A cable-stayed bridge deformation prediction method fusing multi-source response decoupling and physical constraints

Abstract

This invention relates to the field of structural health monitoring and safety early warning technology for large bridges, specifically to a method for predicting the deformation of cable-stayed bridges by integrating multi-source response decoupling and physical constraints. The method acquires the time series of tower displacement and main girder deflection of the cable-stayed bridge through a monitoring system. Wavelet transform is used to decouple the original signals into low-frequency environmental response components and high-frequency live load response components. A dual-branch prediction architecture is then constructed. An LSTM network is used to model the low-frequency response components, while a physical information neural network incorporating the tower-girder deformation coordination relationship is built for the high-frequency response components. A physical constraint residual equation is constructed and transformed into a comprehensive loss function to achieve joint training and optimization of model weights and physical parameters. Finally, the dual-branch output is reconstructed to obtain the deformation prediction result. This invention solves the problems of poor physical interpretability and low generalization accuracy of traditional data-driven models under complex environmental loads, significantly improving the accuracy of key deformation response prediction and the rationality of deformation relationships in cable-stayed bridges.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring and safety early warning technology for large bridges, specifically to a method for predicting the deformation of cable-stayed bridges that integrates multi-source response decoupling and physical constraints. Background Technology

[0002] In engineering practice, the mid-span deflection of the main girder and the horizontal deviation of the bridge tower are core indicators for evaluating the overall stiffness of cable-stayed bridges. Existing monitoring and evaluation methods mostly obtain structural response data by deploying monitoring systems, and often construct time series prediction models for a single measuring point (such as tower deviation or deflection), such as autoregressive models and LSTM. These methods mostly rely on a single structural response index and achieve deformation prediction in a purely data-driven manner.

[0003] With the development of deep learning, existing research has used multivariate time series models to fuse information from multiple measurement points to improve prediction accuracy and has recognized the inherent correlation between tower and beam deformation. However, current technologies mainly rely on statistical analysis or neural network modeling of single or multiple measurement point data such as tower displacement or main beam deflection to predict bridge deformation, which still has the following objective drawbacks:

[0004] Lack of joint modeling capability for the response of key bridge structures: Existing methods usually establish prediction models for single response indicators such as horizontal displacement of bridge towers or mid-span deflection of main beams, which makes it difficult to reflect the coordinated deformation characteristics between different key parts of the bridge at the same time, thus limiting the ability to assess the overall stress state of the bridge.

[0005] Lack of structural coupling constraints in multivariate modeling: Although some existing technologies use multi-point data for joint prediction, they mainly construct input features by simple data splicing and do not introduce the structural force transmission relationship and deformation coupling mechanism between bridge towers and main beams into the model. As a result, the model can only learn statistical correlations and it is difficult to ensure that the prediction results conform to the laws of actual engineering mechanics.

[0006] Pure data-driven models lack physical consistency: Predictive models based on deep learning methods such as Long Short-Term Memory (LSTM) networks usually rely on historical data for training and lack structural physical constraints. This may result in inconsistencies between the predicted results and the actual structural deformation direction. For example, the relationship between the deflection of the main beam and the deflection of the bridge tower does not conform to the actual stress characteristics, thus affecting the reliability of the results.

[0007] The existing methods are not capable of expressing the coupling behavior of bridge structures. They do not model the mechanical transmission process in cable-stayed bridges where the change in main beam deflection causes the change in cable force and further leads to the deflection of the bridge tower. They cannot characterize the coupling deformation law between the bridge tower and the main beam, which limits the interpretability and generalization ability of the prediction model. To this end, a method for predicting the deformation of cable-stayed bridges that integrates multi-source response decoupling and physical constraints is proposed. Summary of the Invention

[0008] To address the technical problems existing in the prior art, the present invention provides a method for predicting the deformation of cable-stayed bridges that integrates multi-source response decoupling and physical constraints.

