange multiplier field representing the interface contact stress is defined.

[0019] Furthermore, the coupling constraint relationship is established by integrating the product of the Lagrange multiplier field with the difference between the structural displacement and the soil displacement over the entire interface, and requiring the integral to be zero.

[0020] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, the surrogate model is a graph neural network model. This graph neural network model has a graph structure that matches the finite element mesh topology of the geotechnical finite element model, and is used to process node feature information and element connection relationships.

[0021] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, the prediction employs a surrogate model, and the prediction uses an asynchronous time step framework, which includes:

[0022] The interface reaction force calculated by the proxy model is updated at a first preset time step;

[0023] Furthermore, during the time interval between two consecutive updates of the interface reaction force, the finite element model of the bridge structure is iteratively solved multiple times with a second preset time step that is smaller than the first preset time step.

[0024] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, the method further includes the following when the surrogate model performs calculations:

[0025] Uncertainty quantification assessment is performed on the predicted interface reaction force output by the proxy model;

[0026] When the assessed uncertainty exceeds the preset threshold, the solver of the geotechnical finite element model is triggered to perform a calculation to calculate the interface reaction force, and the interface reaction force is used to replace the predicted interface reaction force.

[0027] Furthermore, the proxy model is incrementally learned using the {interface reaction force, interface displacement} data pairs generated in this calculation.

[0028] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, the reverse calculation to update the parameters in the geotechnical finite element model includes:

[0029] Based on Bayesian inference theory, the posterior probability distribution of the parameters of the finite element model of soil and rock to be updated is established.

[0030] Furthermore, the Markov chain Monte Carlo sampling method is used to sample the posterior probability distribution to obtain the optimal estimates of the parameters of the soil and rock finite element model.

[0031] As a preferred embodiment of the real-time prediction method for bridge structure deformation based on finite element analysis described in this invention, updating the surrogate model includes:

[0032] Based on the updated geotechnical finite element model, the retraining of the proxy model is triggered.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. By introducing a graph neural network surrogate model to replace the computationally expensive geotechnical finite element model and adopting an asynchronous time step framework, the computational efficiency is greatly improved, enabling the deformation prediction of complex soil-structure interaction models to meet real-time requirements. At the same time, through online uncertainty assessment and incremental learning mechanisms, the prediction accuracy of the surrogate model at critical moments is guaranteed.

[0035] 2. This invention utilizes actual monitoring data of the bridge to invert and update the parameters of the geotechnical model based on Bayesian inference theory, which solves the problems of inaccurate initial geological parameters and model deterioration over time. After the parameters are updated, the surrogate model is automatically retrained, ensuring long-term synchronization between the prediction model and the actual state of the bridge and improving the reliability of long-term prediction. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0037] Figure 1 This is a flowchart illustrating the overall process of a real-time prediction method for bridge structure deformation based on finite element analysis, as described in one embodiment of the present invention. Detailed Implementation

[0038] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0039] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0040] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0041] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0042] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0043] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0044] Example 1

[0045] Reference Figure 1 This is the first embodiment of the present invention, which provides a method for real-time prediction of bridge structure deformation based on finite element analysis, including:

[0046] S1. In computer-aided software, generate a coupled model that includes a bridge structure finite element model and a geotechnical finite element model, and establish coupling constraint relationships on the coupled model.

[0047] It should be noted that the goal of this step is to transform the scattered and inconsistent engineering raw data into structured finite element model input files through an automated process.

[0048] Specifically, this invention creates finite element models using bridge structures and soil and geotechnical engineering, and couples the two models into a single finite element model.

[0049] Furthermore, for the finite element model of the bridge structure, using a Building Information Modeling (BIM) file conforming to the Industry Foundation Class (IFC) standard as the input source, an automated script (a Python-based script using the IfcOpenShell library) is executed to parse the BIM data. This script is used to extract the geometric information (such as the cross-sectional profiles and tension paths of beams, slabs, and piers defined by IfcExtrudedAreaSolid) and material information (such as the concrete grade associated with IfcMaterialLayerSetUsage) of each bridge component. (C50, steel grade HRB400, etc.) Then, the extracted geometric and material information is translated into preprocessor instructions for specific finite element analysis software (such as Abaqus, ANSYS) through this automated script. For example, slender piers and main beams are discretized into beam elements (BEAM188), and large-sized pile caps and pile foundations are discretized into three-dimensional solid elements (SOLID185). In the finite element analysis software, according to the specific material type and property settings, a command flow of material cards and real constants is created, and a model data file containing node coordinates, element connection relationships, material properties and section definitions is output.

