Bridge operation safety early warning method based on windmill bridge vortex vibration coupling system

Abstract

The application relates to the technical field of bridge monitoring, in particular to a bridge operation safety early warning method based on a windmill bridge vortex vibration coupling system, which comprises the following steps: step 1: obtaining static samples of a wind-train-bridge-vortex vibration coupling system, and generating a static sample library; wherein the wind-train-bridge-vortex vibration coupling system is a simulation model of a target bridge, which is used for simulating index data under the action of a train at a specified speed and corresponding wind field; the static samples comprise field state data and index data; the method can comprehensively capture the system state under the wind-car-bridge-vortex vibration coupling effect, and avoid risk omissions and misjudgments caused by insufficient single-source index information.

Description

Technical Field

[0001] This application relates to the field of bridge monitoring technology, and more specifically, to a bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] Bridge early warning systems play a crucial role in the operation of long-span railway bridges. Due to the coupling effects of train loads, environmental wind loads, and structural natural vibrations, the bridge structural response, wheel-rail contact forces, and train operation safety and comfort exhibit nonlinear and time-varying characteristics. Vibration phenomena under specific wind fields, such as vortex-induced vibration, further increase the complexity of the bridge's dynamic behavior. Through real-time monitoring and early warning, abnormal responses can be detected promptly, preventing safety accidents such as derailment, structural fatigue, or damage.

[0004] While traditional numerical methods based on wind-train-bridge coupling analysis can accurately simulate vibration processes, existing bridge early warning methods suffer from high model dimensionality, long solution times, and strong parameter dependencies, making it difficult to meet the real-time requirements of online early warning. To address this, some solutions incorporate deep learning technology, modeling bridge response prediction as a time-series regression problem. A neural network model is trained using historical monitoring data (such as displacement and acceleration) and external load data (such as wind speed and train axle load) to achieve rapid prediction of key responses. This method, acting as a surrogate model for high-fidelity numerical models, significantly improves computational efficiency and has potential applications in real-time early warning scenarios.

[0005] However, deep learning prediction models face the problem of inconsistent distribution between simulation samples and actual bridge monitoring samples in practical engineering applications. Data collected by bridge monitoring systems often suffers from defects such as noise, missing measurements, and uneven distribution of vortex-induced vibration development risks. Meanwhile, the offline simulation data relied upon for model training is based on idealized numerical models and load assumptions, leading to discrepancies in their statistical distribution characteristics. If a purely offline-trained black-box model is directly used for online early warning, the prediction results will exhibit systematic errors due to this distributional shift, resulting in increased false alarm or missed alarm rates in the early warning system. Furthermore, it becomes difficult to guarantee sufficient early warning lead time, thus weakening the reliability and practicality of the early warning system. Summary of the Invention

[0006] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this application propose a bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system to solve the technical problems mentioned in the background section above.

[0008] As a first aspect of this application, some embodiments of this application provide a bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system, including the following steps: Step 1: Obtain static samples of the wind-train-bridge-vortex-induced vibration coupled system and generate a static sample library; Among them, the wind-train-bridge-vortex-induced vibration coupled system is the simulation model of the target bridge, used to simulate the index data of the train under the action of a specified speed and corresponding wind field; the static sample includes field state data and index data; Wind tunnel tests were conducted on the target bridge to construct a database of vortex-induced vibration state parameters for querying vortex-induced vibration state parameters. Step 2: Obtain real-time field data of the target bridge, and preprocess the real-time field data to obtain standard data; Step 3: Convert the standard data into coupled state feature vectors, and stack multiple coupled state feature vectors within the historical time window in chronological order to form a time series input matrix; Step 4: Input the time series input matrix into the physical constraint deep learning prediction model to obtain the target response results within the future prediction time window; The target response results include predicted response feature vectors of key bridge points, predicted vortex-induced vibration state label data, and predicted train operation comfort indices. Step 5: Input the target response result into the residual compensation network for compensation to obtain the real-time response result, and use the real-time response result as the prediction index data; The index data includes the response feature vector of the predicted bridge key points after error compensation, the label data of the predicted vortex-induced vibration state after error compensation, and the index of predicted train running comfort. Step 6: Calculate risk indicators and generate early warning results based on the predicted indicator data, and at the same time build a real sample library for continuous model optimization.

[0009] The technical solution of this application embodiment has at least the following advantages and beneficial effects: (1) By incorporating wind field input data, train operation status data and measured feature vectors of key bridge points into the coupled state feature vector for unified modeling, this method can comprehensively capture the system state under the coupled action of wind-vehicle-bridge-vortex vibration, compared with traditional bridge monitoring methods based on a single data source (such as wind load or vibration response only), thus avoiding risk omissions and misjudgments caused by insufficient information from a single source.

[0010] (2) By first generating a static sample library from the simulation of the wind-train-bridge-vortex-induced vibration coupling system for pre-training, and then constructing an actual sample library from the measured response and early warning results continuously collected during online operation for continuous optimization, the problem of scarce measured samples and difficulty in directly training deep models under extreme working conditions of long-span bridges is solved. It also overcomes the objective deviation between pure simulation models and actual bridges, so that the prediction model has both good generalization ability and on-site accuracy.

[0011] (3) To address the problem that the wheel-rail contact state parameters cannot be directly collected by the sensor under actual working conditions, a field state data matching method based on cosine similarity is adopted to extract and fill the parameter from the static sample library. Compared with simply ignoring the parameter or using a fixed default value, this method can maintain the integrity and representativeness of the coupled state feature vector and improve the accuracy of subsequent predictions.

[0012] (4) Physically constrained deep learning prediction models introduce physical constraints such as bridge structure dynamic balance, vortex-induced vibration locking frequency consistency, aerodynamic-displacement phase relationship and vortex-induced vibration state evolution law while driving data learning. This avoids prediction results that violate the physical laws of bridge dynamics from pure data-driven deep learning models. Compared with traditional neural network methods, it has higher reliability and interpretability.

[0013] (5) By using the residual compensation network to correct the output of the deep learning prediction model in real time, it can dynamically compensate for the systematic deviation between the simulation model and the actual bridge operation. Furthermore, by using cosine similarity matching to determine the error benchmark and using a sliding time window to smooth the prediction error, it further suppresses the influence of observation noise on the correction process. Compared with the fixed threshold method or a single deep learning model, it significantly improves the accuracy and stability of online prediction.

[0014] (6) By integrating the three sub-indicators of bridge structure response risk, train operation safety and comfort risk and vortex-induced vibration development risk into a comprehensive risk index according to their weights, compared with the traditional single-parameter threshold alarm method, it can comprehensively and systematically reflect the bridge operation status and avoid the problem of ignoring other risk dimensions due to changes in the dominant item of a single index.

[0015] (7) By setting four levels of thresholds for comprehensive risk indicators, namely normal monitoring, warning, restriction warning and danger warning, the system provides clear decision-making basis for the operating unit. Compared with the binary judgment method of "whether to alarm", it can finely match different levels of emergency response needs such as speed limit operation, encrypted monitoring and temporary blockade. Attached Figure Description

[0016] Figure 1 This is a flowchart of a bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system.

[0017] Figure 2 This is a simulation diagram of the SIMPACK software.

[0018] Figure 3 The variation of vehicle acceleration with respect to VIV amplitude at different speeds is shown.

