Bridge twin model multi-precision dynamic regulation and health monitoring method

Abstract

This invention belongs to the field of bridge engineering and digital twin technology, and relates to a method for multi-precision dynamic control and health monitoring of bridge twin models. The method includes: acquiring bridge geometric data; establishing geometric relationships between the overall simplified model L0 and the locally refined model clusters, as well as boundary condition data transfer rules; establishing a set of key indicators and generating local refined analysis task instructions; extracting boundary data from the overall simplified model L0 according to the local refined analysis task instructions; constructing a PINN boundary condition corrector for correction; applying the corrected boundary conditions to the corresponding boundary nodes of the L-level refined finite element model for solution; extracting key result fields and recording them in a database; and providing graded early warnings based on key mechanical indicators. This invention solves the technical problems of existing technologies where the large number of elements and nodes leads to extremely high computational load and slow response speed.

Description

Technical Field

[0001] This invention belongs to the field of bridge engineering and digital twin technology, specifically relating to a method for multi-precision dynamic control and health monitoring of bridge twin models. Background Technology

[0002] With the widespread application of digital twin technology in the operation and maintenance of infrastructure such as bridges, building a virtual model that can synchronize with the real bridge in real time and highly restore details has become the mainstream direction for health monitoring. However, in actual engineering, it is often difficult to balance the "computational accuracy" and "analysis efficiency" of the model. This contradiction seriously restricts the large-scale and routine application of digital twin technology in the long-term health monitoring of bridges.

[0003] Currently, the industry mainly relies on the following technical approaches to achieve digital modeling and simulation of bridges, but all of them have obvious limitations: Firstly, to ensure the accuracy of the analysis results, a common practice is to establish a high-precision, detailed finite element model. While such models can reproduce the rich mechanical response details of bridges, the sheer number of elements and nodes results in extremely high computational load and slow response speed. This not only fails to meet the high real-time requirements of daily monitoring but also incurs high long-term computational costs.

[0004] Secondly, to meet the needs of real-time monitoring, another widely adopted approach is to use a highly simplified overall model (such as a spatial beam grid model). While this type of model significantly improves computational efficiency, its simplification-based assumptions prevent it from accurately representing the three-dimensional stress state of key local structures such as the steel-concrete connection section, cable-stayed bridge anchorage zone, and complex nodes. When early signs of damage appear in these critical areas, the simplified model, due to its insufficient accuracy, is highly prone to missed or misjudged damage.

[0005] To attempt to balance the aforementioned contradictions, some studies have proposed constructing a multi-precision model library, which simultaneously maintains a simplified overall model and several pre-defined locally refined models. For example, Chinese invention patent CN118536361A discloses a dual-scale finite element analysis method for steel-concrete composite bridges, employing a hybrid modeling approach of macroscopic beam elements and fine solid elements. However, the division and connection of large-scale and small-scale regions require manual pre-setting, making it impossible to automatically trigger local refined analysis based on real-time monitoring data. Furthermore, existing methods still have shortcomings in key aspects: on the one hand, switching between models of different precisions usually relies on manual judgment, making it impossible to automatically and intelligently trigger based on real-time monitoring data; on the other hand, a more fundamental problem is the fragmentation between various models. The mechanical boundary conditions required for locally refined models are difficult to obtain automatically from the overall model, often requiring manual extraction and application, leading to distorted analysis results. These results also cannot be effectively fed back to the overall model for updating, forming "information silos" and disrupting the essential data consistency and evolutionary continuity of digital twins. In addition, Chinese invention patent CN119720691A proposes a surrogate model and finite element fusion method for bridge safety early warning. Although a multi-scale finite element model is constructed and a surrogate model is introduced, its early warning task triggering depends on the manual setting of preset thresholds, and the boundary condition correction depends on the model parameter correction. It does not solve the problem of automatic transfer of boundary conditions from the overall model to the local model and bidirectional data synchronization.

[0006] In recent years, Physical Information Neural Networks (PINNs) have made significant progress in solving partial differential equations, achieving the fusion of data and physics by embedding the governing equations into a loss function. For example, Chinese invention patent application CN119537878A proposes an intelligent damage identification method for bridge structures based on PINNs, embedding the Euler-Bernoulli beam equation into a loss function for damage identification of a single-precision model. However, its boundary conditions are only used as soft constraints, without addressing the automatic transfer and correction of boundary conditions between multi-precision models. Existing PINN methods primarily focus on solving single-precision models or forward problems, neglecting the automatic transfer and correction of boundary conditions between multi-precision models, and have not been applied to real-time monitoring scenarios of bridge digital twins. Therefore, how to introduce PINN into the collaborative simulation of multi-precision bridge models, achieving automatic and physically consistent boundary transfer from the overall model to local models, and forming a two-way data synchronization mechanism, has become a key bottleneck in multi-precision model collaborative simulation.

[0007] In summary, current technologies have not yet established an intelligent mechanism capable of dynamically and adaptively adjusting the accuracy of different areas of the model based on the actual operational status of the bridge and specific analytical needs. Therefore, there is an urgent need to develop a novel technological approach that enables bridge digital twin systems to achieve automatic, on-demand, high-precision focused analysis of key areas while ensuring overall computational efficiency. This would allow for optimal allocation of computing resources and maximize the effectiveness of bridge health monitoring. Summary of the Invention

[0008] The purpose of this invention is to provide a method for multi-precision dynamic control and health monitoring of bridge twin models that integrates physical information neural networks, so as to solve the technical problems of the existing technology where multi-precision models are mutually isolated and the transfer of boundary conditions depends on manual intervention.

[0009] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is a method for multi-precision dynamic control and health monitoring of bridge twin models, comprising the following steps: S1: Obtain bridge geometric data, establish an overall simplified model L0, identify key parts of the bridge, and build local refined model clusters of L1 to Ln levels for each key part. Utilize the geometric relationship between the overall simplified model L0 and the local refined model clusters to obtain the Ln level refined finite element model. S2: Preprocess and extract features from multi-source sensor data, establish a set of key indicators, determine the set of key indicators, and generate local fine analysis task instructions or secondary tasks. S3: According to the local refinement analysis task instructions, extract boundary data from the overall simplified model L0, construct a boundary condition corrector for correction, apply the corrected boundary conditions to the corresponding boundary nodes of the Ln-level refined finite element model for solution, extract key result fields and record them in the database, synchronize the key result fields to the overall simplified model L0, and update the state information of the corresponding components in the overall simplified model L0. S4: Real-time rendering of the overall bridge status on a 3D visualization platform, drilling and displaying detailed local models, and providing graded early warnings based on key mechanical indicators.

[0010] Furthermore, a simplified overall model L0 for real-time monitoring of the entire bridge is established in S1. The specific steps are as follows: S1.1a: Obtain the three-dimensional geometric data of the actual bridge; S1.2a: Using the three-dimensional geometric data of real bridges, establish an overall L0 model representing the load-bearing components of the main beam and piers; S1.3a: Verify the overall performance of the L0 model from both geometric and mechanical perspectives; S1.4a: Based on the model order reduction technique, the vibration modes in the overall L0 model are extracted, and a simplified overall L0 model is constructed.

[0011] Furthermore, in S1, a cluster of local refinement models at levels L1 to Ln is built for each key part. The specific steps are as follows: S1.1b: Based on bridge design specifications, a list of fatigue-prone parts, and historical inspection reports, identify the key parts that require detailed local analysis; S1.2b: For each key component, establish a corresponding refined analysis sub-model; S1.3b: Perform hierarchical meshing for each refined analysis sub-model to establish a model cluster with multiple precision levels L1 to Ln containing different mesh feature sizes.

[0012] Furthermore, in S1, the geometric relationship between the global simplified model L0 and the local refined model cluster is utilized to obtain the Ln-level refined finite element model. The specific steps are as follows: S1.1c: In the L0 global simplified model, specify the corresponding parent component for each refined analysis sub-model; S1.2c: Using a predefined coordinate transformation, align the origin of the coordinate system of the refined analysis sub-model with a specific control point on the parent component; S1.3c: Using the iterative nearest point algorithm, the point cloud data of the local area of ​​the real bridge collected by the 3D laser scanner is registered with the geometric surface extracted from the surface of the corresponding parent component of the L0 global simplified model.

[0013] Furthermore, S2 specifically refers to: The parallel operation of the rule triggering engine and model triggering engine allows for the separate execution of a fast proxy model for the physical information neural network of key parts, which then makes judgments on the key indicator set. When either the rule-triggered engine or the model-triggered engine determines an anomaly, a local fine-grained analysis task instruction is generated. When the damage index or maximum stress utilization rate predicted by the physical information neural network fast proxy model exceeds a preset threshold, a secondary task is generated. The fast proxy model for physical information of key parts is constructed using a fully connected neural network, and the first total loss function of the model is... As shown in equation (2); (2) in: The first total loss function for training the surrogate model. The weighting coefficient for the first data loss term. This is the first data loss item. The weighting coefficient for the first physical loss term. This is the first physical loss term; The first data loss term is shown in equation (3): (3) in: This is the first data loss item. The number of data points in the finite element simulation. For neural networks in coordinates The predicted displacement vector at that location, This is the displacement vector for the corresponding point in the high-precision finite element simulation. For the sequence number of the finite metadata point; The first physical loss term is shown in equation (4): (4) in: This is the first physical loss term. Let be the index of the collocation point, with values ​​ranging from 1, 2, ... , For the first Spatial coordinates of the points, To train the PINN fast proxy model in the target local region The number of points randomly sampled within the area. For divergence operators, In the allocation point The square of the residuals of the equilibrium equation, For neural networks in coordinates The predicted stress tensor.

[0014] Furthermore, the specific steps for performing local fine-grained analysis tasks in S3 are as follows: S3.1: Based on the boundary condition type in the task instruction, extract the corresponding displacement boundary data or force boundary data from the L0 global simplified model; S3.2: Using the calculation results of the L0 global simplified model as a priori, combined with sparse measured data of local regions and universal physical laws, a PINN boundary condition corrector is constructed and refined boundary conditions are generated for correction. The PINN boundary condition corrector is trained based on a second composite loss function, which is shown in equation (9): (9) in: This is the second composite loss function. For the parameters of the neural network, For the second data loss, For the second physical loss, For boundary deviation loss, The second data weighting coefficient, This is the second physical weighting coefficient. This refers to the boundary deviation weighting coefficient; S3.3: Start the finite element solver, load the finite element analysis input file, set the corresponding solution parameters, perform the calculation, and extract and parse the preset key result fields after the calculation is completed. The key result fields include high-precision stress contour plot and maximum equivalent stress value. Stress concentration factor The fatigue damage index D and the maximum stress utilization rate are calculated based on Miner's linear cumulative damage rule. S3.4: Through predefined application interface calls, key result fields are automatically recorded to the database at a preset frequency and associated with the corresponding parent component attribute records in the L0 overall simplified model to generate a report.