[0050] Furthermore, for geotechnical finite element models, the input source is typically an unstructured geological exploration report in PDF or Word format. First, Natural Language Processing (NLP) technology is used to convert it into editable text. Then, a pre-trained Named Entity Recognition (NER) model, trained with annotations from a large number of geological exploration reports, is deployed. This model identifies and extracts key entities from the text (geological exploration report), such as: {entity: "silty clay", tag: "soil layer name"}, {entity: "15.2m", tag: "bottom elevation"}, {entity: "25 kPa", tag: "Cohesion", etc., and at the same time, the parameters of the same soil layer are structurally correlated. It should be noted that for geotechnical finite element models that require nonlinear analysis, basic physical and mechanical parameters (such as unit weight, cohesion, and internal friction angle) are not enough. It is also necessary to use a built-in engineering knowledge base or a small neural network to automatically infer the parameters required for the constitutive model (such as the modified Cambridge model) based on the soil type and basic parameters. For example, compression index, rebound index, critical state line slope, etc., thereby realizing the automatic conversion from raw text to calculated model parameters.

[0051] Specifically, in one feasible implementation of the present invention, two methods are provided to realize the constitutive model. The first method is to establish a rule-based expert system, which incorporates empirical formulas compiled according to the "Standard for Geotechnical Testing Methods" and related geotechnical engineering handbooks. For example, the compression index is estimated based on the liquid limit and plastic limit of the soil and its natural water content. The second method is to train a multilayer perceptron (MLP) neural network, whose input layer nodes correspond to basic physical and mechanical parameters (such as plasticity index, void ratio, cohesion, etc.), and whose output layer nodes correspond to high-level parameters of the constitutive model (compression index, rebound index, slope of the critical state line, etc.). This neural network is obtained through supervised learning on an existing soil sample database containing a complete set of parameters.

[0052] Furthermore, in order to balance computational accuracy and efficiency, the soil and rock finite element model is divided into two regions: the near-field soil and rock model and the far-field boundary model.

[0053] Specifically, for the regional delineation of the near-field geotechnical model and the far-field boundary model, a three-dimensional space surrounding the outer contour of the bridge substructure (such as a pile foundation) is defined as the near-field geotechnical model, based on the outer contour of the bridge substructure. The area of ​​this model typically extends horizontally to 3 to 5 times the width of the pile foundation, and vertically to a depth of 2 times the pile diameter below the pile bottom. The vast soil mass outside the near-field geotechnical model area is the far-field boundary model.

[0054] Furthermore, in the near-field geotechnical model, unstructured tetrahedral or hexahedral solid elements (such as the C3D8R element in Abaqus) are used and a constitutive model is applied to accurately capture the stress concentration, plastic flow, and failure behavior of the soil around the pile. In the far-field boundary model, a structured mesh with gradually varying sizes from the inside to the outside is used, and a simple linear elastic constitutive model is employed to transfer stress and absorb scattered waves. It is important to note that if the geological exploration data describes reinforced bodies such as geogrids, they are established as embedded elements. This approach eliminates the need to establish shared nodes between the traditional geogrid and the soil. Instead, the nodal degrees of freedom of the embedded elements are constrained within the shape function interpolation field of the host soil element through constraint equations. This efficiently simulates the reinforcing effect of the reinforced body on the soil. Moreover, since the far-field model is outside the near-field model, it can simulate the radiation damping effect of the infinite-domain foundation on the structural dynamic response, preventing false reflections of scattered waves at the artificial boundary.

[0055] Preferably, the artificial boundary is set as a viscoelastic artificial boundary;

[0056] Furthermore, coupling constraint relationships are established on the coupled model, that is, coupling constraint relationships are established between the structure and the soil-rock interface;

[0057] It should be noted that, since the finite element model of bridge structure and the finite element model of soil and rock are usually generated by different software or adopt different meshing strategies, the spatial positions of the nodes on their interface often do not coincide, which can easily form non-coordinated meshes. In order to accurately transmit force and displacement on such interfaces, the present invention adopts a contact algorithm based on the Lagrange multiplier method.

