A bridge service performance evaluation method, system and device

Abstract

The application discloses a bridge service performance evaluation method, system and device, relates to the technical field of bridge service performance evaluation, and comprises the following steps: acquiring vehicle load space-time information and vehicle-induced response by using computer vision technology and bridge response monitoring equipment; reconstructing the vehicle load by means of a data fusion algorithm; establishing a vehicle multi-body dynamics model and a bridge finite element model, solving a vehicle-bridge coupling vibration equation, and obtaining bridge theoretical responses under different vehicle loads; establishing a mapping relationship between actual vehicle loads and vehicle-induced responses based on the reconstructed vehicle loads and vehicle-induced responses; judging the bridge stiffness degradation condition according to the bridge theoretical responses and actual responses; and evaluating the bridge service performance based on the bridge stiffness degradation degree. The bridge service performance evaluation method solves the problems of scattered bridge monitoring data, ineffective collaborative utilization, and poor timeliness and continuity in bridge service performance evaluation.

Description

Technical Field

[0001] This invention relates to the field of bridge service performance evaluation technology, specifically to a bridge service performance evaluation method, system, and device. Background Technology

[0002] Traditional methods primarily employ bridge inspection, on-site material property testing, and load testing to assess the appearance, material properties, and mechanical condition of bridges, and use the analytic hierarchy process (AHP) or theoretical calculations to evaluate the bridge's service condition. Alternatively, a simple assessment of bridge condition can be achieved using bridge response data collected by a health monitoring system.

[0003] However, the existing methods still have the following technical bottlenecks, which lead to many problems that cannot be ignored in practical applications: (1) At the load-response correlation level, although the health monitoring system has multi-source data, the data analysis methods are limited. Various bridge condition early warning methods proposed in recent years only judge anomalies based on a single vibration amplitude threshold, and fail to fully consider the dynamic relationship between key load parameters such as vehicle axle load and driving speed and bridge strain and displacement response, and cannot build a real and effective mapping model. (2) In terms of multi-source data fusion: Most bridge traffic load statistics methods only realize traffic flow statistics through visual recognition technology, without considering vehicle weight information, environmental interference factors and structural response data. The "Review of Bridge Multi-Source Monitoring Data Fusion Technology" explains that visual data and mechanical response data are processed independently, which not only leads to the loss of vehicle weight information, but also the inability to effectively quantify environmental interference factors, resulting in extremely low data utilization rate, which seriously hinders the improvement of assessment accuracy. (3) In terms of timeliness: As an important means of evaluating the bearing capacity of bridges, load testing, taking the static load test method of bridges as an example, not only requires closing traffic and setting up loading devices, but also has a long test cycle. Furthermore, the test load is significantly different from the random traffic load in actual operation, resulting in a relatively delayed evaluation.

[0004] In summary, the existing methods currently used fail to establish a true dynamic mapping relationship between vehicle load parameters and bridge vibration response, which is prone to errors and results in low accuracy of bridge service performance assessment. Summary of the Invention

[0005] To address the shortcomings of existing technologies in failing to establish a true dynamic mapping relationship between vehicle load parameters and bridge vibration response, resulting in low assessment accuracy, this invention proposes a bridge service performance assessment method, system, and device. This method comprehensively utilizes monitoring data on vehicle loads and bridge responses under vehicle loads to dynamically establish a mapping relationship between vehicle load and vehicle-induced response, thereby assessing the bridge's service status. Furthermore, by combining the development trends of vehicle loads and vehicle-induced responses, the method predicts the bridge's service status, thus solving the problems existing in the prior art.

[0006] A method for evaluating the service performance of bridges includes the following steps:

[0007] Historical vehicle monitoring video data and response monitoring data of the bridge under vehicle loads during its service life were acquired. Vehicle load spatiotemporal information and vehicle-induced response were extracted from the vehicle monitoring video data and response monitoring data, respectively. The vehicle load was reconstructed by matching the vehicle load spatiotemporal information with the vehicle-induced response.

[0008] A multibody dynamics model of the vehicle and a finite element model of the bridge are established. The equations of the multibody dynamics model of the vehicle and the dynamics of the bridge are solved simultaneously to obtain the theoretical response of the bridge under different vehicle loads.

[0009] The reconstructed vehicle load is used as input and the vehicle-induced response is used as output. A CNN-LSTM hybrid network with a self-attention mechanism is trained to construct a predictive model that reflects the mapping relationship between vehicle load and vehicle-induced response. During training, the control equations obtained by combining the vehicle multibody dynamics model and the bridge dynamics equations are introduced into the loss function as constraints to measure whether the predicted values ​​of the predictive model satisfy the physical constraints of the control equations. Specifically, the predicted values ​​of the predictive model are used as inputs to the control equations, and the norm of the calculated physical residuals is used as the physical constraints of the control equations.

[0010] The actual vehicle load is input into the prediction model to generate the corresponding vehicle-induced response prediction value; the bridge service performance is evaluated based on the theoretical bridge response under different vehicle loads and the vehicle-induced response prediction value corresponding to the actual vehicle load.