[0015] Furthermore, the construction of the PINN boundary condition corrector and the generation of refined boundary conditions described in S3.2 specifically includes the following steps: S3.2.1: Target local area The boundary Γ is divided into specified displacement boundaries. and specified force boundary Extracting prior displacement boundary conditions from the L0 global simplified model and prior surface force boundary conditions Simultaneously, it acquires the measured displacement values ​​of discrete sensor measurement points within the region or boundary. Or measured strain value Solve the corrected displacement boundary conditions on the specified displacement boundaries. Solve the corrected surface force boundary conditions on the specified force boundary. ; S3.2.2: Construct the PINN boundary condition corrector, which is a fully connected neural network. ,in, Input coordinates for the PINN boundary condition corrector. Let the parameters of the PINN boundary condition corrector be defined, and let the second composite loss function of the PINN boundary condition corrector be defined. Then, based on the displacement measurement points and strain measurement points, the third data loss contributed by the displacement measurement points of the PINN boundary condition corrector is calculated. The fourth data loss contributed by strain measurement points As shown in equations (10) and (11), the second physical loss is calculated. As shown in equation (12), the loss on the displacement boundary is calculated. As shown in equation (13), the loss at the force boundary is calculated. As shown in equation (14), the loss on the displacement boundary is finally utilized. Loss at the force boundary Calculate total boundary loss As shown in equation (16); (10) in: The third data loss contributed to the displacement measurement points. The number of displacement sensors, For neural networks in coordinates The predicted displacement vector at that location, For displacement sensor in Measured displacement vector at the location The number of the displacement measuring point; (11) in: The fourth data loss contributed by strain measurement points. The number of strain sensors, For neural networks in coordinates The predicted strain tensor, For strain sensors in coordinate The measured strain tensor at the location, The square of the Frobenius norm. The number of the strain measurement point; (12) in: The second physical loss is represented by k, where k is the index of the collocation point, and its value ranges from 1, 2, ... , Let K be the spatial coordinates of the kth collocation point. Training the PINN boundary condition corrector in the target local region The number of points randomly sampled within the area. For divergence operators, In the allocation point The square of the residuals of the equilibrium equation; (13) in: For the loss at the displacement boundary, To be at the specified displacement boundary The number of sampling points selected above for calculating the loss, where p is the index of the sampling point. For neural networks at points The predicted displacement vector at that location, Let L0 be the prior displacement vector at the same point, calculated from the global simplified model L0. The known true displacement values ​​on the displacement boundary. The square of the vector's 2-norm. Let p be the spatial coordinates of the p-th sampling point; (14) in: For loss at the force boundary, To be at the specified force boundary The number of sampling points selected above, Let be the spatial coordinates of the q-th sampling point, where q is the index of the boundary sampling point. For neural networks at points The predicted surface force vector at the location, The prior surface force vector provided for the overall simplified model L0; (16) in: Total boundary loss; S3.2.3: Calculate the second composite loss function Relative to network parameters The gradient is used to obtain the trained neural network. This will refine the model boundary locally. The neural network trained by inputting the coordinates of all nodes. Thus, the boundary conditions for enhanced physical information are obtained.

[0016] Furthermore, the specific steps of S4 are as follows: S4.1: Build a 3D visualization and interactive platform, and load the L0 overall simplified model as the basic scene; S4.2: Obtain the latest maximum stress utilization rate of each component and dynamically update the model surface color according to the preset color gradient; S4.3: Displays a locally refined model; S4.4: Execute tiered alerts based on the key result field indicators obtained from S3.3, specifically including: Level I warning is triggered when the maximum stress utilization rate of a component is in the range of [0.8, 0.95) or the fatigue damage index D is in the range of [0.1, 0.5). A notification message is generated within the visualization platform, and maintenance personnel are required to check the feedback within 2 hours. Level II warning is triggered when the maximum stress utilization rate of a component is in the range of [0.95, 1.0) or the fatigue damage index D is in the range of [0.5, 0.8). It automatically sends an SMS alarm to the mobile phone of the pre-set responsible engineer, requiring a response within 30 minutes. Level III warning is triggered when the maximum stress utilization rate of a component is ≥1.0 or the fatigue damage index D is ≥0.8. At the same time, the platform will trigger an audible and visual alarm, send an SMS, and automatically call the operation and maintenance supervisor, requiring the emergency procedure to be initiated within 10 minutes. Each warning is associated with an automatically generated analysis report.

[0017] The beneficial effects of the present invention are as follows: The present invention firstly uses a computing mechanism that allocates data on demand, and under the premise of ensuring the real-time performance of the overall monitoring, it initiates high-precision focused analysis only on the risk area, making it possible to support long-term high-frequency online monitoring of large bridges with limited computing power, and significantly reducing operating costs. By introducing a physical information neural network into the collaborative simulation of a bridge multi-precision model, a composite loss function integrating data loss, physical loss, and boundary deviation loss was constructed. Through this composite loss function, this invention fundamentally solves three major technical challenges: multi-source heterogeneous information fusion, information diffusion from discrete point measurements to continuous boundary fields, and collaborative optimization of multi-objective loss functions. It realizes automatic extraction of boundary conditions, physical information enhancement and correction, and bidirectional data synchronization from the overall simplified model to the local refined model.

[0018] Based on this, the results of local fine analysis are automatically fed back to the overall model, forming a data closed loop from the whole to the local and back to the whole. This ensures the consistency of the state between the overall model and the local model, greatly improves the accuracy and efficiency of local fine analysis, significantly enhances the accuracy of damage identification and early warning capabilities, and provides intuitive and quantitative basis for the investigation and maintenance decisions of hidden diseases.

[0019] Finally, a unified data framework ensures state consistency and evolution continuity among multi-precision models, forming an organic intelligent twin system. Its adaptive strategy can be flexibly configured and continuously optimized, thus enabling its widespread application in various engineering scenarios, from daily inspections to post-event assessments, greatly enhancing the practical value and universality of digital twin technology. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of the method for multi-precision dynamic control and health monitoring of bridge twin models in an embodiment of the present invention; Figure 2 This is a schematic diagram of the multi-precision model hierarchical mapping and data transmission of the bridge twin model multi-precision dynamic control and health monitoring method in the embodiment of the present invention; Figure 3 This is a comparison chart of computational efficiency and accuracy balance between multi-precision dynamic control and health monitoring methods for bridge twin models; Figure 4 This is a graph showing the daily computing cost of a bridge twin model multi-precision dynamic control and health monitoring method over time. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, the bridge twin model multi-precision dynamic control and health monitoring method proposed in this invention is carried out according to the following steps: S1: Acquire bridge geometric data, establish an overall simplified model L0 for real-time monitoring of the entire bridge, identify key parts of the bridge, and build local refined model clusters of L1 to Ln levels for each key part. Utilize the geometric relationship between the overall simplified model L0 and the local refined model clusters to obtain the Ln level refined finite element model and set boundary condition data transfer rules. The specific steps for establishing the overall simplified model L0 for real-time monitoring of the entire bridge in S1 are as follows: S1.1a: Obtain and integrate the design drawings and as-built data of the bridge, and use a total station or 3D laser scanner to collect the 3D geometric data of the actual bridge; S1.2a: Based on the three-dimensional geometric data of real bridges, in professional engineering simulation software (ANSYS (Analysis of Systems), ABAQUS (ABAQUS, finite element software), MIDAS (Mechanical Integrated Design Analysis Software)), through manual modeling or script automation, a spatial beam grid model or a three-dimensional rod system model with shear deformation is selected to establish an overall L0 model representing the main load-bearing components such as the main beam and piers. S1.3a: Verify the overall performance of the L0 model from both geometric and mechanical perspectives; In a specific embodiment of the present invention, the performance of the model proposed in S1.2a is verified from both geometric and mechanical perspectives. The verification method at the geometric level is as follows: the coordinates of the key control points (such as the center line of the pier tower and the intersection of the main beam axis) of the L0 overall simplified model are iteratively fitted with the measured coordinates based on the least squares method, so that the dimensional error of the key components (such as the height of the main beam and the cross section of the pier) is controlled within ≤5%.

[0024] The reason for this approach is that, geometrically, the model must accurately establish the overall structural skeleton of the bridge, that is, accurately reflect how the main load-bearing components (such as main beams and piers) are connected, how forces are transmitted, and how the structure is supported. The specific dimensions of the model must match those of the actual bridge. By iteratively fitting the model coordinates of key control points (such as the intersection of the pier centerline and the main beam axis) with the measured coordinates using the least squares method, the dimensional error of key components (such as the height of the main beam and the cross-sectional dimensions of the piers) can be controlled within ≤5%, thus ensuring that the model's "skeleton" and "shape" are consistent with the actual bridge height.

[0025] The mechanical verification method is as follows: Dynamic calibration of model parameters is performed. By deploying sensors (such as accelerometers), dynamic response data of the bridge under actual operation is collected. After obtaining vibration acceleration time history data of key points, the model material parameters (such as elastic modulus) and boundary conditions are iteratively adjusted by comparing the structural dynamic characteristics (mainly frequency and mode shape) obtained by model calculation and actual measurement until the following calibration criteria are met: the main frequency error is within the allowable range; the mode shape fit MAC (Modal Assurance Criterion) value is ≥0.90.

[0026] The reason for this approach is that, at the mechanical level, the model parameters need to be calibrated. To obtain realistic data for model calibration, sensors (such as accelerometers) are needed to measure the dynamic response of the bridge under actual working conditions. This involves two methods: first, using naturally occurring excitations in the bridge's environment (such as daily vehicle traffic and wind loads) to induce slight vibrations and record the response; second, when conditions permit, conducting specialized tests with controllable loads such as test vehicles to induce vibrations. The typical data collected is the time-varying vibration acceleration record of key bridge points. Based on the measured bridge vibration data, the key parameters of the model are calibrated by comparing the vibration characteristics calculated by the digital model with the actual measurement results. Specifically, with the goal of matching the main vibration frequencies and corresponding mode shapes (i.e., the swaying patterns of the structure) calculated by the model with the measured data, the material parameters (such as the elastic modulus, usually fine-tuned based on the design value) and support boundary conditions of the model are adjusted in the software until the frequency error between the two is within the allowable range and the consistency of the swaying patterns reaches the professional evaluation standard (MAC value ≥ 0.90) or higher, thereby ensuring that the dynamic mechanical behavior of the digital model can truly reflect the state of the physical bridge.

[0027] S1.4a: Using the Krylov subspace or principal component analysis (PCA) method, the 10th to 20th vibration modes in the overall L0 model are extracted, and an equivalent simplified model is constructed to obtain the overall simplified L0 model.

[0028] In the specific implementation of this application, the above approach can solve the pain point of "difficulty in balancing efficiency and accuracy" in the overall model in the prior art. After this processing, the L0 overall simplified model can run on a commercial computer with conventional performance. The calculation time to complete a full bridge static analysis should be ≤10 seconds, and the calculation time to complete a 30-second dynamic process simulation should be ≤30 seconds, thus achieving synergistic optimization of overall efficient calculation and subsequent accurate local analysis.

[0029] In S1, a cluster of local refinement models at levels L1 to Ln is built for each key part. The specific steps are as follows: S1.1b: Based on bridge design specifications, a list of fatigue-prone parts, and historical inspection reports, identify the key parts that require detailed local analysis; Specifically, based on bridge design specifications, a list of fatigue-prone parts, and historical inspection reports, potential weak areas that require special attention are identified, such as the steel-concrete composite section of the bridge tower, the cable-beam anchorage area, the bearing area, and key parts such as typical welding joints. S1.2b: For each key component identified in S1.1b, establish a corresponding refined analysis sub-model in the finite element analysis software (the model type should be selected according to the structural form and stress characteristics of the component). For example, for thick regions that require analysis of internal three-dimensional stress states (such as anchorage areas), solid models (such as hexahedral elements C3D8R, Continuum 3D 8-node Reduced integration element) are preferred. For thin-walled parts that need to withstand in-plane forces (such as steel box girder bridge decks), plate and shell models (such as S4R, Shell 4-node Reduced integration element with finite membrane strain) are preferred.

