CN-122019997-A - Bridge support displacement intelligent prediction method and system based on multi-source structural response

CN122019997ACN 122019997 ACN122019997 ACN 122019997ACN-122019997-A

Abstract

The invention belongs to the field of bridge structure health monitoring, and provides an intelligent bridge support displacement prediction method and system based on multi-source structure response, which are used for preprocessing monitored time sequence data and constructing a structure response graph model; the method comprises the steps of carrying out space-time coding on node characteristics of a structural response graph model by adopting a graph neural network based on attention, introducing a force flow path prior term into corresponding graph meaning force updating for graphs in each time window, constructing a main prediction network for support displacement prediction based on the updated graph meaning force structure, constructing an antagonistic physical constraint, determining physical consistency, and carrying out antagonistic training with the aim of minimizing prediction errors and physical inconsistency during main prediction network training. The invention fundamentally solves the key problems of insufficient obtainable information, lack of physical interpretation of the model, poor generalization capability of the cross-working condition and difficult long-term stable operation of the support sensor in the prior art.

Inventors

DU CONG
ZHONG KAIQI
WANG XIAOCHAO
Tian Weiyang
TIAN YUAN
WANG JIANZHU
WU JIANQING

Assignees

山东大学
山东大学苏州研究院

Dates

Publication Date: 20260512
Application Date: 20260129

Claims (10)

