Bridge Early Warning Method and System Based on Physics-Informed Neural Network and Machine Vision

Abstract

The present invention proposes a bridge warning method and system based on a physics-informed neural network and machine vision, which relates to the technical field of bridge warning, and includes: setting up a visual data acquisition system on the bridge; designing a known load calibration test to collect a preliminary calibration data set; collecting a preliminary random load data set; using an error correction algorithm to obtain a corrected random load data set; establishing a causal association database covering the dynamic response relationship between vehicles and bridges; constructing a physics-informed neural network model, and training the physics-informed neural network based on the causal association database to obtain a trained physics-informed neural network model; collecting bridge dynamic response data in real time, inputting it into the trained physics-informed neural network model, online evaluating the operating state of the bridge, checking for abnormal responses and generating warning signals. The present invention combines physical laws with machine learning algorithms to achieve precise online evaluation and abnormal detection of the operating state of bridges.

Description

Technical Field

[0001] The present invention relates to the technical field of bridge early warning, and particularly to a bridge early warning method and system based on a physics-informed neural network and machine vision. Background Art

[0002] As a key hub in traffic construction, a bridge is not only an important infrastructure for a country's economic and social development, but also a comprehensive reflection of a country's or region's economic strength, scientific and technological level, and productivity development. In the development process of today's economic society, bridges play a crucial role. As major engineering structures, bridges usually have a service life of several decades to a hundred years. However, the combined effects of factors such as environmental erosion, material aging, long-term load action, fatigue effect, and mutation effect will inevitably lead to the accumulation of bridge structure damage and the attenuation of resistance, thereby reducing its ability to resist natural disasters and even normal loads, and even triggering catastrophic accidents in extreme cases.

[0003] These major safety accidents have brought huge losses to the country's economic property and people's lives. To ensure the safe operation of bridges, bridge management units usually adopt the technical means of manual inspection. However, the data obtained by manual inspection lacks continuity and real-time nature, and it is often impossible to detect diseases in time, especially some hidden diseases are difficult to detect, making it difficult to effectively ensure the safety of bridges. Investigations by the Federal Highway Commission of the United States show that "56% of the assessment results obtained by manual visual inspection are inaccurate". For example, although the Mississippi River Bridge in the United States was inspected in 2005 and 2006 respectively and no structural safety hazards were found at that time, it still collapsed in 2007. This series of serious bridge safety accidents highlight the deficiencies of manual detection technology, and also highlight the necessity and importance of bridge dynamic early warning systems.

[0004] Currently, bridge dynamic early warning systems mainly include the following types, but their applications are limited by their respective deficiencies:

[0005] Method based on threshold determination: Traditional bridge health monitoring systems usually rely on methods of fixed threshold determination. These methods judge the monitoring data by presetting safety thresholds, such as deflection, stress, or vibration amplitude. Once the detected value exceeds the threshold range, the system will trigger an alarm. However, this method has obvious defects: Insufficient accuracy: The state of a bridge is usually affected by various factors such as load, environmental temperature, and humidity. A single threshold is difficult to comprehensively reflect the actual health status of the bridge, which may lead to false alarms or missed alarms. Unable to respond dynamically: The threshold setting is usually static, while the operating state of the bridge is highly dynamic. For example, the dynamic response caused by instantaneous overloaded vehicles cannot be captured in time, and important abnormal information is easily missed.

[0006] Method for finite element model updating: The finite element model (FEM) is a commonly used theoretical tool in bridge structure analysis. By establishing a physical model of the bridge, its responses under different loadings can be simulated. However, the following problems exist in the practical application of the finite element model: Time-consuming: The establishment and updating of the model require a large amount of manual participation and experimental data support. Especially when dealing with complex bridge structures or multi-loading conditions, the calculation process is extremely time-consuming and difficult to meet the real-time requirements. Difficulty in integrating multiple factors: The finite element model usually only targets specific mechanical properties (such as statics or dynamics), while the operating state of a bridge involves complex multi-physical field coupling effects, such as vehicle loads, wind loads, temperature changes, etc. These factors are often simplified or ignored in traditional models, resulting in insufficient accuracy and reliability of the analysis results.

[0007] Data-driven methods: In recent years, data-driven bridge monitoring methods have gradually emerged. These methods evaluate the bridge state by analyzing a large amount of sensor data or monitoring historical data and using machine learning or statistical methods. Although these methods have certain flexibility, the following limitations still exist: Inability to integrate physical meaning: Pure data-driven models only focus on the correlations between data and ignore the clear physical laws existing in the bridge structure, such as the load-deflection relationship or vibration mode characteristics. The prediction results of this method lack physical basis and may fail in practical applications. Ignoring causal relationships: Data-driven methods are mostly based on empirical statistics, focusing on result prediction, and do not deeply analyze the causal relationship between input loads and bridge responses, making it difficult to support the root cause diagnosis of bridge anomalies.

[0008] Deep learning methods: Due to their powerful non-linear fitting ability, deep learning techniques have been applied to the field of bridge health monitoring in recent years, attempting to analyze and predict complex data through neural networks. However, this method also has obvious defects: The "black box" problem of the model: The neural network structure is complex, and the internal operating mechanism of the model is difficult for users to understand, resulting in poor interpretability of the results and reducing users' trust in the prediction results, especially prominent in bridge monitoring scenarios with extremely high safety requirements. Difficulty in reflecting physical laws: The training of deep learning models mainly relies on data and lacks physical constraint descriptions of bridge structures and mechanical properties. Although the model performs well in data fitting, it is prone to deviations or even errors when generalized to unseen conditions.

[0009] In summary, traditional bridge health monitoring methods have obvious deficiencies in terms of accuracy, real-time performance, multi-factor analysis, and physical law integration, restricting their application in complex bridge dynamic assessment. Therefore, there is an urgent need for a new method that can combine physical laws with efficient algorithms to make up for the defects of existing methods and improve the reliability and real-time performance of the bridge warning system. Summary of the Invention

[0010] In view of this, the present invention proposes a bridge warning method and system based on physics-informed neural networks and machine vision. By collecting bridge dynamic response data and vehicle trajectory information in real time, and combining physical laws with machine learning algorithms, it realizes accurate online assessment and anomaly detection of the bridge operating state.