[0019] Figure 4 The vortex-induced vibration test was conducted on the main beam segment model.

[0020] Figure 5 The results show the vertical vortex-induced vibration of the main beam. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0022] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0023] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0024] refer to Figure 1 Example 1: The general technical concept of this application is as follows: Bridge operational safety early warning typically employs two methods: simulation and neural network prediction. Simulation methods use software to simulate trains crossing bridges at different speeds and under varying wind conditions, and then assess safety based on the simulation results. However, simulation results are based on virtual computer calculations and may not accurately reflect actual conditions, potentially leading to significant discrepancies.

[0025] Neural network models use current data on trains, bridges, and wind fields to predict the future vibration state of trains and bridges. Their prediction accuracy is highly dependent on the quantity and realism of the training samples. Before a bridge is put into operation, it is difficult to obtain a sufficient number of real samples; if samples from other bridges are used to train the model and then directly applied to the current bridge, accuracy is also difficult to guarantee.

[0026] Therefore, this application first uses a simulation model to generate sufficient static samples for initial training of the neural network model. During bridge operation, real data is continuously collected, and after accumulating a sufficient number of real samples, the neural network model is further trained. In this way, in the early stages of bridge operation, the model trained based on simulation samples can provide basic prediction accuracy; as operation progresses and real data becomes increasingly abundant, the prediction accuracy of the neural network model will gradually improve.

[0027] A bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system includes the following steps: Step 1: Obtain static samples of the wind-train-bridge-vortex-induced vibration coupled system and generate a static sample library; Among them, the wind-train-bridge-vortex-induced vibration coupled system is the simulation model of the target bridge, used to simulate the index data of the train under the action of a specified speed and corresponding wind field; the static sample includes field state data and index data; Wind tunnel tests were conducted on the target bridge to construct a database of vortex-induced vibration state parameters for querying these parameters.

[0028] The static sample library generates a large number of "field state data + index data" samples through the simulation of the wind-train-bridge-vortex-induced vibration coupled system. This provides sufficient offline training data for the physical constraint deep learning prediction model and residual compensation network, solving the problem of the lack of measured samples in the early stage of bridge operation and enabling the model to have basic prediction capabilities.

[0029] Vortex-induced vibration state parameter library: Physical parameters related to vortex-induced vibration (such as locked frequency range, aerodynamic-displacement phase difference, equivalent aerodynamic characterization quantity, etc.) are obtained through wind tunnel tests of target bridges. These parameters are used for physical constraints during training (main frequency consistency, phase relationship, state evolution, etc.) and for querying and judging the vortex-induced vibration state during online inference, ensuring that the model prediction conforms to the actual vibration law.

[0030] Step 2: Obtain real-time field data of the target bridge, and preprocess the real-time field data to obtain standard data; Step 3: Convert the standard data into coupled state feature vectors, and stack multiple coupled state feature vectors within the historical time window in chronological order to form a time series input matrix; Step 3 converts the standard data into a structured format suitable for deep learning models to process.

[0031] Bridge vibration and vortex-induced vibration development exhibit significant time dependence and historical memory effects (such as the inertia of displacement, velocity, and acceleration, and the locking process of vortex-induced vibration). Relying solely on instantaneous data at the current moment is insufficient to accurately predict future responses. The temporal input matrix provides historical context for subsequent temporal feature extraction sub-networks (such as LSTM), enabling the model to learn the evolution of the state over time, thereby improving the prediction accuracy of bridge key point responses, vortex-induced vibration states, and train comfort indicators.

[0032] Step 4: Input the time series input matrix into the physical constraint deep learning prediction model to obtain the target response results within the future prediction time window; The target response results include predicted response feature vectors of key bridge points, predicted vortex-induced vibration state label data, and predicted train operation comfort indices. Step 5: Input the target response result into the residual compensation network for compensation to obtain the real-time response result, and use the real-time response result as the prediction index data; The predictive index data includes the response feature vector of the predicted bridge key points after error compensation, the predicted vortex-induced vibration state label data after error compensation, and the predicted train running comfort index. Step 5 uses a residual compensation network to correct the target response results output by the physical constraint deep learning prediction model online, so as to eliminate the distribution deviation between the simulation samples and the measured data, obtain real-time response results that are closer to the real situation, and use them as indicator data for subsequent risk calculation.

[0033] Step 6: Calculate risk indicators and generate early warning results based on the predicted indicator data, and at the same time build a real sample library for continuous model optimization.

[0034] The predicted index data is the target response result after residual compensation, and the target response result reflects the vibration status of the bridge. Therefore, the predicted index data is the vibration characteristics of the bridge after residual compensation, from which the comprehensive risk index can be calculated.

[0035] The above outlines the general steps and functions of a bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system. Specific implementation details for each step are as follows: Step 1: Obtain static samples of the wind-train-bridge-vortex-induced vibration coupled system and generate a static sample library; Among them, the wind-train-bridge-vortex-induced vibration coupling system is the simulation model of the target bridge, used to simulate the index data of the train under the action of a specified speed and corresponding wind field; the static sample includes field state data and index data.

[0036] The field data includes wind field input data, train operation status data, and measured feature vectors of key bridge points.

[0037] Wind field input data includes average wind speed, wind direction angle, wind angle of attack, and turbulence intensity; Train operation status data includes train speed, car body acceleration, train formation, and wheel-rail contact parameters.

[0038] Wheel-rail contact parameters include derailment coefficient, wheel load reduction rate, and wheel-rail contact force.

[0039] The train formation refers to the number of carriages in the train.

[0040] The measured feature vectors of key bridge points represent the displacement of these key bridge points at the current moment. The key points of a bridge are any one or all of the mid-span, 1 / 4-span, and 3 / 4-span points. In this scheme, the mid-span measuring point is used as an example. Therefore, analyzing the displacement characteristics of this point can describe the overall risk of the bridge. That is, if there is no risk in the displacement of this point, there will naturally be no risk in the other points.

[0041] The index data includes response feature vectors of key bridge points, vortex-induced vibration state label data, and train operation comfort indicators.

[0042] The response feature vector of the bridge key point is the displacement of the bridge key point at the next time step; The vortex-induced vibration state label data is four-dimensional data, including the upper confidence boundary of the amplitude at key bridge points, the dominant frequency of the response, the phase difference, and the phase difference of the lock-in duration; the train running comfort index is one-dimensional dimensionless data, with a value range of 0 to 1, and the smaller the value, the safer the system.

[0043] The above are the data formats for field state data and indicator data. Field state data is data from the current moment to the previous (historical time). Indicator data is data from the current moment onwards.

[0044] In other words, during the bridge simulation process, for the current time t, the wind field input data, train operation status data, and measured feature vectors of key bridge points from the previous H steps are selected as the field state data. The response feature vectors of key bridge points, vortex-induced vibration state label data, and train operation comfort indicators from the H steps after the current time t are selected as the indicator data.

[0045] It is possible to simulate different trains passing over bridges at different speeds under different wind conditions multiple times to obtain a sufficient number of static samples.

[0046] Static sample library , ; in: Indicates the first Field state data of a static sample; Indicates the first Indicator data for a static sample; This represents the total number of samples in the static sample library.