1. The intelligent bridge support displacement prediction method based on the multi-source structural response is characterized by comprising the following steps of: preprocessing time sequence data of girder deflection, girder key section strain/curvature, pier bottom strain and support displacement, and constructing a structural response graph model; space-time coding node characteristics of a structural response graph model by adopting a graph neural network based on attention, and introducing a force flow path prior term in corresponding graph attention updating for graphs in each time window; Based on the updated graph annotation force structure, a main prediction network for supporting seat displacement prediction is constructed, the main prediction network comprises a graph feature encoder, a graph reading layer and a prediction head, the graph feature encoder takes a structural response graph model as input, the updated graph annotation force structure is adopted, the node features of the neighbors are weighted and aggregated through calculating attention weights, the structured propagation of the node features along the true force flow direction is realized, the graph reading layer aggregates the node features output by the node features, and the prediction head obtains the displacement estimated value of the supporting seat, the key cross section curvature and the pier bottom strain according to the aggregated features; constructing countermeasure physical constraint, determining physical consistency according to residual sequences among different types of physical parameters, and performing countermeasure training with the aim of minimizing prediction errors and physical inconsistency during main prediction network training; and obtaining the support displacement prediction based on the target girder deflection, the girder key section strain/curvature and the pier bottom strain data by using the trained main prediction network.
2. The intelligent bridge support displacement prediction method based on the multi-source structural response is characterized in that preprocessing is carried out, the process of constructing a structural response graph model comprises the steps of carrying out time alignment, outlier rejection and filtering processing on various acquired sensing data, constructing a plurality of time windows with the length of T, wherein each time window is regarded as a sample, abstracting key monitoring positions of a bridge into a group of graph nodes in each time window, establishing a directed edge set according to a physical deformation chain of the bridge, determining influence or transfer relations among the nodes, encoding the relations into an adjacent matrix of the graph, weighting the edges according to structural rigidity and spatial distance, and obtaining the structural response graph model.
3. The intelligent bridge support displacement prediction method based on the multi-source structural response according to claim 2, wherein the node comprises a midspan girder deflection measuring point, a girder deflection measuring point near the support, a strain/curvature measuring point of a girder key section, a pier bottom strain measuring point and a support vertical/horizontal displacement; The process of giving weight to the edges according to the structural rigidity and the space distance is as follows: Wherein, the As an element in the adjacency matrix of the figure, Is the geometric distance between two measuring points, Is the equivalent stiffness or influence coefficient of the corresponding path.
4. The intelligent bridge support displacement prediction method based on the multi-source structural response according to claim 1, wherein for the graphs in each time window, the process of introducing the force flow prior term in the corresponding graph annotation force update comprises the steps of taking the physical pre-calculated or set force flow importance as the bias term of attention to participate in calculation, specifically: Wherein: Is an updated node characteristic; Is that Adjacent node features of (a); And Taken from a matrix Weight matrix Sum parameter vector The initial values of all the two are from the interval [ 0.1,0.1] Are randomly generated in the uniform distribution and updated by a gradient descent algorithm in the training process; For the priority of force flow determined by structural mechanics in advance, the sides of the main beam-section-pier body-support on the main path are increased in attention to scoring constant bias a, the weights of the paths are enhanced, the constant bias b is increased in scoring for the auxiliary path with secondary correlation with the support displacement, the bias is not increased for other non-critical sides, and the attention is ensured to be mainly focused on the force flow path with definite physical meaning, and a > b.
5. The intelligent bridge support displacement prediction method based on multi-source structural response according to claim 1, wherein the main prediction network comprises five prediction heads, a support longitudinal displacement regression head, a support transverse displacement regression head, a support corner regression head and two auxiliary prediction heads for predicting key section curvature and pier bottom strain respectively, and Dropout is added into both the prediction heads and the readout layer to realize uncertainty quantification.
6. The intelligent bridge support displacement prediction method based on the multi-source structural response of claim 1, wherein the calculation process of the residual sequence between different types of physical parameters comprises the following steps of calculating deflection-support vertical displacement coordination residual: Wherein, the To predict the vertical displacement of the support; The deflection is predicted for the nodes near the support; The fixed coefficient corresponds to the coordination relation between the deflection of the beam end and the vertical reaction of the support; Calculating a section curvature-support corner consistency residual: Wherein, the The predicted curvature of the key section node; To predict the support rotation angle; is a fixed coefficient; calculating pier bottom strain-support horizontal displacement consistency residual errors: Wherein, the Predicting pier bottom strain; To predict the horizontal displacement of the support; is a fixed coefficient; Combining three types of residuals into a residual matrix: 。
7. The intelligent bridge support displacement prediction method based on multi-source structural response according to claim 1, wherein the process of determining physical consistency comprises the steps of inputting a residual matrix into a time sequence encoder, initially extracting local variation characteristics of residual by utilizing a one-dimensional convolution layer, capturing a correlation mode of the physical residual along with time by utilizing a long-short-period neural network layer to obtain a hidden state at the last moment as the physical consistency characteristic, and calculating a physical consistency score according to the consistency characteristic.
8. The intelligent bridge support displacement prediction method based on multi-source structural response of claim 1, wherein the process of performing countermeasure training aiming at minimizing prediction errors and physical inconsistencies comprises the steps of generating a residual matrix by using a real support displacement sequence, setting a label as 1, generating the residual matrix by using the prediction sequence, setting the label as 0, training a physical consistency discriminator by using a classification cross entropy loss, wherein the physical consistency discriminator is used for determining physical consistency according to the residual sequence among different types of physical parameters; Fixed physical consistency discriminator Training a main prediction network, wherein the loss function of the main prediction network is as follows: ; Wherein, the ; ; Wherein, the As the coefficient of the light-emitting diode, Support displacement sequence output by main prediction network, and alternately training prediction network and support displacement sequence The main prediction network is trained, and the loss function of the main prediction network is as follows until the integral loss converges.
9. The intelligent bridge support displacement prediction method based on the multi-source structural response of claim 1, further comprising the steps of deploying a trained main prediction network as a virtual support displacement sensor for providing support displacement monitoring in the event of a real sensor missing or failure; Specifically, collecting main beam deflection, key section strain, pier bottom strain and working condition parameters in each set period, forming a time window sequence with a fixed length, performing linear normalization on an input sequence according to the normalization parameters of the step one, and generating a structural response diagram model and a working condition vector; Calling a main prediction network, and outputting a support vertical displacement predicted value, a support horizontal displacement predicted value and a support corner predicted value under the condition of the input of the latest fixed value length; Dropout is arranged in a graph reading layer and a regression head of the main prediction network, independent forward propagation is carried out for a plurality of times, statistical characteristics are calculated on each time window, and the statistical characteristics are spliced into working condition characteristic vectors; Outputting a predicted value of the vertical displacement of the support, a predicted value of the horizontal displacement of the support, a predicted value of the rotation angle of the support and an uncertainty index in a set period; And when the real support displacement sensor is offline or abnormal, automatically switching to virtual sensor output.
10. A bridge support displacement intelligent prediction system based on multi-source structural response is characterized by comprising: The structural response graph model construction module is configured to preprocess time sequence data of girder deflection, girder key section strain/curvature, pier bottom strain and support displacement to construct a structural response graph model; The force flow path prior item introduction module is configured to perform space-time coding on node characteristics of the structural response graph model by adopting the graph neural network based on attention, and introduce force flow path prior items in corresponding graph attention updating for graphs in each time window; The main prediction network construction module is configured to construct a main prediction network for support displacement prediction based on the updated graph injection force structure, the main prediction network comprises a graph feature encoder, a graph reading layer and a prediction head, the graph feature encoder takes a structural response graph model as input, the updated graph injection force structure is adopted, the neighbor node features are weighted and aggregated through calculating attention weights, structured propagation of the node features along the direction of a true force flow is realized, the graph reading layer aggregates the node features output by the node features, and the prediction head obtains a displacement estimated value of a support, key cross section curvature and pier bottom strain according to the aggregated features; a physical consistency determiner configured to construct an antagonistic physical constraint, determine physical consistency according to a residual sequence between different types of physical parameters, and perform an antagonistic training with a goal of minimizing a prediction error and physical inconsistency when the main prediction network is trained; the execution module is configured to obtain a support displacement prediction value based on the target girder deflection, the girder key section strain/curvature and pier bottom strain data by using the trained main prediction network.