[0030] S1.3b: For each refined analysis sub-model, hierarchical meshing is performed using mapping mesh or adaptive fine meshing techniques to establish a model cluster with multiple precision levels L1 to Ln containing different mesh feature sizes; To achieve precise matching between different analytical needs and computational resources, the refined sub-models of each key component are divided into hierarchical meshes using mapping mesh or adaptive mesh refinement techniques, establishing model clusters with multiple precision levels (L1 to Ln), as detailed below: L1 level (basic level): The mesh feature size is set to 5~10 mm, which is used for rapid evaluation under normal conditions.

[0031] L2 level (high precision level): The mesh feature size is set to 1~5 mm for fine analysis.

[0032] Ln level (ultra-high precision level): The mesh feature size is set to 0.1~1mm, and methods such as extended finite element method (XFEM) can be introduced into this level of model to simulate special fracture mechanics analysis such as crack initiation and propagation, so as to achieve accurate adaptation to different analysis scenarios and avoid redundancy and insufficiency in precision.

[0033] In S1, the geometric relationship between the global simplified model L0 and the local refined model cluster is used to obtain the Ln-level refined finite element model; The specific implementation method is as follows: S1.1c: For each refined analysis sub-model, explicitly specify its corresponding parent component (i.e., the simplified component representing the local region) in the L0 global simplified model. S1.2c: Align the origin of the coordinate system of the refined analysis sub-model with a specific control point on the parent component through a predefined coordinate transformation; S1.3c: Using the Iterative Closest Point (ICP) algorithm, the point cloud data of the real bridge local area collected by the 3D laser scanner is registered with the geometric surface extracted from the surface of the corresponding parent component of the L0 overall simplified model, and the coordinate deviation of the boundary points is ensured to be ≤2 mm, so as to obtain the Ln level refined finite element model, and the coordinate deviation of the boundary points is ensured to be ≤2 mm. The goal of the ICP algorithm is to minimize the sum of squared Euclidean distances between two sets of points, as shown in equation (1): (1) in: For rotation matrix, It is a translation vector. To find the rotation matrix and translation vector that minimize the total error, This represents the total number of points in the point cloud. For the index of the point, For the local model boundary point cloud, The geometric point cloud is extracted from the surface of the parent component corresponding to the L0 overall simplified model. Point cloud on L0 model According to the rotation matrix The new coordinates obtained after rotation It is the square of the second norm of the vector.

[0034] The ICP algorithm originates from the classic geometric registration method in computer vision (Besl & McKay, 1992). This application provides initial values ​​through predefined coordinate transformations, which accelerates convergence and ensures registration accuracy.

[0035] S1.5: Preset boundary condition data transmission rules; To overcome the shortcomings of low efficiency and inconsistent transmission logic caused by manual intervention, this application predefines the mechanical boundary condition transmission algorithm from the L0 global simplified model to the refined analysis sub-model in the system database or configuration file. Specifically, for each refined analysis sub-model, the boundary condition class to be used in its analysis is clearly defined, for example: a displacement boundary (extracting the average displacement of the nodes of the associated section of the parent component). and average turning angle ) or force boundary (axial force obtained by integrating the stress of the parent member section) Shear force and and bending moment and The specific boundary condition conversion and mapping algorithms (such as shape function interpolation or rigid connection methods) are stored as attribute parameters of the refined analysis sub-model. This means that the completed L0 overall simplified model, as well as the local refined sub-models of different precision levels (L1 to Ln) for each key part, along with their geometric association information and boundary transfer rules, are uniformly stored in the model library. This achieves a seamless transition from overall analysis to local refined analysis. The hierarchical mapping relationship between multi-precision models is as follows: Figure 2 As shown, the bottom layer consists of the actual bridge structure and its sensor system (including 3D laser scanning and other measurement methods). A simplified L0 model (such as a beam grid model) is constructed in the middle layer through geometric topology mapping to support efficient daily health monitoring. On this basis, the system calls the upper-layer local refined model clusters as needed—including the L1 basic refined model for general anomaly analysis, the L2 high-precision model for key part diagnosis, and the Ln ultra-high-precision model for micro-damage assessment. Each local model is precisely connected to the corresponding parent component in the L0 simplified model through a predefined geometric association and boundary condition transfer mechanism. At the same time, the key mechanical results obtained from local analysis (such as stress and damage index) can be fed back to the overall model in real time, realizing two-way data synchronization and status updates, thus forming a collaborative architecture of overall real-time monitoring, local on-demand focusing, and closed-loop data flow.

[0036] The aforementioned multi-precision hierarchical architecture lays the structural foundation for subsequent intelligent triggering and dynamic scheduling. In S1 of this application, a multi-precision hierarchical architecture is constructed, consisting of an overall simplified model (L0) and local refined model clusters (L1~Ln). Compared to traditional single-precision or local refined modeling methods, this forms a model library that can be called upon as needed, achieving efficient collaboration between overall evaluation and local refinement. Based on this architecture, traditional methods rely solely on simple interpolation or rigid connections for boundary condition transfer from the overall model to local models, failing to consider the influence of measured data in local areas and making it difficult to guarantee physical consistency. This unidirectional, open-loop transfer mechanism is the root cause of the fragmentation between multi-precision models. Subsequent steps of this invention will utilize the PINN boundary condition corrector to retain prior information of the overall model while integrating local measured data and physical laws, achieving adaptive refinement of boundary conditions and laying the foundation for bidirectional data synchronization of multi-precision models.

[0037] S2: Preprocess and extract features from multi-source sensor data to establish a set of key indicators; run the rule triggering engine and model triggering engine in parallel, and run the physical information neural network fast proxy model separately to judge the set of key indicators; when the rule triggering engine or model triggering engine judges it as abnormal, generate a local fine analysis task instruction; when the damage index or maximum stress utilization rate predicted by the physical information neural network fast proxy model exceeds the preset threshold, generate a secondary task. In the specific embodiments of this application, the sensor arrangement follows the following principles: Three to five resistance strain gauge or fiber optic strain gauge sensors (range ±1500) are installed at each key location. Accuracy ±1 A total of 1-2 capacitive or piezoelectric accelerometers (range ±2g, frequency response 0.1-200Hz) are used to capture stress changes; simultaneously, 1-2 temperature sensors (range -40℃ to 80℃, accuracy ±0.5℃) are installed at heterogeneous connection points such as steel-concrete joint sections for temperature compensation. Strain sensors are attached to the maximum stress prediction section on the component surface along the main force direction, and accelerometers are installed at vibration mode sensitive points (such as mid-span, 1 / 4 span, and cable-stayed beam anchorage zone). The mapping relationship between sensor installation locations and L0 model component IDs is pre-entered into the database.

[0038] Multiple types of sensors deployed on the bridge structure collect data at different frequencies according to their monitoring objectives: Structural strain sensors (such as vibrating wire or fiber optic strain gauges) acquire data at frequencies of 1-10 Hz, vibration acceleration sensors (triaxial accelerometers) at 50-200 Hz, structural displacement sensors (such as wire displacement gauges or GPS receivers) at 0.1-1 Hz, environmental temperature and humidity sensors at 0.01-0.1 Hz, and wind speed and direction sensors at 1-5 Hz. All sensor data is collected by a unified data acquisition device and pushed to a central processing server in real time via standard communication protocols (such as MQTT or OPCUA).

[0039] The server performs real-time preprocessing on the received raw monitoring data stream. The specific processing method is as follows: For outlier handling, the mean and standard deviation of each sensor signal are calculated in real time within the most recent 60-second time window. The "3σ criterion" is applied to mark data points whose instantaneous values ​​exceed the range of "mean ± 3 times standard deviation" as outliers and remove or smooth them.

[0040] Signal noise reduction is performed separately for different data types. Vibration acceleration data were filtered out using a fourth-order Butterworth bandpass filter (cutoff frequency 0.5-30 Hz) to remove excessively high and low interference frequency components, retaining only the main frequency bands that reflect the true vibration of the bridge. Strain data were denoised using wavelet thresholding (using the db6 wavelet basis and decomposing into 3-5 layers) to remove noise. Displacement data are smoothed using a moving average method (window length of 5-10 sampling points) to eliminate the influence of random fluctuations.

[0041] Multi-dimensional feature extraction calculations are performed in parallel on the preprocessed data to generate a set of key indicators reflecting the immediate state of the structure, specifically: Strain characteristics: Calculate the mean value of the strain signal within a time window of 10-60 seconds. variance And peak factor (peak value / effective value).

[0042] Dynamic characteristics: Perform Fast Fourier Transform (FFT, frequency resolution 0.1-1 Hz) on the vibration acceleration data to identify the top 5 dominant frequencies. The corresponding damping ratio was estimated using the half-power bandwidth method. ; Cable stress variation characteristics: Calculate the rate of change of cable stress within a 5-30 minute time window. ; Displacement cumulative characteristic: Calculates the cumulative deviation of displacement relative to the initial reference state. ; The aforementioned characteristic parameters together constitute a set of key indicators reflecting the real-time state of the bridge structure.

[0043] Establish a set of key indicators, and then run the rule triggering engine, model triggering engine and physical information neural network fast proxy model in parallel to make judgments. The rule triggering engine and model triggering engine judge abnormal events in parallel, generate local fine analysis task instructions, and execute them. Bridge structure monitoring itself faces challenges such as complex environmental interference (wind, temperature drift), diverse anomaly patterns (gradual fatigue / sudden damage), and significant data noise. Existing technologies have long been trapped in a "single-engine optimization" mindset, either relying on rule-based triggering (which only covers known anomalies and has a false negative rate of >30% for unknown / gradual faults) or solely on model-based triggering (accuracy <75% in small-sample anomaly scenarios and lacks interpretable physical logic). Those skilled in the art generally believe that "rules and models have conflicting technical paths," failing to recognize their complementarity in "accurate identification of known anomalies" and "intelligent capture of unknown anomalies." More importantly, the "physical mechanism-driven" nature of rule engines (based on structural mechanics principles)... The "data pattern driven" (based on machine learning algorithms) and the model engine belong to different technical systems. This joint application is not a simple superposition of technologies, but also needs to overcome three major core challenges: the time synchronization problem (the sampling frequency difference of multi-source sensor data is up to 2000 times, and a unified judgment time window needs to be designed), the threshold adaptation problem (the fixedness of the rule threshold and the probability of the model output need to be dynamically matched), and the positioning linkage problem (the accurate mapping between engine alarm and digital twin model components). These problems cannot be solved by conventional technical means and require the design to combine cross-domain knowledge of structural engineering, data communication and machine learning. This invention adopts a dual-engine parallel judgment mechanism to ensure the sensitivity and reliability of triggering.