[0009] To address the aforementioned technical problems, this invention provides the following technical solution: a method for predicting the deformation of cable-stayed bridges that integrates multi-source response decoupling and physical constraints, comprising the following steps:

[0010] S1, Multi-source data acquisition: Through the monitoring system deployed at the top of the bridge tower and the mid-span of the main girder of the cable-stayed bridge, the displacement response time series data of mid-span deflection and bridge tower offset are continuously collected to obtain a multi-source response dataset;

[0011] S2, Data Preprocessing: Process the multi-source response dataset and divide it into training set, validation set and test set;

[0012] S3, Construction of the dual-branch prediction model: Construct a temperature response prediction branch to predict low-frequency deformation response; construct a live load response prediction branch to predict high-frequency deformation response;

[0013] S4, Physical Constraint Construction: Based on the deformation coordination relationship between the tower and beam of a cable-stayed bridge, construct the physical constraint residual equation;

[0014] S5, Joint Model Training: The dual-branch prediction model is trained separately using the training set. The live load response prediction branch constructs a comprehensive loss function by introducing physical constraint residual equations.

[0015] S6, Deformation Prediction and Response Reconstruction: Input the historical sequence data of the time to be predicted into the trained dual-branch prediction model, output the low-frequency prediction value and high-frequency prediction value of the future time respectively, and superimpose the two to reconstruct the final prediction results of bridge tower offset and mid-span deflection.

[0016] Preferably, in step S2, the multi-source response dataset processing steps are as follows:

[0017] S21, missing values ​​are imputed in the multi-source response dataset using linear interpolation. Let the dataset be... exist There is always a missing value. The interpolation expression for the missing value is:

[0018] ;

[0019] in, for Data at any given time; for Data at any given time; This refers to the data collection time interval;

[0020] S22, using wavelet transform, performs multi-scale decomposition and reconstruction on the original displacement response time series data after missing value supplementation, decoupling the original displacement response time series data into low-frequency temperature response components and high-frequency live load response components. The low-frequency temperature response component... , and high-frequency live load response components , Each should be standardized separately.

[0021] Preferably, the wavelet transform employs a wavelet packet decomposition algorithm, setting the number of decomposition layers based on the bridge's natural frequency and the ambient temperature change period, extracting approximate components as low-frequency temperature response components, and extracting detail components as high-frequency live load response components.

[0022] Preferably, in step S22, standardization is performed using Z-score, and the formula is as follows:

[0023] ;

[0024] in, This is the standardized dataset; This is the original dataset; For the original dataset The mean; For the original dataset The variance.

[0025] Preferably, in step S3, the dual-branch prediction model includes a temperature response prediction branch and a live load response prediction branch. The long short-term memory network model in the temperature response prediction branch includes an input layer, a two-layer LSTM layer, and an output layer. The LSTM layer dimension is set to 128, the model learning rate is set to 0.001, the weight decay coefficient is set to 0.001, and the batch size is set to 64. The input layer is used to receive the time series data of the mid-span deflection of the main beam and the low-frequency temperature response components of the tower top displacement. The LSTM layer is used to extract the long-term dependency features in the time series. The output layer is used to output the low-frequency temperature response components of the main beam deflection and the tower displacement at the time to be predicted.

[0026] The physical information long short-term memory network model in the live load response prediction branch includes an input layer, a two-layer LSTM layer, a physical parameter training layer, and an output layer. The LSTM layer has a dimension of 128, the physical parameter training layer has a dimension of 64, the model learning rate is 0.001, the weight decay coefficient is 0.001, and the batch size is 64. The input layer is used to receive time series data of the high-frequency live load response components of the mid-span deflection of the main beam and the tower top offset. The LSTM layer is used to extract long-term dependency features from the time series. The physical parameter training layer is used to train the parameters in the physical constraint equations. The output layer is used to output the high-frequency live load response components of the mid-span deflection of the main beam and the tower top offset at the time to be predicted.

