A rapid prediction method for post-earthquake damage of high-speed railway track bridge systems

Abstract

This invention discloses a rapid post-earthquake damage prediction method for high-speed railway track bridge systems. The method includes acquiring structural and site characteristic data of the target object to determine its site type; constructing a finite element model of the target object; acquiring raw seismic motion data and preprocessing the acquired data; determining the damage state and constructing a "seismic motion-damage state" dataset; using the constructed dataset, training and updating a deep neural network model to construct a deep neural network model for predicting post-earthquake damage state; and, after an actual earthquake, using the constructed deep neural network model for predicting post-earthquake damage state to complete the prediction of the post-earthquake damage state of the high-speed railway track bridge system. This invention adopts a data-driven approach, using the "seismic motion-damage state" dataset to train the deep neural network model, thereby achieving damage state prediction. This invention offers improved efficiency, reduced costs, and convenient operation.

Description

Technical Field

[0001] This invention belongs to the fields of civil engineering and artificial intelligence technology, and specifically relates to a method for rapid prediction of post-earthquake damage to high-speed railway track bridge systems. Background Technology

[0002] To shorten track distances and ensure smooth train operation, high-speed railways utilize a large number of bridges, many of which are located in earthquake-prone areas. Earthquakes can easily damage the high-speed railway track and bridge systems, and the extent of this damage is crucial for post-earthquake emergency response and train safety. Therefore, it is of paramount importance to quickly predict the location and extent of post-earthquake damage to the high-speed railway track and bridge systems.

[0003] Currently, there are generally two methods for predicting structural damage under seismic loading: one is to establish a finite element model based on the structural and site characteristics of the high-speed railway track bridge system, input seismic excitation, obtain the response results, and analyze and process them to obtain the damage state; the other is to rely on monitoring devices to obtain structural vibration response signals, extract damage features through signal identification and processing, thereby achieving the purpose of damage identification.

[0004] The above two methods have the following shortcomings: 1) The post-earthquake finite element analysis process is complex and computationally intensive; 2) It is time-consuming and cannot quickly obtain the damage state, making it difficult to meet the requirements for rapid guidance of post-earthquake emergency response; 3) It requires ensuring the normal operation of the monitoring device, and daily maintenance work is cumbersome; 4) The analysis and processing process is highly dependent on professional knowledge and skills.

[0005] In summary, current methods for predicting post-earthquake damage to high-speed railway track bridge systems cannot quickly and effectively predict the damage status. Summary of the Invention

[0006] The purpose of this invention is to provide a rapid prediction method for post-earthquake damage of high-speed railway track bridge systems that is efficient, cost-effective, and easy to operate.

[0007] The method for rapid prediction of post-earthquake damage to high-speed railway track bridge systems provided by this invention includes the following steps:

[0008] S1. Obtain structural and site characteristic data of the target object to determine the site type of the target object;

[0009] S2. Using the data obtained in step S1, construct a finite element model of the target object;

[0010] S3. Acquire raw ground motion data and perform preprocessing on the acquired data;

[0011] S4. Using the data preprocessed in step S3 and the finite element model constructed in step S2, determine the damage state and construct the "seismic motion-damage state" dataset.

[0012] S5. Using the dataset constructed in step S4, a deep neural network model for predicting post-earthquake damage status is constructed by training and updating the deep neural network model.

[0013] S6. After an actual earthquake occurs, the deep neural network model for predicting post-earthquake damage status constructed in step S5 is used to complete the prediction of post-earthquake damage status of the high-speed railway track bridge system.

[0014] Step S1, which involves obtaining structural and site feature data of the target object and determining the site type of the target object, specifically includes:

[0015] The CRTSII type ballastless track structure high-speed railway simply supported beam bridge is selected as the target object. The high-speed railway track bridge structure includes the bridge structure and the track and structure on the bridge. The bridge structure includes the main beam, bearings, piers, abutments, pile caps, and foundations. The track structure includes the base plate, sliding layer, track slab, CA mortar layer, rails, fasteners, friction plates, shear tooth grooves, shear reinforcement, lateral blocks, and end spikes.

[0016] Obtain the span length, number of spans, total length, foundation depth of the target object, as well as the quantity, spacing, geometric dimensions, material type, and mechanical properties of each structure and component, as structural features of the target object;

[0017] Obtain the geological structure, design characteristic period, and site type of the target object's bridge site as the site characteristics of the target object;

[0018] Step S2, which involves using the data obtained in step S1 to construct a finite element model of the target object, specifically includes:

[0019] Using the data obtained in step S1, define the relevant parameters of the model, including geometric properties, element type, real constants, and material properties. At the same time, create the cross-sectional shape, establish the key elements of each component, divide the nodes, and set the connection and boundary conditions.