[0044] The parallel operation of the rule triggering engine and model triggering engine allows for the separate execution of a fast proxy model for the physical information neural network of key parts, which processes and judges the key indicator set. The rule-triggered engine performs logical judgments on the set of key indicators based on pre-set logical judgment rules based on engineering experience. If the rules are met, the indicator is judged as abnormal. Specifically, if the sliding average value of a certain strain measuring point is greater than 80% to 95% of its design allowable value for 3 to 10 consecutive calculation cycles, an anomaly is triggered. If the first order frequency Relative to the baseline value If the absolute value of the change is greater than 2% to 5% and this state lasts for 5 to 30 minutes, an anomaly is triggered. The specific threshold and duration parameters in the rules can be adjusted according to the bridge characteristics and monitoring requirements.

[0045] The model triggering engine uses a hybrid architecture of one-dimensional convolutional neural network and long short-term memory network to build an anomaly detection model. It uses historical feature data of the bridge under normal operation as negative samples and generates positive samples for supervised training by numerical simulation or by injecting known fault modes into historical data. The model outputs an anomaly probability value. The key indicator set is input into the trained model. When the anomaly probability value reaches the preset trigger threshold, it is judged as an anomaly. Specifically, a hybrid architecture of one-dimensional convolutional neural network and long short-term memory network (1D-CNN-LSTM) is established as an anomaly detection model; Using historical characteristic data (sample size ≥ 50 records) of bridges under normal operating conditions for 1 to 3 months as negative samples (normal state), positive samples (abnormal state) are generated through numerical simulation or by injecting known failure modes into the historical data. The anomaly detection model is then trained under supervision, with the training objective being to make the model output an anomaly probability value between 0 and 1. The model performance requirement is that the area under the receiver operating characteristic curve on its independent test set is not less than 0.95. Real-time feature streams are input into the trained model, and the model calculates the anomaly probability value of the output. When the value reaches or exceeds the preset trigger threshold (0.8 to 0.95), it is judged as abnormal; The specific fast proxy model for physical information neural networks targeting key components is as follows: To improve the physical consistency and response speed of anomaly detection, this invention introduces a fast proxy model of PINN (Physical Information Neural Network) for key parts based on the established real-time data-driven triggering and decision-making mechanism. The pre-trained fast proxy model of PINN for key parts is run in parallel to realize anomaly early warning with enhanced physical information.

[0046] In the offline phase, the PINN rapid proxy model for key components is trained using simulation data from high-precision finite element models (such as L2 level) of the corresponding key components, combined with solid mechanics control equations. This allows the model to learn the mapping relationship from boundary conditions (or equivalent loads directly calculated from sensor data) to high-precision stress fields, strain fields, and damage indices in the local area. When the PINN fast proxy model runs online, the system uses a pre-processed set of real-time key indicators (such as strain and displacement at key points) as input to call the corresponding PINN proxy model in parallel. This model can rapidly deduce the high-precision stress field of a local region at the current moment within milliseconds. and displacement field .

[0047] In a specific embodiment of the present invention, the PINN fast proxy model architecture in the key part adopts a fully connected neural network, and the input is the spatial coordinates of any point in the localized refined model region. ,in, Let x be the coordinate value of the spatial coordinate vector along the x-direction. The coordinate value in the y-direction of the spatial coordinate vector. The coordinates are given by the z-axis of the spatial coordinate vector, and the output is the displacement component. ,in, This represents the displacement component of that point along the x-direction. Let be the displacement component of that point along the y-direction. Let L be the displacement component of this point along the z-direction, and L be the number of network layers and the number of neurons in each layer. The activation function is tanh, and the loss function for training the PINN fast proxy model is shown in equation (2): (2) in: The first total loss function for training the surrogate model. This is the weighting coefficient for the first data loss term, used to balance the importance of the data simulation (usually set according to the data confidence level). The first data loss term measures the deviation between the displacement field predicted by the neural network and the displacement field simulated by the high-precision finite element method. These are the weighting coefficients for the first physical loss term, used to control the strength of the physical constraints. The first physical loss term represents the mean square error of the residuals in the governing equations of solid mechanics (such as equilibrium equations), ensuring that the prediction results conform to physical laws. Its specific form is similar to... Same.

[0048] The first data loss term is shown in equation (3): (3) in: This is the first data loss item. The number of data points in the finite element simulation. For neural networks in coordinates The predicted displacement vector at that location, This is the displacement vector for the corresponding point in the high-precision finite element simulation. For the sequence number of the finite metadata point; The first physical loss term is shown in equation (4): (4) in: This is the first physical loss term. Let be the index of the collocation point, with values ​​ranging from 1, 2, ... , For the first Spatial coordinates of the points, To train the PINN fast proxy model in the target local region The number of coordinate points for internal random sampling. For divergence operators, In the allocation point The square of the residuals of the equilibrium equation, For neural networks in coordinates The predicted stress tensor.

[0049] This loss term is used to train the PINN fast surrogate model, constraining the prediction results of the surrogate model to satisfy the solid mechanics equilibrium equations.

[0050] The design idea of ​​the first total loss function for training the PINN fast surrogate model for key parts is to enable the surrogate model to automatically satisfy the physical equations while approximating the finite element simulation results, thereby improving extrapolation ability and interpretability.

[0051] Here, PINN is a fast surrogate model that replaces high-precision finite element method (FEM) solutions. Its training relies on offline-generated simulation data, and its goal is to quickly approximate the FEM solution while using physical equations to ensure its rationality. The loss function only requires data fitting terms and physical residual terms, without considering boundary deviations from the L0 model, because its input is sensor data rather than boundary conditions.

[0052] To enable the engineering application of dual-engine parallel decision-making, the following mechanism is established: (1) Timing synchronization and adaptive window: The "timestamp alignment + sliding window aggregation" method is adopted. The accelerometer (200Hz) calculates features in a 1-second window, and the temperature and humidity (0.01Hz) are filled by linear interpolation to achieve time alignment of multi-source data. On this basis, the sampling period of the sensor with the lowest sampling frequency is used as the reference window T0. The high-frequency sensor extracts the mean, peak, and root mean square features within T0, and the low-frequency sensor uses linear interpolation for alignment; when the dual engines output the attention status twice in a row ( When the response time is greater than 0.6, the window automatically shrinks to T0 / 2 to improve response speed.

[0053] (2) Spatial positioning and mapping: A correspondence table between each sensor number and its installation coordinates is pre-established, and the spatial range occupied by each bridge component in the L0 model is recorded as a coordinate region. When a single sensor is triggered, the component to which its coordinates belong is directly determined as an abnormal location; when multiple adjacent sensors are triggered simultaneously, the weights are calculated based on the distance of each sensor to a candidate location (the closer the distance, the greater the weight), and the most likely precise coordinates of the abnormality are estimated by weighted average, thereby locating a smaller area (sub-component level) inside the component, such as a section of the main beam or a local area of ​​the anchorage zone.

[0054] (3) Dual-engine arbitration mechanism: based on the alarm status of the rule engine and the output probability of the model engine The combination of these factors categorizes the results into three levels: mild conflict (logging only), moderate conflict (triggering PINN proxy model verification), and high consistency (directly performing Ln-level fine analysis). All conflict cases are stored in the database and used for periodic threshold self-calibration.

[0055] (4) Threshold dynamic adaptation: Establish the mapping relationship between anomaly probability and rule threshold. When the model engine outputs probability... When the value is 0.7~0.8, the judgment threshold of the corresponding rule trigger engine will be reduced by 10% (to improve sensitivity); when At the same time, keep the threshold unchanged to avoid over-response.

[0056] When either the rule-triggered engine or the model-triggered engine determines an anomaly, based on the preset mapping table of sensor installation locations and overall simplified model component identifiers, the anomaly alarm is associated with and located to a specific bridge component, generating a structured local fine-grained analysis task instruction, which is then placed into a queue according to priority for scheduling and execution. When the damage index or maximum stress utilization rate predicted by the physical information neural network fast proxy model exceeds a preset threshold, a secondary task is generated and placed into the task queue, requiring more accurate Ln-level finite element verification. The rule triggering engine and the model triggering engine operate independently in parallel in the background, processing the same real-time feature streams in parallel. As long as either engine outputs an anomaly judgment result, the system is triggered. After triggering, the system immediately associates and locates the anomaly alarm to the specific bridge component based on the preset sensor installation location and L0 overall simplified model component ID mapping table, generating precise positioning information containing the component location and identifier (e.g., "West main tower, segment 15").

[0057] Specifically, this involves creating a task instruction object that contains the following key information; target_component_id: Specifies the corresponding number of the part that needs to be analyzed in detail in the L0 overall model.

[0058] required_fidelity_level: Automatically determines the required fidelity of the model based on the severity of the anomaly (e.g., P_anomaly of 0.8~0.9 is mapped to L1 level, P_anomaly of 0.9~0.95 is mapped to L2 level, and P_anomaly ≥0.95 is mapped to Ln level).

[0059] boundary_condition_spec: Specifies whether displacement boundary conditions or force boundary conditions should be used as input conditions for local analysis.

[0060] Priority & Timeout: Set the urgency level of the task and the maximum allowed computation time. For example, a high-level alarm task requires computation to be completed within 300 seconds. Level III alert (stress utilization rate ≥ 1.0 or D ≥ 0.8) corresponds to the highest priority and has a timeout of 300 seconds; Level II alert (stress utilization rate 0.95~1.0 or D 0.5~0.8) corresponds to the second highest priority and has a timeout of 600 seconds; Level I alert corresponds to the third highest priority and has a timeout of 1800 seconds; Level 2 task (PINN proxy model verification) corresponds to the lowest priority and has a timeout of 300 seconds.

[0061] The generated task instructions will be placed into a priority-ordered task queue according to their set priority, awaiting system resource scheduling and execution of the corresponding local fine-grained simulation analysis. Scheduling employs a priority-preemptive algorithm. When a high-priority task arrives, if the currently executing task has a lower priority than the newly arrived task, the current task is immediately interrupted, the interrupted task's state is saved and returned to the head of the queue, and the high-priority task is then executed. If the current task's priority is not lower than the new task, the new task is inserted into the queue according to its priority. Only one finite element solution task (Ln-level fine-grained analysis) is allowed to run at a time, while the secondary task (PINN verification) and the finite element solution task can be executed in parallel.

[0062] If the damage index or maximum stress utilization rate predicted by the PINN surrogate model exceeds a preset threshold (e.g., D>0.4 or utilization rate>85%), a secondary concern task can be generated even if the rule engine and 1D-CNN-LSTM model are not triggered. This task requires more precise Ln-level finite element verification. Unlike the structured local refinement analysis task instructions generated by the rule trigger engine or model trigger engine, the secondary concern task does not execute the PINN boundary condition correction step in S3.2. Instead, it directly calls the Ln-level refined finite element model in the model library and performs the solution verification according to the steps in S3.3. This enables rapid and accurate verification of the PINN surrogate model's early warning results. This effectively compensates for the shortcomings of the pure data-driven model in terms of physical logic and achieves ultra-fast physical information early warning of potential risks.

[0063] S3: Perform corresponding operations according to the local refinement analysis task instructions or secondary tasks. When the local refinement analysis task instructions are executed, extract boundary data from the overall simplified model L0, construct the PINN boundary condition corrector for correction, apply the corrected boundary conditions to the corresponding boundary nodes of the Ln-level refined finite element model for solution, extract key result fields and record them in the database, generate a report, and realize bidirectional data synchronization between multi-precision models. When performing a level 2 task, the Ln-level refined finite element model is called for solution verification, key result fields are extracted and recorded in the database, and a report is generated. For local refined analysis task instructions, the complete analysis process from S3.1 to S3.4 is executed.