[0047] The field-state data of each static sample can be represented as: ; in: Indicates the first The wind field input data for each static sample includes average wind speed, wind direction angle, wind angle of attack, and turbulence intensity; Indicates the first The train operation status data of a static sample includes train speed, train formation, and wheel-rail contact parameters.

[0048] t represents the time index. Indicates the first Measured feature vectors of key bridge points from a static sample.

[0049] Both are high-dimensional two-dimensional matrix data, with wind field input data. For example.

[0050] The horizontal columns represent average wind speed, wind direction angle, wind attack angle, and turbulence intensity, while the vertical column represents time.

[0051] For example: ; in, , , , These represent the average wind speed, wind direction angle, wind attack angle, and turbulence intensity at time t (the current time). Indicates the length of the time window for data collection.

[0052] in this way, It is a 4× The two-dimensional matrix data, the corresponding train operation status data, and the measured feature vectors of key bridge points are all two-dimensional matrix data.

[0053] Among them, the number of rows for train operation status data and measured feature vectors of key bridge points is [number missing]. Each column is related to the dimensions of its own data.

[0054] Train operation status data includes train speed, train formation, and wheel-rail contact status parameters. Train formation includes 2 data points, and wheel-rail contact status parameters include 3 data points, so the number of columns for train operation status data is 7.

[0055] The number of columns in the measured feature vectors of the corresponding bridge key points is also the same. The number of specific data items in the measured feature vectors of the bridge key points is the same as the number of data items in the response feature vectors of the bridge key points, which will be explained below.

[0056] The metric data for each static sample can be represented as: ; in: Represents the response feature vector of key points of the bridge; This represents the label data indicating the state of vortex-induced vibration; This indicates an indicator of train operating comfort.

[0057] Response feature vectors of key bridge points It is a 6-dimensional vector, the response feature vector. This includes the displacement, velocity, acceleration, mean of displacement and velocity of bridge key points within a historical time window, standard deviation of velocity of bridge key points within a historical time window, and increment of displacement of bridge key points.

[0058] The measured feature vectors and response feature vectors of bridge key points have the same dimension. The measured feature vectors of bridge key points represent the displacement of the bridge center point at the current time t. The response feature vectors of bridge key points represent the displacement of the bridge center point at the next time step (future).

[0059] Specifically, ; in, These represent the displacement, velocity, acceleration, mean of the bridge key point's displacement and velocity within the historical time window, standard deviation of the bridge key point's velocity within the historical time window, and increment of the bridge key point's displacement at time t+1 (future), respectively. t represents the index of the time, and H represents the length of the predicted time window.

[0060] The measured feature vectors of key bridge points have the same dimension as the response feature vectors of key bridge points. The measured feature vectors of key bridge points represent the displacement of the bridge's center point at the current moment, while the response feature vectors represent the displacement of the bridge's center point at the next time step. The measured feature vectors of key bridge points are actually used in conjunction with wind field data and train data to calculate the coupling and vibration relationships of the entire system at the current time step.

[0061] Thus, the resulting response feature vector of the bridge key points is a 6×H two-dimensional matrix. The corresponding measured feature vector of the bridge key points is also 6×H. A two-dimensional matrix.

[0062] The vortex-induced vibration state label data is a 4×H two-dimensional matrix. Each row includes the upper confidence bound of the amplitude at the bridge key point, the dominant frequency of the bridge key point response, the phase difference of the vibration at the bridge key point, and the lock-in duration. The number of rows is H.

[0063] Train operation comfort index This is an H-dimensional column of data used to describe the safety and comfort of the train. It is dimensionless and ranges from 0 to 1. The smaller the value, the safer the entire system.

[0064] In other words, the train running comfort index has only one value at any given time, and therefore can only form a column of data in column H.

[0065] The wind-train-bridge-vortex-induced vibration coupling system is a simulation model of the bridge, used to simulate the index data of the train under specified speed and corresponding wind field.

[0066] The construction method of the wind-train-bridge-vortex-induced vibration coupling system is an existing technology. For the specific construction method, please refer to the applicant's prior patent application CN117890131A, "A method and system for predicting the comfort of trains crossing bridges". The construction process of the wind-train-bridge-vortex-induced vibration coupling system is described in detail in CN117890131A, "A method and system for predicting the comfort of trains crossing bridges".

[0067] Specifically: The wind-train-bridge-vortex-induced vibration coupling system in this solution is based on the wind-train-bridge-vortex-induced vibration coupling system in the first patent application CN117890131A. When vertical vortex-induced vibration occurs, the vibration equation of the wind-train-bridge-vortex-induced vibration coupling system is expressed as: ; ; in: These are the displacement, velocity, and acceleration of the bridge subsystem, respectively. These are the acceleration, velocity, and displacement of the train subsystem, respectively. , , These are the mass matrix, damping matrix, and stiffness matrix of the bridge subsystem, respectively. , , These are the mass matrix, damping matrix, and stiffness matrix of the train subsystem, respectively. The force exerted by the train on the bridge. The reaction force generated by the bridge on the train; These are the wind loads acting on the bridge and the train, respectively, from the wind subsystem. This refers to the vertical vortex-induced force acting on the bridge subsystem.

[0068] After constructing the wind-train-bridge-vortex-induced vibration coupled system, the field state data is input into the system, allowing the extraction of index data from the simulation results. Specifically, after inputting the field state data, the system simulates the bridge's vibration, thereby obtaining response feature vectors of key bridge points and vortex-induced vibration state label data. The train running comfort index is calculated using the method described in the prior patent application CN117890131A, "A Method and System for Predicting Train Crossing Bridge Comfort."

[0069] refer to Figure 2 and Figure 3 , Figure 2 This is a simulation diagram of the wind-train-bridge-vortex-induced vibration coupled system in SIMPACK. Figure 3 The results are from the simulation.

[0070] Wind tunnel tests were conducted on the target bridge to construct a database of vortex-induced vibration state parameters for querying these parameters.

[0071] After obtaining the structural dimensions, cross-sectional shape, and relevant dynamic characteristic parameters of the target bridge, a scaled-down model of the target bridge is made according to the wind tunnel test similarity criteria, and wind tunnel tests are carried out. The vortex-induced vibration state parameters collected in the wind tunnel tests are collected to obtain a vibration sample, thereby establishing a vortex-induced vibration state parameter library.

[0072] The vortex-induced vibration state parameter library includes multiple vibration samples. These samples include field state data and vortex-induced vibration state parameters.

[0073] The vortex-induced vibration state parameters refer to the vortex-induced vibration conditions under corresponding field state data. In other words, the vortex-induced vibration state parameters are retrieved using the field state data.

[0074] The vortex-induced vibration state parameters include the wind load characteristic vector of the vehicle-bridge system, the vortex-induced vibration state characteristic vector, and the equivalent aerodynamic force characterization quantity.

[0075] Among them, the wind load characteristic vector of the vehicle-bridge system includes the bridge static wind load, the bridge buffeting force load, the lateral force of the train, the lift of the train, and the rolling moment of the train. The characteristic vector of vortex-induced vibration includes the equivalent wind speed, the dominant frequency lock-in deviation, the equivalent amplitude, the amplitude growth rate, the displacement characterization at the dominant frequency, the energy concentration coefficient, the amplitude of the first-order vortex-induced force coefficient, the aerodynamic-displacement phase difference, the phase difference change rate, and the dominant frequency of vortex-induced vibration.