[0020] Finite element software was used to establish a finite element model that satisfies the structural and site characteristics of the target object and can accurately simulate the structural response of the target object under seismic motion.

[0021] Step S3 involves acquiring raw ground motion data and preprocessing the acquired data, specifically including:

[0022] (3-1) Obtain the original triaxial ground motion acceleration time history data based on the design acceleration response spectrum of the target object;

[0023] (3-2) Using the raw data obtained in step (3-1), the data is filtered according to the site type, and n seismic waves are retained, specifically including:

[0024] V30 represents the equivalent shear wave velocity within a 30m calculation depth range of the overburden layer. The site types are divided into Class I, Class II, Class III, and Class IV, with corresponding V30 values ​​of above 510m / s, 260m / s to 510m / s, 150m / s to 260m / s, and below 150m / s, respectively.

[0025] Select a seismic wave whose equivalent shear wave velocity of the soil matches the site type of the target object.

[0026] (3-3) Using the seismic waves selected in step (3-2), the peak ground acceleration (PGA) of each seismic wave is compared with the design earthquake peak ground acceleration to obtain the adjustment coefficient of each ground motion. Then, the ground motion data of each ground motion are scaled uniformly according to their respective adjustment coefficients to adjust the energy of the ground motion record to match the fortification intensity of the site where the target object is located, and the preprocessed seismic wave acceleration time history data is obtained.

[0027] Step S4, using the preprocessed data from step S3 and the finite element model constructed in step S2, determines the damage state and constructs a "seismic motion-damage state" dataset, specifically including:

[0028] The n seismic waves retained in step S3 are loaded into the finite element model constructed in step S2, and the corresponding analysis and solution methods are set to perform modal analysis and nonlinear time history analysis on the finite element model to obtain the component response; the damage state is classified according to the damage index.

[0029] Select the damage threshold as the damage index;

[0030] Damage status includes safe, damaged, and failed. When the value of the corresponding index of a component is less than the damage threshold, it is defined as a safe status. When the value of the corresponding index of a component is between the damage threshold and the failure threshold, it is defined as a damaged status. When the value of the corresponding index of a component exceeds the failure threshold, it is defined as a failed status.

[0031] The reference damage threshold values ​​determined for key components of the target object under the design intensity are as follows:

[0032] The component "bridge pier" corresponds to the index "torque," with a critical damage value of 64.5 × 10⁻⁶. 6 N·m, failure threshold value is 79.7×10 6 N·m;

[0033] The component "sliding support" has the corresponding index "displacement", with a damage threshold of 100mm and a failure threshold of 200mm.

[0034] The component "fixed support" corresponds to "displacement", with a damage threshold of 2mm and a failure threshold of 10mm;

[0035] The component "lateral stop" corresponds to "displacement," with a damage threshold of 2mm and a failure threshold of 5mm.

[0036] The component "shear tooth groove" corresponds to "displacement", with a damage threshold of 0.12mm and a failure threshold of 1mm.

[0037] The component "shearing steel bars" corresponds to "displacement," with a damage threshold of 0.08 mm and a failure threshold of 0.7 mm.

[0038] The component "sliding layer" corresponds to "displacement," with a damage threshold of 0.5 mm and a failure threshold of 2 mm.

[0039] The component "CA mortar layer" corresponds to "displacement", with a damage threshold of 0.5mm and a failure threshold of 2mm.

[0040] The component "fastener" corresponds to "displacement," with a damage threshold of 2mm and a failure threshold of 5mm.

[0041] Select n seismic wave data as "features" and use n sets of damage states of each component obtained by finite element model calculation as "labels" to construct a "seismic motion-damage state" dataset;

[0042] Since the CRTS II track system is a longitudinally coupled structure, the impact of transverse earthquakes on the high-speed railway track bridge system is greater than that of longitudinal earthquakes. Therefore, the ground motion was selected with a transverse:vertical ratio of 1:0.65 as the input to the model.

[0043] Step S5, using the dataset constructed in step S4, involves training and updating a deep neural network model to build a deep neural network model for predicting post-earthquake damage status. Specifically, this includes:

[0044] The model of the deep neural network used for the classification task is determined based on the complexity of the dataset constructed in step S4.