[0027] Preferably, in step S4, the specific construction steps are as follows:

[0028] S41, Establish the physical coupling equations describing the deformation coordination relationship between the tower and the beam;

[0029] S42, the live load response prediction branch based on physical information neural network in The dimensionless mid-span high-frequency live load deflection prediction value of the main beam output at any time Predicted values ​​of high-frequency live load displacement of bridge towers Before calculating the physical loss, both are restored to their original state. The calculation formula is as follows:

[0030] ;

[0031] ;

[0032] in, , They are respectively The variance and mean of the dataset; , They are respectively The variance and mean of the dataset; This is the predicted value of high-frequency live load deflection at mid-span of the main beam after restoration; This is the restored predicted value of high-frequency live load displacement of the bridge tower;

[0033] S43, Substitute the restored predicted values ​​into the physical constraint residual equation to obtain The difference between the time-matter neural network prediction results and the physical laws, i.e., the physical residual. :

[0034] .

[0035] Preferably, step S41 includes the following steps:

[0036] S411, based on the three-moment equation, places the main beam near the mid-span. No. 1 cable anchor point High-frequency live load response components at the location High-frequency live load response components at mid-span of the main girder ratio Perform a polynomial expansion:

[0037] ;

[0038] in, All are polynomial coefficients. It is obtained by training in the physical parameter training layer of the neural network; The distance from the anchorage point of the main beam cable to the mid-span;

[0039] S412, placing the bridge tower near the top No. 1 cable anchor point High-frequency response components at the location High-frequency live load response components at the top of the bridge tower ratio Perform a polynomial expansion:

[0040] ;

[0041] in, All are polynomial coefficients. It is obtained by training in the physical parameter training layer integrated into the neural network; This is the distance from the anchor point of the bridge tower cable to the top of the bridge tower.

[0042] S413, when the main girder is subjected to a live load and deflects downwards at mid-span, the stay cables will pull the bridge towers to a horizontal displacement towards mid-span. Based on the deformation compatibility relationship of the stay cables under small deformation conditions, the following formula can be obtained:

[0043] ;

[0044] Rearranging terms, we obtain the physical coupling equation:

[0045] ;

[0046] in, This is the displacement coupling coefficient (for an existing cable-stayed bridge, within the elastic range, there is a displacement coupling coefficient between the anchorage points of the cables at the bridge towers and the main beams). This is a correction item; , All of these were obtained through training in the physical parameter training layer of the neural network.

[0047] Preferably, in step S5, the loss function of the temperature response prediction branch... The structure is as follows:

[0048] ;

[0049] in, For data sequence number; The total number of data sets; Predicted values ​​of the mid-to-low frequency response of the main beam; The measured values ​​of the mid-to-low frequency response of the main beam; This is the predicted value of the low-frequency response at the top of the bridge tower; This is the measured value of the low-frequency response at the top of the bridge tower.

[0050] Preferably, in step S5, the loss function of the live load response prediction branch... The structure is as follows:

[0051] ;

[0052] ;

[0053] in, The mean square error between the predicted value and the measured high-frequency live load response component; These are the physical constraint weighting coefficients; Predicted high-frequency response value at mid-span of the main beam; This is the predicted value of the high-frequency response at the top of the bridge tower.

[0054] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0055] 1. This invention simultaneously collects the time history data of the horizontal deviation of the bridge tower and the time history data of the mid-span deflection of the main beam, and uses a multi-input multi-output LSTM model for joint modeling. This allows the model to learn the evolution law of each time while mining the temporal correlation and coupling characteristics between the two. Compared with the single-variable prediction method, it can more comprehensively reflect the overall deformation state of the bridge, thereby improving the accuracy of deformation prediction in the future.

[0056] 2. This invention introduces a tower-beam coupling constraint term into the loss function, transforming the structural mechanical relationship of "main beam deflection - bridge tower inward deflection" into a calculable physical constraint. This allows the model to not only fit historical data during training but also simultaneously satisfy the stress and deformation laws of the bridge structure, thus significantly improving the physical consistency and engineering credibility of the prediction results.

