Intelligent Analysis Method for Strong Earthquake Response of Bridge Structures Embedded with Physical Information

By constructing a simplified physical model and a multi-level nested prediction model using a deep LSTM network, and combining it with structural dynamics principles, the computational efficiency and accuracy issues of traditional bridge seismic analysis methods were resolved. This enabled efficient and accurate bridge response prediction under near-field fault earthquakes, enhancing the model's engineering reliability.

CN123116388APending Publication Date: 2026-11-17SHIJIAZHUANG TIEDAO UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611125264.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-07-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Traditional bridge seismic analysis methods suffer from high computational costs, low efficiency, and inaccurate results when dealing with nonlinear problems under strong earthquakes in the near field of faults. Furthermore, purely data-driven methods have insufficient prediction accuracy and poor interpretability in strong earthquake scenarios.

Method used

A simplified physical model is constructed and combined with a deep LSTM network. By introducing equilibrium constraints, geometric consistency, and nonlinear constitutive relations at the column ends, a multi-level nested prediction model is formed. The model is trained using structural dynamics laws, which reduces data dependence and improves prediction accuracy and interpretability.

Benefits of technology

It achieves efficient and accurate prediction of bridge nonlinear response under near-field fault earthquakes, avoiding abnormal energy accumulation and displacement inconsistency, and improving the engineering credibility and generalization ability of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116388A_ABST
    Figure CN123116388A_ABST
Patent Text Reader

Abstract

This invention belongs to the field of bridge seismic analysis technology, and provides an intelligent analysis method for the strong earthquake response of bridge structures by embedding physical information. This method integrates equilibrium constraints, geometric consistency constraints, and column-end nonlinear constitutive constraints into the model training process, ensuring that the model's prediction results are always constrained by fundamental physical laws. This effectively avoids physical inconsistencies such as abnormal energy accumulation, incompatible displacement responses, and mismatches between internal forces and deformations that are common in pure data models, significantly improving the reliability of prediction results. The presence of physical constraints endows the model with stronger generalization and extrapolation capabilities. Even when the ground motion amplitude, spectral characteristics, or structural parameters exceed the range of the training samples, the model can still provide reasonable prediction results guided by physical laws. This solves the problem of a sharp drop in prediction accuracy of pure data models in unknown strong earthquake scenarios, and better copes with the strong randomness and extreme nature of near-field ground motions of faults.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of bridge seismic analysis technology, specifically relating to an intelligent analysis method for the strong earthquake response of bridge structures with embedded physical information. Background Technology

[0002] The structural response of bridges near or across fault zones under strong earthquakes exhibits significant uniqueness and complexity. On the one hand, near-field ground motions often contain significant long-period pulses, large velocity peaks, and directional effects, leading to high-energy, short-duration, and non-stationary dynamic inputs on the structure. On the other hand, bridge structures enter a highly nonlinear state under strong earthquakes, manifesting as large deformations in critical components such as piers and bearings, concrete cracking, steel yielding, plastic hinge formation, and even local collapse. Furthermore, fault displacement can cause uncoordinated displacements between bridge spans, resulting in severe damage such as collisions and beam collapses. These response characteristics make the seismic analysis of bridges in fault zones highly nonlinear, path-dependent, and characterized by damage accumulation.

[0003] Traditional bridge seismic analysis methods primarily rely on numerical simulation techniques such as the finite element method (FEM). While these methods have a solid theoretical foundation, they still have a number of limitations when dealing with extreme nonlinear problems under strong near-field earthquakes on faults: First, to accurately capture the nonlinear behaviors of structures such as plastic damage, contact collisions, and large deformations during strong earthquakes, highly refined finite element models are usually required. The modeling process is complex and time-consuming, demanding a high level of expertise from the users. Second, strongly nonlinear dynamic analysis often requires extremely small stepping and complex iterative algorithms, resulting in high computational costs and low efficiency. This is especially true when conducting parameter sensitivity analysis or probabilistic earthquake demand assessments, making it difficult to meet the timeliness requirements of practical engineering projects. Furthermore, numerical models are prone to convergence difficulties under conditions of material softening, contact separation, or large deformation, affecting the continuity of the analysis process and the reliability of the results. Finally, the cyclic constitutive models and damage evolution laws of materials such as concrete and steel often contain many empirical parameters. The uncertainty of these parameters and the difficulty in calibration directly affect the accuracy of the final analysis results.