[0045] The dataset constructed in step S4 is divided into training dataset, validation dataset and test dataset according to the set ratio;

[0046] We selected a two-layer LSTM recurrent neural network and used the mini-batch gradient descent method to divide the training dataset into multiple subsets for batch training of the model.

[0047] The preprocessed training data is selected as the input to the network model, and feature extraction and classification are performed through the LSTM network.

[0048] A two-layer LSTM recurrent neural network is constructed by stacking two LSTM layers as hidden layers. The first layer contains 32 hidden neurons, and the second layer contains 64 hidden neurons. An LSTM is a neural network containing a number of LSTM cells. Each cell contains a memory unit and three gates: a forget gate, an input gate, and an output gate. These are described by the following formula:

[0049] Input Gate:

[0050] i t =σ(W Xi x t +W hi h t-1 +W ci c t-1 +b i )

[0051] Forgotten Gate:

[0052] f t =σ(W Xf x t +W hf h t-1 +W cf c t-1 +b f )

[0053] Cell state:

[0054] c t =f t c t-1 +i t tanh(W Xc x t +W hc h t-1 +b c )

[0055] Output gate:

[0056] o t =σ(W Xo x t +W ho h t-1 +W co c t-1 +b o )

[0057] Hidden layer:

[0058] h t =o t tanh(c t )

[0059] Where, x th represents the input vector at time t; t c represents the hidden layer at time t; t Indicates the cell state at time t; i t f represents the input gate at time t; t Represents the forget gate at time t; o t W represents the output gate at time t; Xi W represents the weight matrix corresponding to the input vector in the input gate; hi W represents the weight matrix corresponding to the hidden states in the input gate. ci W represents the weight matrix corresponding to the cell states in the input gate. Xf W represents the weight matrix corresponding to the input vector in the forget gate. hf W represents the weight matrix corresponding to the hidden states in the forget gate. cf W represents the weight matrix corresponding to the cell states in the forgetting gate. Xc W represents the weight matrix corresponding to the input vector in the cell state. hc W represents the weight matrix corresponding to the hidden states in the cell state. Xo W represents the weight matrix corresponding to the input vector in the output gate. ho W represents the weight matrix corresponding to the hidden states in the output gate. co b represents the weight matrix corresponding to the cell states in the output gate; i b represents the bias vector of the input gate; f b represents the bias vector of the forget gate; c The bias vector representing the cell state; b o σ represents the bias vector of the output gate; σ(·) represents the activation function sigmoid; tanh(·) represents the activation function tanh.

[0060] Add a Dropout operation between two LSTM layers to prevent the model from overfitting.

[0061] The conditional probability value of each damage category is output using a normalized exponential function (softmax). The label corresponding to the maximum value is the predicted component damage state of the high-speed railway track bridge system. The number of categories is determined based on the selected damage index.

[0062] The principle of the softmax function can be expressed by the following formula:

[0063]

[0064] Among them, y i y j Each represents an element in the vector y;

[0065] The cross-entropy function is used to evaluate the error between the true label of a sample and the model's prediction. The formula for calculating the loss function is shown below:

[0066]

[0067] Where N represents the total number of samples; c represents the category code, which has three types: 1 for safe, 2 for damaged, and 3 for failed; p ic This represents the predicted probability that sample i belongs to category c;

[0068] The Adaptive Moment Estimator (Adam) is selected to update the model parameters. By fitting the training set data, the convergence of the model is accelerated, memory requirements are reduced, and the value of the loss function L is continuously decreased. After each iteration, the model is evaluated using a validation dataset. When the number of iterations reaches a set limit, the model training ends, and all evaluation results are selected. The model with the best result is selected as the chosen two-layer LSTM network model. The selected two-layer LSTM network model is evaluated using a test dataset to assess its generalization ability. The final evaluated two-layer LSTM network model is retained as a deep neural network model for predicting post-earthquake damage status.

[0069] Step S6 describes the process of using the deep neural network model for predicting post-earthquake damage status, constructed in step S5, to predict the post-earthquake damage status of the high-speed railway track bridge system after an actual earthquake. Specifically, this includes:

[0070] (6-1) After the earthquake, the seismic waves collected by the stations in the area where the target object is located are obtained. Step S3 is used to preprocess the seismic wave data except for site type selection.