[0064] S3.1: Extract the corresponding displacement boundary or force boundary data from the L0 global simplified model according to the boundary condition type in the task instruction; Specifically, after the task scheduler retrieves the detailed analysis task to be executed from the queue, it first accesses the running L0 global simplified model (or reads its latest result file) to obtain the latest mechanical state of the target component. Based on "boundary_condition_spec", it calls the predefined interface function. Then, based on the boundary_condition_spec in the task instruction, it calls the predefined interface function to extract the corresponding boundary data. If the task requires displacement boundaries, it extracts the displacement data of all nodes on the associated cross-section of the target component and calculates its average displacement vector. and average rotation vector If the task requires force boundaries, then extract the stress data of that section and calculate the internal force vector borne by that section through integration. As shown in equation (5): (5) in: The vector of internal forces acting on the cross section. For axial force, The shear force is located in the y-axis direction of the local coordinate system of the cross section. Let be the shear force along the z-axis of the local coordinate system of the cross section. The bending moment about the y-axis, The bending moment about the z-axis, For transpose, The normal stress is along the x-axis. This refers to the shear stress acting on the x-normal plane along the y-direction. Let z be the shear stress acting on the x-normal plane along the z-direction. To obtain the bending moment about the y-axis by integrating over the cross-sectional area, To obtain the bending moment about the z-axis by integrating over the cross-sectional area, It is a micro-area element.

[0065] Equation (5) is derived from the relationship between internal force and stress in mechanics of materials and is a direct application of Saint-Venant's principle.

[0066] S3.2: Construct the PINN boundary condition corrector and generate refined boundary conditions. Use the calculation results of the L0 global simplified model as a priori, and combine the sparse measured data of the local region with universal physical laws to perform correction, thus solving the problem of automatic transfer and correction of boundary conditions for multi-precision models.

[0067] Applying Physical Information Neural Network (PINN) to the boundary condition correction of bridge multi-precision models requires simultaneously solving three major technical challenges: multi-source heterogeneous information fusion, information diffusion from discrete point measurements to continuous boundary fields, and collaborative optimization of multi-objective loss functions. To address these challenges, this invention designs a composite loss function (Equation (9)) that integrates data loss, physical loss, and boundary deviation loss. Through dynamically adjustable weight coefficients, it achieves an adaptive balance among the three, thereby organically unifying sparse measured data with the prior information of the overall model while satisfying the solid mechanics control equations, generating physically consistent and precision-controllable boundary conditions.

[0068] This step automatically generates high-precision boundary conditions, solving the problem in traditional methods where local model boundary conditions are difficult to obtain automatically from the overall model and lack precision. It is the core link in realizing "two-way data synchronization" between multiple precision models.

[0069] S3.2.1: Define the correction problem as follows: For the target local area (Corresponding to the Ln-level model), its boundary Γ consists of two parts, as shown in equation (6): (6) in: For the target local area The boundary, To specify the displacement boundary, To specify the force boundary, It is a union.

[0070] The prior boundary conditions automatically extracted from the L0 global simplified model can be expressed as: In superior, ,in, For displacement field, The displacement vector at boundary point x extracted from the L0 global simplified model, in superior, , For stress tensor, The unit outward normal vector, The a priori surface force values ​​are provided by the overall simplified model L0.

[0071] Here, PINN is a boundary condition refinement model. It does not replace the finite element solution, but provides more accurate boundary conditions for subsequent Ln-level finite element analysis. Therefore, in addition to the data term (matching the measured values) and the physical term (satisfying the equilibrium equations), a boundary deviation term must also be introduced to maintain a reasonable degree of deviation from the L0 prior (avoiding overcorrection).

[0072] In the target local area There are m discrete sensor measurement points inside or on the boundary, and their measured values ​​are denoted as... (Displacement) or (Strain), where i is the index of the displacement sensor measuring point and j is the index of the strain sensor measuring point, the goal is to find a set of corrected boundary conditions. and ,in, Corrected displacement boundary conditions The corrected surface force boundary conditions ensure that in the local region of the target... It satisfies the laws of physics, approximates the measured value at the measurement point, and does not significantly deviate from the prior value of the L0 global simplified model.

[0073] In this invention, the first data loss and the second data loss belong to two different neural network modules. The first data loss is used for offline training of the PINN fast surrogate model, and its supervision signal comes from high-precision finite element simulation data, aiming to enable the surrogate model to quickly approximate the simulation solution in the online stage. The second data loss is used for online correction of the PINN boundary condition corrector, and its supervision signal comes from sensor measured data, aiming to make the corrected boundary conditions reflect the local real mechanical state.

[0074] S3.2.2: Construct the PINN boundary condition corrector, which is a fully connected neural network. ,in, Input coordinates for the PINN boundary condition corrector. For the parameters of the PINN boundary condition corrector, this step uses a fully connected neural network as the basic architecture of the PINN boundary condition corrector. The principle behind this is that fully connected networks have a universal function approximation capability, able to approximate continuous functions with arbitrary precision, and are suitable for nonlinear mappings from spatial coordinates to displacement fields. Through automatic differentiation techniques, the network's output displacement field can automatically calculate the strain and stress fields, thereby constructing the physical loss term, allowing the network to learn both data distribution and physical laws during training.

[0075] The network is designed to approximate a local area of ​​the target. The actual displacement field within the region can be used to obtain strain and stress through automatic differentiation using a fully connected neural network. Predicted strain tensor As shown in equation (7): (7) in: The strain tensor predicted by the fully connected neural network. displacement gradient This is the transpose of the displacement gradient.

[0076] Fully connected neural networks Predicted stress tensor As shown in equation (8): (8) in: Let C be the stress tensor predicted by the neural network, and C be the elastic stiffness tensor derived from the generalized Huke's law.

[0077] The goal of training the PINN boundary condition corrector is to minimize the following composite loss function (this structure borrows from the general framework of PINN, but introduces a dynamic weighting mechanism and boundary deviation constraints for the boundary condition correction problem of bridge multi-precision models), as shown in Equation (9): (9) in: This is the second composite loss function. For the parameters (weights and biases) of a neural network, training aims to find an optimal set. To minimize the total loss, The second data loss is used to force the network to predict values ​​at sensor measurement points that match the actual measured values, measuring the degree of agreement between the network predictions and the measured data. The second physical loss measures the degree to which the network's predictions violate the laws of physics. Boundary loss measures the degree to which the network predictions deviate from the prior boundary of the L0 model. The second data weighting coefficient, This is the second physical weighting coefficient. The three weighting coefficients mentioned above, representing boundary bias weighting coefficients, are used to balance the importance of the three losses. They can be dynamically adjusted based on the confidence level of each component and the overall model accuracy to achieve multi-objective synergistic optimization. For example, when sensor data is highly reliable, the weighting coefficient of the second data component can be increased. When there is insufficient confidence in the accuracy of the L0 model in a specific region, the boundary bias weighting coefficient can be reduced. Weighting coefficients The range of values ​​is determined based on the magnitude differences of different loss terms in the PINN loss function. Studies have shown that physical loss (balance equation residuals) involves higher-order differential operators, and its magnitude is typically 10 times larger than that of data loss. 1 ~10 3 The weights are multiples of the loss, so smaller weights are needed as soft constraints; while the data loss and boundary bias loss are of similar magnitude, and their weights are around 1.0. (Value range is 0.5~2.0) (Value range is 0.01~0.5) (Values ​​range from 0.01 to 1.0) are derived from academic consensus and engineering practice experience in the PINN field.

[0078] The third data loss contributed by the displacement measuring point is calculated. As shown in equation (10): (10) in: The third data loss contributed to the displacement measurement points. For displacement, For data, The number of displacement sensors, For neural networks in coordinates The predicted displacement vector at that location, For displacement sensors in The measured displacement vector at the location, This refers to the serial number of the displacement measuring point. Let i be the spatial coordinates of the i-th displacement measurement point; The fourth data loss for calculating the contribution of strain measurement points. As shown in equation (11): (11) in: The fourth data loss contributed by strain measurement points. The number of strain sensors, For neural networks in coordinates The predicted strain tensor (obtained from the displacement by automatic differentiation). For strain sensors in The measured strain tensor at the location The square of the Frobenius norm, for a matrix (strain tensor), is the square root of the sum of the squares of all its elements, equivalent to the matrix version of the 2-norm. This is the serial number of the strain measurement point.

[0079] This part forces the network to approximate the measured value at the sensor measurement point, but the measured data is only discrete "point" information and cannot directly constrain the entire boundary. Therefore, it is necessary to introduce physical loss and boundary deviation loss to diffuse the "point" information to the entire boundary through physical laws.

[0080] Calculate the second physical loss (Measure the degree of violation of the laws of physics) as shown in equation (12): (12) in: For the second physical loss, k is the index of the collocation point (from 1 to ...). ), Training the PINN boundary condition corrector in the target local region The number of random sampling points (these points may not necessarily have sensors, but are used to force the physical equations to be satisfied). For neural networks in coordinates The predicted stress tensor (obtained from strain through constitutive equations) is given. The divergence of the stress tensor corresponds to the equilibrium equation in solid mechanics (derived from the law of conservation of momentum, when there are no body forces). ), For divergence operators, In the allocation point The square of the residual of the equilibrium equation is embedded in the network in the form of the residual of the equilibrium equation, which forces the prediction field to satisfy the mechanical laws throughout the region, thus solving the problem that pure data-driven methods are prone to producing physically discontinuous or uninterpretable results.

[0081] This loss term is used to train the PINN boundary condition corrector, and the corrected boundary conditions satisfy the solid mechanics equilibrium equations.

[0082] Displacement boundary ( Losses As shown in equation (13): (13) in: The loss is at the displacement boundary, where p is the index of the sampling point (from 1 to p). ), In order to be in The number of sampling points selected above for calculating the loss (these points can be boundary nodes or uniform sampling points). Let p be the spatial coordinates of the p-th sampling point. For neural networks at points The predicted displacement vector at that location, The prior displacement vector at the same point is calculated from the overall simplified model L0 (obtained by interpolation from the L0 beam element results). These are the known true displacement values ​​on the displacement boundary, i.e., the prior values. Used to measure the magnitude of the deviation between the predicted value and the prior value.

[0083] Force boundary ( Losses on As shown in equation (14): (14) in: For loss at the force boundary, In order to be in The number of sampling points selected above, Let q be the spatial coordinates of the q-th sampling point. The index of the boundary sampling point. For neural networks at points The predicted surface force vector at the location, The prior surface force vector is provided by the global simplified model L0 (usually obtained by stress recovery of the internal forces of the L0 section). (15) in: Let n(x) be the stress tensor predicted by the neural network, and n(x) be the unit outward normal vector at point x on the boundary. Utilizing losses on displacement boundaries Losses at the boundary Calculate total boundary loss .

[0084] = + (16) This part ensures that the corrected boundary conditions do not deviate significantly from the L0 prior, avoiding distortion caused by overfitting noisy data.