[0011] The technical solution of the present invention is implemented as follows:

[0012] On the one hand, the present invention provides a bridge warning method based on physics-informed neural networks and machine vision, including:

[0013] S1. Set up a visual data acquisition system on the bridge, including a bridge dynamic displacement acquisition device and a trajectory vision acquisition device;

[0014] S2. Design a known load calibration test, including selecting a standardized specific load vehicle and designing a predetermined trajectory. The visual data acquisition system collects dynamic response data under known load conditions to form a preliminary calibration data set;

[0015] S3. Collect vehicle trajectories, vehicle types and corresponding bridge response data under natural traffic flow to form a preliminary random load data set;

[0016] S4. Conduct the known load calibration test again. The visual data acquisition system collects the bridge dynamic response to form a second calibration data set. Compare the preliminary calibration data set and the second calibration data set to identify systematic errors and random errors, and use an error correction algorithm to correct the preliminary random load data set to obtain a corrected random load data set;

[0017] S5. Through the corrected random load data set, the preliminary calibration data set and the second calibration data set, establish a causal association database covering the dynamic response relationship between vehicles and bridges;

[0018] S6. Construct a physics-informed neural network model. Based on the causal association database, use a supervised learning method to train the physics-informed neural network and optimize the network parameters to obtain a trained physics-informed neural network model;

[0019] S7. Collect bridge dynamic response data in real time, input it into the trained physics-informed neural network model, conduct an online assessment of the bridge operating state, check for abnormal responses and generate warning signals.

[0020] On the basis of the above technical solution, preferably, the bridge dynamic displacement acquisition device includes a digital image correlation acquisition device deployed under the bridge, which has a multi-scale image acquisition unit and a real-time optical flow algorithm module for non-contact monitoring of the deflection changes of key parts of the bridge under different frequency loads; the trajectory vision acquisition device includes a high-resolution camera array and a multi-view synchronous collector deployed on the bridge deck and at the bridgehead. The camera array obtains vehicle shape, speed, and driving trajectory information through the Yolo v8 object detection model, and fuses these visual data with the data output by the bridge dynamic displacement acquisition device under the same time reference.

[0021] On the basis of the above technical solution, preferably, the error correction algorithm in step S4 is as follows:

[0022] Step 1: Define the initial error vector , where is the initial system error, is the initial random error, is the initial environmental error;

[0023] Step 2: At the k-th iteration correction, calculate the intermediate correction function , where, is the preliminary calibration data set, is the second calibration data set, is the adaptive convergence factor, Based on the multi-dimensional non-linear mapping, the previous iteration result and , are coupled and calculated to output a new correction vector ;

[0024] Step 3: Substitute the intermediate correction result into the correction formula:

[0025] ,

[0026] In the formula, is the exponential penalty factor, represents the vector norm, is the quadratic coupling correction function, which performs non-linear compensation according to the distribution characteristics of vehicle load effects and environmental disturbance characteristics;

[0027] Step 4: After several iterations, when or is lower than the predetermined threshold, output the final correction result , and is applied to the preliminary random load data set to obtain the corrected random load data set.

[0028] Based on the above technical solution, preferably, step S5 includes:

[0029] S51. Integrate the corrected random load data set, the preliminary calibration data set, and the second calibration data set, and associate the vehicle load parameters and the bridge response parameters according to the time stamp to obtain synchronous records of multi-source data;

[0030] S52. Perform multi-level feature extraction on the above synchronous records to obtain the data layer, the feature layer, and the association layer respectively; where the data layer contains the original data of the vehicle node, the response node, and the environment node, the feature layer includes frequency domain and time domain features, and the association layer is used to record the direct or indirect causal relationships between nodes;

[0031] S53. Extract the main causal chains based on the causal strength to form the direct and indirect causal relationships between the vehicle load and the bridge response; use an optimization algorithm to optimize the accuracy and integrity of the causal chain;

[0032] S54. Enter the constructed causal chain and association relationship into the graph database to form a causal association database covering a multi-level data structure.

[0033] Based on the above technical solution, preferably, the causal strength calculation formula is:

[0034] ,

[0035] In the formula, represents the causal strength from node i to node j; is the mutual information between node i and node j; is the direction consistency from node i to node j; is the non-linear parameter for adjusting the causal strength;

[0036] Use the Bayesian optimization algorithm to adjust the weights of the extracted causal chain, and the formula is defined as:

[0037] ,

[0038] Among them, is the final causal weight from node i to node j; m is the total number of nodes that have a causal relationship with node i.

[0039] Based on the above technical solution, preferably, the physics-informed neural network model is a multi-branch hybrid structure, including:

[0040] Physical branch: includes an input layer, two fully connected hidden layers, and an output layer; each hidden layer applies a physical constraint activation function; the physical branch is used to process the embedded bridge mechanical equations and output the physical constraint results;

[0041] Data branch: It includes an input layer, a multi-scale convolutional layer, an attention mechanism layer, and a fully-connected hidden layer; the multi-scale convolutional layer is used to extract multi-scale features; the attention mechanism layer is used to autonomously learn the importance of different features; the data branch is used to process high-frequency visual data and extract deep feature information;

[0042] Fusion layer: It includes an adaptive weight fusion unit that dynamically weights and fuses the outputs of the physical branch and the data branch; the fusion layer realizes the effective integration of information by learning weight parameters;

[0043] Output layer: It synthesizes the fusion results of the physical branch and the data branch and outputs the evaluation of the bridge operation state.

[0044] On the basis of the above technical solution, preferably, the bridge mechanical equations in the physical branch include static constraints and dynamic constraints, where:

[0045] The static constraint formula is:

[0046] ,

[0047] In the formula, x is the spatial coordinate position, that is, the position parameter of each measuring point in the bridge; represents the mass matrix of the bridge; represents the damping matrix; represents the stiffness matrix; represents the displacement vector of the bridge; is the external load vector; represents the second-order partial derivative with respect to time; represents the first-order partial derivative with respect to time;

[0048] The dynamic constraint formula is:

[0049] ,

[0050] In the formula, represents the natural frequency of the bridge; represents the vibration mode vector of the bridge; is the stiffness matrix.