[0071] (6-2) The post-earthquake damage state prediction deep neural network model constructed in step S5 is used to obtain the post-earthquake damage state of the target object under the seismic wave.

[0072] The method provided by this invention for rapid prediction of post-earthquake damage in high-speed railway track bridge systems adopts a data-driven approach, using a "seismic motion-damage state" dataset to train a deep neural network model, thereby predicting the damage state. This invention offers improved efficiency, reduced costs, and ease of operation. Attached Figure Description

[0073] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0074] Figure 2 This is a schematic diagram of the finite element model of a CRTSII type slab trackless track structure high-speed railway simply supported beam bridge in the method of this invention.

[0075] Figure 3 This is a schematic diagram of the process for selecting and preprocessing seismic waves in the method of the present invention.

[0076] Figure 4 This is a schematic diagram of the two-layer LSTM recurrent neural network structure in the method of the present invention.

[0077] Figure 5 This is a schematic diagram of the LSTM cell structure in the method of the present invention. Detailed Implementation

[0078] like Figure 1 The diagram shown is a flowchart of the method of the present invention. The method of the present invention is particularly applicable to CRTSII type ballastless track structure high-speed railway simply supported beam bridges. The rapid prediction method for post-earthquake damage of high-speed railway track bridge systems provided by the present invention includes the following steps:

[0079] S1. Obtain structural and site characteristic data of the target object to determine its site type; specifically including:

[0080] The method of this invention selects a Windows 10 system as the software environment, combines the MATLAB platform and ANSYS software, and uses a 3.90GHz Intel Core i5-11600K processor, an NVIDIA GeForce RTX3080 Ti graphics card and 32GB of RAM;

[0081] like Figure 2 The diagram shows a finite element model of a CRTSII type ballastless track structure high-speed railway simply supported beam bridge. The CRTSII type ballastless track structure high-speed railway simply supported beam bridge is selected as the target object. The high-speed railway track bridge structure includes the bridge structure and the track and structure on the bridge. The bridge structure includes the main beam, bearings, piers, abutments, pile caps, and foundations. The track structure includes the base plate, sliding layer, track slab, CA mortar layer, rails, fasteners, friction plates, shear grooves, shear reinforcement, lateral blocks, and end spikes.

[0082] Obtain the span length, number of spans, total length, foundation depth of the target object, as well as the quantity, spacing, geometric dimensions, material type, and mechanical properties of each structure and component, as structural features of the target object;

[0083] Obtain the geological structure, design characteristic period, and site type of the target object's bridge site as the site characteristics of the target object;

[0084] S2. Using the data obtained in step S1, construct a finite element model of the target object; specifically including:

[0085] Using the data obtained in step S1, define the relevant parameters of the model, including geometric properties, element type, real constants, and material properties. At the same time, create the cross-sectional shape, establish the key elements of each component, divide the nodes, and set the connection and boundary conditions.

[0086] Finite element software was used to establish a finite element model that satisfies the structural and site characteristics of the target object and can accurately simulate the structural response of the target object under seismic motion.

[0087] The finite element method selected for this invention is ANSYS 19.2;

[0088] S3. Acquire raw ground motion data and preprocess the acquired data; specifically including:

[0089] like Figure 3 The diagram shows the process flow for selecting and preprocessing seismic waves.

[0090] (3-1) Using the MATLAB programming language, the original triaxial ground motion acceleration time history data are obtained based on the design acceleration response spectrum of the target object;

[0091] The raw seismic wave data obtained in the method of this invention comes from the PEER NGA database in the United States;

[0092] (3-2) Using the raw data obtained in step (3-1), the data is filtered according to the site type, and n seismic waves are retained, specifically including:

[0093] V30 represents the equivalent shear wave velocity within a 30m calculation depth range of the overburden layer. The site types are divided into Class I, Class II, Class III, and Class IV, with corresponding V30 values ​​of above 510m / s, 260m / s to 510m / s, 150m / s to 260m / s, and below 150m / s, respectively.

[0094] Select a seismic wave whose equivalent shear wave velocity of the soil matches the site type of the target object.

[0095] (3-3) Using the seismic waves selected in step (3-2), the peak ground acceleration (PGA) of each seismic wave is compared with the designed peak ground acceleration to obtain the adjustment coefficient of each ground motion. Then, the ground motion data of each ground motion are scaled uniformly according to their respective adjustment coefficients to adjust the energy of the ground motion record to match the fortification intensity of the site where the target object is located, and the preprocessed seismic wave acceleration time history data is obtained.

