Bridge support displacement intelligent prediction method and system based on multi-source structure response

Abstract

The invention belongs to the field of bridge structure health monitoring, and provides a bridge support displacement intelligent prediction method and system based on multi-source structure response, and the method comprises the steps: carrying out the preprocessing of monitored time sequence data, and constructing a structure response diagram model; performing space-time coding on node features of the structure response graph model by adopting an attention-based graph neural network, and introducing a force flow path priori item into corresponding graph attention updating for graphs in each time window; and on the basis of the updated graph attention structure, constructing a main prediction network for support displacement prediction, constructing adversarial physical constraints, determining physical consistency, and during main prediction network training, performing adversarial training by taking minimization of prediction errors and physical inconsistency as targets. The key problems that in the prior art, available information is insufficient, a model lacks physical interpretation, the cross-working-condition generalization ability is poor, and a support sensor cannot stably work for a long time are fundamentally solved.

Description

Technical Field

[0001] This invention belongs to the field of bridge structural health monitoring, specifically relating to an intelligent prediction method and system for bridge support displacement based on multi-source structural response. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] In the field of bridge structural health monitoring, bearing displacement is a crucial indicator reflecting the boundary constraint state of a bridge, bearing functional degradation, and overall structural safety. However, the bearing area is characterized by confined spaces, intense vibrations, large temperature differences, and severe pollution. Traditional displacement gauges, vision systems, or fiber optic sensors are difficult to operate stably and accurately over long periods, making bearing displacement one of the most challenging physical quantities to directly obtain in structural monitoring. Furthermore, bearing displacement is a result of the coupling of multiple factors, including main beam deflection changes, redistribution of internal forces at critical section boundaries, pier bending deformation, and temperature conditions. Its variation patterns are complex, highly nonlinear, and vary significantly across different bridge types and load conditions, making it difficult for existing data-driven methods to capture this essential structural coupling pattern.

[0004] While existing deep learning models can learn certain correlations from data, they lack constraints on the laws of structural mechanics, are not interpretable, have poor stability across different working conditions, and cannot guarantee that the predicted results conform to the deformation transmission path of the actual structure. Furthermore, most existing models are based on simple sequential inputs, failing to reflect the unique deformation chain structure of bridges—"main beam—pier—support"—and are ill-suited to complex, multi-source, and unstable bridge monitoring scenarios. In addition, existing methods cannot automatically adjust model behavior based on different temperatures, vehicle loads, and environmental factors, nor can they assess the reliability of model predictions or sensor anomalies, and they cannot automatically provide alternative monitoring capabilities when support sensors are damaged. Summary of the Invention

[0005] To address the aforementioned problems, this invention proposes an intelligent prediction method and system for bridge bearing displacement based on multi-source structural response. This invention fundamentally solves the key problems in existing technologies, such as insufficient available information, lack of physical interpretability of models, poor generalization ability across working conditions, and difficulty in long-term stable operation of bearing sensors.

[0006] According to some embodiments, the present invention adopts the following technical solution: A smart prediction method for bridge support displacement based on multi-source structural response includes the following steps: The time series data of main beam deflection, strain / curvature of key sections of main beam, pier bottom strain and support displacement are preprocessed to construct a structural response diagram model; Attention-based graph neural networks are used to spatiotemporally encode the node features of the structural response graph model. For the graph within each time window, a priori terms of force flow paths are introduced in the corresponding graph attention update. Based on the updated graph attention structure, a main prediction network for support displacement prediction is constructed. The main prediction network includes a graph feature encoder, a graph readout layer, and a prediction head. The graph feature encoder takes the structural response graph model as input and adopts the updated graph attention structure. It calculates attention weights to perform weighted aggregation of neighbor node features, realizing the structured propagation of node features along the real force flow direction. The graph readout layer aggregates its output node features. The prediction head obtains the displacement estimate of the support, the curvature of the key section, and the strain at the bottom of the pier based on the aggregated features. Construct adversarial physical constraints, determine physical consistency based on the residual sequence between different types of physical parameters, and conduct adversarial training during the training of the main prediction network with the goal of minimizing prediction error and physical inconsistency; Using the trained master prediction network, the predicted support displacement is obtained based on the target main beam deflection, the strain / curvature of the key section of the main beam, and the strain at the pier bottom.