[0085] Three losses passed , , Dynamic equilibrium solves the problem of multi-objective collaborative optimization. S3.2.3: Train and solve the PINN boundary condition corrector to obtain physically enhanced boundary conditions. ; The key to training the PINN boundary condition corrector lies in solving a multi-objective optimization problem consisting of data, physical, and boundary terms. Since the physical loss involves high-order differential operators, and the three loss terms have different dimensions and scales, direct optimization can easily lead to gradient imbalance or slow convergence. Therefore, this invention employs the Adam adaptive optimization algorithm and dynamically adjusts the weight coefficients. , , To balance the gradient contributions of various losses, ensuring that the network converges stably to the optimal solution that is physically consistent, data-approximate, and bounded by the boundary constraints.

[0086] Specifically, using differential techniques to calculate Relative to network parameters The gradient is calculated, and the Adam optimization algorithm is used to iteratively update the network weights. Hyperparameter settings are as follows: initial learning rate is 0.001, multiplied by a decay factor of 0.9 every 1000 iterations; batch size is set according to the total number of points, typically 256~1024. The convergence condition is: the absolute value of the total loss change is less than 10 over 200 consecutive iterations. -6 The maximum number of iterations can be reached, up to 20,000 (adjustable depending on computing resources and problem size). After training, the model boundaries will be locally refined. Input the coordinates of all nodes in the trained program The boundary conditions for physical information enhancement can then be obtained, denoted as... .

[0087] For the dynamic determination of weight coefficients, data weights The setting is adaptively configured based on the signal-to-noise ratio (SNP) of the sensor's measured signal: when SNR > 30dB, Take a value of 1.5~2.0; when the SNR is between 20~30dB, Take a value of 1.0~1.5; when SNR < 20dB, Set the value to 0.5~1.0 (to reduce dependence on noise). Physical weights. Weights of boundary deviation Based on the pre-calibration of the confidence level of the L0 global simplified model in the target local region: if the frequency error of the L0 model after dynamic calibration in this region... 2%, MAC 0.95, then Use 0.5~1.0 (to retain more prior information); otherwise, use 0.05~0.2. It is usually fixed in the range of 0.05 to 0.2 to balance the magnitude difference of the residuals in the equation.

[0088] Technical Performance Description of PINN Boundary Condition Corrector This step utilizes physical information technology to automatically correct boundary conditions, offering the following significant advantages: (1) The solid mechanics control equation (equilibrium equation) is embedded into the network training in the form of a loss function to ensure that the corrected boundary conditions satisfy the equilibrium equation. Even in the boundary region where there is no sensor measured data, the correction results still conform to the laws of mechanics, avoiding the physical uninterpretability problem caused by the pure data-driven method, thus enhancing physical consistency. (2) By simultaneously utilizing the L0 global model prior (providing the overall trend of the boundary), sparse sensor measured data (providing the local real state), and physical laws (providing field equation constraints), information diffusion from discrete "point" measurements to "boundary" distribution is achieved. In the boundary region close to the sensor measurement point, the correction result is mainly based on the measured value, while in the region far from the measurement point, it is mainly based on physical constraints and L0 prior, thus overcoming the contradiction between the sparsity of measured data and the insufficient accuracy of the global model. (3) The boundary condition correction process does not require manual intervention. The system automatically completes the entire process of "L0 prior extraction, measured data reading, network training, and boundary condition generation", which solves the problem of low efficiency of traditional methods that rely on manual reading from the overall model and manual application of boundary conditions. Moreover, after training, the boundary condition correction can be completed in milliseconds, supporting real-time online analysis. (4) The corrected boundary conditions are used for local fine-grained solution. The solution results are automatically fed back to the L0 global model through API, and the corresponding parent construction attribute records are updated. This enables the global model to not only provide boundary conditions for the local model, but also to receive the fine-grained analysis results of the local model in real time, forming a data closed loop of "global to local and back to global". This achieves state consistency and evolution continuity between multi-precision models and bidirectional data synchronization, solving the problem of multi-precision models being fragmented and forming information islands in the existing technology.

[0089] S3.3: Start the finite element solver, load the finite element analysis input file, set the corresponding solution parameters, perform calculations, and extract and parse the preset key result fields after the calculations are completed.

[0090] The system automatically reads the corrected boundary conditions. Using predefined coordinate transformation matrices and shape functions, these are precisely applied to the corresponding boundary nodes of the Ln-level refined finite element model retrieved from the model library, generating the final finite element analysis input file. A professional finite element solver (such as Abaqus / Standard or ANSYS Mechanical) is then started, and the corresponding solution parameters are set according to the analysis type (for static analysis, the convergence criterion is displacement tolerance). ~ The calculation is performed using a transient dynamic analysis time integration step of 0.001~0.01 seconds (m). After the calculation is completed, the preset key result fields are extracted and parsed from the result file, including: high-precision stress contour map (spatial resolution ≤1 mm), maximum equivalent stress value, etc. (Reading error ≤ 5 MPa), stress concentration factor (Typical range 1.2~5.0), fatigue damage index calculated based on Miner's linear cumulative damage rule. (Range 0~1), maximum stress utilization rate, where the stress concentration factor formula comes from fracture mechanics, the fatigue damage index formula comes from the Palmgren-Miner criterion, and the dimension of both is 1. The stress spectrum required for calculation is obtained from local time history analysis, and the damage accumulation step is set to 1~5 minutes.

[0091] S3.4: Through predefined application interface calls, the extracted key result fields are automatically recorded to the central database at a preset frequency and associated with the corresponding parent component attribute records in the L0 overall model.

[0092] Specifically, synchronization is triggered immediately after each successful local refinement solution. The maximum stress utilization rate and fatigue damage index D calculated by the local model directly replace the original attributes of the corresponding parent component in the L0 model. The high-precision stress cloud map and stress concentration coefficient are only stored in the local result library for visualization drilling and do not replace the overall evaluation indicators of the L0 model. The frequency of the data synchronization process is consistent with the solution cycle of the local refinement model (usually once every 1 to 10 minutes), which is determined by the task queue scheduling. When the attributes of the parent component of the L0 model are directly modified by other processes (such as manual calibration or another parallel local task) during two synchronizations, the local high-precision solution result is used to directly overwrite it. If multiple local tasks of the same parent component complete the synchronization request, the result of the later completed task is used, and the conflict overwrite event is recorded in the log. Subsequently, the system generates a structured local analysis report, including: the abnormal trigger time, the target area, the Ln model level used, a summary of the applied boundary conditions, the results of key mechanical indicators, a preliminary assessment of damage risk, and links to relevant cloud map or visualization result screenshots.

[0093] For the second-level task, skip the boundary extraction and correction steps in S3.1 and S3.2, directly call the Ln-level refined finite element model in the model library, solve and verify according to the steps in S3.3, and extract the key result fields and record them in the database according to S3.4.

[0094] S4: Real-time rendering of the overall bridge status on a 3D visualization platform, drilling and displaying detailed local models, and providing graded early warnings based on key mechanical indicators.

[0095] S4.1: Build a WebGL-based 3D visualization and interactive platform, and load the L0 overall simplified model as the basic scene; This invention develops a 3D visualization and interactive platform based on WebGL technology (such as using Three.js or the Cesium framework). The platform loads a simplified overall model L0 as the basic scene and realizes real-time rendering of the global model state.

[0096] S4.2: Obtain the latest maximum stress utilization rate of each component and dynamically update the model surface color according to the preset color gradient; The system reads the latest "maximum stress utilization rate" (i.e., the ratio of current stress to allowable stress) attribute value of each component from the database and renders it according to the preset color gradient rules. The specific rules are as follows: when the ratio is ≤60%, it is judged as a safe state and displayed in blue; when the ratio is between 60% and 90%, it is judged as a state that needs attention and displayed in yellow; when the ratio is >90%, it is judged as a dangerous state and displayed in red.

[0097] S4.3: Displays a locally refined model; When a user clicks on an abnormal component that is highlighted (such as red or yellow) in the global view, the front end will send a request to the server. The server will retrieve the corresponding Ln-level fine model file and the latest stress cloud map result data, and transmit it to the front end. The front end engine will then seamlessly switch to that local view to display the fine geometry and high-resolution stress distribution of that area, supporting zooming to 1 mm precision to observe details.

[0098] The 3D visualization interactive platform interface provides a timeline control, allowing users to view snapshots of the model's structural state at any point within the past 1 to 72 hours. It supports comparing the overall results of the L0 model with the local results of the Ln model in a semi-transparent overlay or split-screen manner, facilitating cross-scale verification and in-depth analysis.

[0099] S4.4: Execute tiered early warnings based on the key result field indicators obtained from S3.4; Level I (Attention / Yellow): Triggered when the maximum stress utilization rate (Max_Stress_Utilization) of a component is in the range [0.8, 0.95) (based on engineering experience and design safety margin, when the stress utilization rate reaches 0.8, the structure enters a high stress state) or the fatigue damage index (D_fatigue) is in the range [0.1, 0.5) (according to engineering experience, when D≥0.1, it indicates that the structure has entered a detectable fatigue accumulation stage). The system generates a notification message within the visualization platform and requires maintenance personnel to check the feedback within 2 hours.

[0100] Level II (Warning / Orange): Triggered when the maximum stress utilization rate (Max_Stress_Utilization) of a component is in the range [0.95, 1.0) (based on engineering experience and design safety margins, a stress utilization rate of 0.95 indicates that the current stress is very close to the maximum value allowed by the code) or the fatigue damage index (D_fatigue) is in the range [0.5, 0.8) (according to engineering experience, when D ≥ 0.5, it indicates that more than half of the structure's fatigue life has been consumed). The system automatically sends an SMS alert to the pre-set responsible engineer's mobile phone, requiring a response within 30 minutes.

[0101] Level III (Alert / Red): Triggered when the maximum stress utilization rate (Max_Stress_Utilization) of a component is ≥1.0 (according to the "Highway Bridge Maintenance Specification" and related design standards, a stress utilization rate ≥1.0 indicates a clear overload condition) or the fatigue damage index (D_fatigue) is ≥0.8 (based on engineering experience, D ≥0.8 indicates that the structure is close to failure). The system simultaneously triggers an audible and visual alarm on the platform, sends an SMS message, and automatically calls the operations and maintenance supervisor, requiring the emergency response process to be initiated within 10 minutes. Each alert is associated with an automatically generated analysis report for decision-making reference.

[0102] Example: This example uses the anchorage zone of a cable-stayed bridge as a key component to verify the training effect and online inference performance of the physical information neural network fast proxy model for key components in this invention. A 6-layer fully connected neural network is constructed as the physical information neural network fast proxy model, with 128 neurons in each layer. The activation function is tanh. The input layer receives the equivalent load (including axial force, shear force, and bending moment) calculated from sensor data. 2000 sets of L2-level finite element simulation results under different load conditions are used as training data. The test results show that when the prediction results of the PINN model are compared with the high-precision finite element simulation results under the corresponding load conditions, the average relative error is 3.2%, which meets the engineering accuracy requirements (allowable error ≤ 5%). During online testing, the equivalent load calculated from the current sensor is input, and the model outputs the maximum principal stress of the anchorage zone within 0.05 seconds.