[0096] S4. Using the preprocessed data from step S3 and the finite element model constructed in step S2, determine the damage state and construct a "seismic motion-damage state" dataset; specifically including:

[0097] The n seismic waves retained in step S3 are loaded into the finite element model constructed in step S2, and the corresponding analysis and solution methods are set to perform modal analysis and nonlinear time history analysis on the finite element model to obtain the component response; the damage state is classified according to the damage index.

[0098] Select the damage threshold as the damage index;

[0099] Damage status includes safe, damaged, and failed. When the value of the corresponding index of a component is less than the damage threshold, it is defined as a safe status. When the value of the corresponding index of a component is between the damage threshold and the failure threshold, it is defined as a damaged status. When the value of the corresponding index of a component exceeds the failure threshold, it is defined as a failed status.

[0100] The reference damage threshold values ​​determined for key components of the target object under the design intensity are as follows:

[0101] The component "bridge pier" corresponds to the index "torque," with a critical damage value of 64.5 × 10⁻⁶. 6 N·m, failure threshold value is 79.7×10 6 N·m;

[0102] The component "sliding support" has the corresponding index "displacement", with a damage threshold of 100mm and a failure threshold of 200mm.

[0103] The component "fixed support" corresponds to "displacement", with a damage threshold of 2mm and a failure threshold of 10mm;

[0104] The component "lateral stop" corresponds to "displacement," with a damage threshold of 2mm and a failure threshold of 5mm.

[0105] The component "shear tooth groove" corresponds to "displacement", with a damage threshold of 0.12mm and a failure threshold of 1mm.

[0106] The component "shearing steel bars" corresponds to "displacement," with a damage threshold of 0.08 mm and a failure threshold of 0.7 mm.

[0107] The component "sliding layer" corresponds to "displacement," with a damage threshold of 0.5 mm and a failure threshold of 2 mm.

[0108] The component "CA mortar layer" corresponds to "displacement", with a damage threshold of 0.5mm and a failure threshold of 2mm.

[0109] The component "fastener" corresponds to "displacement," with a damage threshold of 2mm and a failure threshold of 5mm.

[0110] Select n seismic wave data as "features" and use n sets of damage states of each component obtained by finite element model calculation as "labels" to construct a "seismic motion-damage state" dataset;

[0111] Since the CRTS II track system is a longitudinally coupled structure, the impact of transverse earthquakes on the high-speed railway track bridge system is greater than that of longitudinal earthquakes. Therefore, the ground motion was selected with a transverse:vertical ratio of 1:0.65 as the input to the model.

[0112] S5. Using the dataset constructed in step S4, a deep neural network model for predicting post-earthquake damage status is built by training and updating the deep neural network model; specifically including:

[0113] The model of the deep neural network used for the classification task is determined based on the complexity of the dataset constructed in step S4.

[0114] The dataset constructed in step S4 is divided into training dataset, validation dataset and test dataset according to the set ratio;

[0115] In the method of this invention, 80% of the data in the dataset is selected to construct the training dataset, 10% of the data is selected to construct the validation dataset, and 10% of the data is selected to construct the test dataset.

[0116] We selected a two-layer LSTM recurrent neural network and used the mini-batch gradient descent method to divide the training dataset into multiple subsets for batch training of the model.

[0117] In the method of this invention, the mini-batch gradient descent method used has a batch size of 32, a training iteration count of 200 epochs, and an initial learning rate of LR of 0.0067.

[0118] The preprocessed training data is selected as the input to the network model, and feature extraction and classification are performed through the LSTM network.

[0119] like Figure 4 The diagram shown is a schematic of a two-layer LSTM recurrent neural network structure.

[0120] A two-layer LSTM recurrent neural network stacks two LSTM layers as hidden layers. The first layer contains 32 hidden neurons, and the second layer contains 64 hidden neurons. An LSTM is a neural network containing several LSTM cells. Each cell contains a memory unit and three gates: a forget gate, an input gate, and an output gate. Figure 5 The diagram shows the LSTM cell structure; it is described using the following formula:

[0121] Input Gate:

[0122] i t =σ(WXi x t +W hi h t-1 +W ci c t-1 +b i )

[0123] Forgotten Gate:

[0124] f t =σ(W Xf x t +W hf h t-1 +W cf c t-1 +b f )

[0125] Cell state:

[0126] c t =f t c t-1 +i t tanh(W Xc x t +W hc h t-1 +b c )

[0127] Output gate:

[0128] o t =σ(W Xo x t +W ho h t-1 +W co c t-1 +b o )

[0129] Hidden layer:

[0130] h t =o t tanh(c t )

[0131] Where, x t h represents the input vector at time t; t c represents the hidden layer at time t; t Indicates the cell state at time t; i t f represents the input gate at time t; t Represents the forget gate at time t; o t W represents the output gate at time t; Xi W represents the weight matrix corresponding to the input vector in the input gate; hi W represents the weight matrix corresponding to the hidden states in the input gate. ciW represents the weight matrix corresponding to the cell states in the input gate. Xf W represents the weight matrix corresponding to the input vector in the forget gate. hf W represents the weight matrix corresponding to the hidden states in the forget gate. cf W represents the weight matrix corresponding to the cell states in the forgetting gate. Xc W represents the weight matrix corresponding to the input vector in the cell state. hc W represents the weight matrix corresponding to the hidden states in the cell state. Xo W represents the weight matrix corresponding to the input vector in the output gate. ho W represents the weight matrix corresponding to the hidden states in the output gate. co b represents the weight matrix corresponding to the cell states in the output gate; i b represents the bias vector of the input gate; f b represents the bias vector of the forget gate; c The bias vector representing the cell state; b o σ represents the bias vector of the output gate; σ(·) represents the activation function sigmoid; tanh(·) represents the activation function tanh.

[0132] Add a Dropout operation between two LSTM layers to prevent the model from overfitting.

[0133] In the method of this invention, a Dropout rate of 0.5 is selected;

[0134] The conditional probability value of each damage category is output using a normalized exponential function (softmax). The label corresponding to the maximum value is the predicted component damage state of the high-speed railway track bridge system. The number of categories is determined based on the selected damage index.

[0135] The principle of the softmax function can be expressed by the following formula:

[0136]

[0137] Among them, y i y j Each represents an element in the vector y;

[0138] The cross-entropy function is used to evaluate the error between the true label of a sample and the model's prediction. The formula for calculating the loss function is shown below:

[0139]

[0140] Where N represents the total number of samples; c represents the category code, which has three types: 1 for safe, 2 for damaged, and 3 for failed; p icThis represents the predicted probability that sample i belongs to category c;

[0141] The Adaptive Moment Estimator (Adam) is selected to update the model parameters. By fitting the training set data, the convergence of the model is accelerated, memory requirements are reduced, and the value of the loss function L is continuously decreased. After each iteration, the model is evaluated using a validation dataset. When the number of iterations reaches a set limit, the model training ends, and all evaluation results are selected. The model with the best result is selected as the chosen two-layer LSTM network model. The selected two-layer LSTM network model is evaluated using a test dataset to assess its generalization ability. The final evaluated two-layer LSTM network model is retained as a deep neural network model for predicting post-earthquake damage status.

[0142] S6. After an actual earthquake, the deep neural network model for predicting post-earthquake damage status, constructed in step S5, is used to predict the post-earthquake damage status of the high-speed railway track bridge system; specifically including:

[0143] (6-1) After the earthquake, the seismic waves collected by the stations in the area where the target object is located are obtained. Step S3 is used to preprocess the seismic wave data except for site type selection.

[0144] (6-2) The post-earthquake damage state prediction deep neural network model constructed in step S5 is used to obtain the post-earthquake damage state of the target object under the seismic wave.