[0011] Furthermore, the reconstructed vehicle load is used as input and the vehicle-induced response is used as output to train a CNN-LSTM hybrid network with a self-attention mechanism to construct a predictive model reflecting the mapping relationship between vehicle load and vehicle-induced response. This specifically includes the following steps:

[0012] The reconstructed vehicle load and vehicle-induced response are input into a CNN-LSTM hybrid network with a self-attention mechanism. One-dimensional convolution operations are performed on the reconstructed vehicle load to capture its local spatial feature sequence. A multi-layer LSTM unit with a memory-gated enhancement mechanism processes the local spatial feature sequence sequentially time-step to learn the long-term temporal dependency between the reconstructed vehicle load and the vehicle-induced response, outputting the intermediate hidden state at each time step. Based on the intermediate hidden state at each time step, the importance weight of each time step for the current prediction task is calculated. The intermediate hidden states of all time steps are weighted and summed according to their importance weights to generate a context vector. The vehicle-induced response value is output by concatenating the context vector with the intermediate hidden state of the last time step. This process trains a prediction model reflecting the mapping relationship between vehicle load and vehicle-induced response.

[0013] Furthermore, the establishment of the vehicle multibody dynamics model and the bridge finite element model specifically includes the following steps:

[0014] Based on the finite element model, the bridge structure is discretized into beam elements, plate elements, or solid elements; based on the beam elements, plate elements, or solid elements, the element stiffness matrix is ​​established based on elasticity and the principle of virtual work; by assembling the element stiffness matrix, a finite element model of the bridge is formed to describe the displacement, stress, and strain of the bridge under load.

[0015] Based on multibody system dynamics, the vehicle is discretized into body, wheels and suspension components; the motion equations of each component are established based on Newton's second law, and a multibody dynamics model of the vehicle is constructed by considering the mass, spring stiffness and damping parameters of each component.

[0016] Furthermore, adopt Newmark-β The implicit dynamics method is used to solve the control equations of the vehicle multibody dynamics model and the bridge dynamics equations to obtain the theoretical response of the bridge under different vehicle loads.

[0017] Furthermore, before training the CNN-LSTM hybrid network with self-attention mechanism, the reconstructed vehicle load is used as input and the vehicle-induced response is used as output. The Fast Dynamic Time Warping (FastDTW) algorithm is then used to align the time series of the reconstructed vehicle load and the vehicle-induced response.

[0018] Furthermore, computer vision methods are used to analyze vehicle monitoring video data and extract spatiotemporal information about vehicle loads. This includes the following steps:

[0019] Use the OpenCV library to parse vehicle monitoring video data into continuous image frames;

[0020] After normalizing each frame image and adding batch dimensions, it is input into the pre-trained YOLOv5 model for forward propagation to obtain multiple bounding boxes, class confidence scores, and class labels for the vehicle.

[0021] By setting confidence thresholds and intersection-union ratio (IU) thresholds, non-maximum suppression is performed to filter out detection boxes with confidence levels below the thresholds and overlapping regions above the thresholds.

[0022] The DeepSORT algorithm is used for vehicle trajectory tracking, and the YOLOv5 backbone network is used to extract the feature vectors of vehicles within the filtered detection box.

[0023] The cascaded matching strategy associates the feature vector of the current frame with the trajectory of the previous frame, and uses Kalman filtering to predict the vehicle's state parameters in subsequent frames, thus obtaining the spatiotemporal information of the vehicle load.

[0024] Furthermore, intelligent time-series data processing methods are used to process the response monitoring data and extract the vehicle-induced response; specifically, the following steps are included:

[0025] The response monitoring data is cleaned, including filling in missing values, detecting anomalies, and smoothing and reducing noise.

[0026] The processed response monitoring data is input into a time-series data component separation algorithm based on wavelet transform or empirical mode decomposition to extract the time features and time-series patterns of the processed response monitoring data and generate vehicle-induced responses.

[0027] Furthermore, the evaluation of bridge service performance based on the theoretical bridge response under different vehicle loads and the predicted vehicle-induced response corresponding to actual vehicle loads is specifically expressed as follows:

[0028] ;

[0029] Among them, R 实际 and R 理论 These are the actual response and the theoretical response, respectively; if η > η 阈值 To determine the degradation of bridge stiffness, η 阈值 To set a threshold.

[0030] The present invention also includes a bridge service performance evaluation system, comprising:

[0031] The acquisition module is used to acquire historical vehicle monitoring video data on the bridge during its service period and response monitoring data generated by the bridge under vehicle loads. It extracts the spatiotemporal information of vehicle loads and vehicle-induced responses from the vehicle monitoring video data and response monitoring data, respectively. The vehicle loads are reconstructed by matching the spatiotemporal information of vehicle loads with the vehicle-induced responses.

[0032] The theoretical response calculation module is used to establish a vehicle multibody dynamics model and a bridge finite element model, and to solve the vehicle multibody dynamics model and the bridge dynamics equations simultaneously to obtain the theoretical response of the bridge under different vehicle loads.