[0051] On the basis of the above technical solution, preferably, the loss function used when training the physical information neural network model is:

[0052] ,

[0053] ,

[0054] ,

[0055] ,

[0056] In the formula, is the total loss function; is the data-driven loss; is the physical constraint loss; is the regularization loss; , , are adaptive weight coefficients; represents the s-th sample; is the number of samples; represents the actual value of the sample, that is, the true response data of the bridge at the s-th position; represents the model prediction value, that is, the bridge response data predicted by the model; is the adjustment parameter; , , , , are the physical parameters corresponding to sample s; is the normalization parameter; is the regularization coefficient; is the weight parameter of the -th layer in the model; Q is the number of network layers.

[0057] Based on the above technical solution, preferably, step S7 includes:

[0058] S71. Real-time acquisition of the dynamic response data generated by the bridge during operation through the visual data acquisition system, and performing time synchronization and preliminary preprocessing on the acquired dynamic response data;

[0059] S72. Convert the preprocessed dynamic response data into the input format required by the physics-informed neural network model, and input it into the trained physics-informed neural network model. Calculate the current operating state parameters of the bridge through the forward propagation process. The model outputs the health assessment results of the bridge according to the embedded physical constraints and causal association database, including the current deflection, vibration mode, and potential structural anomaly indicators;

[0060] S73. Based on the health assessment results output by the model, compare with the preset safety standards and thresholds to determine whether the bridge operating state is normal, and use a multi-level assessment mechanism to distinguish between minor anomalies, obvious anomalies, and serious anomalies;

[0061] S74. Generate warning signals of corresponding levels according to the abnormal evaluation results, including first-level warning, second-level warning, and third-level warning, corresponding to slight abnormality, obvious abnormality, and severe abnormality respectively. Each level of warning signal includes specific abnormal parameter values, abnormal categories, and recommended response measures, and transmits the warning signals to the bridge management and operation and maintenance system in real time through wireless communication or wired network.

[0062] On the other hand, the present invention also provides a bridge warning system based on physics-informed neural network and machine vision. The system is used to execute the method described in any one of the above, and the system includes:

[0063] A data acquisition module, integrating a bridge dynamic displacement acquisition device and a trajectory vision acquisition device, for real-time acquisition of bridge dynamic response data and vehicle trajectory data;

[0064] A dual calibration module, performing at least two known load calibration tests and natural traffic flow acquisitions, evaluating the difference of random load data, and calling an error correction algorithm for dynamic correction;

[0065] A causal association database module, based on a graph database structure containing vehicle nodes, bridge response nodes, and environmental nodes, storing multi-source heterogeneous data, constructing causal relationships, and performing iterative updates;

[0066] A physics-informed neural network module, containing a physics-informed neural network model with a multi-branch hybrid structure, for evaluating the bridge operation state;

[0067] A real-time evaluation and warning module, inputting dynamic response data into the trained neural network model in real time; generating multi-level warning signals according to the model output, and performing signal transmission and feedback through a warning setting unit;

[0068] A communication and storage module, for transmitting the warning signals to the bridge management and operation and maintenance system;

[0069] A user interface and monitoring module, providing a visual monitoring interface for the bridge operation state; displaying real-time data, warning information, and historical records.

[0070] The present invention has the following beneficial effects compared with the prior art:

[0071] (1) By combining the physics-informed neural network and machine vision technologies, the present invention constructs a bridge warning method for multi-source data fusion. This method uses a dual calibration mechanism and an error correction algorithm to improve data quality, stores a multi-level data structure through a causal association database, and uses a physics-informed neural network model with a multi-branch hybrid structure for state evaluation, realizing accurate monitoring and timely warning of the bridge operation state, and effectively overcoming the deficiencies of traditional warning methods in terms of accuracy, real-time performance, and physical interpretability;

[0072] (2) The visual data acquisition system of the present invention consists of a digital image correlation acquisition device and a high-resolution camera array. Combined with the Yolo v8 object detection model, it realizes non-contact real-time monitoring of the dynamic displacement of the bridge and the vehicle trajectory, avoiding the problem of difficult installation and maintenance of traditional contact sensors, and improving the accuracy and reliability of data acquisition at the same time;

[0073] (3) The error correction algorithm designed in the present invention defines a three-dimensional error vector of systematic error, random error and environmental error, and uses multi-dimensional non-linear mapping and quadratic coupling correction function for iterative optimization, effectively eliminating various errors in the data acquisition process and significantly improving the accuracy of monitoring data;

[0074] (4) The causal association database constructed in the present invention adopts a graph database structure, calculates the causal strength through mutual information and direction consistency, and uses Bayesian optimization for weight adjustment, realizing the accurate expression and efficient storage of the complex causal relationship between vehicle load and bridge response;

[0075] (5) The physical information neural network model of the present invention adopts a multi-branch hybrid structure of physical branch and data branch. By embedding the bridge mechanical equation as a physical constraint and combining multi-scale convolution and attention mechanism to process visual data, it realizes the effective fusion of physical laws and data features, and improves the prediction accuracy and generalization ability of the model;

[0076] (6) The present invention adopts a multi-level evaluation mechanism and a hierarchical early warning strategy, which can accurately distinguish minor anomalies, obvious anomalies and serious anomalies, and generate early warning signals of corresponding levels, including specific anomaly parameter values, categories and recommended measures, providing a scientific decision-making basis for bridge management departments. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained according to these drawings.