[0004] In recent years, purely data-driven methods, represented by machine learning, have shown some potential in structural response prediction, but they still face significant shortcomings when applied to the nonlinear analysis of bridges under strong earthquakes. On the one hand, methods such as deep learning typically rely on large amounts of high-quality training data, while nonlinear response data of real structures under strong earthquakes, especially damage data with near-field fault characteristics, are extremely scarce, making it difficult to support sufficient model training. On the other hand, purely data-driven models may lead to physically unreasonable predictions, such as abnormal energy accumulation or inconsistent displacement responses. Furthermore, these models have weak extrapolation capabilities; when ground motion characteristics or structural parameters exceed the range of training samples, their prediction accuracy and reliability decrease significantly, limiting their practical application in unknown strong earthquake scenarios. Simultaneously, the internal structure of these models is often a "black box," making it difficult to reveal physical processes crucial for seismic design and assessment, such as the development sequence of plastic hinges and damage accumulation mechanisms, resulting in insufficient interpretability and engineering credibility. Summary of the Invention

[0005] This invention provides an intelligent analysis method for the strong earthquake response of bridge structures by embedding physical information, aiming to solve the technical problems described in the background art above.

[0006] To achieve the above objectives, the technical solution adopted by this invention is: to provide an intelligent analysis method for the strong earthquake response of bridge structures with embedded physical information, comprising the following steps: S1. Construct a simplified physical model of the bridge structure, and establish equilibrium constraint equations, geometric consistency constraint equations, and column end nonlinear constitutive relation equations based on the simplified physical model. S2, perform time-series sampling processing on the seismic acceleration time history to obtain time-series input samples adapted to LSTM; S3, construct a single-level deep LSTM basic network; S4. Based on the single-level deep LSTM basic network, a multi-level LSTM nested prediction model of "state-geometry-constitutive" is constructed to form a hierarchical progressive prediction link from the seismic motion time history input to the displacement / velocity, column end rotation / curvature and bending moment time history of the bridge structure. S5. Based on the completeness of physical information in actual engineering, select one of the following two paths to train the multi-level nested prediction model: Fully Physically Constrained Path: The constraint terms in the loss function are constructed using the equilibrium constraint equation, the geometric consistency constraint equation, and the column end nonlinear constitutive relation equation in S1. The constraint terms are then weighted and summed to form the fully constrained total loss function, which drives the multi-level nested prediction model to be trained. Partial physical constraint path: A preset number of high-precision finite element results of bridges under strong earthquakes are introduced as a calibration dataset. The geometric consistency constraint equation, the column end nonlinear constitutive relation equation, and the calibration dataset are selected to construct the constraint terms in the loss function. The constraint terms are weighted and summed to form a partial constraint total loss function, which drives the multi-level nested prediction model to be trained. S6. After training is completed, the ground motion time history is input into the multi-level nested prediction model. The multi-level nested prediction model directly outputs the displacement / velocity, column end rotation / curvature and bending moment time history of the bridge structure, realizing rapid analysis and prediction of the nonlinear response of the bridge under strong earthquakes.

[0007] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, the simplified physical model equates the mass of the bridge span to the concentrated mass at the top of the fixed support. The total structural deformation Decomposed into horizontal deformation components of the support Elastic-plastic bending deformation component of the bridge pier itself With the basic elastic deformation component The basic elastic deformation component includes the basic translational deformation component. With the basic rotational deformation component The equivalent height of the bridge piers is .

[0008] It should be noted that treating the span mass as equivalent to concentrated mass significantly simplifies the derivation and solution of the structure's dynamic equations, while also aligning with the stress characteristics of bridge spans under strong earthquakes, where rigid body motion is the dominant force. The multi-component decomposition of the total structural deformation clearly distinguishes the deformation contributions from different locations.

[0009] By employing reasonable mass equivalence and deformation decomposition, the complexity of the physical model is significantly reduced while ensuring that the core mechanical behavior is not lost. This also reduces the difficulty of solving subsequent constraint equations and provides a clear physical basis for the independent prediction and physical constraints of responses at different levels.

[0010] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, the equilibrium constraint equations include structural dynamic equilibrium constraint equations and column end moment equilibrium constraint equations.

[0011] It should be noted that the structural dynamics equilibrium constraint equations are based on Newton's second law and are used to constrain the relationship between the motion state of concentrated mass points and external forces; the column end moment equilibrium constraint equations are used to ensure the balance between the internal forces at the pier column ends and the external loads and structural inertial forces, and are the key constraints connecting macroscopic motion and local internal forces.