[0033] The prediction model construction module is used to train a CNN-LSTM hybrid network with a self-attention mechanism, taking the reconstructed vehicle load as input and the vehicle-induced response as output, to construct a prediction model that reflects the mapping relationship between vehicle load and vehicle-induced response. Simultaneously, during training, the control equations, obtained by combining the vehicle multibody dynamics model and the bridge dynamics equations, are introduced as constraints into the loss function to measure whether the predicted values ​​of the prediction model satisfy the physical constraints of the control equations. Specifically, the predicted values ​​of the prediction model are used as inputs to the control equations, and the norm of the calculated physical residuals is used as the physical constraints of the control equations.

[0034] The evaluation module is used to input actual vehicle loads into the prediction model and generate corresponding vehicle-induced response prediction values; based on the theoretical response of the bridge under different vehicle loads and the vehicle-induced response prediction values ​​corresponding to actual vehicle loads, the service performance of the bridge is evaluated.

[0035] This invention provides a method for evaluating the service performance of bridges, which has the following beneficial effects:

[0036] This invention addresses the problems of missing vehicle weight information and difficulty in quantifying environmental interference in traditional methods by reconstructing vehicle loads and extracting vehicle-induced responses, thereby improving the completeness and accuracy of load information and thus enhancing the accuracy of bridge service performance assessment. Based on the reconstructed actual load and response data, a CNN-LSTM hybrid network combined with a self-attention mechanism is used to construct an actual load-response mapping model, capturing nonlinearity and temporal dependence to improve prediction accuracy. Furthermore, the control equations obtained by combining the vehicle multibody dynamics model and the bridge dynamics equations are used as constraints to achieve "dual-track verification" of the bridge response, enhancing the reliability of the assessment. This method is based on the analysis of the changes in the mapping relationship between vehicle load and vehicle-induced response, and performs bridge service performance assessment under vehicle loads and under actual operating vehicle load conditions. By improving the completeness and accuracy of load information and the prediction accuracy of actual vehicle-induced responses, it improves the accuracy of bridge service performance assessment, solving the problem that the ineffective coordination between vehicle load parameters and bridge vibration response easily leads to errors, resulting in low accuracy of bridge service performance assessment. Attached Figure Description

[0037] Figure 1 This is a general framework diagram of the bridge service performance evaluation method in this embodiment of the invention;

[0038] Figure 2This is a flowchart illustrating the vehicle load reconfiguration process in an embodiment of the present invention.

[0039] Figure 3 This is a schematic diagram illustrating the online evaluation and comparison of bridge service status in an embodiment of the present invention;

[0040] Figure 4 This is a schematic diagram illustrating the prediction of bridge service status in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0042] This invention proposes a method for evaluating bridge service performance. It employs computer vision technology and artificial intelligence algorithms to process video and bridge response time-series data, reconstructing vehicle loads and extracting vehicle-induced responses, and synchronously associating them to achieve a complete load-response information chain. Based on vehicle-bridge coupled computation theory, a theoretical vehicle load-vehicle-induced response mapping relationship is established. Through multi-source monitoring data, a regression algorithm is used to establish an actual vehicle load-vehicle-induced response mapping relationship. Based on the analysis of changes in the mapping relationship between vehicle load and vehicle-induced response, bridge service performance is evaluated under design vehicle loads and under actual operational vehicle load conditions. Furthermore, bridge service performance is predicted based on time-series prediction and extreme value extrapolation algorithms. This method addresses the problems of scattered bridge monitoring data, ineffective collaborative utilization, and weak timeliness and continuity in bridge service performance evaluation, thus contributing to precise and intelligent bridge maintenance decisions to support long-term service.

[0043] like Figure 1 As shown, the method specifically includes the following steps:

[0044] S1. Acquire vehicle monitoring video data and response monitoring data on the bridge; response monitoring data is physical quantity data reflecting the dynamic behavior of the bridge under vehicle load, collected by sensors (such as strain gauges, accelerometers, displacement gauges, etc.) installed on the bridge.

[0045] S2. Extract the spatiotemporal information of vehicle load and vehicle-induced response.

[0046] S2.1 Analyze bridge monitoring videos using computer vision (YOLOv5, DeepSORT) to obtain vehicle monitoring video data, including spatiotemporal information such as vehicle type, speed, and wheelbase; specifically including:

[0047] Video parsing: Use the OpenCV library to parse the video into consecutive image frames.

[0048] Frame preprocessing: (1) According to the input requirements of YOLOv5, the image frames with excessively high resolution are adjusted to a suitable size by downsampling. (2) Histogram equalization, CLAHE and other methods are used to enhance image contrast, and Gaussian filtering is used to remove noise and improve image quality. (3) The BGR format (OpenCV default) is converted to RGB format to meet the input requirements of the YOLOv5 model.

[0049] Vehicle detection: Load the pre-trained YOLOv5 model, perform preprocessing such as normalization and adding batch dimension for each frame image, and input it into the model for forward propagation to obtain multiple bounding boxes, class confidence and class label of the vehicle; then, by setting confidence threshold and intersection-over-union (IoU) threshold, perform non-maximum suppression to filter out detection boxes with low confidence and excessive overlap.