[0078] Figure 1 It is the flowchart of the method of the present invention;

[0079] Figure 2 It is the model structure diagram of the present invention;

[0080] Figure 3 It is the system framework diagram of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0081] Next, in combination with the embodiments of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0082] As Figure 1 shown, the present invention provides a bridge warning method based on a physics-informed neural network and machine vision, including:

[0083] S1. Set up a visual data acquisition system on the bridge, including a bridge dynamic displacement acquisition device and a trajectory vision acquisition device;

[0084] S2. Design a known load calibration test, including selecting a standardized specific load vehicle and designing a predetermined trajectory. The visual data acquisition system acquires dynamic response data under known load conditions to form a preliminary calibration data set;

[0085] S3. Acquire vehicle trajectories, vehicle types, and corresponding bridge response data under natural traffic flow to form a preliminary random load data set;

[0086] S4. Conduct the known load calibration test again. The visual data acquisition system acquires the bridge dynamic response to form a second calibration data set. Compare the preliminary calibration data set and the second calibration data set, identify systematic errors and random errors, and use an error correction algorithm to correct the preliminary random load data set to obtain a corrected random load data set;

[0087] S5. Establish a causal association database covering the dynamic response relationship between vehicles and bridges through the corrected random load data set, the preliminary calibration data set, and the second calibration data set;

[0088] S6. Construct a physics-informed neural network model. Based on the causal association database, use the supervised learning method to train the physics-informed neural network, optimize the network parameters, and obtain a trained physics-informed neural network model;

[0089] S7. Real-time collect bridge dynamic response data, input it into the trained physics-informed neural network model, conduct an online assessment of the bridge operation state, check for abnormal responses, and generate warning signals.

[0090] Specifically, in an embodiment of the present invention, the bridge dynamic displacement acquisition device includes a digital image correlation acquisition device deployed under the bridge. This device has a multi-scale image acquisition unit and a real-time optical flow algorithm module for non-contact monitoring of the deflection changes of key parts of the bridge under different frequency loads; the trajectory visual acquisition device includes a high-resolution camera array and a multi-view synchronous collector deployed on the bridge deck and at the bridgehead. The camera array obtains vehicle shape, speed, and driving trajectory information through the Yolov8 object detection model, and fuses this visual data with the data output by the bridge dynamic displacement acquisition device under the same time reference.

[0091] Specifically, the bridge dynamic displacement acquisition device is deployed at key parts of the bridge to non-contact monitor the displacement changes of the bridge under different load conditions. Among them: Select the digital image correlation acquisition device (Digital Image Correlation, DIC) as the dynamic displacement acquisition device. According to the structural characteristics and force distribution of the bridge, the multi-scale image acquisition units (Multi-Scale Image Acquisition Units) are evenly arranged at the key nodes under the bridge to ensure coverage of all main force-bearing parts of the bridge. Each image acquisition unit is equipped with a high-resolution camera, which can obtain images of the bridge surface at different magnification ratios to achieve precise monitoring of the subtle displacements of the bridge. Integrate the real-time optical flow algorithm module (Optical Flow Algorithm Module) into the dynamic displacement acquisition device to process the captured image sequence and calculate the displacement of each monitoring point of the bridge in real time.

[0092] The trajectory visual acquisition device is used to monitor the dynamic information of vehicles on the bridge, including the shape, speed, and driving trajectory of the vehicles. Specifically, install multiple high-resolution cameras at the bridge deck and at the bridgehead to form a camera array. The camera array should cover the entire bridge deck to ensure comprehensive monitoring of all moving vehicles. The cameras should be equipped with wide-angle lenses and high frame rates to capture images of fast-moving vehicles. Set multi-angle synchronous collectors (Multi-Angle Synchronized Collectors) between the cameras to achieve synchronous acquisition of data from different perspectives. Embed the Yolo v8 object detection model into the trajectory visual acquisition device to process the captured vehicle images in real time. Synchronously fuse the data collected by the bridge dynamic displacement acquisition device and the trajectory visual acquisition device according to a unified time reference. Ensure the consistency of the bridge structure response data and the vehicle load data in time and space.

[0093] Specifically, in an embodiment of the present invention, step S2 includes:

[0094] Select test vehicles that meet industry standards, such as standardized trucks or heavy vehicles, and ensure that their load characteristics match those of actual traffic vehicles. Determine the static and dynamic load parameters of the vehicle, including total mass, axle load distribution, vehicle dimensions, and tire pressure, etc. Equip necessary sensor devices, such as load sensors and position sensors, to monitor the actual load conditions and driving status of the vehicle in real time.

[0095] According to the bridge structure characteristics and the results of force analysis, plan the driving path of the vehicle on the bridge to ensure that the main stressed parts of the bridge are covered. Set the driving speed and acceleration of the vehicle to simulate the dynamic load conditions during different traffic flows and when vehicles pass by. Determine the spacing and repetition times of vehicle driving to ensure the sufficiency and representativeness of data collection.

[0096] Install the bridge dynamic displacement acquisition device and the trajectory vision acquisition device correctly according to the implementation method in step S1. Through preliminary tests under known load conditions, calibrate the vision data acquisition system, and adjust the focal length, exposure parameters, and image processing algorithm parameters of the camera. Conduct data synchronization tests to ensure the synchronization of bridge dynamic response data and vehicle trajectory data in time and space.

[0097] Set up standardized load vehicles on the bridge to ensure that the vehicle state meets the preset load parameters and driving trajectories. Start the vision data acquisition system and record the dynamic response data generated during the vehicle's driving on the bridge, including parameters such as the displacement, stress, and vibration of the bridge. Record and store the dynamic response data of the bridge under different load conditions in real time. According to the design requirements, repeat multiple calibration tests covering different load conditions and driving speeds to obtain diverse calibration data.

[0098] Filter out possible noise and abnormal data during the acquisition process. Label the collected dynamic response data and record information such as the corresponding load parameters, vehicle types, and driving status. Store the processed calibration data in the preliminary calibration dataset in a unified format.

[0099] Specifically, in an embodiment of the present invention, step S3 includes:

[0100] Reasonably arrange multiple high-resolution cameras on the bridge deck and at both ends, and use the Yolo v8 object detection model to process the real-time collected images to accurately identify and classify different types of vehicles (such as cars, trucks, heavy vehicles, etc.). Subsequently, apply a multi-object tracking algorithm (such as the SORT algorithm) to track the trajectories of the identified vehicles and record their dynamic parameters such as driving paths, speeds, and accelerations on the bridge.

[0101] Based on the output of the Yolo v8 model, a deep learning classification algorithm is further applied to finely classify vehicles. The classification model is trained to identify the type, size, and load characteristics of vehicles. The collected data is manually annotated and verified. For misidentified vehicle types, secondary detection or manual correction is performed.