[0012] By introducing two types of equilibrium constraint equations, the motion state and internal force response predicted by the model can conform to the basic laws of mechanics, effectively avoiding physical inconsistencies such as abnormal energy accumulation and mismatch between internal force and displacement that may occur in purely data-driven models, and improving the reliability of the prediction results.

[0013] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, the geometric consistency constraint equation includes the mapping relationship equation between lumped mass displacement and pier top plastic displacement and the geometric relationship equation between pier top plastic displacement and column end rotation / curvature.

[0014] It should be noted that the geometric consistency constraint equation ensures the geometric coordination between different levels of response, solves the displacement response inconsistency problem that may occur in pure data-driven models, and provides a physical basis for predicting local geometric response from macroscopic motion state, thereby improving the model's ability to predict local damage characteristics.

[0015] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, the nonlinear constitutive equation at the column end is based on the bending-rotation / curvature relationship in the plastic zone at the column end, combined with the skeleton curve and hysteresis rule, to form a constitutive equation with the column end rotation / curvature time history as input, considering the nonlinear historical state, and the column end bending moment as output.

[0016] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, the time-series sampling process includes: cutting the seismic acceleration time history using a sliding time window of fixed length, setting an overlapping area between adjacent time windows, and constructing standardized time-series input samples.

[0017] It should be noted that by introducing nonlinear constitutive equations that take into account historical states, the plastic damage evolution process of bridge piers under strong seismic cyclic loading can be accurately simulated, avoiding the problems of difficult parameter calibration and large uncertainty in traditional constitutive models. At the same time, it provides physical constraints for the prediction of column end bending moments and improves the accuracy of internal force response prediction.

[0018] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, the single-level deep LSTM base network uses two stacked LSTM layers to extract high-order temporal features, sets residual connections between layers, and performs tensor concatenation between the hidden state of the first LSTM layer and the output of the second LSTM layer, and maps it to the output space through two fully connected layers.

[0019] It should be noted that the two stacked LSTM layers can progressively extract low-order and high-order temporal features of the ground motion time history; the inter-layer residual connections can alleviate the gradient vanishing problem during the training process of deep networks, and improve the training depth and convergence speed of the network; the tensor splicing operation can fuse feature information from different levels and retain more temporal details; the two fully connected layers are used to map the extracted temporal features to the corresponding response output space.

[0020] Through the above basic network structure design, the network's ability to extract complex temporal features and its training stability have been significantly improved. It can better capture the special features of near-field ground motion of faults, such as long-period pulses and large velocity peaks, laying the foundation for accurate prediction of subsequent multi-level nested models.

[0021] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by this invention, the multi-level nested prediction model includes: The motion state prediction module takes the seismic acceleration time series sample as input and outputs the displacement and velocity time series data of the concentrated mass points of the bridge. The geometric state mapping module takes the output of the motion state prediction module as input and outputs the timing data of the geometric response of the bridge column end rotation angle / curvature. The plastic zone constitutive prediction module takes the output of the geometric state mapping module as input and outputs the time series data of bending moment and internal force at the bridge column ends.

[0022] It should be noted that the three modules are nested sequentially according to the physical causal relationship of "motion state - geometric deformation - internal force response". The output of the previous module serves as the input of the next module. Each module is built on a single-level deep LSTM base network and has its own clearly defined physical constraints. The hierarchical progression between modules is consistent with the response process of structural dynamics, and the output of each module has a clear physical meaning.

[0023] By using a modular, multi-level nested structure, the complex nonlinear response prediction is decomposed into multiple simple sub-tasks, reducing the training difficulty of the model. At the same time, the clear physical hierarchy makes the model's prediction process interpretable, revealing physical processes such as plastic hinge development and damage accumulation, improving the model's engineering credibility, and solving the problem of the uninterpretable black box of purely data-driven models.

[0024] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, in step S5, the preset number of the calibration dataset is determined by covering the bridge structure parameters and seismic motion input parameters under representative working conditions, and by gradually expanding the finite element samples according to a preset gradient and combining it with independent validation set evaluation.

[0025] It should be noted that gradually expanding the finite element sample size can avoid the computational waste caused by generating a large number of samples at once.

[0026] By using the above method to determine the number of calibration datasets, the number of finite element samples generated can be reasonably controlled while ensuring the model's generalization ability. This reduces the computational cost during data preparation and avoids problems such as model overfitting due to insufficient samples or computational inefficiency due to excessive samples.