[0007] As an alternative implementation method, the preprocessing and structural response graph model construction process includes: performing time alignment, outlier removal and filtering on the acquired various sensor data; constructing multiple time windows of length T, with each time window considered as a sample; within each time window, abstracting the key monitoring locations of the bridge into a set of graph nodes; establishing a directed edge set based on the physical deformation chain of the bridge to determine the influence or transmission relationship between each node; encoding the relationship into an adjacency matrix of the graph; and assigning weights to the edges based on structural stiffness and spatial distance to obtain the structural response graph model.

[0008] As a further defined implementation, the nodes include the deflection measuring point of the main beam at mid-span, the deflection measuring point of the main beam near the support, the strain / curvature measuring point of the key section of the main beam, the strain measuring point at the bottom of the pier, and the vertical / horizontal displacement of the support. The process of assigning weights to edges based on structural stiffness and spatial distance is as follows:

[0009] in, Let be the elements in the adjacency matrix of the graph. The geometric distance between the two measuring points is... This represents the equivalent stiffness or influence coefficient of the corresponding path.

[0010] As an alternative implementation, for the graph within each time window, the process of introducing force flow path priors in the corresponding graph attention update includes: using physically pre-calculated or predefined force flow importance as a bias term in the attention calculation, specifically:

[0011] in, Features of the node being updated; for The characteristics of neighboring nodes; and Taken from matrix Weight matrix and parameter vector The initial values ​​are all from the interval [ The values ​​are randomly generated from a uniform distribution of [0.1, 0.1] and updated during training using the gradient descent algorithm. To prioritize the force flow based on prior knowledge of structural mechanics, the edges on the main path from main beam to section to pier to support are given an additional scoring constant bias 'a' to enhance their weight. For auxiliary paths that have a secondary correlation with support displacement, their scores are given an additional scoring constant bias 'b'. No additional bias is given to other non-critical edges to ensure that attention is mainly focused on force flow paths with clear physical meaning, where a > b.

[0012] As an alternative implementation, the main prediction network includes five prediction heads: a support longitudinal displacement regression head, a support lateral displacement regression head, a support rotation angle regression head, and two auxiliary prediction heads for predicting the curvature of key sections and the strain at the bottom of the pier, respectively. Dropout is added to both the prediction heads and the readout layer to achieve uncertainty quantification.

[0013] As an alternative implementation, the calculation process for the residual sequence between different types of physical parameters includes: calculating the deflection-support vertical displacement compatibility residual.

[0014] in, To predict the vertical displacement of the support; The predicted deflection of the nodes near the support; This is a fixed coefficient, corresponding to the coordination relationship between beam end deflection and support vertical response; Calculate the cross-sectional curvature-support rotation consistency residual:

[0015] in, Predicted curvature for key cross-section nodes; To predict the support rotation angle; It is a fixed coefficient; Calculate the consistency residual between pier bottom strain and support horizontal displacement:

[0016] in, To predict the strain at the bottom of the pier; To predict the horizontal displacement of the support; It is a fixed coefficient; Combine the three types of residuals into a residual matrix: .

[0017] As an alternative implementation, the process of determining physical consistency includes: inputting the residual matrix into a time-series encoder, using a one-dimensional convolutional layer therein to initially extract local variation features of the residuals; using a long short-term neural network layer therein to capture the correlation patterns of the physical residuals over time, obtaining the hidden state at the last moment as a physical consistency feature; and calculating a physical consistency score based on the consistency features.

[0018] As an alternative implementation, the process of adversarial training aimed at minimizing prediction error and physical inconsistency includes: generating a residual matrix using the real support shift sequence and setting the label to 1; generating a residual matrix using the prediction sequence and setting the label to 0; and training a physical consistency discriminator using binary cross-entropy loss, wherein the physical consistency discriminator is used to determine physical consistency based on the residual sequence between physical parameters of different classes. Fixed physical consistency discriminator Train the master prediction network. The loss function of the master prediction network is: ; in, ;

[0019] in, For coefficients, The support shift sequence output by the main prediction network is used to train the prediction network alternately with... Train the master prediction network, with the loss function of the master prediction network being , until the overall loss converges.