[0103] To further clarify the specific implementation process and effect of the PINN-based boundary condition corrector in this invention, a detailed explanation is provided using a steel-concrete composite section of a cable-stayed bridge tower as an example. This composite section is a critical load-bearing component connecting the concrete bridge tower and the steel main beam, and accurate simulation of its mechanical behavior is crucial for the overall bridge structural safety assessment. In this embodiment, the overall simplified model L0 is modeled using beam elements, and the average displacement of the composite section cross-section is extracted from the L0 model as the prior boundary condition. Simultaneously, three triaxial strain gauges are embedded within the concrete portion of the composite section to acquire measured strain data for the local area. Based on these prior and measured data, the PINN boundary condition corrector proposed in this invention will be used to correct the boundary conditions provided by the L0 model, generating refined boundary conditions that better reflect the actual local mechanical state, and verifying its effectiveness in improving analysis accuracy.

[0104] Taking the steel-concrete composite section of a cable-stayed bridge tower as an example, its L0 model is a beam element, and the average displacement extracted on the cross section of the composite section is... As shown in equation (17), (17) Three triaxial strain gauges were embedded inside the concrete section of the joint, and the measured strain tensors are shown in equations (18) to (20): (18) in: The first triaxial strain gauge, corresponding to the measured tensor; (19) in: For the second triaxial strain gauge, the corresponding measured tensor; (20) in: The measured tensor for the third triaxial strain gauge; Then, a fully connected network with 5 layers and 256 neurons per layer is constructed as... The activation function is tanh, and 10,000 collocation points are randomly sampled within the binding segment region for calculation. ,Will As a boundary prior, the prior displacements of all nodes on the boundary are obtained through shape function interpolation. Set weights =1.0, =0.1, =0.01 (assuming high trust in sensor data, and giving certain constraints on the prior of L0 but allowing for correction).

[0105] On a server equipped with an NVIDIA V100 GPU, the loss converged after training the PINN network for approximately 15 minutes (20,000 epochs). The coordinates of the combined segment boundary nodes were then input into the trained network to obtain the corrected boundary displacement. The results showed that in the boundary region near the sensor measurement point, and The deviation is approximately 5%-12%, while in areas far from the measurement points and with stronger constraints, the deviation is less than 3%. This indicates that PINN successfully diffuses sparse point measurement information across the entire boundary using physical laws, generating a globally better boundary condition.

[0106] To objectively and quantitatively evaluate the performance of the method of this invention, a comparative verification process based on a high-precision simulation benchmark was designed. First, a widely accepted and refined finite element model was established as the ground truth. Then, based on this benchmark, the performance of traditional methods and this invention was compared. A highly representative engineering example, a 500m main span double-tower, double-cable-stayed steel box girder bridge, was selected as the analysis object, with its critical vulnerable section—the steel-concrete composite section of the bridge towers—as the focus of verification. Using the general-purpose commercial finite element software ABAQUS, an ultra-high-precision three-dimensional solid finite element model of the steel-concrete composite section of the bridge towers was established. This model strictly followed the design drawings, using hexahedral reduced integral elements (C3D8R) for mesh generation. The mesh size was refined to 1 mm in the region of interest, and the total number of elements in the model exceeded 1.5 million. Its material constitutive relations, contact settings, and boundary conditions were carefully calibrated, and it can be considered the gold standard reflecting the true mechanical response of this part.

[0107] Three parallel analysis schemes were established for comparison: Scheme A (traditional global high-precision): A refined plate-shell-solid hybrid model of the entire bridge (approximately 5 million elements) was established as a representative of the traditional high-precision method. Scheme B (traditional global simplification): A simplified model of the entire bridge lattice system was established (approximately 12,000 elements) as a representative of the traditional high-efficiency method. Scheme C (the method of this invention): The system of this invention was deployed, including an L0 global model equivalent to Scheme B, and an L2-level locally refined sub-model (approximately 850,000 elements) derived from the baseline model for the joint section.

[0108] (1) Generation of measured data: On the benchmark model, standard vehicle load, wind load and temperature load conditions based on the specifications are applied, and the stress and strain time history data of the key points of the joint section are calculated. This data is then used as the measured data of the virtual sensor and input to each comparison scheme.

[0109] (2) Abnormal scenario simulation: By artificially modifying the material properties of local areas in the benchmark model (such as reducing the elastic modulus by 5%-15%) or introducing microcracks, typical damage is simulated, and test data streams containing abnormal modes are generated.

[0110] (3) Performance index calculation: Each scheme was run on the same hardware platform (Intel Xeon Gold 6248R, 256GB RAM), and its performance index was calculated within 72 hours of continuous operation (simulating 30 days of monitoring).

[0111] Table 1 summarizes the average values ​​of the main results obtained from the comparative tests based on the above simulation benchmarks.

[0112] Table 1. Performance Comparison and Verification Results of Three Bridge Digital Twin Modeling Schemes

[0113] The following conclusions can be drawn from Table 1: (1) In terms of computational efficiency, Scheme C (the present invention) takes only 10-15 seconds per cycle under normal conditions, which is on the same order of magnitude as Scheme B (traditional simplified model, 8-12 seconds) and is far superior to Scheme A (global high-precision model, 180-240 seconds). When local fine analysis is triggered, the time taken by Scheme C increases to 40-60 seconds. Although it is higher than normal, it is still significantly lower than the global high-precision calculation time of Scheme A. In terms of the accuracy of local stress analysis, the relative error range of Scheme C is 3.2%-4.8%, which is better than 15.3%-22.8% of Scheme B and is close to 2.1%-3.5% of Scheme A. In terms of the reliability of anomaly detection, the anomaly omission rate of Scheme C is 0.5% and the misjudgment rate is 1.0%, which is between Scheme A (0% omission, 1.2% misjudgment) and Scheme B (8.5% omission, 0.8% misjudgment), indicating that it controls the omission risk at a low level while ensuring a low misjudgment rate. In terms of computing power cost, Scheme C requires 10-15 hours of CPU per day, higher than Scheme B (5-8 hours), but only about one-tenth of Scheme A (120-150 hours), demonstrating good engineering economics. Considering the above indicators, Scheme C maintains a computational efficiency similar to the simplified model under normal conditions. While the time consumption increases when detailed analysis is required, it is still far lower than the global high-precision model. Simultaneously, its stress calculation accuracy is significantly better than the simplified model and approaches the level of the high-precision model. The control of missed and false positives in anomaly detection is also well-balanced. Overall, Scheme C achieves a better balance between computational efficiency, analysis accuracy, detection reliability, and computing power cost.

[0114] also, Figure 3 This reflects the position of the three schemes in the time consumption-error coordinate system. The method proposed in this invention is close to the ideal region of low time consumption and low accuracy.

[0115] Figure 4 This reflects the fluctuations in computing power consumption of each scheme during the simulation test period. The consumption curve of the method proposed in this invention is at a low to medium level, which significantly reduces the operating cost of long-term, online, and refined health monitoring of large bridges, making it feasible for large-scale promotion.

[0116] In summary, this simulation benchmark verification fully demonstrates that the multi-precision dynamic control method proposed in this invention fundamentally solves the contradiction between computational accuracy and efficiency in digital twin models, realizes intelligent optimization of computing resources, and provides a practical and advanced technical solution for the digital and intelligent operation and maintenance of large-scale infrastructure such as bridges.

[0117] To verify the advantages of using the ICP algorithm for model registration in this invention, a comparative experiment was designed as follows. This application selects the steel-concrete composite section of the towers of a 500m main span double-tower double-cable-stayed steel box girder bridge as the experimental object. In the L0 overall simplified model, the corresponding parent component is a spatial beam grid model, and its surface geometric data is obtained by integrating design drawing analysis and actual bridge measurement data. The local refined model is an L2 level solid model, and its boundary region point cloud data is collected by a 3D laser scanner with a scanning density of 100 points / mm. 2 The total number of effective point cloud acquisitions is ≥5 million, covering the entire registration boundary. This application employs the Iterative Closest Point (ICP) algorithm combined with a predefined coordinate transformation for registration, setting the number of iterations to 50 and the convergence threshold to 1×10⁻⁶. -6 m, iteratively registers the point cloud data of the local model boundary region with the geometric surfaces extracted from the surface of the parent component of the L0 model until the convergence condition is met.

[0118] Traditional registration methods employ the industry-standard feature point matching-based registration method, manually marking 10 feature points (selecting easily identifiable points such as the boundary corner of the steel-concrete composite section of the bridge tower and the center of the weld), and using the least squares method to fit the marked feature points to complete the registration between the two.

[0119] 100 sampling points were evenly selected in the registration boundary area. The comprehensive coordinate deviation of each sampling point in the X, Y, and Z axes was counted. The average deviation, the maximum deviation, and the percentage of sampling points with a deviation ≤ 2 mm were calculated.

[0120] The accuracy indices of the two registration methods measured in the experiment are shown in Figure 2: Table 2. Accuracy Indicators of the Two Registration Methods

[0121] As shown in Table 2, the ICP algorithm used in this invention for registration has an average boundary point coordinate deviation of only 0.87 mm and a maximum of 1.73 mm. The coordinate deviation of all sampling points is ≤2 mm, which fully meets the preset accuracy requirements. In contrast, the average boundary point coordinate deviation of the traditional feature point matching method is 4.21 mm and the maximum is 8.56 mm. Only 23% of the sampling points have a deviation ≤2 mm, while more than 70% of the sampling points have a deviation exceeding 2 mm. The comparison shows that the registration accuracy of this invention is far superior to that of the traditional method, which can ensure the accurate transmission of overall mechanical properties to local areas and provide a reliable geometric basis for the collaborative simulation of multi-precision models. The experimental data directly prove that the boundary point coordinate deviation is ≤2 mm, which is far superior to the registration accuracy of the traditional method.

[0122] The above uses a cable-stayed bridge as an example to illustrate the specific implementation of the present invention, but this method is also applicable to beam bridges, arch bridges, and suspension bridges, and the adaptation rules are as follows: For beam bridges, the key areas are the support region and the mid-span section. The support region uses force boundary conditions (extracting the support reaction force from the L0 model as the boundary input), while the mid-span section uses displacement boundary conditions (extracting the mid-span deflection from the L0 model as the forced displacement). The local model accuracy levels L1 to Ln are divided according to the same criteria as for cable-stayed bridges, and the mesh feature size is adjusted proportionally according to the cross-sectional dimensions (usually 1 / 20 to 1 / 50 of the cross-sectional height).

[0123] For arch bridges, the key components are the arch abutments and the hanger anchorage zone. Fixed-end displacement boundary conditions are used for the arch abutments (extracting the displacement and rotation of the arch abutment nodes in the L0 model), while force boundary conditions are used for the hanger anchorage zone (extracting the hanger axial force as surface force input). Geometric nonlinearity must be considered during analysis, employing the arc-length method (a built-in function in finite element software). The initial arc-length increment is set to 0.01~0.05, and the maximum arc-length step is set to 200~500. Large deformation effects are implemented in Abaqus by setting the geometric nonlinearity (Nlgeom) switch of the analysis step to ON, and in ANSYS by activating the Nlgeom command.

[0124] For suspension bridges, the key parts are the main cable clamps and the cable saddle area. The cable saddle area adopts displacement boundary conditions (extracting the precise coordinates of the main cable after deformation in the L0 model as forced displacement applied to the cable clamp boundary). Due to the large flexibility of the main cable, the local model introduces an updated Lagrangian scheme at the L1 level to handle the problem of large displacement and small strain (in Abaqus, the Nlgeom switch is set to ON, and in ANSYS, the Nlgeom command is activated).