[0058] Specifically, at the non-coordinated mesh interface between the structure and the soil, an independent Lagrange multiplier field representing the interfacial contact stress is introduced. This field is used to establish a weak form of displacement compatibility by forcing it through integral equations across the entire non-coordinated mesh interface. The core constraint equation can be expressed as:

[0059] ;

[0060] in, Represents the entire structure-soil coupling interface region. It is defined in The Lagrange multiplier vector field on the interface represents the contact stress or contact force density at the interface. It is the displacement vector field on the interface belonging to one side of the bridge structure. It is the displacement vector field on the rock and soil side of the interface;

[0061] It should be noted that this integral equation requires that the inconsistency between contact force and displacement be zero under weighted integration across the entire structure-soil interface, thereby ensuring the accurate transmission of force and displacement on both sides of the structure-soil interface on a macroscopic level.

[0062] S2. A proxy model is used for prediction. The proxy model is used to reproduce the mechanical response of the geotechnical finite element model. At the same time, when the proxy model makes real-time predictions, it calls the proxy model to calculate the interface reaction force based on the interface state of the bridge structure finite element model, and obtains the deformation state of the bridge structure.

[0063] It should be noted that this step aims to replace the computationally expensive geotechnical finite element model with a surrogate model that has extremely fast computation speed, thereby breaking through the real-time bottleneck of fully coupled analysis. In order to effectively reproduce the complex nonlinear mechanical behavior of the geotechnical finite element model, this invention adopts a neural network model that is isomorphic to the finite element problem—Graph Neural Network (GNN)—as the surrogate model.

[0064] Furthermore, a graph neural network is constructed that perfectly matches the mesh topology of the geotechnical finite element model. In this graph neural network, each node of the geotechnical finite element model corresponds to a vertex in the graph topology, and the element connection relationship between nodes corresponds to the edge in the graph topology.

[0065] Specifically, the input to this graph neural network is the displacement field applied to the structure-soil coupling interface. The output is the reaction force field fed back to the interface by the finite element model of the soil and rock under this displacement. In other words, the surrogate model actually learns a highly nonlinear mapping relationship from interface displacement to interface reaction force. ;

[0066] In addition, the proxy model needs to be trained offline before it is officially deployed for prediction.

[0067] Specifically, a large number of representative, randomly generated interface displacement load samples are applied to the constructed geotechnical finite element model. And run the finite element solver to calculate the corresponding interface reaction force. The obtained interface displacement load sample-interface reaction force data are used for This is the training dataset for the graph neural network. Using this dataset, the network weights of the surrogate model are optimized by minimizing the loss function (such as mean squared error loss) between the predicted reaction force and the actual reaction force, and by employing backpropagation and gradient descent algorithms (Adam optimizer) until the model reaches the preset accuracy requirement on the validation set.

[0068] It is important to note that for randomly generated interface displacement load samples, quasi-Monte Carlo methods such as Latin Hypercube Sampling (LHS) or Sobol sequences can be used to generate uniformly distributed sample points in the estimated multidimensional space of interface displacement, ensuring effective exploration of the entire possible working range. In addition, the amplitude range of the samples can be determined based on the interface displacement response range obtained through preliminary analysis of the bridge under typical loads (such as design traffic flow, extreme wind load, and temperature gradient), to ensure that the training data covers the situations that the model is most likely to encounter in practical applications.

[0069] Furthermore, in the real-time prediction stage of the proxy model, the coupled model in S1 is decoupled, and the bridge structure finite element model is retained. Then, the geotechnical finite element model is replaced by the trained proxy model. In order to further improve the processing efficiency of this replacement process, this invention adopts an asynchronous time step solution framework.

[0070] Specifically, within this framework, we set two different time steps: one is a first preset time step used to update the soil-rock reaction force, denoted as... Another method is to use a second preset time step to solve the dynamic response of a bridge structure, denoted as... ,and The solution process is as follows:

[0071] First preset time step At the beginning, the displacement state of the bridge structure at the structure-soil coupling interface is transferred to the surrogate model using the finite element model of the bridge structure. At the same time, the surrogate model performs a fast forward propagation calculation and instantaneously outputs the predicted interface reaction force. And in the following whole Within a time period (i.e., from time step) arrive The predicted interfacial reaction force A constant boundary condition is applied to the finite element model of the bridge structure; it should be noted that during this period, the finite element model of the bridge structure uses the second preset time step required for its own stability. Perform multiple times ( (Time intervals) of independent iterative solutions are used to calculate the deformation state under the combined action of external loads (such as traffic loads and wind loads) and the reaction force of the fixed interface. When time reaches... Repeat this process when necessary;

[0072] It should be noted that the solution framework using asynchronous time steps effectively avoids calling the model to solve at each second preset time step, which greatly reduces the amount of computation.