[0020] As an alternative implementation, the method further includes the following step: deploying the trained master prediction network as a virtual support displacement sensor to provide support displacement monitoring when the real sensor is missing or malfunctioning.

[0021] As a further implementation method, the main beam deflection, key section strain, pier bottom strain and working condition parameters are collected in each set period to form a time window sequence of fixed length. The input sequence is linearly normalized according to the normalization parameters in step one, and a structural response diagram model and working condition vector are generated. The main prediction network is invoked, and the predicted values ​​of the vertical displacement, horizontal displacement, and rotation angle of the support are output, based on the input with the most recent fixed value length as a condition. Dropout is set in both the graph readout layer and the regression head of the main prediction network, and multiple independent forward propagations are performed. Statistical features are calculated in each time window and then concatenated into a working condition feature vector. The system outputs predicted values ​​of vertical displacement, horizontal displacement, and rotation angle of the support, along with uncertainty indicators, at a set period. When the real support displacement sensor is offline or malfunctioning, it automatically switches to virtual sensor output.

[0022] A bridge support displacement intelligent prediction system based on multi-source structural response includes: The structural response diagram model building module is configured to preprocess the time series data of main beam deflection, strain / curvature of key sections of main beam, pier bottom strain and support displacement to build a structural response diagram model. The force flow path prior term introduction module is configured to use an attention-based graph neural network to spatiotemporally encode the node features of the structural response graph model. For the graph within each time window, the force flow path prior term is introduced in the corresponding graph attention update. The main prediction network construction module is configured to construct a main prediction network for support displacement prediction based on the updated graph attention structure. The main prediction network includes a graph feature encoder, a graph readout layer, and a prediction head. The graph feature encoder takes the structural response graph model as input and adopts the updated graph attention structure. It calculates attention weights to perform weighted aggregation of neighbor node features to achieve structured propagation of node features along the real force flow direction. The graph readout layer aggregates its output node features. The prediction head obtains the displacement estimate of the support, the curvature of the key section, and the strain at the bottom of the pier based on the aggregated features. The physical consistency determiner is configured to construct adversarial physical constraints and determine physical consistency based on the residual sequence between different classes of physical parameters. During the training of the main prediction network, adversarial training is performed with the goal of minimizing prediction error and physical inconsistency. The execution module is configured to use the trained master prediction network to obtain the support displacement prediction based on the target main beam deflection, the strain / curvature of the key section of the main beam and the strain at the bottom of the pier.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention improves prediction accuracy and stability: By constructing a structural response map with actual force flow direction and introducing force flow path priors into the graph attention, this invention enables the model to automatically strengthen the structural nodes that have the greatest impact on support displacement during the feature extraction stage, thereby significantly improving prediction accuracy. Compared with traditional LSTM or pure data-driven models based on single sequence input, this invention exhibits higher stability under different load scenarios and seasonal conditions.

[0024] (2) Ensuring that the prediction results conform to the laws of structural mechanics and improving the reliability of the project: This invention introduces a physical consistency discriminator, which constrains the predicted values ​​through the discrimination mechanism of physical residuals such as deflection coordination, curvature-rotation consistency, and pier bottom strain-horizontal displacement coupling, so that the model can simultaneously meet the requirements of numerical accuracy and physical rationality during the training process. The prediction results after this adversarial training can reliably reflect the deformation mode of the real structure and improve the reliability of the model in engineering monitoring.

[0025] (3) Achieving adaptive prediction capability across working conditions: By modulating the feature channels inside the network with working condition vectors (temperature, wind speed, traffic flow) of a fixed structure, the present invention can automatically adjust the feature weights of the model under different environmental and load conditions, so that the prediction results have excellent cross-seasonal, cross-climate and cross-time period adaptability, avoiding the accuracy decay of existing models when working conditions change.

[0026] (4) Ensuring continuous monitoring capability in the event of missing or malfunctioning support sensors: This invention deploys the trained prediction network as a virtual support displacement sensor, which can continuously provide estimates of support displacement, horizontal displacement, and rotation angle even when real sensors are offline, damaged, or cannot be deployed for a long period of time. At the same time, combined with the uncertainty output of Monte Carlo Dropout, the monitoring system can obtain prediction confidence and realize reliable alternative monitoring and early warning functions, solving the problem of "unavoidable monitoring gaps" in existing support monitoring systems.