[0102] The DIC device deployed at the critical parts of the bridge uses the multi-scale image acquisition unit to continuously monitor the displacement changes of the bridge under natural traffic loads. The DIC device can accurately capture the minute displacements of the bridge under different frequency load conditions and record key parameters such as deflection, stress, and vibration amplitude. The real-time optical flow algorithm module integrated in the DIC device processes the collected image sequences and calculates the displacement amounts of each monitoring point of the bridge.

[0103] The vehicle type, trajectory parameters, and bridge response data are associated and annotated to form structured data records. Each record contains multi-dimensional parameters such as vehicle type, driving speed, acceleration, bridge displacement, stress, and vibration. The processed data is stored in the preliminary random load dataset in a unified format (such as CSV, JSON, or database table).

[0104] Specifically, in an embodiment of the present invention, step S4 includes:

[0105] To ensure data consistency and comparability, standardized specific load vehicles of the same type as in step S2 are selected. And the same vehicle driving trajectory as in step S2. Set up the standardized load vehicle on the bridge, start the visual data acquisition system, and record the dynamic response data generated during the vehicle's driving on the bridge, including parameters such as bridge displacement, stress, and vibration. Continuously record and store the dynamic response data of the bridge under different load conditions. According to the design requirements, repeat multiple calibration tests covering the same load conditions and driving speeds to obtain the second calibration dataset.

[0106] By comparing the preliminary calibration dataset and the second calibration dataset, systematic errors and random errors are identified. Systematic errors: Analyze the consistency of the bridge response data in the two calibration tests and identify systematic deviations caused by equipment installation, environmental condition changes, etc. Random errors: Identify the random fluctuations and noises in the two calibration data through statistical analysis and evaluate the random uncertainties in the data acquisition process.

[0107] According to the identified systematic errors and random errors, an error correction algorithm is applied to correct the preliminary random load dataset. The specific steps include:

[0108] Step 1. Define the initial error vector , where is the initial systematic error, is the initial random error, is the initial environmental error;

[0109] Step 2: During the k-th iteration correction, calculate the intermediate correction function , where is the preliminary calibration data set, is the second calibration data set, is the adaptive convergence factor, Based on multi-dimensional non-linear mapping, the previous iteration result and , are coupled and calculated to output a new correction vector ;

[0110] Step 3: Substitute the intermediate correction result into the correction formula:

[0111] ,

[0112] In the formula, is the exponential penalty factor, represents the vector norm, is the quadratic coupling correction function, which performs non-linear compensation according to the distribution characteristics of the vehicle load effect and the environmental disturbance characteristics;

[0113] Step 4: After several iterations, when or is lower than the predetermined threshold, output the final correction result , and is applied to the preliminary random load data set to obtain the corrected random load data set.

[0114] The present invention accurately identifies systematic errors and random errors by performing a second known load calibration test and comparing the preliminary calibration data set with the second calibration data set. The quadratic coupling correction function using multi-dimensional non-linear mapping is used for iterative optimization to effectively eliminate systematic biases and random fluctuations generated during the data acquisition process.

[0115] Specifically, in an embodiment of the present invention, step S5 includes:

[0116] S51: Integrate the corrected random load data set, the preliminary calibration data set, and the second calibration data set, and associate the vehicle load parameters and the bridge response parameters according to the time stamp to obtain a synchronous record of multi-source data.

[0117] Specifically, the corrected random load dataset, the preliminary calibration dataset, and the second calibration dataset are merged to form a unified data source. Based on the unified time reference, vehicle load parameters (such as vehicle type, speed, trajectory) and bridge response parameters (such as displacement, stress, vibration) are precisely matched according to timestamps to form synchronized multi-source data records. A time synchronization mechanism is used to ensure the consistency of data from different data sources on the time axis. Data synchronization tools are employed to process and transmit data in real time. The synchronized data records are stored in a data warehouse according to a unified structure, and relational databases or time series databases are used to manage multi-source data.

[0118] S52. Perform multi-level feature extraction on the above synchronized records to obtain the data layer, the feature layer, and the association layer respectively; where the data layer contains the original data of vehicle nodes, response nodes, and environmental nodes, the feature layer includes frequency domain and time domain features, and the association layer is used to record the direct or indirect causal relationships between nodes.

[0119] Specifically, the data structure is divided into a data layer, a feature layer, and an association layer. The data layer is the original data from vehicle nodes, response nodes, and environmental nodes, and the data is divided into vehicle nodes (recording vehicle load parameters), response nodes (recording bridge dynamic response parameters), and environmental nodes (recording environmental impact factors such as temperature and humidity). The feature layer is time domain features and frequency domain features, specifically including statistical features: calculating statistics such as mean, variance, kurtosis, and skewness. Dynamic features: extracting dynamic parameters such as the maximum value, minimum value, and amplitude in the displacement time series. Fourier transform: performing Fourier transform on the time domain signal to extract the main frequency components and amplitudes in the spectrum. Power spectral density: calculating the power spectral density of the signal to analyze the frequency distribution characteristics. The association layer includes: using statistical methods and machine learning algorithms (such as random forest) to identify the direct or indirect causal relationships between vehicle loads and bridge responses. Recording the direct and indirect associations between nodes, constructing a multi-level causal relationship network to reflect the complex interactions between different nodes.

[0120] S53. Extract the main causal chains based on causal strength to form the direct and indirect causal relationships between vehicle loads and bridge responses; use optimization algorithms to optimize the accuracy and integrity of the causal chains.

[0121] The formula for causal strength is:

[0122] ,

[0123] In the formula, represents the causal strength from node i to node j; is the mutual information between node i and node j; is the direction consistency from node i to node j; is a non - linear parameter for adjusting causal strength.

[0124] Based on the calculated causal strength, extract the main causal chains, that is, those node pairs with higher causal strength, to form the direct and indirect causal relationships between vehicle loads and bridge responses.

[0125] Use the Bayesian optimization algorithm to adjust the weights of the extracted causal chains. The formula is as follows:

[0126] ,

[0127] where, is the final causal weight from node i to node j; m is the total number of nodes that have causal relationships with node i.