[0057] 3. By introducing a structural coupling mechanism, this invention constrains the model training process to the inherent mechanical laws of the bridge, effectively suppressing abnormal prediction results and avoiding large prediction deviations in cable-stayed bridges under conditions such as temperature changes, wind loads, and vehicle loads. This improves the stability of the model under different operating conditions. At the same time, this physical mechanism has strong universality, thus enhancing the model's generalization ability under different bridge types or different monitoring cycles.

[0058] 4. This invention can achieve joint prediction of the key deformation response of a bridge based on only two key measuring points: the horizontal displacement of the bridge tower and the mid-span deflection of the main girder. This can reduce the number of monitoring points while ensuring the prediction effect, thereby reducing the engineering implementation cost and improving the feasibility of practical application. Furthermore, by jointly predicting the bridge tower displacement and mid-span deflection, the operating status of the two key stress-bearing parts, the bridge tower and the main girder, can be evaluated simultaneously, elevating the monitoring from local monitoring to overall structural safety assessment. This provides more reliable technical support for long-term health monitoring and early warning of bridges. Attached Figure Description

[0059] Figure 1 This is a flowchart of the bridge deformation prediction method of the present invention;

[0060] Figure 2 This is a schematic diagram of the LSTM model structure of the present invention. Detailed Implementation

[0061] The present invention will be further described below with reference to the accompanying drawings and embodiments, which illustrate the above and other technical features and advantages of the present invention. However, the following embodiments are merely preferred embodiments of the present invention and are not exhaustive.

[0062] like Figure 1-2 As shown, this invention provides a method for predicting the deformation of cable-stayed bridges that integrates multi-source response decoupling and physical constraints, comprising the following steps:

[0063] S1, Multi-source data acquisition: Through the monitoring system deployed at the top of the bridge tower and the mid-span of the main girder of the cable-stayed bridge, the displacement response time series data of mid-span deflection and bridge tower offset are continuously collected to obtain a multi-source response dataset;

[0064] S2, Data Preprocessing: Process the multi-source response dataset and divide it into training set, validation set and test set;

[0065] S3, Construction of the dual-branch prediction model: Construct a temperature response prediction branch to predict low-frequency deformation response; construct a live load response prediction branch to predict high-frequency deformation response;

[0066] S4, Physical Constraint Construction: Based on the deformation coordination relationship between the tower and beam of a cable-stayed bridge, construct the physical constraint residual equation;

[0067] S5, Joint Model Training: The dual-branch prediction model is trained separately using the training set. The live load response prediction branch constructs a comprehensive loss function by introducing physical constraint residual equations.

[0068] S6, Deformation Prediction and Response Reconstruction: Input the historical sequence data of the time to be predicted into the trained dual-branch prediction model, output the low-frequency prediction value and high-frequency prediction value of the future time respectively, and superimpose the two to reconstruct the final prediction results of bridge tower offset and mid-span deflection.

[0069] In this embodiment, the multi-source response dataset processing steps in step S2 are as follows:

[0070] S21, missing values ​​are imputed in the multi-source response dataset using linear interpolation. Let the dataset be... exist There is always a missing value. The interpolation expression for the missing value is:

[0071] ;

[0072] in, for Data at any given time; for Data at any given time; This refers to the data collection time interval;

[0073] S22, using wavelet transform, performs multi-scale decomposition and reconstruction on the original displacement response time series data after missing value supplementation, decoupling the original displacement response time series data into low-frequency temperature response components and high-frequency live load response components. The low-frequency temperature response component... , and high-frequency live load response components , Each should be standardized separately.

[0074] In this embodiment, the wavelet transform uses the wavelet packet decomposition algorithm. The number of decomposition layers is set according to the natural frequency of the bridge and the period of change of ambient temperature. Approximate components are extracted as low-frequency temperature response components, and detail components are extracted as high-frequency live load response components.