[0076] The field state data in the static sample library can be used as the working condition input for wind tunnel tests. Under the wind field conditions corresponding to each field state data, the vortex-induced vibration state parameters of the target bridge can be collected.

[0077] Meanwhile, the vortex-induced vibration locking frequency range was tested in the wind tunnel.

[0078] Step 2: Obtain real-time field data of the target bridge, and preprocess the real-time field data to obtain standard data; Real-time field data refers to the measured data collected in real time during the operation of the target bridge. Field data in the static sample library refers to the data input into the simulation software for the target bridge.

[0079] Real-time field-state data has high fidelity, but may have data gaps. Static sample databases have lower fidelity, but the data is more complete.

[0080] Real-time field data includes real-time wind field input data, real-time train operation status data, and measured feature vectors of key bridge points.

[0081] Real-time wind field input data includes average wind speed, wind direction angle, wind attack angle, and turbulence intensity measured in real time.

[0082] Real-time wind farm input data is acquired through sensors installed on the bridge, and real-time train operation status data is obtained directly from the train's travel information by communicating with the train.

[0083] Real-time train operation status data includes train speed, train formation, train, acceleration, and wheel-rail contact status parameters.

[0084] The real-time response feature vector of a bridge's key points is calculated by acquiring the displacement of the center point using sensors installed on the bridge. The measured feature vector of the real-time bridge key points is the actual measured feature vector of the bridge's key points based on data collected in real-time by displacement sensors installed on the bridge. How to calculate speed and acceleration based on the displacement of the center point is a current technology.

[0085] Preprocessing includes denoising, missing value compensation, temporal alignment, outlier removal, and normalization.

[0086] Denoising, missing value compensation, temporal alignment, outlier removal, and normalization are existing technologies, and their specific methods will not be described further here.

[0087] In practice, real-time field data is complete; the main issue is that "wheel-rail contact state parameters" are unavailable. These parameters need to be extracted from a static sample library.

[0088] The extraction method is as follows: Calculate the real-time field data and the static sample library. The cosine similarity of each field state data is used to extract the field state data that is closest to the real-time field state data. The wheel-rail contact state parameters of the field state data are then filled into the missing "wheel-rail contact state parameters" item in the real-time field state data, thus completing the data supplementation.

[0089] Specifically: This embodiment supplements this parameter in the following way: (1) Calculate real-time field data and static sample library Cosine similarity of the field state data in the middle; (2) Extract the field state data with the highest cosine similarity from the static sample library; (3) Fill the missing wheel-rail contact state parameter items in the real-time field data with the wheel-rail contact state parameters extracted from the field data, thereby completing the data supplementation of wheel-rail contact state parameters.

[0090] The method for calculating cosine similarity is existing technology. Field state data is a high-dimensional vector, and the similarity between high-dimensional vectors can be directly calculated using the cosine similarity formula.

[0091] The reason for this approach is that wheel-rail contact state parameters cannot be directly obtained from sensors; they can only be extracted during the simulation of the wind-train-bridge-vortex-induced vibration coupled system. Therefore, when monitoring field data in real time, since this parameter cannot be obtained, cosine similarity matching is used to match the field state data from a static sample library that most closely matches the real-time field state data, thus completing the data filling.

[0092] After the preprocessing in step 2, the real-time field data has the same dimensions and format as the field data in the static sample library.

[0093] Step 3: Convert the standard data into coupled state feature vectors, and stack multiple coupled state feature vectors within the historical time window in chronological order to form a time-series input matrix.

[0094] The coupled state feature vector is composed of wind field feature vector, train operation state vector, measured feature vector of bridge key points, wind load feature vector of vehicle-bridge system, feature vector of vortex-induced vibration state, and statistical feature vector of historical time window, which are sequentially spliced ​​together.

[0095] Specifically, acquiring time Standard data and the former Standard data for each historical moment Generate the time series input matrix at time t. .

[0096] ; express The coupling state feature vector at consecutive time points, This represents the dimension of the feature vector of the coupling state. Indicates the length of the historical time window. Indicates time The time-series input matrix, Represents the dimension symbol.

[0097] Furthermore, the coupling state feature vector for: ; Represents the wind field eigenvector. Represents the train's operating state vector. The measured feature vectors representing key points of the bridge. This represents the eigenvector of wind load on the vehicle-bridge system. The eigenvector representing the vortex-induced vibration state, represents the statistical feature vector of the historical time window, used to characterize the evolution trend of the system within the previous time window, and T represents the matrix transpose.

[0098] ; Indicates time average wind speed, Indicates time The wind angle, Indicates time Angle of attack of the wind, Indicates time The turbulence intensity, where T represents the matrix transpose.

[0099] ; Indicates time The train speed, Indicates the train formation. Indicates the vehicle's acceleration. Indicates the derailment coefficient. Indicates the wheel load reduction rate. This represents the wheel-rail contact force, and T represents the matrix transpose.

[0100] ; Indicates the displacement of key points on the bridge. Indicates the speed at key points of the bridge. This represents the acceleration at key points of the bridge. This represents the average displacement velocity of key bridge points within a historical time window. This represents the standard deviation of the velocity at key bridge points within a historical time window. , represents the increment of displacement at a key point on the bridge.

[0101] ; Indicates the static wind load on the bridge. Indicates the bridge flutter load. This represents the lateral force on the train. Indicates the lift of the train. This indicates the rolling moment of the train.

[0102] ; This indicates the converted wind speed. Indicates the frequency lock-in deviation. Indicates the converted amplitude. Indicates the amplitude growth rate. The displacement at the dominant frequency is represented by T, which represents the matrix transpose. Indicates the energy concentration factor; This represents the amplitude of the first-order vortex excitation coefficient; This represents the phase difference between aerodynamic force and displacement; Indicates the rate of change of phase difference; Indicates the dominant frequency of vortex-induced vibration; Vehicle-bridge system wind load characteristic vector eigenvectors of vortex-induced vibration state The parameters are obtained by querying the vortex-induced vibration state parameter database. This involves comparing the similarity of the wind field feature vector, train operation state vector, and measured feature vector of key bridge points with the field state data in the vortex-induced vibration state parameter database to find the vortex-induced vibration state parameters corresponding to the closest field state data. This yields the wind load feature vector and the vortex-induced vibration state feature vector of the bridge system.

[0103] refer to Figure 4 and Figure 5 , Figure 4 and Figure 5 The images show the effect of wind tunnel testing and the relationship between wind speed and vertical amplitude during wind tunnel testing.

[0104] ; Indicates the preceding The mean of the average wind speed within each time step. Indicates the preceding The average train speed within each time step Indicates the preceding The average displacement of key bridge points within a given time step. Indicates the preceding The average amplitude converted within each time step. Indicates the preceding The range of displacements of key bridge points within a time step is used to characterize the system's evolution trend. The value represents the length of the historical time window, and T represents the matrix transpose.

[0105] The coupled state eigenvectors are essentially the data from the real-time field data at the same time t, combined according to the above format. After obtaining L coupled state eigenvectors at time t, they are directly used to form the time-series input matrix X. t .

[0106] Whether it's real-time field data or field data from a static sample library, the time-series input matrix X can be constructed in the manner described above. t .