[0128] Through the Bayesian optimization algorithm, dynamically adjust the weight parameters in the causal chain to improve the accuracy and integrity of the causal chain.

[0129] S54. Enter the constructed causal chains and association relationships into the graph database to form a causal association database covering a multi - level data structure.

[0130] Select a graph database suitable for storing complex causal relationships. Define nodes (vehicle nodes, response nodes, environmental nodes) and edges (causal relationship edges) in the graph database, and assign corresponding attributes (such as causal strength, weight). Construct a multi - level data structure covering the data layer, feature layer, and association layer. Take the optimized main causal chains as the core part of the graph database to record the direct and indirect influence paths of vehicle loads on bridge responses.

[0131] Specifically, in an embodiment of the present invention, step S6 includes:

[0132] Construct a physical information neural network model. This model adopts a multi - branch hybrid structure. The input data of the model consists of two parts: vehicle load parameters, including vehicle type, weight, speed, passing trajectory, etc.; bridge dynamic response data, including displacement, stress, vibration, etc. As Figure 2 shown, the model structure includes:

[0133] Physical branch: includes an input layer, two fully - connected hidden layers, and an output layer; each hidden layer applies a physical constraint activation function; the physical branch is used to process the embedded bridge mechanical equations and output physical constraint results.

[0134] Specifically, the input layer of the physical branch receives the parameters of the bridge mechanical model (such as material properties and structural configuration), and uses the BN layer for feature normalization. The first hidden layer uses the LeakyReLU activation function, is initialized with He, followed by a batch normalization layer, and a Dropout layer is added during training with a dropout rate set to 0.3 and the number of nodes set to h1, which depends on the number of input features; the second hidden layer uses the ReLU activation function and the number of nodes is set to h2. Each hidden layer is finally followed by a physical constraint layer, which sets a custom physical constraint activation function. The specific physical constraints include static constraints and dynamic constraints, where:

[0135] The static constraint formula is:

[0136] ,

[0137] In the formula, x is the spatial coordinate position, that is, the position parameters of each measurement point in the bridge; represents the mass matrix of the bridge; represents the damping matrix; represents the stiffness matrix; represents the displacement vector of the bridge; is the external load vector; represents the second-order partial derivative with respect to time; represents the first-order partial derivative with respect to time;

[0138] The dynamic constraint formula is:

[0139] ,

[0140] In the formula, represents the natural frequency of the bridge; represents the vibration mode vector of the bridge; is the stiffness matrix.

[0141] Data branch: includes an input layer, a multi-scale convolutional layer, an attention mechanism layer, and a fully connected hidden layer; the multi-scale convolutional layer is used to extract multi-scale features; the attention mechanism layer is used to autonomously learn the importance of different features; the data branch is used to process high-frequency visual data and extract deep feature information.

[0142] Specifically, the data branch sequentially includes, in the order of data flow:

[0143] Input layer, which receives visual data, that is, dynamic response data.

[0144] The multi-scale convolution module is equipped with a small-scale convolution branch, a medium-scale convolution branch, and a large-scale convolution branch. The convolution kernel of the small-scale convolution branch is 3*3, the stride is 1, the activation function is ReLU, and it is followed by a BN layer; the convolution kernel of the medium-scale convolution branch is 5*5, the stride is 1, the activation function is ReLU, and it is followed by a BN layer; the convolution kernel of the large-scale convolution branch is 7*7, the stride is 1, the activation function is ReLU, and it is followed by a BN layer. Finally, a fusion layer is provided to splice the output features of different scales in the channel dimension.

[0145] The attention mechanism layer is divided into channel attention and spatial attention; the spatial attention has two layers of 1*1 convolution, uses Sigmoid activation, and its output is multiplied by the input; the channel attention includes global average pooling and three fully-connected layers, uses Sigmoid activation, and its output is multiplied by the input.

[0146] The max pooling layer has a pooling kernel of 2*2 and a stride of 2.

[0147] The global average pooling layer.

[0148] The fully-connected layer uses the ReLU activation function.

[0149] The fusion layer: includes an adaptive weight fusion unit to dynamically weight and fuse the outputs of the physical branch and the data branch; the fusion layer realizes the effective integration of information by learning weight parameters.

[0150] Specifically, first splice the outputs of the physical branch and the data branch, and then output two weight coefficients α and β through two fully-connected layers FC1 and FC2. FC1: 128→64, ReLU activation; FC2: 64→2, Softmax activation. Use the weight coefficients α and β to dynamically weight and fuse the outputs of the physical branch and the data branch.

[0151] The output layer: synthesizes the fusion results of the physical branch and the data branch and outputs the evaluation of the bridge operation state.

[0152] The output layer contains two fully-connected layers. The first fully-connected layer is used for feature integration, and the second fully-connected layer is used for prediction.

[0153] The model training process is as follows:

[0154] Divide the dataset from the causal association database into a training set, a validation set, and a test set, with a ratio of 70% for training and 30% for validation. The dataset includes vehicle load parameters, bridge dynamic response parameters, and related environmental factors. Using the supervised learning method, utilize the labeled bridge health state data (such as normal, slightly abnormal, significantly abnormal, severely abnormal) as the supervision signal to guide the model to learn the relationship between the bridge dynamic response and the health state. Use the gradient descent method and its variants (such as the Adam optimizer) to optimize the network parameters and minimize the loss function. The loss function is designed as follows:

[0155] ,

[0156] ,

[0157] ,

[0158] ,

[0159] In the formula, is the total loss function; is the data-driven loss; is the physical constraint loss; is the regularization loss; , , are adaptive weight coefficients; represents the sth sample; is the number of samples; represents the actual value of the sample, that is, the true response data of the bridge at the sth position; represents the model prediction value, that is, the bridge response data predicted by the model; is the adjustment parameter; , , , , are the physical parameters corresponding to sample s; is the normalization parameter; is the regularization coefficient; is the weight parameter of the th layer in the model; Q is the number of network layers.

[0160] The model training steps are as follows:

[0161] Initialization: Randomly initialize the network parameters, including the weights of the physical branch and the data branch.