[0027] (5) It has good deployability and engineering application value: The present invention is based entirely on existing conventional bridge monitoring data (deflection, strain, working condition parameters) to build the model, without the need for additional sensor deployment. The model has a light computational load and can run in real time on edge computing devices, field servers or operation and maintenance platforms. It is suitable for long-term monitoring of highway bridges, urban rail bridges and extra-large span bridges, which helps to reduce operation and maintenance costs and improve the operational safety of bridges.

[0028] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0029] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0030] Figure 1 This is a schematic diagram of the model training phase in one embodiment. Figure 2 This is a schematic diagram illustrating model call prediction in one embodiment. Detailed Implementation

[0031] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0032] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0034] Where there is no conflict, the embodiments and features described in this application may be combined with each other.

[0035] Example 1 As described in the background section, existing deep learning methods generally input structural responses in a single-sequence format, failing to represent the hierarchical deformation transmission chain of a bridge—"main girder—pier—support"—nor can they establish a correlation that conforms to the laws of force flow between multi-source structural responses. Furthermore, existing models lack mechanisms to determine physical consistency, making them prone to predictions that violate mechanical laws under noise, abnormal operating conditions, or sensor data drift. Simultaneously, these models are sensitive to environmental changes such as temperature, load, and traffic flow, exhibiting insufficient stability across operating conditions. More critically, when support monitoring equipment is damaged or missing, existing technologies cannot provide alternative means of estimating support displacement.

[0036] Therefore, the following technical issues need to be addressed: First, construct a structural response diagram model that can express the characteristics of the deformation chain, so that the support displacement prediction is based on the real structural relationship rather than black box correlation; second, establish an attention mechanism and physical consistency discrimination module consistent with the internal force transmission law to improve the physical reliability of the prediction; and third, achieve highly stable support displacement prediction across working conditions, across time periods, and under the condition of sensor absence through working condition adaptation and virtual support displacement sensor mechanism.

[0037] This embodiment provides a bridge support displacement intelligent prediction method based on multi-source structural response, including the following steps: Construction of the S1 structural response diagram model: Based on the time-series data of main beam deflection, strain / curvature of key sections of main beam, pier bottom strain and support displacement, time-aligned, outlier-removed and filtered data of each sensor are processed to construct a time window of length T, and each time window is regarded as a sample.

[0038] Within each time window, this embodiment abstracts the key monitoring locations of the bridge into a set of graph nodes: Node v1: Deflection measuring point of the main beam at mid-span Node v2: Deflection measuring point of the main beam near the support Node v3: Strain / curvature measurement point at key sections of the main beam (such as beam ends and negative bending moment zones). Node v4: Strain measuring point at the bottom of the pier Node v5: Vertical / Horizontal Displacement of Support Next, a directed edge set is established based on the physical deformation chain of the bridge. The meanings of some of the edges are as follows: v1→v3: Influence of overall deflection of the main beam on internal forces at key sections v2→v5: Influence of deflection near the support on the vertical displacement of the support v3→v4: Bending moment / shear force transfer from key sections to the pier body v4→v5: Influence of pier bending on horizontal displacement / rotation of supports The above relationships are encoded as an adjacency matrix of the graph. Furthermore, weights are assigned to the edges based on structural stiffness and spatial distance:

[0039] in The geometric distance between the two measuring points is... This represents the equivalent stiffness or influence coefficient of the corresponding path.

[0040] The node feature matrix of the output structural response graph in this step is denoted as X, and the weighted adjacency matrix is ​​denoted as... .