[0107] Step 4: Construct a physical constraint deep learning prediction model. Input the time series input matrix into the physical constraint deep learning prediction model to obtain the target response results and equivalent aerodynamic force characterization within the future prediction time window. The target response results include predicted response feature vectors for key bridge points, predicted vortex-induced vibration state label data, and predicted train ride comfort indices. Maintain consistency with the structure of the indicator data in the static sample library: ; in, This represents the response feature vector for predicting key points of a bridge. This represents the label data for predicting vortex-induced vibration state; This indicates an index predicting train operating comfort.

[0108] A physical constraint deep learning prediction model includes at least a temporal feature extraction subnetwork, a response reconstruction subnetwork, and an equivalent aerodynamic force-assisted output subnetwork. These three subnetworks are cascaded as follows: The temporal feature extraction subnetwork is located at the network inlet and performs time-related modeling on the temporal input matrix; its output hidden feature vector is simultaneously distributed to two parallel branches: the response reconstruction subnetwork and the equivalent aerodynamic auxiliary output subnetwork.

[0109] The temporal feature extraction subnetwork uses an LSTM architecture to extract time-related features from the temporal input matrix. : ; This represents the temporal feature extraction subnetwork. This represents the training parameters of the temporal feature extraction subnetwork. This represents the time-related features output by the time-series feature extraction subnetwork. Represents the timing input matrix; The response reconfiguration subnetwork is used to output the future. Target response results within each time step: ; This indicates a response to reconstruct the subnetwork. In response to the reconstructed subnetwork training parameters, This indicates the target response result within the future prediction time window. This represents the time-related features output by the temporal feature extraction subnetwork; Specifically, the response reconstruction subnetwork consists of three parallel output heads, corresponding to three types of predictable quantities: bridge structural response, vortex-induced vibration state, and train running comfort index.

[0110] Each output head is implemented using a network structure of stacked fully connected layers or temporal convolutional layers, and the parameters are not shared between the output heads; Bridge response prediction head: ; in, Indicates the bridge response prediction header. This represents the parameters to be trained in the bridge response prediction head. This represents the response feature vector for predicting key points of a bridge. This represents the time-related features output by the temporal feature extraction subnetwork; Vortex-induced vibration state prediction head: ; This indicates the prediction head for vortex-induced vibration state. This represents the parameters to be trained for the vortex-induced vibration state prediction head. This represents the label data for predicting vortex-induced vibration state. This represents the time-related features output by the temporal feature extraction subnetwork; Train running comfort index prediction head: ; This indicates the prediction head for vortex-induced vibration state. This represents the parameters to be trained for the vortex-induced vibration state prediction head. This indicates a predicted indicator of train running comfort. This represents the time-related features output by the temporal feature extraction subnetwork; The equivalent aerodynamic auxiliary output subnetwork is used to output the equivalent aerodynamic characterization quantities related to vortex-induced vibration: ; This represents the equivalent aerodynamic auxiliary output subnetwork. The parameters to be trained in the equivalent aerodynamic-assisted output subnetwork are... Indicates the length of the future forecast time window. It represents the equivalent aerodynamic force characterization quantity.

[0111] Step 5: Construct a residual compensation network, compensate the target response result based on the residual compensation network to obtain the real-time response result, and use the real-time response result as the prediction index data; Step 5 includes the following steps: Step 51: Calculate the cosine similarity between the real-time field data and each field data in the static sample library. Extract the field data with the highest cosine similarity from the static sample library. Use the index data corresponding to the extracted field data as the error benchmark data. Calculate the difference between the error benchmark data and the target response result as the online prediction error. ; Indicates time Error baseline data, This represents the target response result at time t; Indicates time The prediction error.

[0112] The indicator data in the static sample library has the same structure and is highly consistent with the target response results. Therefore, subtracting the two can yield a prediction error with the same dimension or data structure.

[0113] Step 52: Obtain online prediction error in real time The online prediction error is updated using a sliding time window recursive statistical method to obtain a smoothed error estimate. ; ; Indicates time The smoothing error estimator, Indicates time The smoothing error estimator, Denotes the smoothing factor, satisfying .

[0114] Step 53: Construct the residual compensation network The input to the residual compensation network is a time-series input matrix. Smoothing error estimator The output is the residual compensation amount. : ; Step 54: Correct the target response result using the residual compensation amount to obtain the real-time response result: ; This represents the target response result after error compensation at time t; This represents the target response result before error compensation at time t; This represents the time-decaying corrected gain matrix; This represents the relative time step from the current time t within the future prediction time window, where T = 1, 2, ..., H.

[0115] ; Indicates the first The attenuation coefficient corresponding to each element in the real-time response result. The total number of elements in the real-time response result; diag This represents a diagonal matrix.

[0116] Real-time response results The real-time response results are highly consistent with the indicator data in the static sample. Therefore, the total number of elements in the real-time response results can be obtained from the arrangement of the indicator data in the static sample. The specific content of each item in the indicator data has been provided above.

[0117] By introducing a time-decaying correction gain matrix, predictions closer to the current time are corrected more significantly, while predictions farther from the current time are corrected less significantly, thus balancing online correction capability and long-term prediction stability.

[0118] The final real-time response result after error compensation The structure of the indicator data used in the next step is consistent with that of the indicator data in the static sample library: ; This represents the response feature vector of the predicted bridge key points after error compensation. This represents the label data of the predicted vortex-induced vibration state after error compensation. This indicates an index predicting train operating comfort.

[0119] Step 6: Calculate risk indicators and generate early warning results based on the predicted indicator data, and at the same time build a real sample library for continuous model optimization.

[0120] Step 6 includes the following steps: Step 61: Normalize the real-time response results; Real-time response results It is a high-dimensional vector containing information in many dimensions. The units or dimensions of these dimensions are not uniform, so each dimension needs to be normalized separately. The normalization process is omitted here. In this way, each vector (element) can be unified to a value between 0 and 1, achieving a uniformity of dimensions.

[0121] Step 62: Calculate the sub-indices of bridge structural response risk based on the normalized results. Train operation safety risk sub-indicators Sub-indicators of vortex-induced vibration development risk Bridge structural response risk sub-indicators Train operation safety risk sub-indicators Sub-indicators of vortex-induced vibration development risk The weighted coefficients are used to merge the results into a comprehensive index. ; in: ; ; in, These represent the weights of the first bridge, the second bridge, and the third bridge, respectively. express The sum of all elements in the same row after normalization is used to describe the vibration risk of the bridge. express The sum of all elements in the expression.

[0122] express The sum of the first derivatives of all elements in the same row is used to describe whether the vibration at the bridge's center point is increasing rapidly. express The sum of the first derivatives of all elements in the expression.

[0123] express The number of times all elements in the same row exceed their respective preset values ​​bU is used to describe the duration of abnormal conditions affecting bridge vibration risk.

[0124] express The sum of the number of all elements that exceed their respective preset values ​​bU.

[0125] For example, ; In this scheme, the predicted time length is H, so There are 6 elements in one row, and there are H rows in total. In the calculation... In this case, simply sum all the elements of the corresponding time row. For example, It is All elements in column t+1 are summed after normalization. The purpose of normalization is to unify the units of measurement.