[0050] Multi-target tracking and trajectory generation: The DeepSORT algorithm is used for vehicle tracking, leveraging the YOLOv5 backbone network to extract feature vectors of vehicles within the detection box. A cascaded matching strategy is used to correlate the detection results of the current frame with the trajectories of previous frames, and Kalman filtering is used to predict and update the vehicle's position, speed, and other state parameters in subsequent frames. Vehicle parameters are extracted; the vehicle type is directly obtained from the YOLOv5 detection category label. Vehicle speed calculation requires first performing geometric calibration using camera mounting parameters to establish a mapping relationship between image pixels and actual physical distances; then, based on the actual coordinates of trajectory points of the same vehicle in two consecutive frames, the longitudinal displacement and time interval are calculated, and the vehicle speed is calculated using a formula. Wheelbase measurement involves locating the pixel coordinates of the front and rear axles of the vehicle, calculating the pixel distance, and then converting it to the actual wheelbase using a homography matrix. Finally, the extracted vehicle type, speed, and wheelbase information are output, combined with sensor data from the health monitoring system, to calibrate vehicle weight and axle load information, ultimately obtaining the vehicle's spatiotemporal distribution information.

[0051] S2.2 Analyze the response monitoring data using intelligent time-series data processing methods to extract vehicle-induced response signals. Specifically, this includes the following steps: cleaning the response monitoring data, including filling missing values, detecting anomalies, and smoothing and reducing noise; inputting the preprocessed data into a time-series data component separation algorithm based on wavelet transform or empirical mode decomposition to extract vehicle-induced response signals.

[0052] S3. Reconstruct vehicle loads and simulate traffic flow.

[0053] S3.1 Vehicle load reconfiguration:

[0054] Time synchronization: Time deviations are eliminated through spatiotemporal synchronization calibration (hardware clock synchronization, software interpolation) to ensure that the timestamps of video frames and sensor data are in the same time coordinate system.

[0055] Homography is used to achieve spatial mapping from pixel coordinates to bridge physical coordinates: converting the pixel coordinates of vehicles in the image into physical world coordinates on the bridge surface. Specifically, the homography matrix is ​​a projection transformation matrix used to describe the mapping relationship between the camera's pixel coordinate plane and the actual three-dimensional coordinates on the bridge structure. Pixel coordinates: refer to the two-dimensional coordinates (u, v) of the vehicle in the image captured by the camera, in pixels; bridge physical coordinates: refer to the actual three-dimensional coordinates (x, y, z) on the bridge structure, typically in meters. Mapping process: By calibrating the camera parameters and the bridge's geometric model, the homography matrix can map the pixels in the image (vehicle positions, etc.) to the bridge's physical coordinate system (the two-dimensional plane of the bridge surface xy, generally ignoring the z-axis height). This step is crucial for spatial synchronization, ensuring that the visual data is aligned with the bridge's physical space.

[0056] Cross-domain correlation of features: (1) Based on the vehicle-bridge coupling theory, a physical model is constructed. The vehicle information (position, speed, axle load) at a certain moment is input. The physical model can calculate the theoretical response that should be generated at a certain sensor position on the bridge. (2) Combine convolutional neural network to mine spatiotemporal nonlinear correlation: Spatiotemporal nonlinear correlation refers to time correlation and spatial correlation. Among them, time correlation is the dynamic matching of vehicle trajectory (position changes with time) and bridge response (strain changes with time); spatial correlation is the coupling relationship between the vehicle's position (x, y) on the bridge and the response at different positions of the bridge (such as the strain distribution of multiple sensor measurement points). Through convolutional neural network, the complex nonlinear mapping between vehicle load distribution and bridge response (how vehicle weight, speed, and position affect the reading of specific sensors) is learned from massive data, thereby making up for the error caused by the simplification of traditional physical model. (3) Multi-level data fusion (integrating trajectory, sensor, and structural parameter data) to achieve accurate matching between vehicle spatiotemporal trajectory and vehicle-induced bridge response, and complete the high-precision reconstruction of load distribution. The trajectory data includes the vehicle's spatiotemporal location information (x, y, t sequences acquired through visual tracking or GPS); sensor data includes physical quantities (strain, acceleration time history curves, displacement time history curves, and vibration frequencies, etc.) collected in real time by monitoring equipment on the bridge (strain gauges, accelerometers, displacement gauges); and structural parameter data includes the bridge's geometric parameters: span, cross-sectional dimensions, support locations, etc., material parameters: elastic modulus, density, etc., and boundary conditions: support stiffness, constraint type, etc. By fusing dynamic loads (vehicle trajectory), structural response (sensor data), and the bridge's inherent characteristics (structural parameters), a high-precision load distribution is reconstructed using data-driven methods (such as Kalman filtering and deep learning). Figure 2 As shown.

[0057] S3.2 In order to capture the dynamic interaction patterns of vehicles on the bridge more comprehensively and provide a more realistic dynamic load basis for the assessment of the bridge's service status, after completing the high-precision reconstruction of vehicle loads, it is necessary to use multi-dimensional data fusion and dynamic simulation algorithms to realize traffic flow projection.