Claims

1. A method for rapid prediction of post-earthquake damage to high-speed railway track bridge systems, comprising the following steps: S1. Obtain structural and site characteristic data of the target object to determine the site type of the target object; S2. Using the data obtained in step S1, construct a finite element model of the target object; S3. Acquire raw ground motion data and perform preprocessing on the acquired data; specifically including: (3-1) Obtain the original triaxial ground motion acceleration time history data based on the design acceleration response spectrum of the target object; (3-2) Using the raw data obtained in step (3-1), filter and process the data according to the site type, retaining... The seismic waves specifically include: V30 represents the equivalent shear wave velocity over a 30m calculation depth of the overburden layer. Site types are categorized into Class I, Class II, Class III, and Class IV, with corresponding V30 value ranges as follows: above, to , to , the following; Select a seismic wave whose equivalent shear wave velocity of the soil matches the site type of the target object. (3-3) Using the seismic waves selected in step (3-2), the peak ground acceleration of each seismic wave is compared with the designed peak ground acceleration to obtain the adjustment coefficient of each ground motion. Then, the ground motion data of each ground motion are uniformly scaled according to the corresponding adjustment coefficient, so as to adjust the energy of the ground motion record to match the fortification intensity of the site where the target object is located, and obtain the preprocessed seismic wave acceleration time history data. S4. Using the preprocessed data from step S3 and the finite element model constructed in step S2, determine the damage state and construct a "seismic motion-damage state" dataset; specifically including the following steps: The data retained in step S3 Each seismic wave is loaded into the finite element model constructed in step S2, and the corresponding analysis and solution methods are set to perform modal analysis and nonlinear time history analysis on the finite element model to obtain the component response; the damage state is classified according to the damage index. Select the damage threshold as the damage index; Damage status includes safe, damaged, and failed. When the value of the corresponding index of a component is less than the damage threshold, it is defined as a safe status. When the value of the corresponding index of a component is between the damage threshold and the failure threshold, it is defined as a damaged status. When the value of the corresponding index of a component exceeds the failure threshold, it is defined as a failed status. S5. Using the dataset constructed in step S4, a deep neural network model for predicting post-earthquake damage status is constructed by training and updating the deep neural network model. S6. After an actual earthquake occurs, the deep neural network model for predicting post-earthquake damage status constructed in step S5 is used to complete the prediction of post-earthquake damage status of the high-speed railway track bridge system.

2. The method for rapid prediction of post-earthquake damage to a high-speed railway track bridge system according to claim 1, characterized in that... Step S1, which involves obtaining structural and site feature data of the target object and determining the site type of the target object, specifically includes: The CRTSII type ballastless track structure high-speed railway simply supported beam bridge is selected as the target object. The high-speed railway track bridge structure includes the bridge structure and the track and structure on the bridge. The bridge structure includes the main beam, bearings, piers, abutments, pile caps, and foundations. The track structure includes the base plate, sliding layer, track slab, CA mortar layer, rails, fasteners, friction plates, shear tooth grooves, shear reinforcement, lateral blocks, and end spikes. Obtain the span length, number of spans, total length, foundation depth of the target object, as well as the quantity, spacing, geometric dimensions, material type, and mechanical properties of each structure and component, as structural features of the target object; Obtain the geological structure, design characteristic period, and site type of the target object's bridge site as the site characteristics of the target object.

3. The method for rapid prediction of post-earthquake damage to a high-speed railway track bridge system according to claim 2, characterized in that... Step S2, which involves using the data obtained in step S1 to construct a finite element model of the target object, specifically includes: Using the data obtained in step S1, define the relevant parameters of the model, including geometric properties, element type, real constants, and material properties. At the same time, create the cross-sectional shape, establish the key elements of each component, divide the nodes, and set the connection and boundary conditions. Finite element software is used to establish a finite element model that satisfies the structural and site characteristics of the target object and can accurately simulate the structural response of the target object under seismic motion.

4. The method for rapid prediction of post-earthquake damage to a high-speed railway track bridge system according to claim 3, characterized in that... Step S4, using the preprocessed data from step S3 and the finite element model constructed in step S2, determines the damage state and constructs a "seismic motion-damage state" dataset, specifically including: The reference damage threshold values ​​determined for key components of the target object under the design intensity are as follows: The component "bridge pier" corresponds to the index "torque," and the critical value for damage is... The failure threshold value is ; The component "sliding support" has the corresponding index "displacement", with a damage threshold of 100mm and a failure threshold of 200mm. The component "fixed support" corresponds to "displacement", with a damage threshold of 2mm and a failure threshold of 10mm. The component "lateral stop" corresponds to "displacement," with a damage threshold of 2mm and a failure threshold of 5mm. The component "shear tooth groove" corresponds to "displacement", with a damage threshold of 0.12mm and a failure threshold of 1mm. The component "shearing steel bars" corresponds to "displacement," with a damage threshold of 0.08 mm and a failure threshold of 0.7 mm. The component "sliding layer" corresponds to "displacement," with a damage threshold of 0.5 mm and a failure threshold of 2 mm. The component "CA mortar layer" corresponds to "displacement", with a damage threshold of 0.5mm and a failure threshold of 2mm. The component "fastener" corresponds to "displacement," with a damage threshold of 2mm and a failure threshold of 5mm. choose Seismic wave data, used as "features," were obtained through calculations using a finite element model. The damage status of each component is used as a "label" to construct a "seismic motion-damage status" dataset; Because the CRTS II track system is a longitudinally coupled structure, lateral earthquakes have a greater impact on the high-speed railway track and bridge system than longitudinal earthquakes. Therefore, the selected ground motion is based on a lateral:vertical ratio. , as input to the model.