[0162] Forward propagation: Input the training data, pass through the physical branch and the data branch, and output the prediction result through the fusion layer.

[0163] Calculate the loss: Based on the actual output and the predicted output, calculate the data-driven loss, the physical constraint loss, and the regularization loss, and comprehensively obtain the total loss. 。

[0164] Backpropagation: Calculate the gradients of each parameter through the backpropagation algorithm.

[0165] Parameter update: Use the optimization algorithm to update the network parameters and minimize the total loss.

[0166] Iterative training: Repeatedly execute forward propagation, loss calculation, backpropagation, and parameter update until the loss function converges or reaches the preset number of training epochs.

[0167] Model validation: Evaluate the model performance on the validation set, and adjust the model structure or training parameters to improve the generalization ability of the model.

[0168] The physics-informed neural network model provided by the present invention effectively integrates the bridge mechanical equation and the high-frequency visual data features, not only improving the accuracy of bridge operation state evaluation, but also enhancing the generalization ability and robustness of the model.

[0169] Specifically, in one embodiment of the present invention, step S7 includes:

[0170] S71. Real-time acquire the dynamic response data generated by the bridge during operation through the visual data acquisition system, and perform time synchronization and preliminary preprocessing on the acquired dynamic response data.

[0171] The time synchronization is implemented by using the NIP protocol and a unified time server, and a timestamp is added to each frame of acquired data. The preliminary preprocessing includes Kalman filter denoising, linear interpolation to fill in missing data, and normalization processing of the dynamic response data to unify data with different dimensions to the [0,1] interval.

[0172] S72. Convert the preprocessed dynamic response data into the input format required by the physics-informed neural network model, and input it into the trained physics-informed neural network model. Calculate the current operation state parameters of the bridge through the forward propagation process. The model outputs the health assessment results of the bridge according to the embedded physical constraints and causal association database, including the current deflection, vibration mode, and potential structural anomaly indicators.

[0173] Integrate the real-time acquired vehicle load parameters (such as vehicle type, weight, speed, trajectory) and bridge dynamic response parameters (such as displacement, stress, vibration) into a unified feature vector. Assemble the data within a continuously acquired time window into batches.

[0174] Load the parameters of the trained physics-informed neural network model, input the dynamically collected response data in real time into the model, and generate the current operating state parameters of the bridge through physical branch processing, data branch input, fusion layer integration, and output layer evaluation, including deflection, vibration mode, and potential structural anomaly indicators.

[0175] Among them, the model output refers to the causal relationship in the reference causal association database to ensure that the model output meets the physical constraints of the bridge mechanical equation.

[0176] Finally, output the health assessment results, including quantitative parameters and structural indicators. The quantitative parameters are the key operating parameters such as the deflection value and vibration amplitude of the current bridge. The structural indicators refer to evaluating the structural health status of the bridge and identifying potential structural anomalies (such as cracks, fatigue damage, etc.).

[0177] S73. Compare the health assessment results based on the model output with the preset safety standards and thresholds to determine whether the bridge operating state is normal, and use a multi-level assessment mechanism to distinguish minor anomalies, obvious anomalies, and severe anomalies.

[0178] According to the relevant bridge safety specifications in the industry, set the safety thresholds for various operating parameters (such as the maximum allowable deflection, vibration frequency, etc.). Optimize and adjust the preset thresholds based on historical monitoring data and bridge design parameters.

[0179] If all key operating parameters (such as deflection, vibration amplitude) are lower than the preset safety thresholds, it is determined that the bridge operating state is normal.

[0180] Otherwise, it is determined as abnormal. Specifically:

[0181] Minor anomaly: Some parameters slightly exceed the safety thresholds, which need attention but no immediate measures are required.

[0182] Obvious anomaly: Multiple parameters significantly exceed the safety thresholds, and repair or reinforcement measures need to be taken in a timely manner.

[0183] Severe anomaly: Key parameters seriously exceed the standard, which may lead to structural damage or failure of the bridge, and emergency measures such as shutdown and repair need to be taken.

[0184] S74. Generate warning signals of corresponding levels according to the abnormal assessment results, including level-one warning, level-two warning, and level-three warning, corresponding to minor anomaly, obvious anomaly, and severe anomaly respectively. Each level of warning signal contains specific abnormal parameter values, abnormal categories, and recommended response measures, and transmits the warning signals to the bridge management and operation and maintenance system in real time through wireless communication or wired network.

[0185] Set the first-level warning, second-level warning, and third-level warning according to minor anomalies, obvious anomalies, and severe anomalies. The classification basis includes the degree of parameter overrun, the duration of anomalies, and the comprehensive evaluation of multiple parameters. After determining the level, generate a warning signal. The content of the warning signal includes: Abnormal parameter value: Specifically list the exceeded operating parameters and their actual values. Abnormal category: Clearly marked as minor anomaly, obvious anomaly, or severe anomaly. Recommended response measures: Give corresponding maintenance or repair suggestions according to the warning level. The warning signal format adopts the JSON format, including the timestamp, bridge ID, anomaly category, abnormal parameters, and recommended measures.

[0186] Transmit the warning signal to the bridge management and operation and maintenance system by wireless or wired means.

[0187] In addition, as Figure 3 shown, the present invention also provides a bridge warning system based on the physics-informed neural network and machine vision. The system is used to execute the method described in any one of the above, and the system includes:

[0188] The data acquisition module integrates bridge dynamic displacement acquisition devices and trajectory vision acquisition devices, and is used to collect bridge dynamic response data and vehicle trajectory data in real time;

[0189] The dual calibration module performs at least two known load calibration tests and natural traffic flow acquisitions, evaluates the differences in random load data, and calls the error correction algorithm for dynamic correction;

[0190] The causal association database module stores, constructs causal relationships, and iteratively updates multi-source heterogeneous data based on the graph database structure containing vehicle nodes, bridge response nodes, and environmental nodes;

[0191] The physics-informed neural network module contains a physics-informed neural network model with a multi-branch hybrid structure for evaluating the operating state of the bridge;

[0192] The real-time evaluation and warning module inputs dynamic response data into the trained neural network model in real time; generates multi-level warning signals according to the model output, and transmits and feedback the signals through the warning setting unit;

[0193] The communication and storage module is used to transmit the warning signal to the bridge management and operation and maintenance system;

[0194] The user interface and monitoring module provides a visual monitoring interface for the operating state of the bridge; displays real-time data, warning information, and historical records.