[0027] In one possible implementation of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by the present invention, when the multi-level nested prediction model reaches convergence in both the prediction error of key response quantities and the physical consistency residual, and the performance brought by continuing to increase the number of samples is lower than the preset threshold, the corresponding number of samples is the preset number of the calibration dataset.

[0028] It should be noted that the prediction errors of key response quantities include displacement error, velocity error, column end rotation angle error, and column end bending moment error of concentrated mass points in the bridge; physical consistency residuals include equilibrium constraint residuals, geometric consistency constraint residuals, and constitutive relation constraint residuals; the preset thresholds can be adjusted according to the engineering accuracy requirements.

[0029] By simultaneously considering both prediction error and physical consistency residual as convergence conditions, the problem of insufficient physical rationality of the model caused by using prediction error as the sole criterion can be avoided. This allows for the scientific determination of the minimum effective sample size, further improving the efficiency of model training and the reliability of prediction results, and meeting the accuracy requirements of engineering applications.

[0030] The beneficial effects of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by this invention are as follows: Compared with the prior art, the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided by this invention has the following beneficial effects: The fully physically constrained path utilizes only the fundamental physical laws of structural dynamics to construct a loss function to drive model training, without relying on any measured earthquake damage data or finite element calibration data, effectively alleviating the constraints of data scarcity on model application. The partially physically constrained path, on the other hand, can further improve prediction accuracy by combining data with physical constraints when a small amount of high-precision finite element data is available, maximizing data utilization efficiency.

[0031] By embedding equilibrium constraints, geometric consistency constraints, and nonlinear constitutive constraints at the column ends into the model training process, the model's prediction results are always constrained by fundamental physical laws. This effectively avoids physical inconsistencies such as abnormal energy accumulation, incompatible displacement responses, and mismatches between internal forces and deformations that are prone to occur in pure data models, significantly improving the reliability of the prediction results.

[0032] The existence of physical constraints gives the model a stronger ability to generalize and extrapolate. When the ground motion amplitude, spectral characteristics or structural parameters exceed the range of the training samples, the model can still give reasonable prediction results under the guidance of physical laws. This solves the problem of the sharp drop in prediction accuracy of pure data models in unknown strong earthquake scenarios and can better cope with the strong randomness and extreme nature of near-field ground motion of faults.

[0033] The constructed "state-geometry-constitutive" multi-level nested prediction model strictly follows the physical causal logic of structural dynamics. The model's prediction process is completely consistent with the actual response process of bridges under strong earthquakes. That is, the seismic input first causes changes in the macroscopic motion state of the structure, which then translates into local geometric deformation of the piers, and finally generates internal force response through material constitutive relations. The output of each module has a clear physical meaning. Engineers can not only obtain the final response results, but also clearly observe physical mechanisms that are crucial to seismic design, such as the development sequence of plastic hinges and the damage accumulation process. This significantly improves the engineering credibility and practical value of the model, and can provide scientific and reliable technical support for the seismic design, performance evaluation, and reinforcement of bridges in fault zones. Attached Figure Description

[0034] Figure 1 A flowchart illustrating the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided in this embodiment of the invention; Detailed Implementation To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0035] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. The following description of at least one exemplary embodiment is actually illustrative only and is in no way intended to limit this application or its application or use. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0036] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this application. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0037] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application. It should also be understood that, for ease of description, the dimensions of the various parts shown in the drawings are not drawn to actual scale. Techniques, methods, and devices known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and devices should be considered part of the specification. In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values. It should be noted that similar reference numerals and letters in the following drawings denote similar items; therefore, once an item is defined in one drawing, it need not be further discussed in subsequent drawings.

[0038] In the description of this application, it should be understood that the orientation or positional relationship indicated by directional terms such as "front, back, up, down, left, right", "horizontal, vertical, horizontal" and "top, bottom" is usually based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this application and simplifying the description. Unless otherwise stated, these directional terms 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, and therefore should not be construed as a limitation on the scope of protection of this application; the directional terms "inner" and "outer" refer to the inner and outer contours relative to the outline of each component itself.

[0039] For ease of description, spatial relative terms such as "above," "over," "on the upper surface of," "above," etc., are used herein to describe the spatial positional relationship of a device or feature as shown in the figures to other devices or features. It should be understood that spatial relative terms are intended to encompass different orientations in use or operation beyond the orientation of the device as described in the figures. For example, if the device in the figures were inverted, a device described as "above" or "above" other devices or structures would subsequently be positioned as "below" or "under" other devices or structures. Thus, the exemplary term "above" can include both "above" and "below." The device may also be positioned in other different ways, and the spatial relative descriptions used herein will be interpreted accordingly.