[0058] Traffic flow simulation principle: Based on the statistical characteristics of actual traffic flow (vehicle type ratio, traffic density, and speed patterns), combined with bridge spatial constraints (length, number of lanes, and speed limit), vehicle driving behavior is transformed into a dynamic distribution of "time-longitudinal position" through spatiotemporal discretization modeling and stochastic process joint simulation. This generates a traffic flow evolution process that matches the actual traffic load, providing time-varying load input for vehicle-bridge response analysis.

[0059] S3.2.1 Basic data collection and preprocessing, including traffic flow statistics: vehicle type ratio, traffic flow time period characteristics; bridge structure data: bridge deck longitudinal length, number of lanes, design speed limit, bridge deck alignment; spatiotemporal reference calibration: through hardware clock synchronization and FastDTW algorithm, the time reference of traffic flow monitoring and bridge response monitoring is unified to ensure data time sequence alignment.

[0060] S3.2.2. Constructing dynamic traffic flow using microscopic traffic flow models and stochastic simulation algorithms: Vehicle generation and driving simulation: Based on the Monte Carlo method, vehicle arrival sequences are randomly generated according to traffic flow statistics (simulating traffic flow randomness). Vehicle speeds are calculated using a following model (such as an intelligent driving model IDM), and the change in vehicle longitudinal position over time is simulated using a spatiotemporal recursive formula.

[0061] ;

[0062] in for t Time of the first i The longitudinal position of the vehicle To correspond to vehicle speed, acceleration and deceleration are considered. Traffic flow conflict verification: Collision avoidance rules are introduced, with a minimum safe distance. Adjust vehicle speed in real time.

[0063] S3.2.3. Output of Derivation Results and Calibration Iterative Calculation: Scatter plots of vehicle longitudinal position-time distribution traffic flow before and after the derivation are generated, intuitively presenting the dynamic distribution of vehicles on the bridge at different times. By comparing with actual traffic flow monitoring data, Kalman filtering is used to calibrate the model parameters online, ensuring the consistency between the derivation traffic flow and the actual traffic load. This provides a real, continuous, time-varying load input, including the dynamic driving trajectory of vehicles, the traffic flow distribution at different times, and the characteristics of load changes, for establishing the theoretical vehicle load-vehicle response mapping relationship, making the establishment of the theoretical mapping relationship more realistic, targeted, and accurate.

[0064] S4. Establish the theoretical vehicle load-vehicle-induced response mapping relationship: Based on the bridge finite element model and vehicle dynamics model, the theoretical response of the bridge under vehicle load is obtained by using the vehicle-bridge coupling calculation theory, and the theoretical vehicle load-vehicle-induced response mapping relationship is established; the bridge finite element calculation model and vehicle load dynamics model include models from one-dimensional to three-dimensional and from simple to refined.

[0065] (1) Establishing a bridge finite element model: The bridge was modeled with high precision using both the independently developed vehicle-bridge coupled dynamic analysis software and the large-scale commercial finite element software ANSYS. The high-precision models were compared and verified. In the model construction stage, the bridge finite element model discretized the bridge structure into beam elements, plate elements or solid elements. Based on the principles of elasticity and virtual work, the element stiffness matrix was established. The mechanical properties (stiffness, mass, damping matrix) of each discrete element (beam element, plate element, solid element, etc.) were assembled according to the actual structural connection relationship to form the dynamic equation of the global structure, so as to describe the displacement, stress and strain of the bridge under load.

[0066] (2) Establishing a multibody dynamics model of the vehicle: Based on multibody system dynamics, the vehicle is simplified into a multi-degree-of-freedom system, including the body, suspension, and wheels. The motion equations for each component are established based on Newton's second law, considering mass, spring stiffness, and damping parameters to characterize the vehicle's vertical, pitch, and lateral vibration characteristics. Specifically, the vehicle is discretized into a body, wheel, and suspension system, considering the degrees of freedom, stiffness, and damping between each component; the motion assumptions only consider the vehicle's vertical vibration, pitch vibration, and lateral and roll motions, ignoring longitudinal vibration. Motion equations are established for each degree of freedom based on Newton's second law or Lagrange's equations, ultimately yielding the vehicle's motion equations.

[0067] (3) Establish vehicle-bridge coupling relationship: In the process of vehicle-bridge coupling calculation, the coupling relationship between the vehicle and the bridge is established based on displacement coordination and force balance conditions. By assuming that the wheels and the bridge deck always keep in contact and that the contact force and the relative displacement are linearly related, the vehicle load is applied to the corresponding position of the bridge finite element model in the form of nodal forces.

[0068] (4) The Newmark-β method is used to solve the vehicle-bridge coupled vibration equation, calculate the theoretical response (strain, acceleration, etc.) of the bridge under different vehicle loads, and then establish the theoretical vehicle load-vehicle response mapping relationship.

[0069] S5. Establish the actual vehicle load-vehicle response mapping relationship: Based on the reconstructed vehicle load information and the extracted vehicle response information, use regression algorithms (including linear regression algorithms and various nonlinear regression algorithms) to establish the actual vehicle load-vehicle response mapping relationship.