[0126] Correspondingly, During the calculation, the real-time response result at the current time t is used. (Normalized value) and the real-time response result at the previous time t-1 By comparing the normalized values, the change in each element is calculated, thus obtaining the first derivative. The first derivative reflects the rate of change of each element during the prediction process.

[0127] Correspondingly, In terms of time, it is... Each element in the table (actually only 6 elements in one row need to be set) is assigned a different preset value bU. If the number of elements exceeding the preset value bU is sufficiently large, it indicates a high probability of risk and volatility. The summation calculation indicates a prolonged period of high vibration risk, necessitating a higher warning level. This method effectively filters outomas; that is, only when a small range of values ​​exceed the risk threshold is an outoma detected. It won't be very big.

[0128] Specifically, the value of the preset value U is determined according to the "Design Specification for High-Speed ​​Railway (TB 10621-2014)" and relevant bridge operation verification standards.

[0129] Train operation comfort risk sub-indicators It is used to characterize the risk to passenger comfort when a train is subjected to crosswinds, bridge vibrations, track irregularities, and changes in vehicle posture during the crossing of a bridge.

[0130] ; ; in, These represent the weights of the first train, the second train, and the third train, respectively. express The sum of all elements after normalization is used to describe the vibration risk of the train; express The sum of the first derivatives of all elements in the equation is used to describe whether the train vibration is increasing rapidly. express The number of times all elements exceed their respective preset values ​​vU is used to describe the duration of abnormal situations that pose a risk to train operation comfort.

[0131] The calculation method and The calculation method is the same, because There is only one element at any given moment, therefore... The calculation is relatively simple.

[0132] Sub-indicators of vortex-induced vibration development risk Used to characterize the degree of vortex-induced vibration development and operational impact risks; Under normal conditions, the risk sub-index for vortex-induced vibration development is directly set to 0, and the equivalent aerodynamic force characterization is monitored in real time. When the similarity between the equivalent aerodynamic characterization quantity and the pre-saved aerodynamic characterization quantity that generates vortex-induced vibration exceeds a threshold, it indicates that vortex-induced vibration will occur in the next time period, and then the risk sub-index of vortex-induced vibration development needs to be calculated according to the following formula.

[0133] The sub-index for vortex-induced vibration development risk is based on the predicted vortex-induced vibration state label data. calculate.

[0134] Sub-indicators of vortex-induced vibration development risk yes: ; ; This indicates the risk of the converted amplitude exceeding the limit; ; This represents the predicted amplitude of vortex-induced vibration. This is the amplitude (preset value). To compare functions, select the maximum value from them; This indicates a risk of frequency lock-in deviation. ; Indicates the predicted response frequency. Indicates the modal frequency limit. This represents the unit of frequency change. To compare functions, select the maximum value from them; This indicates the risk of aerodynamic-displacement phase difference; ; This represents the predicted phase difference. Indicates the phase difference limit. To compare functions, select the maximum value from them; ; This indicates a risk associated with the duration of the lockout; This indicates the predicted lockout duration. Indicates the lock duration limit. To compare functions, select the maximum value from them; The risk sub-indicator of vortex-induced vibration development is actually a comparison of the relationship between vortex-induced vibration amplitude, response dominant frequency, phase difference, lock-in time and each preset limit. When it is less than the preset value of each part, it is 0.

[0135] These represent amplitude weight, main frequency weight, shift weight, and duration weight, respectively.

[0136] The comprehensive risk index is used to comprehensively reflect the risks of bridge structural response, train operation safety, train operation comfort, and vortex-induced vibration development. Based on the relationship between the comprehensive risk index and preset classification thresholds, a corresponding early warning level is output. The comprehensive risk index is expressed by the following formula: ; ; in, For bridge structural response risk sub-indicators; Sub-indicators for train operation comfort risk; This is a sub-indicator for the development risk of vortex-induced vibration. These are the structural response weight, train operation comfort weight, and vortex-induced vibration development weight, respectively.

[0137] Step 63: Set the first threshold value respectively Second threshold Third threshold An alarm level is generated based on the comprehensive threshold and the first, second, and third thresholds. .

[0138] ; Preset the first, second, and third thresholds, and increase them; When the comprehensive risk index is less than the first threshold, the normal monitoring status is output. When the comprehensive risk index is greater than or equal to the first threshold and less than the second threshold, an alert will be issued. When the comprehensive risk index is greater than or equal to the second threshold and less than the third threshold, a restriction warning is output. When the comprehensive risk index is greater than or equal to the third threshold, a danger warning is issued.

[0139] Example 2: Example 2 provides a training method for a physically constrained deep learning prediction model based on Example 1.

[0140] The loss function used during training is: ; in, These are the weighting coefficients for each loss term. For the bridge response prediction error term, This is the error term for predicting vortex-induced vibration. These are structural dynamic equilibrium constraints. This is a constraint term related to the aerodynamic-displacement phase relationship. This is a constraint term for the consistency of the dominant frequency of vortex-induced vibration. This is a constraint term for the state evolution of vortex-induced vibration. To address the smoothness or time continuity constraints, the weighting coefficients are selected based on the following principles: The primary optimization objective is to select larger values ​​to ensure prediction accuracy. The physical constraint weights can be appropriately increased in the early stages of training to guide the model to learn reasonable physical mapping relationships, and then gradually reduced in the later stages of training to improve the data fitting accuracy. This can be determined through cross-validation, multi-objective optimization methods, or engineering experience.

[0141] Bridge response prediction error term: ; in, The predicted value of the response feature vector of the key points of the bridge ( ), Here, k represents the measured feature vector (measured value) of the key points of the bridge, and H represents the prediction step size. express The L2 norm.

[0142] Error term for prediction of vortex-induced vibration state: ; in, The predicted value of the vortex-induced vibration state vector ( ), The actual value of the vortex-induced vibration state vector is obtained from the static parameter library through similarity matching. express The L2 norm of the prediction, k represents the future time, and H represents the prediction step size.

[0143] The structural dynamic equilibrium constraint term is used to limit the deviation between the predicted bridge response and the structural dynamic equilibrium relationship. This constraint requires that the bridge prediction response output by the model must satisfy the structural dynamic equation expressed by the mass matrix, damping matrix and stiffness matrix, so as to make the prediction results physically consistent.

[0144] Structural dynamic equilibrium constraints: ; Let represent the residual term, t represent the current time, k represent the future time, and H represent the prediction step size. express The square of the L1 norm; ; Where M, C, and K are the mass matrix, damping matrix, and stiffness matrix of the bridge structure, respectively; To predict the wind load vector; To predict the train load vector; To predict the equivalent vortex-induced force vector, , , These are the predicted values ​​of bridge displacement, velocity, and acceleration at key points output by the model, extracted from the response feature vectors of the predicted bridge key points. , The characteristic vectors of wind load and vortex-induced vibration state are extracted from the time-series input matrix. Extract the wheel-rail contact force from the timing input matrix of this input.

[0145] Aerodynamic-displacement phase relationship constraint terms : ; ; Among them, among them, To predict the equivalent aerodynamic characteristics at the dominant frequency; To predict the displacement at the dominant frequency; This indicates the phase angle; the subscript dom indicates the dominant frequency. The phase difference is given by the vortex-induced vibration state parameter library of the bridge main girder, where t represents the current time, k represents the future time, and H represents the prediction step size. This indicates the phase difference reference value. To predict the displacement characteristics at the dominant frequency, the displacement response at the predicted key points of the bridge is extracted by Fourier transform.