[0040] Furthermore, it should be noted that the use of terms such as "first" and "second" to define components is merely for the purpose of distinguishing the corresponding components. Unless otherwise stated, the above terms have no special meaning and therefore should not be construed as limiting the scope of protection of this invention.

[0041] Please refer to the following: Figure 1The present invention will now describe the intelligent analysis method for strong earthquake response of bridge structures based on embedded physical information. The intelligent analysis method for strong earthquake response of bridge structures based on embedded physical information includes the following steps: S1. Construct a simplified physical model of the bridge structure, equating the mass of the bridge span to the concentrated mass at the top of the fixed support. The total structural deformation Decomposed into horizontal deformation components of the support Elastic-plastic bending deformation component of the bridge pier itself With the basic elastic deformation component The basic elastic deformation component includes the basic translational deformation component. With the basic rotational deformation component The equivalent height of the bridge piers is Based on the simplified physical model, equilibrium constraint equations, geometric consistency constraint equations, and nonlinear constitutive relation equations at the column ends are established. The equilibrium constraint equations include: Structural dynamic equilibrium constraint equations: ;and Column end moment equilibrium constraint equations: .

[0042] The geometric consistency constraint equations include: Equation relating plastic displacement at the pier top to concentrated mass displacement: ;and Geometric constraint equations for pier top plastic displacement - column end rotation / curvature: .

[0043] Nonlinear constitutive relation equations at the column ends: .

[0044] Among them, among them, For relative velocity, For relative acceleration, The damping coefficient is... For equivalent restoring force, This refers to the horizontal acceleration of the ground. Time history of the concentrated mass horizontal displacement of the bridge. The time history of plastic displacement at the pier top; H is the geometrically calculated height of the pier, representing the vertical distance between the control section at the bottom of the pier and the deformation control point at the top of the pier; The length of the plastic hinge; Pick or ; For the column end bending moment, For the column end rotation angle time history, Time history for controlling the curvature of the section at the column base or column end. These are internal variables that describe nonlinear historical states.

[0045] The nonlinear constitutive equation at the column end is based on the bending-rotation / curvature relationship in the plastic zone of the column end, combined with the skeleton curve and hysteresis rule, to form a constitutive equation that takes the column end rotation / curvature time history as input, considers the nonlinear historical state, and takes the column end bending moment as output. The bending-rotation / curvature relationship of the plastic zone at the column end: Using the column end rotation angle as input: ; Inputting the column end curvature: ; in, For the column end bending moment, This refers to the change in the rotation angle of the column top or end node over time, reflecting the dynamic response of the component end rotation over time. This refers to the change in curvature of the column base or end section over time. These are internal variables that describe nonlinear historical states.

[0046] The skeleton curve (envelope) and hysteresis rule: ,here and These represent the positive and negative skeleton curves, respectively.

[0047] The form of a piecewise function is: ; in, Desirable or , The instantaneous rotation angle of the column top or end node. The instantaneous curvature is the base or tip of the column.

[0048] S2, perform time-series sampling processing on the seismic acceleration time history to obtain time-series input samples adapted to LSTM (Long Short-Term Memory Network). The time-series sampling processing specifically involves: using a sliding time window of fixed length to cut the seismic acceleration time history, setting an overlapping area between adjacent time windows, and constructing standardized time-series input samples.

[0049] S3, construct a single-level deep LSTM base network; the single-level deep LSTM base network uses two stacked LSTM layers to extract high-order temporal features, sets residual connections between layers, and concatenates the hidden state of the first LSTM layer with the output of the second LSTM layer using tensors, and finally maps to the output space through two fully connected layers.

[0050] The single-level deep LSTM base network includes a first-layer LSTM, a second-layer LSTM, residual connections, a tensor stitching module, a first fully connected layer, and a second fully connected layer. Depending on its position in the multi-level nested prediction model, the inputs to this base network are time series samples of ground motion acceleration, displacement and velocity time series of concentrated mass points on the bridge, or time series of column end rotation angles or curvatures.

[0051] The first LSTM layer processes all time steps of the input sequence, extracting basic temporal features, and then inputs the resulting first temporal feature sequence completely into the second LSTM layer. The second LSTM layer further extracts the temporal correlations between ground motion and lumped mass motion, lumped mass motion and column end geometric response, and column end geometric response and column end bending moment. Both LSTM layers use the same number of hidden units, and their outputs have the same number of time steps and feature dimensions.