[0041] S2's graph attention mechanism based on force flow path alignment: After obtaining the structural response map, to avoid inconsistencies between attention and force flow, the invention employs an attention-based graph neural network to spatiotemporally encode node features. For a graph within a certain time window, in the graph attention update layer, this invention introduces a force flow path prior term on top of the traditional GAT, using the physically pre-calculated or predefined "force flow importance" as a bias term for attention in the calculation. Specifically, it can be written as:

[0042] in: Features of the node being updated; for The characteristics of neighboring nodes; and Taken from matrix Weight matrix and parameter vector The initial values ​​are all from the interval [ The values ​​are randomly generated from a uniform distribution of [0.1, 0.1] and updated during training using the gradient descent algorithm. Based on the "force flow priority" determined a priori by structural mechanics, this design sets the attention scoring constant bias of 1.0 for edges on the "main beam-section-pier-support" path, thereby systematically enhancing the weight of these paths. These edges include: v1→v3, v2→v5, v3→v4, and v4→v5. For auxiliary paths that have a secondary correlation with support displacement, the scoring constant bias is increased by 0.5 to prevent them from being completely ignored. These edges include: v1→v2, v1→v5, and v3→v5. No bias is added for other non-critical edges to ensure that attention is mainly focused on force flow paths with clear physical meaning.

[0043] Construction and training of the S3 prediction network: like Figure 1 As shown, after constructing a node update mechanism based on the graph attention layer of force flow path alignment, this invention uses this graph attention structure as the core feature extraction module to construct a main prediction network for support displacement prediction. The network consists of three parts: (1) Graph feature encoder The graph feature encoder is the first part of the prediction network. Its input is the structural response graph constructed in step one, including the node feature matrix. Adjacency matrix weighted by physical priors The encoder employs the graph attention layer based on force flow path alignment constructed in step two, and calculates attention weights. It also performs weighted aggregation of neighbor node features to achieve structured propagation of node features along the actual force flow direction; (2) Readout layer The node feature matrix output by the graph feature encoder It is a graph-based high-dimensional representation, where each node represents a physical measurement point. To obtain the overall structural state for final prediction, this invention employs a graph readout layer to aggregate node features, forming a fixed-length global structural state vector. .

[0044] (3) Prediction Head: The support displacement regression head is the third part of the prediction network and the final location for generating the prediction output. It contains five prediction heads: the support longitudinal displacement regression head, the support lateral displacement regression head, the support rotation angle regression head, and two auxiliary prediction heads used to predict the curvature of the key section and the strain at the pier bottom, respectively. Dropout is added to both the prediction head and the readout layer to quantify uncertainty. This module uses the global state vector of the readout layer. As input, output the displacement estimate of the target support. Key section curvature and pier bottom strain .

[0045] S4 introduces a physical consistency discriminant to construct adversarial physical constraints: In prediction networks Based on this, the present invention adds a physical consistency discriminator. This is used to determine physical plausibility. The physical consistency discriminator of this invention... It adopts a fixed structure and consists of three parts: a physical residual calculation module, a timing coding module, and a consistency discrimination output layer.

[0046] Physical residual calculation module: Given the structural response diagram and operating conditions for a certain time window, predict the support displacement sequence output by the network:

[0047] in For multi-source node input, This is the working condition vector.

[0048] This invention constructs three types of physical residual sequences based on the deformation compatibility and internal force transmission relationships determined in structural mechanics, which are used as inputs to the discriminator: (1) Deflection-support vertical displacement compatibility residual:

[0049] in, To predict the vertical displacement of the support; The predicted deflection of the nodes near the support; As a fixed coefficient, it is set to be in this invention. =1.0, which corresponds to the coordination relationship between beam end deflection and support vertical response.

[0050] (2) Cross-section curvature-support rotation consistency residual

[0051] in, Predicted curvature for key cross-section nodes; To predict the support rotation angle; As a fixed coefficient, it is set to be in this invention. =1.0.

[0052] (3) Consistency residual of pier bottom strain-support horizontal displacement

[0053] in, To predict the strain at the bottom of the pier; To predict the horizontal displacement of the support; As a fixed coefficient, it is set to be in this invention. =1.0.

[0054] This invention combines three types of residuals into a residual matrix: .

[0055] Timing coding module The residual matrix R is fed into a temporal encoder consisting of a one-dimensional convolutional layer and an LSTM layer: the one-dimensional convolutional layer is used to initially extract the local variation features of the residual; the LSTM layer is used to capture the correlation patterns of the physical residual over time; this module outputs the hidden state at the final time step. As a physical consistency feature.