[0146] Consistency constraints for eddy-induced vibration dominant frequency: ; in, This refers to the lock-in frequency range of vortex-induced vibration determined by wind tunnel testing; the subscript "lock" indicates the lock-in range. H represents the predicted dominant frequency at time t+k, where t represents the current time, k represents the future time, and H represents the prediction step size.

[0147] Constraints on the state evolution of vortex-induced vibration : ; in, This is a state transition operator established based on the vortex-induced vibration state parameter library of the main girder of the bridge. This is the normalized vector of the input field state data. express The square of the L2 norm.

[0148] Response smoothness or time continuity constraints: ; Where: t represents the current time, k represents the future time, and H represents the prediction step size. This represents the target response result at time t+k+1. This represents the target response result at time t+k+2. This represents the target response result at time t+k.

[0149] The bridge response prediction error term is used to limit the deviation between the model output and the true label. The vortex-induced vibration state prediction error term is used to constrain the deviation between the predicted vortex-induced vibration state parameters and the sample labels. The structural dynamic equilibrium constraint term is used to limit the deviation between the predicted response result and the dynamic equilibrium relationship of the bridge structure. This constraint ensures that the predicted response satisfies the basic equilibrium relationship of bridge dynamics, meaning that the model cannot output predicted results that violate the laws of structural dynamics.

[0150] The aerodynamic-displacement phase relationship constraint term is used to limit the degree to which the phase difference between the predicted equivalent aerodynamic force and the predicted displacement at the dominant frequency deviates from the phase relationship.

[0151] The eddy-induced vibration dominant frequency consistency constraint term is used to limit the degree to which the predicted response dominant frequency deviates from the locked frequency or locked frequency range; The vortex-induced vibration state evolution constraint term is used to limit the degree to which the predicted results deviate from the evolution law during the locking development, stabilization limiting, or unlocking process; Response smoothness or temporal continuity constraints are used to suppress non-physical abrupt changes in prediction results between adjacent time steps.

[0152] The training process for a physically constrained deep learning prediction model follows these logical steps: S1: Read the static sample library established in step 1, and perform unified preprocessing on the time-series data of each simulation condition according to the same preprocessing rules as in step 2. The normalized extrema used in the preprocessing process are consistent with those in step 2 to ensure the consistency of the input data distribution between offline training and online inference. Based on this, generate coupled state feature vectors step by step according to the method in step 3, and slice the feature vector sequence of each simulation condition using a sliding time window method; construct training sample pairs, each training sample pair is a time-series input matrix formed by stacking the coupled state feature vectors of the previous several time steps in chronological order as the model input, and the real index data corresponding to the next several time steps is used as the model supervision signal.

[0153] All training sample pairs are divided into training set, validation set and test set according to a preset ratio, and the labels of different working conditions are evenly distributed in the three datasets to avoid excessive concentration of one working condition.

[0154] S2: Randomly initialize all trainable parameters of the three parallel output heads of the temporal feature extraction subnetwork, the response reconstruction subnetwork, and the equivalent aerodynamic auxiliary output subnetwork. Simultaneously set training hyperparameters such as optimizer type (e.g., Adam), initial learning rate, batch size, and maximum number of training epochs.

[0155] S3: Settings .

[0156] In the early stages of training Set it to the maximum value to allow the model to first obtain a basic fit to the index data on the static sample database; gradually increase the value as the number of training epochs increases. This allows the model to gradually approximate physical laws while maintaining the accuracy of data fitting.

[0157] S4: In each training step, a mini-batch of samples is randomly selected from the training set. For each sample in the mini-batch, the temporal feature extraction, multi-head parallel prediction, and component concatenation are completed sequentially according to the forward inference process in step 4 to obtain the target response result; at the same time, the hidden feature vector is fed into the equivalent aerodynamic auxiliary output sub-network to obtain the equivalent aerodynamic auxiliary output.

[0158] S5: The total loss value for this mini-batch is obtained by weighted summation based on the current weight coefficients. The total loss value is then propagated backward along the computation graph using the backpropagation algorithm, calculating the gradient of each parameter to be trained layer by layer. The optimizer then updates all parameters to be trained synchronously based on the current learning rate.

[0159] S6: Repeat S4 and S5 until the maximum number of training rounds is reached, and load the archived optimal model parameters as the pre-training parameters of the physical constraint deep learning prediction model.

[0160] Furthermore, the residual compensation network can be trained independently or simultaneously with the physical constraint deep learning prediction model. In this scheme, the residual compensation network and the physical constraint deep learning prediction model are combined into a neural network and trained simultaneously in the manner described above.

[0161] Furthermore, during the online monitoring of the bridge, more actual static samples will be continuously generated, and these actual static samples will be added to the static sample library of step 1.

[0162] The actual static samples include actual field state data and actual index data. The actual field state data are the wind field input data, train operation status data and measured feature vectors of key bridge points collected at time t.

[0163] Actual indicator data refers to the actual indicator data collected at time t+1, including the response feature vectors of key bridge points, vortex-induced vibration state label data, and train operation comfort indicators. The response feature vectors of key bridge points, vortex-induced vibration state label data, and train operation comfort indicators can be directly monitored and obtained through sensors deployed on trains and bridges.

[0164] Specifically, the response feature vectors of key bridge points can be obtained from data collected by displacement, velocity, or acceleration sensors installed on the bridge; vortex-induced vibration state label data can be obtained from the bridge response time history through frequency domain analysis and phase analysis; and train running comfort indicators can be calculated from the acceleration installed on the train, that is, the train running comfort indicators can be directly converted from acceleration.

[0165] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A bridge operation safety early warning method based on a wind turbine bridge vortex-vibration coupling system, characterized in that, Includes the following steps: Step 1: Obtain static samples of the wind-train-bridge-vortex-induced vibration coupled system and generate a static sample library; Among them, the wind-train-bridge-vortex-induced vibration coupled system is the simulation model of the target bridge, used to simulate the index data of the train under the action of a specified speed and corresponding wind field; the static sample includes field state data and index data; Wind tunnel tests were conducted on the target bridge to construct a database of vortex-induced vibration state parameters for querying vortex-induced vibration state parameters. Step 2: Obtain real-time field data of the target bridge, and preprocess the real-time field data to obtain standard data; Step 3: Convert the standard data into coupled state feature vectors, and stack multiple coupled state feature vectors within the historical time window in chronological order to form a time series input matrix; Step 4: Input the time series input matrix into the physical constraint deep learning prediction model to obtain the target response results within the future prediction time window; The target response results include predicted response feature vectors of key bridge points, predicted vortex-induced vibration state label data, and predicted train operation comfort indices. Step 5: Input the target response result into the residual compensation network for compensation to obtain the real-time response result, and use the real-time response result as the prediction index data; The predictive index data includes the response feature vector of the predicted bridge key points after error compensation, the predicted vortex-induced vibration state label data after error compensation, and the predicted train running comfort index. Step 6: Calculate risk indicators and generate early warning results based on the predicted indicator data, and at the same time build a real sample library for continuous model optimization; In step 4, the physical constraint deep learning prediction model includes a temporal feature extraction subnetwork, a response reconstruction subnetwork, and an equivalent aerodynamic force-assisted output subnetwork. The temporal feature extraction subnetwork adopts an LSTM architecture to extract the temporal features of the temporal input matrix and output hidden feature vectors. The response reconstruction subnetwork and the equivalent aerodynamic auxiliary output subnetwork are two parallel branches that receive hidden feature vectors respectively. The response reconstruction subnetwork includes a bridge response prediction head, a vortex-induced vibration state prediction head, and a train operation comfort index prediction head. The three output heads do not share parameters and output the response feature vector of the predicted bridge key points, the predicted vortex-induced vibration state label data, and the predicted train running comfort index, respectively. The equivalent aerodynamic auxiliary output subnetwork is used to output the equivalent aerodynamic characterization quantities related to vortex-induced vibration; Step 5 includes the following steps: Step 51: Calculate the cosine similarity between the real-time field data and each field data in the static sample library. Extract the field data with the highest cosine similarity from the static sample library. Use the index data corresponding to the extracted field data as the error benchmark data. Calculate the difference between the error benchmark data and the target response result as the online prediction error. Step 52: Update the online prediction error using a sliding time window recursive statistical method to obtain the smoothed error estimate; Step 53: Input the time series input matrix and the smoothing error estimate into the residual compensation network, and output the residual compensation amount; Step 54: Use the time decay type modified gain matrix to weight the residual compensation amount, and then add the weighted residual compensation amount to the target response result to obtain the real-time response result.

2. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 1, characterized in that, In step 1, the field data includes wind field input data, train operation status data, and measured feature vectors of key bridge points; Wind field input data includes average wind speed, wind direction angle, wind angle of attack, and turbulence intensity; Train operation status data includes train speed, car body acceleration, train formation, and wheel-rail contact parameters; The measured feature vectors of key bridge points represent the displacement of these key bridge points at the current moment. The indicator data includes response feature vectors of key bridge points, vortex-induced vibration state label data, and train operation comfort indicators. The response characteristics of the bridge's key points are the displacement of the key points at the next time step. The vortex-induced vibration state label data includes the amplitude of vortex-induced vibration at key points of the bridge, the dominant frequency of the response, the phase difference, and the lock-in duration. The train running comfort index is a one-dimensional dimensionless data, with a value range from 0 to 1. The smaller the value, the safer the system.

3. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 1, characterized in that, In step 2, the real-time field data includes real-time wind field input data, real-time train operation status data, and measured feature vectors of real-time bridge key points; The wheel-rail contact status parameters in real-time train operation status data are supplemented in the following way: Calculate the cosine similarity between the real-time field data and each field data in the static sample library. Extract the field data with the highest cosine similarity from the static sample library and fill the missing wheel-rail contact state parameter items in the real-time field data with the wheel-rail contact state parameters from the extracted field data.

4. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 1, characterized in that, In step 3, the coupled state feature vector is composed of the wind field feature vector, the train operation state vector, the measured feature vector of the bridge key points, the wind load feature vector of the vehicle-bridge system, the feature vector of the vortex-induced vibration state, and the historical time window statistical feature vector. The wind field characteristic vector includes the time-averaged wind speed, wind direction angle, wind angle of attack, and turbulence intensity; The train operation state vector includes train speed, train formation, car body acceleration, derailment coefficient, wheel load reduction rate, and wheel-rail contact force; The wind load characteristic vector of the vehicle-bridge system includes the bridge static wind load, the bridge buffeting force load, the lateral force of the train, the lift of the train, and the rolling moment of the train. The characteristic vector of vortex-induced vibration state includes the equivalent wind speed, the dominant frequency lock-in deviation, the equivalent amplitude, the amplitude growth rate, the displacement characterization at the dominant frequency, the energy concentration factor, the amplitude of the first-order vortex-induced force coefficient, the aerodynamic-displacement phase difference, the phase difference change rate, and the dominant frequency of vortex-induced vibration. The historical time window statistical feature vector includes the average wind speed, average train speed, average bridge key point displacement, average converted amplitude, and range of bridge key point displacement within the previous several time steps.

5. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 1, characterized in that, The time-decaying correction gain matrix is ​​a diagonal matrix, and its diagonal elements decay exponentially with the relative time steps from the current time within the future prediction time window by a preset decay coefficient.

6. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 1, characterized in that, Step 6 involves calculating the comprehensive risk index based on the forecast indicator data, specifically including the following steps: Step 6 includes the following steps: Step 61: Normalize the real-time response results; Step 62: Calculate the sub-indices of bridge structural response risk based on the normalized results. Train operation safety risk sub-indicators Sub-indicators of vortex-induced vibration development risk Bridge structural response risk sub-indicators Train operation safety risk sub-indicators Sub-indicators of vortex-induced vibration development risk The weighted coefficients are used to merge the results into a comprehensive index. ; ； ； in, For bridge structural response risk sub-indicators; Sub-indicators for train operation comfort risk; This is a sub-indicator for the development risk of vortex-induced vibration. These are the structural response weight, train running comfort weight, and vortex-induced vibration development weight, respectively. Step 63: Set the first threshold value respectively Second threshold Third threshold An alarm level is generated based on the comprehensive threshold and the first, second, and third thresholds. .

7. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 6, characterized in that, Step 63 generates an early warning result based on comprehensive risk indicators, specifically including: The first threshold, the second threshold, and the third threshold increase sequentially. When the comprehensive risk index is less than the first threshold, the normal monitoring status is output. When the comprehensive risk index is greater than or equal to the first threshold and less than the second threshold, an alert will be issued. When the comprehensive risk index is greater than or equal to the second threshold and less than the third threshold, a restriction warning is output. When the comprehensive risk index is greater than or equal to the third threshold, a danger warning is issued.

8. The bridge operation safety early warning method based on the vortex-vibration coupling system of a wind turbine bridge according to claim 1, characterized in that, The training loss function of the physical constraint deep learning prediction model includes the bridge response prediction error term, the vortex-induced vibration state prediction error term, the structural dynamic balance constraint term, the aerodynamic-displacement phase relationship constraint term, the vortex-induced vibration dominant frequency consistency constraint term, the vortex-induced vibration state evolution constraint term, and the response smoothness or time continuity constraint term. The bridge response prediction error term is used to limit the deviation between the model output and the true label. The vortex-induced vibration state prediction error term is used to constrain the deviation between the predicted vortex-induced vibration state parameters and the sample labels. The structural dynamic equilibrium constraint term is used to limit the deviation between the predicted response result and the dynamic equilibrium relationship of the bridge structure. This constraint can make the predicted response satisfy the basic equilibrium relationship of bridge dynamics, that is, the model cannot output the predicted result that violates the laws of structural dynamics. The aerodynamic-displacement phase relationship constraint term is used to limit the degree to which the phase difference between the predicted equivalent aerodynamic force and the predicted displacement deviates from the phase relationship at the dominant frequency; The eddy-induced vibration dominant frequency consistency constraint term is used to limit the degree to which the predicted response dominant frequency deviates from the locked frequency or locked frequency range; The vortex-induced vibration state evolution constraint term is used to limit the degree to which the predicted results deviate from the evolution law during the locking development, stabilization limiting, or unlocking process; Response smoothness or temporal continuity constraints are used to suppress non-physical abrupt changes in prediction results between adjacent time steps.

Images