[0052] A residual connection is set between the outputs of the first and second LSTM layers. The values ​​at the same time step and the same feature position in the two temporal feature sequences are added together to form a residual-enhanced temporal feature sequence. Subsequently, the first temporal feature sequence and the residual-enhanced temporal feature sequence are concatenated step by step along the feature dimension, while keeping the number of samples, the number of time steps, and the time order unchanged during the concatenation process.

[0053] The spliced ​​and fused temporal features are sequentially input into two fully connected layers. The first fully connected layer uses the ReLU activation function (i.e., a linear rectified function) for nonlinear transformation, while the second fully connected layer uses a linear output method to map each time step. Thus, the motion state prediction module outputs the displacement and velocity time series of the bridge's concentrated mass points, the geometric state mapping module outputs the column end rotation angle or curvature time series, and the plastic zone constitutive prediction module outputs the column end bending moment time series.

[0054] S4. Based on the single-level deep LSTM basic network, a multi-level LSTM nested prediction model of "state-geometry-constitutive" is constructed to form a hierarchical progressive prediction link from the seismic motion time history input to the displacement / velocity, column end rotation / curvature and bending moment time history of the bridge structure. The multi-level nested prediction model includes: The motion state prediction module takes the seismic acceleration time series sample as input and outputs the displacement and velocity time series data of the concentrated mass points of the bridge. The geometric state mapping module takes the output of the motion state prediction module as input and outputs the timing data of the geometric response of the bridge column end rotation angle / curvature. The plastic zone constitutive prediction module takes the output of the geometric state mapping module as input and outputs the time series data of bending moment and internal force at the bridge column ends.

[0055] S5. Based on the completeness of physical information in actual engineering, select one of the following two paths to train the multi-level nested prediction model: Fully Physically Constrained Path: The constraint terms in the loss function are constructed using the equilibrium constraint equation, the geometric consistency constraint equation, and the column end nonlinear constitutive relation equation in S1. The constraint terms are then weighted and summed to form the fully constrained total loss function, which drives the multi-level nested prediction model to be trained. Partial physical constraint path: A preset number of high-precision finite element results of bridges under strong earthquakes are introduced as a calibration dataset. The geometric consistency constraint equation, the column end nonlinear constitutive relation equation, and the calibration dataset are selected to construct the constraint terms in the loss function. The constraint terms are weighted and summed to form a partial constraint total loss function, which drives the multi-level nested prediction model to be trained.

[0056] Specifically, when the physical information of the bridge is complete—that is, when it is possible to establish equilibrium constraint equations, geometric consistency constraint equations, and column-end nonlinear constitutive relation equations based on the existing physical information—a fully physical constraint path is selected for training. When the physical information of the bridge only supports the establishment of geometric consistency constraint equations and column-end nonlinear constitutive relation equations, a partial physical constraint path is selected for training.

[0057] The preset number of the calibration dataset is determined by covering representative working conditions of bridge structural parameters and seismic input parameters, and by gradually expanding the finite element samples and combining them with independent validation set evaluation. When the multi-level nested prediction model converges in both the prediction error of key response quantities and the physical consistency residual, and the performance brought by continuing to increase the number of samples is lower than the preset threshold, the corresponding number of samples is the preset number of the calibration dataset.

[0058] Specifically, the method for determining the calibration dataset is as follows: 01. Determine the range of variables that the finite element sample should cover, specifically including three categories: (1) Bridge structural parameters: pier elastic modulus, column cross-sectional area, moment of inertia, span mass distribution, beam cross-sectional properties (equivalent bending stiffness, equivalent moment of inertia, material grade), total bridge mass, span, column end plastic parameters (yield bending moment, ultimate rotation angle, plastic hinge length, hysteresis curve parameters); (2) Support and foundation parameters: horizontal elasticity of the support, friction of the support, and flexible deformation of the foundation (translational stiffness and rotational stiffness); (3) Seismic motion input parameters: acceleration time history.

[0059] 02. Select representative working conditions from the parameter space to form an initial high-precision finite element analysis sample set. The representative working conditions are a set of working conditions that have been verified by finite element analysis, cover the range of key structural parameters, can reflect the nonlinear response mode of the structure, have physical feasibility, and have uniform samples.

[0060] The key structural parameters range includes bridge structural parameters, bearing and foundation parameters, and seismic input parameters, ensuring that the minimum, maximum, and key range of each parameter are reflected in the finite element sample.

[0061] Complete structural response characteristics: It can trigger the nonlinear behavior of bridge structures, covering typical and extreme response states.