[0070] Its establishment process includes the following steps:

[0071] (1) The Fast Dynamic Time Warping (FastDTW) algorithm is used to process the time series of measured load and response data. FastDTW divides the original time series into multiple subsequences, reduces the computational complexity by using downsampling technology, and uses Euclidean distance to measure the similarity between subsequences. By constructing a path matrix, the optimal alignment path is found, which solves the problem of time series misalignment caused by changes in vehicle speed and differences in sensor sampling frequency, and achieves accurate alignment of load and response data in the time dimension.

[0072] (2) A load-response nonlinear mapping model is established using deep learning technology. First, the load and response data after spatiotemporal alignment are normalized and divided into training, validation, and test sets. Multi-dimensional innovations are made on the basic LSTM network architecture: the self-attention mechanism is introduced to enhance the feature extraction of key time steps, making up for the shortcomings of traditional LSTM in processing all time step input data equally. In the load-response relationship, some key time points (such as the peak strain when the vehicle crosses the bridge) have a greater impact on the prediction. Combining 1D-CNN to capture the local spatial features of load data, a CNN-LSTM hybrid network is constructed to simultaneously model spatiotemporal dependence. Although CNN and LSTM are common models in deep learning, this invention combines them and introduces a self-attention mechanism to innovate and improve the efficiency of spatiotemporal feature extraction. At the same time, the control equations in the vehicle-bridge coupled vibration analysis theory are embedded as physical constraints into the loss function to realize a physical information-driven deep learning framework. This is used to measure whether the predicted values ​​of the prediction model satisfy the physical constraints of the governing equations; wherein, by using the predicted values ​​of the prediction model as inputs to the governing equations, the physical residuals are calculated, and the norm of the physical residuals is used as the physical constraints of the governing equations.

[0073] The model employs multi-layer LSTM units to capture the long-term dependencies of load sequences, maps response predictions through fully connected layers, uses MSE as the loss function, employs an Adam optimizer for parameter updates, and utilizes a validation set to adjust hyperparameters to prevent overfitting. Furthermore, wavelet transform is used to perform time-frequency decomposition of non-stationary signals, followed by frequency-band modeling to enhance the representation of transient impact loads. Finally, after multiple rounds of training, an intelligent load-response prediction model with both theoretical innovation and engineering practicality is output. After training, actual vehicle load sequences are input into the model to generate corresponding predicted response sequences, thus establishing a nonlinear mapping relationship between load and response. Post-processing analysis of the model output (combined with physical constraints) quantifies key relationships and applies them to a bridge monitoring system to achieve real-time prediction and risk assessment, forming a complete closed-loop application.

[0074] S6. Bridge Service Condition Assessment:

[0075] By comparing the theoretical and practical vehicle load-vehicle response mapping relationships, the degradation status of bridge stiffness is determined. The stiffness degradation index is shown in the following formula. The service status of the bridge under design vehicle loads and actual vehicle load conditions is also evaluated. Figure 3 As shown.

[0076] ;

[0077] Among them, R 实际 and R 理论 These represent the actual and theoretical responses, respectively, η 阈值 To set a threshold; if the stiffness degradation index η > η 阈值 The bridge stiffness degradation is determined, and the bridge status is classified based on the degree of degradation (e.g., normal, slightly degraded, severely degraded).

[0078] S7. Bridge Service Status Prediction:

[0079] Based on historical vehicle load data, future traffic volume is predicted, and future traffic flow is simulated using sampling methods. Through comprehensive analysis of historical traffic data, real-time monitoring information, and traffic planning parameters, probability statistics and stochastic process theory are applied to generate traffic flow data that conforms to the characteristics of future scenarios. During simulation, Monte Carlo simulation can be used to extract the probability distribution functions of key parameters (vehicle volume, vehicle type ratio, vehicle speed distribution, etc.) from historical traffic flow data, and possible traffic flow scenarios can be generated through random sampling. Alternatively, a Markov chain model can be used to simulate the dynamic evolution of traffic flow under different states (smooth flow, congestion, etc.) based on the traffic state transition probability matrix. The service performance of bridges is predicted by combining future traffic flow and predicted vehicle-induced responses. Figure 4 As shown.

[0080] It predicts future traffic volume and uses sampling methods to simulate future traffic flow. The prediction methods include simple logistic regression prediction models and machine learning-based algorithms, and the sampling algorithms include MCMC, LHS and other sampling algorithms.

[0081] Historical vehicle response time series data is used to predict future vehicle response, and the prediction methods include LSTM, Bi-LSTM, and Transformer.

[0082] The effects that this invention can achieve are: (1) By jointly analyzing vehicle monitoring videos and response data, the complete information chain from load to response can be grasped, significantly improving the ability to perceive the service status of bridges; (2) The bridge service performance evaluation method established in this invention can be applied to various types of bridge structures and has broad application prospects; (3) The bridge service performance prediction method proposed in this invention accurately analyzes the development trend of vehicle load and vehicle-induced response, and can accurately grasp the future service status of bridges; (4) The bridge service performance evaluation and prediction method proposed in this invention can provide important support for the rapid and accurate formulation of bridge structure maintenance strategies.