[0073] In addition, to ensure the reliability of the surrogate model in long-term prediction and to prevent error accumulation, this invention also introduces an online self-evaluation and correction mechanism.

[0074] Specifically, a Bayesian Graph Neural Network (BGNN) is used to update the original graph neural network, or a probability distribution layer is added after the output layer of a standard graph neural network. This allows the model to not only predict the expected value of the interface reaction force, but also simultaneously output the uncertainty of its prediction (as expressed by the variance of the predicted value). (to quantify)

[0075] Specifically, if the Bayesian graph neural network is used, the uncertainty is quantified by applying a probability distribution to the weights of the original graph neural network and performing multiple random samplings during prediction (such as using Monte Carlo Dropout). The dispersion (variance) of the multiple prediction results is statistically analyzed. If a probability distribution layer is added after the graph neural network, the output layer is designed to output a Gaussian distribution of parameters (i.e., mean and variance). The training objective of the graph neural network is to maximize the log-likelihood of the real interface reaction force, thereby directly obtaining the predicted value (mean) and its uncertainty (variance).

[0076] Specifically, during prediction at each first preset time step, the uncertainty of the surrogate model output is evaluated, and a preset uncertainty threshold is set. When the prediction variance If the surrogate model lacks confidence in its current predictions, this usually occurs when the input displacement pattern exceeds the distribution range of the offline training data. In this case, by actively triggering a call to the solver of the original soil and rock finite element model, the interface reaction force calculated by the solver can be obtained. To replace the unreliable predictions of the current proxy model And use this value to complete the calculation of the current time step;

[0077] Specifically, the methods for determining this preset uncertainty threshold include, but are not limited to, the following:

[0078] Based on experience: During the offline training phase of the surrogate model, the prediction variance of the model under different inputs is evaluated using the validation dataset, and its distribution pattern is statistically analyzed. The 95th percentile or mean of the prediction variance of all validation samples plus three times the standard deviation can be set as the initial uncertainty threshold.

[0079] Dynamic threshold based on relative error: This method correlates the uncertainty threshold with the magnitude of the predicted reaction force, setting it as the square of a certain percentage of the expected value of the predicted reaction force. ,in, This is a preset scaling factor (such as 0.05 or 0.1). To predict the expected value of the interface reaction force, this method can be adapted to reaction forces of different magnitudes;

[0080] Based on the engineering safety requirements: According to the bridge structure safety assessment specifications, the upper limit of the allowable interface reaction force error is calculated and converted into the uncertainty threshold of the surrogate model prediction result. That is, if a 5% error in the reaction force will cause the structural response to exceed the upper limit of the reaction force error, the uncertainty threshold is set at a level that can guarantee the error is within 5% with a high probability.

[0081] It should be noted that by actively invoking the data, not only can a single prediction error be corrected, but a new data pair can also be generated online. ;

[0082] Furthermore, this new data is used to perform incremental learning directly on the agent model;

[0083] Specifically, this incremental learning includes, but is not limited to, the following methods:

[0084] Online fine-tuning: The newly generated data (interface reaction force, interface displacement) is treated as a mini-batch sample, and gradient descent is applied to the current surrogate model (graph neural network) for several cycles. To prevent the model from overfitting to the new data and forgetting old knowledge (i.e., catastrophic forgetting), a learning rate much smaller than that used in offline training is typically employed; for example, the offline training learning rate is 10. -3 The online fine-tuning learning rate can then be set to 10. -7 Or 10 -10 ;

[0085] Experience replay: By establishing an experience pool to store historical "high-value" data pairs (the "high-value" data pairs refer to the data pairs that triggered the solver's calculation), a small portion of old data is randomly extracted from this experience pool during incremental learning and paired with the newly generated data to form a training batch. The gradient of the surrogate model (graph neural network) is then updated to effectively mitigate the catastrophic forgetting problem.

[0086] Training Error Correction: Keeping the surrogate model unchanged, an additional lightweight error correction surrogate model is trained. The input of this error correction model is the interface displacement, and the output is the prediction error of the surrogate model under that input. When the solver is triggered, the newly generated data is used to correct the error. This is used to train the error correction model. It should be noted that if this method is used, the final prediction result in subsequent predictions will be the sum of the predictions from the surrogate model and the predictions from the error correction model.

[0087] S3. Receive the actual monitored deformation data of the bridge structure, and based on the difference between the actual monitored deformation data and the model predicted deformation data, perform reverse calculation to update the parameters in the geotechnical finite element model, as well as the surrogate model.