[0062] Physical feasibility: Each operating condition must meet physical constraints.

[0063] Uniform samples: The samples are evenly distributed in the parameter space and response space, without repetition or excessive similarity.

[0064] 03. Divide the initial high-precision finite element analysis sample set into a training set, a validation set, and a test set; the training set is used to train the model, and the validation set and test set are used to validate and test the training results, respectively.

[0065] 04. When training the model, increase the number of samples in batches according to the preset gradient based on experience or actual situation. Each time a new batch of samples is added, repeat the training of the model and record the prediction error of key response quantities and the physical consistency residual. The prediction error of key response quantities includes the prediction error of displacement and velocity of bridge concentrated mass points, bridge column end rotation / curvature and bridge column end bending moment. The physical consistency residual includes equilibrium constraint residual, geometric consistency constraint residual and constitutive relation constraint residual.

[0066] 05. After both the prediction error and physical consistency residual of the key response quantities converge, and the performance improvement from further increasing the number of samples is lower than the preset threshold, the corresponding number of samples is the preset number of the calibration dataset. Here, performance specifically refers to the reduction in the prediction error and physical consistency residual of the key response quantities on the validation set after further increasing the number of finite element samples. The prediction error of the key response quantities includes the displacement error, velocity error, column end rotation angle error, and column end bending moment error of the concentrated mass points in the bridge. The physical consistency residual includes the equilibrium constraint residual, the geometric consistency constraint residual, and the constitutive relation constraint residual.

[0067] Constraint loss function for calibration dataset: ; To calibrate the displacement in the dataset.

[0068] The equilibrium constraint loss function includes: column end moment equilibrium constraint loss function: ;and Structural dynamic equilibrium constraint loss function: .

[0069] The geometric consistency constraint loss function includes: Mapping constraint loss function between lumped mass displacement and pier top plastic displacement: ;and Geometric constraint loss function for pier top plastic displacement - column end rotation / curvature: .

[0070] Constitutive consistency constraint loss function: ; In the formula, Indicates based on p The loss function of the norm.

[0071] Fully constrained total loss function: ; Partially constrained total loss function: .

[0072] S6. After training is completed, the ground motion time history is input into the multi-level nested prediction model. The multi-level nested prediction model directly outputs the displacement / velocity, column end rotation / curvature and bending moment time history of the bridge structure, realizing rapid analysis and prediction of the nonlinear response of the bridge under strong earthquakes.

[0073] It should be noted that steps S1 and S2-S4 are parallel and have no sequential relationship.

[0074] The beneficial effects of the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided in this embodiment of the invention are as follows: Compared with the prior art, the intelligent analysis method for strong earthquake response of bridge structures with embedded physical information provided in this embodiment of the invention has the following beneficial effects: The fully physically constrained path utilizes only the fundamental physical laws of structural dynamics to construct a loss function to drive model training, without relying on any measured earthquake damage data or finite element calibration data, effectively alleviating the constraints of data scarcity on model application. The partially physically constrained path, on the other hand, can further improve prediction accuracy by combining data with physical constraints when a small amount of high-precision finite element data is available, maximizing data utilization efficiency.

[0075] By embedding equilibrium constraints, geometric consistency constraints, and nonlinear constitutive constraints at the column ends into the model training process, the model's prediction results are always constrained by fundamental physical laws. This effectively avoids physical inconsistencies such as abnormal energy accumulation, incompatible displacement responses, and mismatches between internal forces and deformations that are prone to occur in pure data models, significantly improving the reliability of the prediction results.

[0076] The existence of physical constraints gives the model a stronger ability to generalize and extrapolate. When the ground motion amplitude, spectral characteristics or structural parameters exceed the range of the training samples, the model can still give reasonable prediction results under the guidance of physical laws. This solves the problem of the sharp drop in prediction accuracy of pure data models in unknown strong earthquake scenarios and can better cope with the strong randomness and extreme nature of near-field ground motion of faults.

[0077] The constructed "state-geometry-constitutive" multi-level nested prediction model strictly follows the physical causal logic of structural dynamics. The model's prediction process is completely consistent with the actual response process of bridges under strong earthquakes. That is, the seismic input first causes changes in the macroscopic motion state of the structure, which then translates into local geometric deformation of the piers, and finally generates internal force response through material constitutive relations. The output of each module has a clear physical meaning. Engineers can not only obtain the final response results, but also clearly observe physical mechanisms that are crucial to seismic design, such as the development sequence of plastic hinges and the damage accumulation process. This significantly improves the engineering credibility and practical value of the model, and can provide scientific and reliable technical support for the seismic design, performance evaluation, and reinforcement of bridges in fault zones.