[0083] Based on the same inventive concept, this invention also proposes a bridge service performance evaluation system, comprising:

[0084] The acquisition module is used to acquire historical vehicle monitoring video data on the bridge during its service period and response monitoring data generated by the bridge under vehicle loads. It extracts the spatiotemporal information of vehicle loads and vehicle-induced responses from the vehicle monitoring video data and response monitoring data, respectively. The vehicle loads are reconstructed by matching the spatiotemporal information of vehicle loads with the vehicle-induced responses.

[0085] The theoretical response calculation module is used to establish a multibody dynamics model of the vehicle and a finite element model of the bridge, and to solve the equations of the multibody dynamics model of the vehicle and the dynamics of the bridge simultaneously to obtain the theoretical response of the bridge under different vehicle loads.

[0086] The prediction model construction module is used to train a CNN-LSTM hybrid network with a self-attention mechanism, taking the reconstructed vehicle load as input and the vehicle-induced response as output, to construct a prediction model that reflects the mapping relationship between vehicle load and vehicle-induced response. Simultaneously, during training, the control equations, obtained by combining the vehicle multibody dynamics model and the bridge dynamics equations, are introduced into the loss function as constraints to measure whether the predicted values ​​of the prediction model satisfy the physical constraints of the control equations. Specifically, the predicted values ​​of the prediction model are used as inputs to the control equations, and the norm of the calculated physical residuals is used as the physical constraints of the control equations.

[0087] The evaluation module is used to input actual vehicle loads into the prediction model and generate corresponding vehicle-induced response prediction values; based on the theoretical response of the bridge under different vehicle loads and the vehicle-induced response prediction values ​​corresponding to actual vehicle loads, the service performance of the bridge is evaluated.

[0088] The present invention also proposes a computer device for evaluating the service performance of bridges, comprising: a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the bridge service performance evaluation method.

[0089] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for evaluating the service performance of bridges, characterized in that, Includes the following steps: Historical vehicle monitoring video data and response monitoring data of the bridge under vehicle loads during its service life were acquired. Vehicle load spatiotemporal information and vehicle-induced response were extracted from the vehicle monitoring video data and response monitoring data, respectively. The vehicle load was reconstructed by matching the vehicle load spatiotemporal information with the vehicle-induced response. A multibody dynamics model of the vehicle and a finite element model of the bridge are established. The equations of the multibody dynamics model of the vehicle and the dynamics of the bridge are solved simultaneously to obtain the theoretical response of the bridge under different vehicle loads. The establishment of the vehicle multibody dynamics model and the bridge finite element model specifically includes the following steps: discretizing the bridge structure into beam elements, plate elements, or solid elements based on the finite element model; establishing element stiffness matrices based on elasticity and the principle of virtual work, according to the beam elements, plate elements, or solid elements; forming a bridge finite element model to describe the displacement, stress, and strain of the bridge under load by assembling the element stiffness matrices; discretizing the vehicle into body, wheels, and suspension components based on multibody system dynamics; establishing the equations of motion for each component based on Newton's second law; and constructing the vehicle multibody dynamics model by considering the mass, spring stiffness, and damping parameters of each component; and using... Newmark-β The implicit dynamics method is used to solve the control equations after combining the vehicle multibody dynamics model and the bridge dynamics equations to obtain the theoretical response of the bridge under different vehicle loads. The reconstructed vehicle load is used as input and the vehicle-induced response is used as output. A CNN-LSTM hybrid network with a self-attention mechanism is trained to construct a predictive model that reflects the mapping relationship between vehicle load and vehicle-induced response. During training, the control equations obtained by combining the vehicle multibody dynamics model and the bridge dynamics equations are introduced into the loss function as constraints to measure whether the predicted values ​​of the predictive model satisfy the physical constraints of the control equations. Specifically, the predicted values ​​of the predictive model are used as inputs to the control equations, and the norm of the calculated physical residuals is used as the physical constraints of the control equations. The actual vehicle load is input into the prediction model to generate the corresponding vehicle-induced response prediction value; the bridge service performance is evaluated based on the theoretical bridge response under different vehicle loads and the vehicle-induced response prediction value corresponding to the actual vehicle load.

2. The bridge service performance evaluation method according to claim 1, characterized in that, The process involves using the reconstructed vehicle load as input and the vehicle-induced response as output to train a CNN-LSTM hybrid network with a self-attention mechanism to construct a predictive model that reflects the mapping relationship between vehicle load and vehicle-induced response. This process specifically includes the following steps: The reconstructed vehicle load and vehicle-induced response are input into a CNN-LSTM hybrid network with a self-attention mechanism. One-dimensional convolution operations are performed on the reconstructed vehicle load to capture its local spatial feature sequence. A multi-layer LSTM unit with a memory-gated enhancement mechanism processes the local spatial feature sequence sequentially time-step to learn the long-term temporal dependency between the reconstructed vehicle load and the vehicle-induced response, outputting the intermediate hidden state at each time step. Based on the intermediate hidden state at each time step, the importance weight of each time step for the current prediction task is calculated. The intermediate hidden states of all time steps are weighted and summed according to their importance weights to generate a context vector. The vehicle-induced response value is output by concatenating the context vector with the intermediate hidden state of the last time step. This process trains a prediction model reflecting the mapping relationship between vehicle load and vehicle-induced response.