Description

Bridge support displacement intelligent prediction method and system based on multi-source structural response Technical Field The invention belongs to the field of bridge structure health monitoring, and particularly relates to an intelligent bridge support displacement prediction method and system based on multi-source structure response. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. In the field of bridge structure health monitoring, support displacement is an important index reflecting the boundary constraint state of a bridge, the functional degradation of a support and the safety of the whole structure. However, the space at the support is narrow, the vibration is severe, the temperature difference is large, the pollution is serious, the traditional displacement meter, the vision system or the optical fiber sensor are difficult to work for a long time, stably and accurately, and the displacement of the support is the physical quantity which is most difficult to obtain directly in the structural monitoring. In addition, the support displacement is formed by coupling multiple factors such as girder deflection change, force redistribution in a critical section boundary, pier body bending deformation, temperature working conditions and the like, the change mode is complex, nonlinearity is strong, and differences between different bridge types and different load conditions are obvious, so that the conventional data driving method is difficult to capture the intrinsic structural coupling rule. The existing deep learning model can learn a certain relevance from data, but lacks the constraint of structural mechanics law, has no interpretability, has poor stability under the working condition, cannot ensure that a predicted result accords with a deformation transmission path of a real structure, and simultaneously, the existing model is mostly based on simple sequence input, cannot embody a deformation chain structure special for a bridge, namely a main beam, a pier body and a support, and cannot adapt to complex, multi-source and unstable bridge monitoring scenes. In addition, the existing method cannot automatically adjust the model behavior according to different temperatures, vehicle loads and environmental factors, cannot judge the model prediction reliability or abnormal conditions of the sensor, and cannot automatically provide alternative monitoring capability when the support sensor is damaged. Disclosure of Invention In order to solve the problems, the invention provides an intelligent bridge support displacement prediction method and system based on multi-source structural response, which fundamentally solve the key problems that the available information is insufficient, the model lacks of physical interpretation, the generalization capability of the cross-working condition is poor and the support sensor is difficult to stably work for a long time in the prior art. According to some embodiments, the present invention employs the following technical solutions: a bridge support displacement intelligent prediction method based on multi-source structural response comprises the following steps: preprocessing time sequence data of girder deflection, girder key section strain/curvature, pier bottom strain and support displacement, and constructing a structural response graph model; space-time coding node characteristics of a structural response graph model by adopting a graph neural network based on attention, and introducing a force flow path prior term in corresponding graph attention updating for graphs in each time window; Based on the updated graph annotation force structure, a main prediction network for supporting seat displacement prediction is constructed, the main prediction network comprises a graph feature encoder, a graph reading layer and a prediction head, the graph feature encoder takes a structural response graph model as input, the updated graph annotation force structure is adopted, the node features of the neighbors are weighted and aggregated through calculating attention weights, the structured propagation of the node features along the true force flow direction is realized, the graph reading layer aggregates the node features output by the node features, and the prediction head obtains the displacement estimated value of the supporting seat, the key cross section curvature and the pier bottom strain according to the aggregated features; constructing countermeasure physical constraint, determining physical consistency according to residual sequences among different types of physical parameters, and performing countermeasure training with the aim of minimizing prediction errors and physical inconsistency during main prediction network training; and obtaining the support displacement prediction based on the target girder deflection, the girder key section strain/curvature a