[0088] It should be noted that this step aims to dynamically calibrate the geotechnical finite element model and surrogate model by utilizing actual monitoring data collected from sensors deployed on the bridge (such as GPS, hydrostatic level, strain gauges, etc.), thereby addressing the issues of model parameter uncertainty and model degradation over time, and ensuring the long-term reliability of the prediction results. Meanwhile, due to the complexity and inherent uncertainty of geological conditions, the initial parameters (prior information) obtained from the geological survey report in step S1 often deviate from the actual engineering conditions. Therefore, this invention employs a Bayesian inference framework to transform the parameter inversion problem into a problem of solving the posterior probability distribution of the parameters.

[0089] Furthermore, it should be emphasized that the real-time prediction of this invention is mainly reflected in the use of an efficient surrogate model for rapid deformation prediction in step S2. The parameter back calculation and model update process in step S3 is a relatively computationally intensive background calibration task. It is not performed synchronously with each deformation prediction, but is executed once every long time period (e.g., daily, weekly, or monthly).

[0090] Furthermore, according to Bayes' theorem, the posterior probability distribution of the parameters of the soil finite element model to be updated (such as the elastic modulus, internal friction angle, cohesion, etc. of each soil layer) can be expressed as:

[0091] ;

[0092] in, It is the posterior probability distribution, representing the probability distribution after obtaining actual monitoring data. Then, the model parameters are The probability is the final objective to be solved, which integrates information from prior knowledge and actual observation data; A vector representing one or more sets of parameters of the soil-rock finite element model to be inverted, for example, , These represent the elastic modulus, cohesion, and internal friction angle of different soil layers, respectively. This represents the actual deformation dataset collected from specific monitoring points on the bridge structure, for example... , This represents the deformation or displacement data actually monitored by physical sensors at the first monitoring point of the bridge structure, and so on. It is the likelihood function, representing the likelihood of a model with parameters of... Under these conditions, predict the observed data. It's important to note that calculating this function requires running the original coupled model, i.e., for a given set of parameters... By constructing finite element models of the bridge structure and geotechnical structures, the results were calculated in relation to the monitoring data. Corresponding model predicted deformation value , Representative in Under identical time, load, and environmental conditions, the deformation or displacement data predicted by calculation at the same monitoring point can be defined as follows: Since the likelihood function typically assumes that the prediction error follows a Gaussian distribution, its form can be defined as:

[0093] ;

[0094] in, It is a prior probability distribution, representing the probability distribution of parameters before any monitoring data is obtained. The initial beliefs or knowledge, which can be derived from the parameter values ​​and their empirical uncertainty range extracted from the geological exploration report in step S1, can usually be assumed to be uniformly distributed or normally distributed; It is the evidence factor, also known as the marginal likelihood, which plays a normalization role to ensure that the posterior probability integral is 1. However, in actual calculation, this factor is usually difficult to solve for its optimal value directly due to its complexity and high dimensionality. Therefore, this invention uses the Markov Chain Monte Carlo (MCMC) sampling method to sample the posterior probability distribution.

[0095] Specifically, by constructing a Markov chain whose stationary distribution is exactly the target posterior probability distribution. By performing a "random walk" on the posterior probability distribution for a sufficiently long time, a large number of parameter samples that follow the posterior distribution are generated. After obtaining a large number of parameter samples, the final values ​​of the parameters are inferred by statistical analysis of these samples.

[0096] Specifically, this statistical analysis method uses the mean, median, or mode of all samples as the optimal estimates of the parameters of the geotechnical finite element model. In addition, the distribution of all samples (such as variance and confidence interval) also provides a quantitative basis for assessing the uncertainty of the parameters.

[0097] It should be noted that when the key parameters in the geotechnical finite element model (such as...) After being updated through the above inversion process, it means that the constitutive relationship of the soil and rock materials has changed. Therefore, the original surrogate model (graph neural network) was trained based on training data generated from the old soil and rock model. The "interface displacement-interface reaction force" mapping relationship it learned is no longer accurate and cannot represent the updated soil and rock mechanical behavior. Therefore, the surrogate model must be updated. The specific steps are as follows:

[0098] Once the parameters of the geotechnical finite element model are successfully updated, the retraining process for the surrogate model is triggered.