3. The bridge service performance evaluation method according to claim 1, characterized in that, Before training the CNN-LSTM hybrid network with self-attention mechanism, the reconstructed vehicle load is used as input and the vehicle-induced response is used as output. The Fast Dynamic Time Warping (FastDTW) algorithm is then used to align the time series of the reconstructed vehicle load and vehicle-induced response.

4. The bridge service performance evaluation method according to claim 1, characterized in that, The analysis of vehicle monitoring video data using computer vision methods to extract spatiotemporal information about vehicle load includes the following steps: Use the OpenCV library to parse vehicle monitoring video data into continuous image frames; After normalizing each frame image and adding batch dimensions, it is input into the pre-trained YOLOv5 model for forward propagation to obtain multiple bounding boxes, class confidence scores, and class labels for the vehicle. By setting confidence thresholds and intersection-union ratio (IU) thresholds, non-maximum suppression is performed to filter out detection boxes with confidence levels below the thresholds and overlapping regions above the thresholds. The DeepSORT algorithm is used for vehicle trajectory tracking, and the YOLOv5 backbone network is used to extract the feature vectors of vehicles within the filtered detection box. The cascaded matching strategy associates the feature vector of the current frame with the trajectory of the previous frame, and uses Kalman filtering to predict the vehicle's state parameters in subsequent frames, thus obtaining the spatiotemporal information of the vehicle load.

5. The bridge service performance evaluation method according to claim 1, characterized in that, The response monitoring data is processed using intelligent time-series data processing methods to extract the vehicle-induced response; specifically, the following steps are included: The response monitoring data is cleaned, including filling in missing values, detecting anomalies, and smoothing and reducing noise. The processed response monitoring data is input into a time-series data component separation algorithm based on wavelet transform or empirical mode decomposition to extract the time features and time-series patterns of the processed response monitoring data and generate vehicle-induced responses.

6. The bridge service performance evaluation method according to claim 1, characterized in that, The bridge service performance is evaluated based on the theoretical response of the bridge under different vehicle loads and the predicted vehicle-induced response corresponding to actual vehicle loads. Specifically, this is expressed as follows: ； Among them, R 实际 and R 理论 These are the actual response and the theoretical response, respectively; if η > η 阈值 To determine the degradation of bridge stiffness, η 阈值 To set a threshold.

7. A bridge service performance evaluation system, characterized in that, include: The acquisition module is used to acquire historical vehicle monitoring video data on the bridge during its service period and response monitoring data generated by the bridge under vehicle loads, and extract vehicle load spatiotemporal information and vehicle-induced response from the vehicle monitoring video data and response monitoring data, respectively. The vehicle load is reconstructed by matching the spatiotemporal information of the vehicle load with the vehicle-induced response. The theoretical response calculation module is used to establish a vehicle multibody dynamics model and a bridge finite element model, and to solve the vehicle multibody dynamics model and the bridge dynamics equations simultaneously to obtain the theoretical response of the bridge under different vehicle loads. The establishment of the vehicle multibody dynamics model and the bridge finite element model specifically includes the following steps: discretizing the bridge structure into beam elements, plate elements, or solid elements based on the finite element model; establishing element stiffness matrices based on elasticity and the principle of virtual work, according to the beam elements, plate elements, or solid elements; forming a bridge finite element model to describe the displacement, stress, and strain of the bridge under load by assembling the element stiffness matrices; discretizing the vehicle into body, wheels, and suspension components based on multibody system dynamics; establishing the equations of motion for each component based on Newton's second law; and constructing the vehicle multibody dynamics model by considering the mass, spring stiffness, and damping parameters of each component; and using... Newmark-β The implicit dynamics method is used to solve the control equations after combining the vehicle multibody dynamics model and the bridge dynamics equations to obtain the theoretical response of the bridge under different vehicle loads. The prediction model construction module is used to train a CNN-LSTM hybrid network with a self-attention mechanism, taking the reconstructed vehicle load as input and the vehicle-induced response as output, to construct a prediction model that reflects the mapping relationship between vehicle load and vehicle-induced response. Simultaneously, during training, the control equations, obtained by combining the vehicle multibody dynamics model and the bridge dynamics equations, are introduced as constraints into the loss function to measure whether the predicted values ​​of the prediction model satisfy the physical constraints of the control equations. Specifically, the predicted values ​​of the prediction model are used as inputs to the control equations, and the norm of the calculated physical residuals is used as the physical constraints of the control equations. The evaluation module is used to input actual vehicle loads into the prediction model and generate corresponding vehicle-induced response prediction values; based on the theoretical response of the bridge under different vehicle loads and the vehicle-induced response prediction values ​​corresponding to actual vehicle loads, the service performance of the bridge is evaluated.

8. A computer device for evaluating the service performance of bridges, characterized in that, include: A memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the bridge service performance evaluation method according to any one of claims 1-6.

Images