[0056] Consistency discrimination output layer The last layer of the discriminator is a fully connected network, with the input being... The output is a physical consistency score. :

[0057] in ( ) is a Sigmoid function that makes the score s∈[0,1]. The higher the value, the more the predicted sequence conforms to the mechanical laws of bridge deformation.

[0058] adversarial training process This invention employs a fixed adversarial training process, enabling the prediction network to simultaneously minimize prediction error and physical inconsistency during training.

[0059] The training steps include: (1) Fixed Training the discriminator The residual matrix Rtrue is generated using the real support displacement sequence (partial real support displacement can still be obtained in the historical or short-term deployment phases for training the discriminator; virtual support monitoring is used in the long-term prediction phase), and the label is set to 1; the residual matrix Rpred is generated using the predicted sequence, and the label is set to 0; binary cross-entropy loss is used for training. .

[0060] (2) Fixed Training the prediction network : The loss function of the prediction network is:

[0061] in:

[0062]

[0063] coefficient It is fixed at 0.5.

[0064] (3) Train the prediction network Fθ and the discriminator Dφ alternately until the overall loss converges. Construction and deployment of S5 virtual support displacement sensor: This invention will train the prediction network. Deployed as a virtual support displacement sensor, it provides support displacement monitoring when the real sensor is missing or malfunctions. The virtual sensor consists of fixed input processing, predictive inference, uncertainty quantification, and output modules.

[0065] The system collects main beam deflection, key section strain, pier bottom strain, and working parameters every 0.1 seconds, forming a time window sequence of 600 points (60 seconds). The input sequence is then linearly normalized using the normalization parameters from step one, and a structural response diagram is generated. With working condition vector .

[0066] System call prediction network Based on the input of the most recent 600 points, the system outputs three types of support displacement predictions: support vertical displacement prediction, support horizontal displacement prediction, and support rotation angle prediction.

[0067] like Figure 2 As shown, uncertainty quantification employs a fixed-structure Monte Carlo Dropout method. Dropout is applied to both the graph readout layer and the regression head of the prediction network, with a ratio of 0.1. The system performs 20 independent forward propagations.

[0068]

[0069] in, This is the complete "structural response map node feature sequence" constructed in step 1 for the past 60 seconds; temperature T(t), wind speed Vw(t), and traffic flow q(t) variables are extracted from the monitoring system, and statistical characteristics are calculated for each time window: average temperature Peak traffic volume per unit time ,average value These features are then combined into a working condition feature vector: .

[0070] The system outputs the following predicted values ​​μt for vertical and horizontal displacement and rotation angle of the support at 0.1s cycles; the uncertainty index σt, CI 95%, is calculated as follows:

[0071]

[0072]

[0073] When the real support displacement sensor is offline or malfunctioning, the system automatically switches to virtual sensor output.

[0074] This embodiment can significantly improve the reliability, continuity, and engineering applicability of bridge bearing displacement monitoring.

[0075] In other embodiments, the structure of the force flow path aligned attention layer remains unchanged in the graph feature encoder, but the specific form of the graph convolution can be any of the following fixed implementations: spectral domain graph convolution based on Chebyshev polynomials; inductive graph convolution based on GraphSAGE; or Diffusion GCN based on a fixed diffusion kernel.

[0076] The alternative convolutional structure only changes the way node features are aggregated, and does not change the prior coefficients p of the force flow. ij The rules for determining the value and how it functions.

[0077] Example 2 A bridge support displacement intelligent prediction system based on multi-source structural response includes: The structural response diagram model building module is configured to preprocess the time series data of main beam deflection, strain / curvature of key sections of main beam, pier bottom strain and support displacement to build a structural response diagram model. The force flow path prior term introduction module is configured to use an attention-based graph neural network to spatiotemporally encode the node features of the structural response graph model. For the graph within each time window, the force flow path prior term is introduced in the corresponding graph attention update. The main prediction network construction module is configured to construct a main prediction network for support displacement prediction based on the updated graph attention structure. The main prediction network includes a graph feature encoder, a graph readout layer, and a prediction head. The graph feature encoder takes the structural response graph model as input and adopts the updated graph attention structure. It calculates attention weights to perform weighted aggregation of neighbor node features to achieve structured propagation of node features along the real force flow direction. The graph readout layer aggregates its output node features. The prediction head obtains the displacement estimate of the support, the curvature of the key section, and the strain at the bottom of the pier based on the aggregated features. The physical consistency determiner is configured to construct adversarial physical constraints and determine physical consistency based on the residual sequence between different classes of physical parameters. During the training of the main prediction network, adversarial training is performed with the goal of minimizing prediction error and physical inconsistency. The execution module is configured to use the trained master prediction network to obtain the support displacement prediction based on the target main beam deflection, the strain / curvature of the key section of the main beam and the strain at the pier bottom.