[0125] Through the above adaptation rules, this invention can be widely applied to the health monitoring of various bridges without requiring substantial modifications to the core algorithm (dual-engine parallel judgment, PINN boundary condition correction, and multi-precision dynamic scheduling).

[0126] In summary, this invention first utilizes an on-demand computing mechanism to initiate high-energy-consuming analyses only for risk areas while ensuring the real-time performance of overall monitoring. This makes it possible to support long-term, high-frequency online monitoring of large bridges with limited computing power, significantly reducing operating costs. Specifically, in anomaly triggering, a dual-engine parallel judgment mechanism monitors the bridge status in real time. Only when an anomaly or risk signal is detected is the subsequent high-precision analysis task triggered. After an anomaly is triggered, the required analysis precision level is automatically determined based on the severity of the anomaly, and a local refined analysis task instruction is generated, rather than continuously performing high-precision calculations on the entire bridge. Based on the task instruction, only the local refined model of the corresponding risk area is loaded, and the boundary conditions of that area are extracted from the overall L0 model, avoiding the computational burden of high-precision modeling of the entire bridge. High-precision finite element analysis is performed only on the risk area to extract key indicators such as stress and damage, achieving "precise focused analysis," while the overall model remains in low-precision, high-efficiency operation.

[0127] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for multi-precision dynamic control and health monitoring of bridge twin models, characterized in that, Includes the following steps: S1: Obtain bridge geometric data, establish an overall simplified model L0, identify key parts of the bridge, and build local refined model clusters of L1 to Ln levels for each key part. Utilize the geometric relationship between the overall simplified model L0 and the local refined model clusters to obtain the Ln level refined finite element model. S2: Preprocess and extract features from multi-source sensor data, establish a set of key indicators, determine the set of key indicators, and generate local fine analysis task instructions or secondary tasks. S3: According to the local refinement analysis task instructions, extract boundary data from the overall simplified model L0, construct a boundary condition corrector for correction, apply the corrected boundary conditions to the corresponding boundary nodes of the Ln-level refined finite element model for solution, extract key result fields and record them in the database, synchronize the key result fields to the overall simplified model L0, and update the state information of the corresponding components in the overall simplified model L0. S4: Real-time rendering of the overall bridge status on a 3D visualization platform, drilling and displaying detailed local models, and providing graded early warnings based on key mechanical indicators.

2. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 1, characterized in that, In S1, a simplified overall model L0 is established for real-time monitoring of the entire bridge. The specific steps are as follows: S1.1a: Obtain the three-dimensional geometric data of the actual bridge; S1.2a: Using the three-dimensional geometric data of real bridges, establish an overall L0 model representing the load-bearing components of the main beam and piers; S1.3a: Verify the overall performance of the L0 model from both geometric and mechanical perspectives; S1.4a: Based on the model order reduction technique, the vibration modes in the overall L0 model are extracted, and a simplified overall L0 model is constructed.

3. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 1, characterized in that, In S1, a cluster of local refinement models at levels L1 to Ln is built for each key part. The specific steps are as follows: S1.1b: Based on bridge design specifications, a list of fatigue-prone parts, and historical inspection reports, identify the key parts that require detailed local analysis; S1.2b: For each key component, establish a corresponding refined analysis sub-model; S1.3b: Perform hierarchical meshing for each refined analysis sub-model to establish a model cluster with multiple precision levels L1 to Ln containing different mesh feature sizes.

4. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 3, characterized in that, In S1, the geometric relationship between the global simplified model L0 and the local refined model cluster is used to obtain the Ln-level refined finite element model. The specific steps are as follows: S1.1c: In the L0 global simplified model, specify the corresponding parent component for each refined analysis sub-model; S1.2c: Using a predefined coordinate transformation, align the origin of the coordinate system of the refined analysis sub-model with a specific control point on the parent component; S1.3c: Using the iterative nearest point algorithm, the point cloud data of the local area of ​​the real bridge collected by the 3D laser scanner is registered with the geometric surface extracted from the surface of the corresponding parent component of the L0 global simplified model.

5. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 1, characterized in that, S2 specifically refers to: The parallel operation of the rule triggering engine and model triggering engine allows for the separate execution of a fast proxy model for the physical information neural network of key parts, which then makes judgments on the key indicator set. When either the rule-triggered engine or the model-triggered engine determines an anomaly, a local fine-grained analysis task instruction is generated. When the damage index or maximum stress utilization rate predicted by the physical information neural network fast proxy model exceeds a preset threshold, a secondary task is generated. The fast proxy model for physical information of key parts is constructed using a fully connected neural network, and the first total loss function of the model is... As shown in equation (1); （1） in: The first total loss function for training the surrogate model. The weighting coefficient for the first data loss term. This is the first data loss item. The weighting coefficient for the first physical loss term. This is the first physical loss term; The first data loss term is shown in equation (2): （2） in: This is the first data loss item. The number of data points in the finite element simulation. For neural networks in coordinates The predicted displacement vector at that location, This is the displacement vector for the corresponding point in the high-precision finite element simulation. For the sequence number of the finite metadata point; The first physical loss term is shown in equation (3): （3） in: This is the first physical loss term. Let be the index of the collocation point, with values ​​ranging from 1, 2, ... , For the first Spatial coordinates of the points, To train the PINN fast proxy model in the target local region The number of points randomly sampled within the area. For divergence operators, In the allocation point The square of the residuals of the equilibrium equation, For neural networks in coordinates The predicted stress tensor.

6. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 1, characterized in that, The specific steps for performing a local fine-grained analysis task in S3 are as follows: S3.1: Based on the boundary condition type in the task instruction, extract the corresponding displacement boundary data or force boundary data from the L0 global simplified model; S3.2: Using the calculation results of the L0 global simplified model as a priori, combined with sparse measured data of local regions and universal physical laws, a PINN boundary condition corrector is constructed and refined boundary conditions are generated for correction. The PINN boundary condition corrector is trained based on a second composite loss function, which is shown in equation (4): （4） in: This is the second composite loss function. For the parameters of the neural network, For the second data loss, For the second physical loss, For boundary deviation loss, The second data weighting coefficient, This is the second physical weighting coefficient. This refers to the boundary deviation weighting coefficient; S3.3: Start the finite element solver, load the finite element analysis input file, set the corresponding solution parameters, perform the calculation, and extract and parse the preset key result fields after the calculation is completed. The key result fields include high-precision stress contour plot and maximum equivalent stress value. Stress concentration factor The fatigue damage index D and the maximum stress utilization rate are calculated based on Miner's linear cumulative damage rule. S3.4: Through predefined application interface calls, key result fields are automatically recorded to the database at a preset frequency and associated with the corresponding parent component attribute records in the L0 overall simplified model to generate a report.

7. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 6, characterized in that, S3.2 The construction of the PINN boundary condition corrector and the generation of refined boundary conditions specifically includes the following steps: S3.2.1: Target local area The boundary Γ is divided into specified displacement boundaries. and specified force boundary Extracting prior displacement boundary conditions from the L0 global simplified model and prior surface force boundary conditions Simultaneously, it acquires the measured displacement values ​​of discrete sensor measurement points within the region or boundary. Or measured strain value Solve the corrected displacement boundary conditions on the specified displacement boundaries. Solve the corrected surface force boundary conditions on the specified force boundary. ; S3.2.2: Construct the PINN boundary condition corrector, which is a fully connected neural network. ,in, Input coordinates for the PINN boundary condition corrector. Let the parameters of the PINN boundary condition corrector be defined, and let the second composite loss function of the PINN boundary condition corrector be defined. Then, based on the displacement measurement points and strain measurement points, the third data loss contributed by the displacement measurement points of the PINN boundary condition corrector is calculated. The fourth data loss contributed by strain measurement points As shown in equations (5) and (6), the second physical loss is calculated. As shown in equation (7), the loss on the displacement boundary is calculated. As shown in equation (8), the loss at the force boundary is calculated. As shown in equation (9), the loss on the displacement boundary is finally utilized. Loss at the force boundary Calculate total boundary loss As shown in equation (10); （5） in: The third data loss contributed to the displacement measurement points. The number of displacement sensors, For neural networks in coordinates The predicted displacement vector at that location, For displacement sensors in Measured displacement vector at the location The number of the displacement measuring point; （6） in: The fourth data loss contributed by strain measurement points. The number of strain sensors, For neural networks in coordinates The predicted strain tensor, For strain sensors in coordinate The measured strain tensor at the location The square of the Frobenius norm. The number of the strain measurement point; （7） in: The second physical loss is represented by k, where k is the index of the collocation point, and its value ranges from 1, 2, ... , Let K be the spatial coordinates of the kth collocation point. Training the PINN boundary condition corrector in the target local region The number of points randomly sampled within the area. For divergence operators, In the allocation point The square of the residuals of the equilibrium equation; （8） in: For the loss at the displacement boundary, To be at the specified displacement boundary The number of sampling points selected above for calculating the loss, where p is the index of the sampling point. For neural networks at points The predicted displacement vector at that location, Let L0 be the prior displacement vector at the same point calculated from the global simplified model L0. The known true displacement values ​​on the displacement boundary. The square of the vector's 2-norm. Let p be the spatial coordinates of the p-th sampling point; （9） in: For loss at the force boundary, The number of sampling points selected above, Let q be the spatial coordinates of the q-th sampling point. The index of the boundary sampling point. For neural networks at points The predicted surface force vector at the location, The prior surface force vector provided for the overall simplified model L0; = + （10） in: Total boundary loss; S3.2.3: Calculate the second composite loss function Relative to network parameters The gradient is used to obtain the trained neural network. This will refine the model boundary locally. The neural network trained by inputting the coordinates of all nodes. Thus, the boundary conditions for enhanced physical information are obtained.

8. The method for multi-precision dynamic control and health monitoring of bridge twin models according to claim 1, characterized in that, The specific steps for S4 are as follows: S4.1: Build a 3D visualization and interactive platform, and load the L0 overall simplified model as the basic scene; S4.2: Obtain the latest maximum stress utilization rate of each component and dynamically update the model surface color according to the preset color gradient; S4.3: Displays a locally refined model; S4.4: Execute tiered alerts based on the key result field indicators obtained from S3.3, specifically including: Level I warning is triggered when the maximum stress utilization rate of a component is in the range of [0.8, 0.95) or the fatigue damage index D is in the range of [0.1, 0.5). A notification message is generated within the visualization platform, and maintenance personnel are required to check the feedback within 2 hours. Level II warning is triggered when the maximum stress utilization rate of the component is in the range of [0.95, 1.0) or the fatigue damage index D is in the range of [0.5, 0.8). It automatically sends an SMS alarm to the mobile phone of the preset responsible engineer, requiring a response within 30 minutes. Level III warning is triggered when the maximum stress utilization rate of a component is ≥1.0 or the fatigue damage index D is ≥0.

8. At the same time, the platform will trigger an audible and visual alarm, send an SMS, and automatically call the operation and maintenance supervisor, requiring the emergency procedure to be initiated within 10 minutes. Each warning is associated with an automatically generated analysis report.

Images