[0078] The above description is merely 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 within the protection scope of the present invention.

Claims

1. A method for intelligent analysis of the strong earthquake response of bridge structures by embedding physical information, characterized in that, Includes the following steps: S1. Construct a simplified physical model of the bridge structure, and establish equilibrium constraint equations, geometric consistency constraint equations, and column end nonlinear constitutive relation equations based on the simplified physical model. S2, perform time-series sampling processing on the seismic acceleration time history to obtain time-series input samples adapted to LSTM; S3, construct a single-level deep LSTM basic network; S4. Based on the single-level deep LSTM basic network, a multi-level LSTM nested prediction model of "state-geometry-constitutive" is constructed to form a hierarchical progressive prediction link from the seismic motion time history input to the displacement / velocity, column end rotation / curvature and bending moment time history of the bridge structure. S5. Based on the completeness of physical information in actual engineering, select one of the following two paths to train the multi-level nested prediction model: Fully Physically Constrained Path: The constraint terms in the loss function are constructed using the equilibrium constraint equation, the geometric consistency constraint equation, and the column end nonlinear constitutive relation equation in S1. The constraint terms are then weighted and summed to form the fully constrained total loss function, which drives the multi-level nested prediction model to be trained. Partial physical constraint path: A preset number of high-precision finite element results of bridges under strong earthquakes are introduced as a calibration dataset. The geometric consistency constraint equation, the column end nonlinear constitutive relation equation, and the calibration dataset are selected to construct the constraint terms in the loss function. The constraint terms are weighted and summed to form a partial constraint total loss function, which drives the multi-level nested prediction model to be trained. S6. After training is completed, the ground motion time history is input into the multi-level nested prediction model. The multi-level nested prediction model outputs the displacement / velocity, column end rotation / curvature and bending moment time history of the bridge structure, realizing rapid analysis and prediction of the nonlinear response of the bridge under strong earthquake.

2. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The simplified physical model equates the mass of a bridge span to the concentrated mass at the top of a fixed support. The total structural deformation Decomposed into horizontal deformation components of the support Elastic-plastic bending deformation component of the bridge pier itself With the basic elastic deformation component ; The basic elastic deformation component includes the basic translational deformation component. With the basic rotational deformation component The equivalent height of the bridge piers is .

3. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The equilibrium constraint equations include structural dynamic equilibrium constraint equations and column end moment equilibrium constraint equations.

4. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The geometric consistency constraint equations include the mapping relationship equation between the lumped mass displacement and the pier top plastic displacement and the geometric relationship equation between the pier top plastic displacement and the column end rotation / curvature.

5. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The nonlinear constitutive equation for the column end is based on the bending-rotation / curvature relationship in the plastic zone of the column end, combined with the skeleton curve and hysteresis rule, to form a constitutive equation that takes the column end rotation / curvature time history as input, considers the nonlinear historical state, and takes the column end bending moment as output.

6. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The time-series sampling process includes: cutting the seismic acceleration time history using a fixed-length sliding time window, setting an overlapping area between adjacent time windows, and constructing standardized time-series input samples.

7. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The single-level deep LSTM base network uses two stacked LSTM layers to extract high-order temporal features, with residual connections between the layers; and the hidden state of the first LSTM layer is concatenated with the output of the second LSTM layer using tensors, and then mapped to the output space through two fully connected layers.

8. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, The multi-level nested prediction model includes: The motion state prediction module takes the seismic acceleration time series sample as input and outputs the displacement and velocity time series data of the concentrated mass points of the bridge. The geometric state mapping module takes the output of the motion state prediction module as input and outputs the timing data of the geometric response of the bridge column end rotation angle / curvature. The plastic zone constitutive prediction module takes the output of the geometric state mapping module as input and outputs the time series data of bending moment and internal force at the bridge column ends.

9. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 1, characterized in that, In step S5, the preset number of the calibration dataset is determined by covering representative working conditions of bridge structural parameters and seismic input parameters, and by gradually expanding the finite element samples according to a preset gradient and combining it with independent validation set evaluation.

10. The intelligent analysis method for strong earthquake response of bridge structures with embedded physical information as described in claim 9, characterized in that, When the multi-level nested prediction model converges in both the prediction error of key response quantities and the physical consistency residual, and the performance brought by continuing to increase the number of samples is lower than the preset threshold, the corresponding number of samples is the preset number of the calibration dataset.