[0075] In this embodiment, in step S22, standardization is performed using Z-score, and the formula is as follows:

[0076] ;

[0077] in, This is the standardized dataset; This is the original dataset; For the original dataset The mean; For the original dataset The variance.

[0078] In this embodiment, in step S3, the dual-branch prediction model includes a temperature response prediction branch and a live load response prediction branch. The long short-term memory network model in the temperature response prediction branch includes an input layer, a two-layer LSTM layer, and an output layer. The LSTM layer dimension is set to 128, the model learning rate is set to 0.001, the weight decay coefficient is set to 0.001, and the batch size is set to 64. The input layer is used to receive the time series data of the mid-span deflection of the main beam and the low-frequency temperature response components of the tower top displacement. The LSTM layer is used to extract the long-term dependency features in the time series. The output layer is used to output the low-frequency temperature response components of the main beam deflection and the tower displacement at the time to be predicted.

[0079] The physical information long short-term memory network model in the live load response prediction branch includes an input layer, a two-layer LSTM layer, a physical parameter training layer, and an output layer. The LSTM layer has a dimension of 128, the physical parameter training layer has a dimension of 64, the model learning rate is 0.001, the weight decay coefficient is 0.001, and the batch size is 64. The input layer receives the time series data of the high-frequency live load response components of the main beam mid-span deflection and tower top offset. The LSTM layer extracts the long-term dependency features from the time series. The physical parameter training layer trains the parameters in the physical constraint equations. The output layer outputs the high-frequency live load response components of the main beam mid-span deflection and tower top offset at the time to be predicted.

[0080] In this embodiment, the specific construction steps in step S4 are as follows:

[0081] S41. Establish the physical coupling equations describing the deformation compatibility relationship between the tower and the beam. The specific steps are as follows:

[0082] S411, based on the three-moment equation, places the main beam near the mid-span. No. 1 cable anchor point High-frequency live load response components at the location High-frequency live load response components at mid-span of the main girder ratio Perform a polynomial expansion:

[0083] ;

[0084] in, All are polynomial coefficients. It is obtained by training in the physical parameter training layer of the neural network; The distance from the anchorage point of the main beam cable to the mid-span;

[0085] S412, placing the bridge tower near the top No. 1 cable anchor point High-frequency response components at the location High-frequency live load response components at the top of the bridge tower ratio Perform a polynomial expansion:

[0086] ;

[0087] in, All are polynomial coefficients. It is obtained by training in the physical parameter training layer integrated into the neural network; This is the distance from the anchor point of the bridge tower cable to the top of the bridge tower.

[0088] S413, when the main girder is subjected to a live load and deflects downwards at mid-span, the stay cables will pull the bridge towers to a horizontal displacement towards mid-span. Based on the deformation compatibility relationship of the stay cables under small deformation conditions, the following formula can be obtained:

[0089] ;

[0090] Rearranging terms, we obtain the physical coupling equation:

[0091] ;

[0092] in, This is the displacement coupling coefficient (for an existing cable-stayed bridge, within the elastic range, there is a displacement coupling coefficient between the anchorage points of the cables at the bridge towers and the main beams). This is a correction item; , All of these were obtained through training in the physical parameter training layer of the neural network.

[0093] S42, the live load response prediction branch based on physical information neural network in The dimensionless mid-span high-frequency live load deflection prediction value of the main beam output at any time Predicted values ​​of high-frequency live load displacement of bridge towers Before calculating the physical loss, both are restored to their original state. The calculation formula is as follows:

[0094] ;

[0095] ;

[0096] in, , They are respectively The variance and mean of the dataset; , They are respectively The variance and mean of the dataset; This is the predicted value of high-frequency live load deflection at mid-span of the main beam after restoration; This is the restored predicted value of high-frequency live load displacement of the bridge tower;

[0097] S43, Substitute the restored predicted values ​​into the physical constraint residual equation to obtain The difference between the time-matter neural network prediction results and the physical laws, i.e., the physical residual. :

[0098] .