[0078] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0079] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0080] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0081] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent prediction of bridge support displacement based on multi-source structural response, characterized in that, Includes the following steps: The time series data of main beam deflection, strain / curvature of key sections of main beam, pier bottom strain and support displacement are preprocessed to construct a structural response diagram model; Attention-based graph neural networks are used to spatiotemporally encode the node features of the structural response graph model. For the graph within each time window, a priori terms of force flow paths are introduced in the corresponding graph attention update. Based on the updated graph attention structure, a main prediction network for support displacement prediction is constructed. The main prediction network includes a graph feature encoder, a graph readout layer, and a prediction head. The graph feature encoder takes the structural response graph model as input and adopts the updated graph attention structure. It calculates attention weights to perform weighted aggregation of neighbor node features, realizing the structured propagation of node features along the real force flow direction. The graph readout layer aggregates its output node features. The prediction head obtains the displacement estimate of the support, the curvature of the key section, and the strain at the bottom of the pier based on the aggregated features. Construct adversarial physical constraints, determine physical consistency based on the residual sequence between different types of physical parameters, and conduct adversarial training during the training of the main prediction network with the goal of minimizing prediction error and physical inconsistency; Using the trained master prediction network, the predicted support displacement is obtained based on the target main beam deflection, the strain / curvature of the key section of the main beam, and the strain at the pier bottom.

2. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, The preprocessing and structural response graph model construction process includes: time alignment, outlier removal and filtering of various types of sensor data; construction of multiple time windows of length T, with each time window considered as a sample; abstracting the key monitoring locations of the bridge into a set of graph nodes within each time window; establishing a directed edge set based on the physical deformation chain of the bridge to determine the influence or transmission relationship between nodes; encoding the relationship into an adjacency matrix of the graph; and assigning weights to the edges based on structural stiffness and spatial distance to obtain the structural response graph model.

3. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 2, characterized in that, The nodes include the deflection measuring point of the main beam at mid-span, the deflection measuring point of the main beam near the support, the strain / curvature measuring point of the key section of the main beam, the strain measuring point at the bottom of the pier, and the vertical / horizontal displacement of the support; The process of assigning weights to edges based on structural stiffness and spatial distance is as follows: in, Let be the elements in the adjacency matrix of the graph. The geometric distance between the two measuring points is... This represents the equivalent stiffness or influence coefficient of the corresponding path.

4. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, For the graph within each time window, the process of introducing force flow path priors in the corresponding graph attention update includes: using the physically pre-calculated or predefined force flow importance as a bias term in the attention calculation, specifically: in: Features of the node being updated; for The characteristics of neighboring nodes; and Taken from matrix Weight matrix and parameter vector The initial values ​​are all from the interval [ The values ​​are randomly generated from a uniform distribution of [0.1, 0.1] and updated during training using the gradient descent algorithm. To prioritize the force flow based on prior knowledge of structural mechanics, the edges on the main path from main beam to section to pier to support are given an additional scoring constant bias 'a' to enhance their weight. For auxiliary paths that have a secondary correlation with support displacement, their scores are given an additional scoring constant bias 'b'. No additional bias is given to other non-critical edges to ensure that attention is mainly focused on force flow paths with clear physical meaning, where a > b.

5. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, The main prediction network includes five prediction heads: a longitudinal displacement regression head for the support, a lateral displacement regression head for the support, a rotation angle regression head for the support, and two auxiliary prediction heads for predicting the curvature of key sections and the strain at the bottom of the pier, respectively. Dropout is added to both the prediction heads and the readout layer to quantify uncertainty.

6. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, The calculation process of the residual sequence between different types of physical parameters includes: calculating the deflection-support vertical displacement compatibility residual. in, To predict the vertical displacement of the support; The predicted deflection of the nodes near the support; This is a fixed coefficient, corresponding to the coordination relationship between beam end deflection and support vertical response; Calculate the cross-sectional curvature-support rotation consistency residual: in, Predicted curvature for key cross-section nodes; To predict the support rotation angle; It is a fixed coefficient; Calculate the consistency residual between pier bottom strain and support horizontal displacement: in, To predict the strain at the bottom of the pier; To predict the horizontal displacement of the support; It is a fixed coefficient; Combine the three types of residuals into a residual matrix: .

7. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, The process of determining physical consistency includes: inputting the residual matrix into the temporal encoder, using the one-dimensional convolutional layer to initially extract the local variation features of the residual; using the long short-term neural network layer to capture the correlation pattern of the physical residual over time, obtaining the hidden state at the last moment as the physical consistency feature; and calculating the physical consistency score based on the consistency feature.

8. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, The adversarial training process aimed at minimizing prediction error and physical inconsistency includes: generating a residual matrix using the real support shift sequence and setting the label to 1; generating a residual matrix using the prediction sequence and setting the label to 0; and training a physical consistency discriminator using binary cross-entropy loss. The physical consistency discriminator is used to determine physical consistency based on the residual sequences between physical parameters of different classes. Fixed physical consistency discriminator Train the master prediction network. The loss function of the master prediction network is: ; in, ; ; in, For coefficients, The support shift sequence output by the main prediction network is used to train the prediction network alternately with... Train the master prediction network, with the loss function of the master prediction network being , until the overall loss converges.

9. The intelligent prediction method for bridge support displacement based on multi-source structural response as described in claim 1, characterized in that, It also includes the following steps: The trained master prediction network is deployed as a virtual support displacement sensor to provide support displacement monitoring when the real sensor is missing or malfunctions. Specifically, the deflection of the main beam, the strain of key sections, the strain at the bottom of the pier, and the working parameters are collected in each set period to form a time window sequence of fixed length. The input sequence is linearly normalized according to the normalization parameters in step one, and a structural response diagram model and a working vector are generated. The main prediction network is invoked, and the predicted values ​​of the vertical displacement, horizontal displacement, and rotation angle of the support are output, based on the input with the most recent fixed value length as a condition. Dropout is set in both the graph readout layer and the regression head of the main prediction network, and multiple independent forward propagations are performed. Statistical features are calculated in each time window and then concatenated into a working condition feature vector. The system outputs predicted values ​​of vertical displacement, horizontal displacement, and rotation angle of the support, along with uncertainty indicators, at a set period. When the real support displacement sensor is offline or malfunctioning, it automatically switches to virtual sensor output.

10. A bridge support displacement intelligent prediction system based on multi-source structural response, characterized in that, include: The structural response diagram model building module is configured to preprocess the time series data of main beam deflection, strain / curvature of key sections of main beam, pier bottom strain and support displacement to build a structural response diagram model. The force flow path prior term introduction module is configured to use an attention-based graph neural network to spatiotemporally encode the node features of the structural response graph model. For the graph within each time window, the force flow path prior term is introduced in the corresponding graph attention update. The main prediction network construction module is configured to construct a main prediction network for support displacement prediction based on the updated graph attention structure. The main prediction network includes a graph feature encoder, a graph readout layer, and a prediction head. The graph feature encoder takes the structural response graph model as input and adopts the updated graph attention structure. It calculates attention weights to perform weighted aggregation of neighbor node features to achieve structured propagation of node features along the real force flow direction. The graph readout layer aggregates its output node features. The prediction head obtains the displacement estimate of the support, the curvature of the key section, and the strain at the bottom of the pier based on the aggregated features. The physical consistency determiner is configured to construct adversarial physical constraints and determine physical consistency based on the residual sequence between different classes of physical parameters. During the training of the main prediction network, adversarial training is performed with the goal of minimizing prediction error and physical inconsistency. The execution module is configured to use the trained master prediction network to obtain the support displacement prediction based on the target main beam deflection, the strain / curvature of the key section of the main beam and the strain at the bottom of the pier.

Images