[0099] Using the updated geotechnical finite element model as the "true value" model, similar to the offline training process in step S2, a large number of representative, randomly generated interface displacement load samples are applied to the model again, and the finite element solver is run to calculate the corresponding new interface reaction force. Thus, new {interface reaction force, interface displacement} data pairs that can reflect the latest geotechnical properties are generated, forming a new training dataset.

[0100] Using this new training dataset, the surrogate model (graph neural network) built in step S2 is retrained, and the network weights are optimized through backpropagation and gradient descent algorithms until the surrogate model can reproduce the mechanical response of the updated geotechnical finite element model with sufficiently high accuracy.

[0101] After retraining, the new version of the proxy model is deployed to the real-time prediction system, replacing the old proxy model.

[0102] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0103] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0104] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0105] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0106] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0107] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for real-time prediction of bridge structural deformation based on finite element analysis, characterized in that, include: In computer-aided software, a coupled model containing a bridge structure finite element model and a geotechnical finite element model is generated, and coupling constraint relationships are established on this coupled model. The generated coupling model includes: Analyze the architectural information model data and geological exploration data of the bridge; Furthermore, the soil and rock finite element model is divided into a near-field soil and rock model surrounding the substructure of the bridge and a far-field boundary model outside of it. Establishing the coupling constraint relationship includes: On the non-coordinated mesh interface between the bridge structure finite element model and the soil finite element model, a Lagrange multiplier field representing the interface contact stress is defined. Furthermore, the coupling constraint relationship is established by integrating the product of the Lagrange multiplier field with the difference between the structural displacement and the soil displacement over the entire interface, and requiring the integral to be zero. A surrogate model is used for prediction. The surrogate model is used to reproduce the mechanical response of the soil and rock finite element model. At the same time, when the surrogate model makes real-time predictions, it calls the surrogate model to calculate the interface reaction force based on the interface state of the bridge structure finite element model, and obtains the deformation state of the bridge structure. The proxy model is a graph neural network model, which has a graph structure that matches the finite element mesh topology of the geotechnical finite element model, and is used to process node feature information and element connection relationships. The system receives actual monitored deformation data of the bridge structure and, based on the difference between the actual monitored deformation data and the model predicted deformation data, performs reverse calculations to update the parameters in the geotechnical finite element model and the surrogate model.

2. The method for real-time prediction of bridge structure deformation based on finite element analysis as described in claim 1, characterized in that, The analysis of the geological exploration data includes: Using natural language processing technology, soil layer information and basic physical and mechanical parameters are automatically identified from unstructured geological exploration report text, and constitutive model parameters for nonlinear analysis of the near-field soil and rock model are inferred from the basic physical and mechanical parameters.

3. The method for real-time prediction of bridge structure deformation based on finite element analysis as described in claim 1, characterized in that, Also includes: The geogrid described in the exploration data is established as an embedded element, and the nodal degrees of freedom of the embedded element are constrained within the shape function interpolation field of the host soil element.

4. The method for real-time prediction of bridge structure deformation based on finite element analysis as described in claim 1, characterized in that, The prediction is performed using a surrogate model, and the prediction employs an asynchronous time-step framework, which includes: The interface reaction force calculated by the proxy model is updated at a first preset time step; Furthermore, during the time interval between two consecutive updates of the interface reaction force, the finite element model of the bridge structure is iteratively solved multiple times with a second preset time step that is smaller than the first preset time step.

5. The method for real-time prediction of bridge structure deformation based on finite element analysis as described in claim 1, characterized in that, The calculation process of the proxy model also includes: Uncertainty quantification assessment is performed on the predicted interface reaction force output by the proxy model; When the assessed uncertainty exceeds the preset threshold, the solver of the geotechnical finite element model is triggered to perform a calculation to calculate the interface reaction force, and the interface reaction force is used to replace the predicted interface reaction force. Furthermore, the proxy model is incrementally learned using the {interface reaction force, interface displacement} data pairs generated in this calculation.

6. The method for real-time prediction of bridge structure deformation based on finite element analysis as described in claim 1, characterized in that, The reverse calculation to update the parameters in the geotechnical finite element model includes: Based on Bayesian inference theory, the posterior probability distribution of the parameters of the finite element model of soil and rock to be updated is established. Furthermore, the Markov chain Monte Carlo sampling method is used to sample the posterior probability distribution to obtain the optimal estimates of the parameters of the soil and rock finite element model.

7. The method for real-time prediction of bridge structure deformation based on finite element analysis as described in claim 1 or 6, characterized in that, Updating the agent model includes: Based on the updated geotechnical finite element model, the retraining of the proxy model is triggered.

Images