[0099] In this embodiment, in step S5, the loss function of the temperature response prediction branch... The structure is as follows:

[0100] ;

[0101] in, For data sequence number; The total number of data sets; Predicted values ​​of the mid-to-low frequency response of the main beam; The measured values ​​of the mid-to-low frequency response of the main beam; This is the predicted value of the low-frequency response at the top of the bridge tower; This is the measured value of the low-frequency response at the top of the bridge tower.

[0102] In this embodiment, in step S5, the loss function of the live load response prediction branch... The structure is as follows:

[0103] ;

[0104] ;

[0105] in, The mean square error between the predicted value and the measured high-frequency live load response component; These are the physical constraint weighting coefficients; Predicted high-frequency response value at mid-span of the main beam; This is the predicted value of the high-frequency response at the top of the bridge tower.

[0106] The above are merely preferred embodiments of the present invention and are illustrative in nature, not restrictive. Those skilled in the art will understand that many changes, modifications, and even equivalents can be made within the spirit and scope defined by the claims of the present invention, all of which will fall within the protection scope of the present invention.

Claims

1. A method for predicting the deformation of cable-stayed bridges that integrates multi-source response decoupling and physical constraints, characterized in that, Includes the following steps: S1, Multi-source data acquisition: Through the monitoring system deployed at the top of the bridge tower and the mid-span of the main girder of the cable-stayed bridge, the displacement response time series data of mid-span deflection and bridge tower offset are continuously collected to obtain a multi-source response dataset; S2, Data Preprocessing: Process the multi-source response dataset and divide it into training set, validation set and test set; S3, Construction of the dual-branch prediction model: Construct a temperature response prediction branch to predict low-frequency deformation response; construct a live load response prediction branch to predict high-frequency deformation response; S4, Physical Constraint Construction: Based on the deformation coordination relationship between the tower and beam of a cable-stayed bridge, construct the physical constraint residual equation; S5, Joint Model Training: The dual-branch prediction model is trained separately using the training set. The live load response prediction branch constructs a comprehensive loss function by introducing physical constraint residual equations. S6, Deformation Prediction and Response Reconstruction: Input the historical sequence data of the time to be predicted into the trained dual-branch prediction model, output the low-frequency prediction value and high-frequency prediction value of the future time respectively, and superimpose the two to reconstruct the final prediction results of bridge tower offset and mid-span deflection.

2. The method for predicting the deformation of a cable-stayed bridge that integrates multi-source response decoupling and physical constraints as described in claim 1, characterized in that, In step S2, the multi-source response dataset processing steps are as follows: S21, missing values ​​are imputed in the multi-source response dataset using linear interpolation. Let the dataset be... exist There is always a missing value. The interpolation expression for the missing value is: ； in, for Data at any given time; for Data at any given time; This refers to the data collection time interval; S22, using wavelet transform, performs multi-scale decomposition and reconstruction on the original displacement response time series data after missing value supplementation, decoupling the original displacement response time series data into low-frequency temperature response components and high-frequency live load response components. The low-frequency temperature response component... , and high-frequency live load response components , Each should be standardized separately.

3. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 2, characterized in that, The wavelet transform employs a wavelet packet decomposition algorithm. The number of decomposition layers is set according to the natural frequency of the bridge and the period of change in ambient temperature. Approximate components are extracted as low-frequency temperature response components, and detail components are extracted as high-frequency live load response components.

4. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 2, characterized in that, In step S22, standardization is performed using Z-score, and the formula is as follows: ； in, This is the standardized dataset; This is the original dataset; For the original dataset The mean; For the original dataset The variance.

5. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 1, characterized in that, In step S3, the long short-term memory network model in the temperature response prediction branch includes an input layer, a two-layer LSTM layer, and an output layer. The LSTM layer dimension is set to 128, the model learning rate is set to 0.001, the weight decay coefficient is set to 0.001, and the batch size is set to 64. The input layer is used to receive the time series data of the mid-span deflection of the main beam and the low-frequency temperature response components of the tower top displacement. The LSTM layer is used to extract the long-term dependency features in the time series. The output layer is used to output the low-frequency temperature response components of the main beam deflection and the tower displacement at the time to be predicted. The physical information long short-term memory network model in the live load response prediction branch includes an input layer, a two-layer LSTM layer, a physical parameter training layer, and an output layer. The LSTM layer has a dimension of 128, the physical parameter training layer has a dimension of 64, the model learning rate is 0.001, the weight decay coefficient is 0.001, and the batch size is 64. The input layer is used to receive time series data of the high-frequency live load response components of the mid-span deflection of the main beam and the tower top offset. The LSTM layer is used to extract long-term dependency features from the time series. The physical parameter training layer is used to train the parameters in the physical constraint equations. The output layer is used to output the high-frequency live load response components of the mid-span deflection of the main beam and the tower top offset at the time to be predicted.

6. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 1, characterized in that, In step S4, the specific steps for constructing the physical constraints are as follows: S41, Establish the physical coupling equations describing the deformation coordination relationship between the tower and the beam; S42, the live load response prediction branch based on physical information neural network in The dimensionless mid-span high-frequency live load deflection prediction value of the main beam output at any time Predicted values ​​of high-frequency live load displacement of bridge towers Before calculating the physical loss, both are restored to their original state. The calculation formula is as follows: ； ； in, , They are respectively The variance and mean of the dataset; , They are respectively The variance and mean of the dataset; This is the predicted value of high-frequency live load deflection at mid-span of the main beam after restoration; This is the restored predicted value of high-frequency live load displacement of the bridge tower; S43, Substitute the restored predicted values ​​into the physical constraint residual equation to obtain The difference between the time-matter neural network prediction results and the physical laws, i.e., the physical residual. : 。 7. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 6, characterized in that, Step S41 includes the following steps: S411, based on the three-moment equation, places the main beam near the mid-span. No. 1 cable anchor point High-frequency live load response components at the location High-frequency live load response components at mid-span of the main girder ratio Perform a polynomial expansion: ； in, All are polynomial coefficients. It is obtained by training in the physical parameter training layer of the neural network; The distance from the anchorage point of the main beam cable to the mid-span; S412, placing the bridge tower near the top No. 1 cable anchor point High-frequency response components at the location High-frequency live load response components at the top of the bridge tower ratio Perform a polynomial expansion: ； in, All are polynomial coefficients. It is obtained by training in the physical parameter training layer integrated into the neural network; This is the distance from the anchor point of the bridge tower cable to the top of the bridge tower. S413, when the main girder is subjected to a live load and deflects downwards at mid-span, the stay cables will pull the bridge towers to a horizontal displacement towards mid-span. Based on the deformation compatibility relationship of the stay cables under small deformation conditions, the following formula can be obtained: ； Rearranging terms, we obtain the physical coupling equation: ； in, The displacement coupling coefficient; This is a correction item; , All of these were obtained through training in the physical parameter training layer of the neural network.

8. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 1, characterized in that, In step S5, the loss function of the temperature response prediction branch The structure is as follows: ； in, For data sequence number; The total number of data sets; Predicted values ​​of the mid-to-low frequency response of the main beam; The measured values ​​of the mid-to-low frequency response of the main beam; This is the predicted value of the low-frequency response at the top of the bridge tower; This is the measured value of the low-frequency response at the top of the bridge tower.

9. The method for predicting the deformation of a cable-stayed bridge by integrating multi-source response decoupling and physical constraints according to claim 1, characterized in that, In step S5, the loss function of the live load response prediction branch The structure is as follows: ； ； in, The mean square error between the predicted value and the measured high-frequency live load response component; These are the physical constraint weighting coefficients; Predicted high-frequency response value at mid-span of the main beam; This is the predicted value of the high-frequency response at the top of the bridge tower.

Images