5. A method for rapid prediction of post-earthquake damage to a high-speed railway track bridge system according to claim 4, characterized in that... Step S5, using the dataset constructed in step S4, involves training and updating a deep neural network model to build a deep neural network model for predicting post-earthquake damage status. Specifically, this includes: The model of the deep neural network used for the classification task is determined based on the complexity of the dataset constructed in step S4. The dataset constructed in step S4 is divided into training dataset, validation dataset and test dataset according to the set ratio; We selected a two-layer LSTM recurrent neural network and used the mini-batch gradient descent method to divide the training dataset into multiple subsets for batch training of the model. The preprocessed training data is selected as the input to the network model, and feature extraction and classification are performed through the LSTM network. A two-layer LSTM recurrent neural network is constructed by stacking two LSTM layers as hidden layers. The first layer contains 32 hidden neurons, and the second layer contains 64 hidden neurons. An LSTM is a neural network containing a number of LSTM cells. Each cell contains a memory unit and three gates: a forget gate, an input gate, and an output gate. These are described by the following formula: Input Gate: Forgotten Gate: Cell state: Output gate: Hidden layer: in, express The input vector at time step; express Hidden layers of time; express Cellular state at any given moment; express The time input gate; express The Gate of Forgetting Time; express The output gate at any given time; This represents the weight matrix corresponding to the input vector in the input gate; This represents the weight matrix corresponding to the hidden states in the input gate; This represents the weight matrix corresponding to the cell states in the input gate; This represents the weight matrix corresponding to the input vector in the forget gate; This represents the weight matrix corresponding to the hidden states in the forget gate; This represents the weight matrix corresponding to the cell states in the forgetting gate; This represents the weight matrix corresponding to the input vector in the cell state; This represents the weight matrix corresponding to the hidden states in the cell state; This represents the weight matrix corresponding to the input vector in the output gate; This represents the weight matrix corresponding to the hidden states in the output gate; This represents the weight matrix corresponding to the cell states in the output gate; This represents the bias vector of the input gate; The bias vector representing the forget gate; A bias vector representing the cell state; This represents the bias vector of the output gate; This represents the activation function sigmoid; This represents the activation function tanh; Add a Dropout operation between two LSTM layers to prevent the model from overfitting. The conditional probability value of each damage category is output using a normalized exponential function, and the label corresponding to the maximum value is the predicted component damage state of the high-speed railway track bridge system; the number of categories is determined according to the selected damage index. The principle of the softmax function can be expressed by the following formula: in, , Both represent vectors One of the elements; The classification cross-entropy is used as the loss function to evaluate the error between the true label of the sample and the model's prediction. The formula for calculating the loss function is shown below: in, Indicates the total sample size; This indicates the category code, which is divided into three types: 1 for safety, 2 for damage, and 3 for failure. Indicates sample Category The predicted probability; Choosing an adaptive moment estimator optimizer to update the model parameters accelerates model convergence, reduces memory requirements, and improves the loss function by fitting the training set data. The value of is continuously reduced; after each iteration, the model is evaluated using a validation dataset; when the number of iterations reaches a set limit, the model ends training, and all evaluation results are selected, with the model with the best result being selected as the chosen two-layer LSTM network model; the chosen two-layer LSTM network model is evaluated using a test dataset to assess its generalization ability; the final evaluated two-layer LSTM network model is retained as a deep neural network model for predicting post-earthquake damage status, and is used for predicting post-earthquake damage status.

6. The method for rapid prediction of post-earthquake damage to a high-speed railway track bridge system according to claim 5, characterized in that... Step S6 describes the process of using the deep neural network model for predicting post-earthquake damage status, constructed in step S5, to predict the post-earthquake damage status of the high-speed railway track bridge system after an actual earthquake. Specifically, this includes: (6-1) After the earthquake, the seismic waves collected by the stations in the area where the target object is located are obtained. Step S3 is used to preprocess the seismic wave data except for site type selection. (6-2) The post-earthquake damage state prediction deep neural network model constructed in step S5 is used to obtain the post-earthquake damage state of the target object under the seismic wave.

Images