[0195] The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A bridge early warning method based on physical information neural network and machine vision, characterized in that: include: S1. Setting up a visual data acquisition system on the bridge, including bridge dynamic displacement acquisition equipment and trajectory visual acquisition equipment; S2. Design a known load calibration test, including selecting a standardized specific load vehicle and designing a predetermined trajectory, and the visual data acquisition system collects dynamic response data under known load conditions to form a preliminary calibration data set; S3, collecting vehicle trajectories, vehicle types and corresponding bridge response data under natural traffic flow to form a preliminary random load data set; S4, conduct a known load calibration test again, collect the dynamic response of the bridge by the visual data acquisition system, form a second calibration data set, compare the preliminary calibration data set with the second calibration data set, identify systematic errors and random errors, and use the error correction algorithm to correct the preliminary random load data set to obtain a corrected random load data set; S5. Establish a causal correlation database covering the dynamic response relationship between the vehicle and the bridge through the corrected random load data set, the preliminary calibration data set and the second calibration data set; S6. Construct a physical information neural network model, train the physical information neural network using a supervised learning method based on a causal association database, optimize network parameters, and obtain a trained physical information neural network model; S7. Collect bridge dynamic response data in real time and input it into the trained physical information neural network model to conduct online evaluation of the bridge operation status, check abnormal responses and generate early warning signals.

2. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 1 is characterized in that: The bridge dynamic displacement acquisition equipment includes a digital image correlation acquisition device deployed under the bridge, which has a multi-scale image acquisition unit and a real-time optical flow algorithm module for non-contact monitoring of the deflection changes of key parts of the bridge under different frequency loads; the trajectory visual acquisition equipment includes a high-resolution camera array and a multi-view synchronous collector deployed on the bridge deck and bridge head. The camera array obtains vehicle shape, speed and driving trajectory information through the Yolo v8 target detection model, and fuses these visual data with the data output by the bridge dynamic displacement acquisition equipment on the same time basis.

3. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 1 is characterized in that: The error correction algorithm in step S4 is as follows: Step 1: Define the initial error vector ,in is the initial system error, is the initial random error, is the initial environmental error; Step 2: Calculate the intermediate correction function at the kth iteration correction ,in, For the preliminary calibration dataset, is the second calibration data set, is the adaptive convergence factor, Based on multidimensional nonlinear mapping, the previous iteration result and , After coupling operation, a new correction vector is output ; Step 3: Substitute the intermediate correction result into the correction formula: ， In the formula, is the exponential penalty factor, represents the vector norm, It is a quadratic coupling correction function, which performs nonlinear compensation according to the distribution characteristics of vehicle load effects and environmental disturbance characteristics; Step 4: After several iterations, when or Output the final correction result when it is lower than the predetermined threshold , and Applied to the preliminary random load data set, a revised random load data set is obtained.

4. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 1 is characterized in that: Step S5 includes: S51, integrating the corrected random load data set, the preliminary calibration data set and the second calibration data set, associating the vehicle load parameters with the bridge response parameters according to the timestamps, and obtaining a synchronous record of multi-source data; S52, performing multi-level feature extraction on the above synchronous records to obtain a data layer, a feature layer and an association layer respectively; wherein the data layer includes the original data of the vehicle node, the response node and the environment node, the feature layer includes frequency domain and time domain features, and the association layer is used to record the direct or indirect causal relationship between the nodes; S53. Extract the main causal chain based on the causal strength to form the direct and indirect causal relationship between vehicle load and bridge response; use optimization algorithm to optimize the accuracy and completeness of the causal chain; S54. Enter the constructed causal chain and association relationship into the graph database to form a causal association database covering a multi-level data structure.

5. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 4 is characterized in that: The formula for calculating causal strength is: ， In the formula, represents the causal strength from node i to node j; is the mutual information between node i and node j; is the directional consistency from node i to node j; A nonlinear parameter to adjust the causal strength; The Bayesian optimization algorithm is used to adjust the weight of the extracted causal chain, and the formula is defined as: ， in, is the final causal weight from node i to node j; m is the total number of nodes that have a causal relationship with node i.

6. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 1, characterized in that: The physical information neural network model is a multi-branch hybrid structure, including: Physical branch: It includes an input layer, two fully connected hidden layers, and an output layer. Each hidden layer applies a physical constraint activation function. The physical branch is used to process the embedded bridge mechanics equations and output the physical constraint results. Data branch: includes input layer, multi-scale convolution layer, attention mechanism layer and fully connected hidden layer; multi-scale convolution layer is used to extract multi-scale features; attention mechanism layer is used to autonomously learn the importance of different features; data branch is used to process high-frequency visual data and extract deep feature information; Fusion layer: contains an adaptive weight fusion unit, which dynamically weights and fuses the outputs of the physical branch and the data branch. The fusion layer learns weight parameters to achieve effective information integration. Output layer: integrates the fusion results of the physical branch and the data branch, and outputs the bridge operation status assessment.

7. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 6, characterized in that: The bridge mechanics equations in the physics branch include static constraints and dynamic constraints, among which: The static constraint formula is: ， In the formula, x is the spatial coordinate position, that is, the position parameter of each measuring point in the bridge; represents the mass matrix of the bridge; represents the damping matrix; represents the stiffness matrix; represents the displacement vector of the bridge; is the external load vector; express Second-order partial derivative with respect to time; express First-order partial derivative with respect to time; The dynamic constraint formula is: ， In the formula, represents the natural frequency of the bridge; represents the vibration mode vector of the bridge; is the stiffness matrix.

8. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 7, characterized in that: The loss function used in the training of the physical information neural network model is: ， ， ， ， In the formula, is the total loss function; For data driven losses; Loss of physical restraint; is the regularization loss; , , is the adaptive weight coefficient; represents the sth sample; is the number of samples; represents the actual value of the sample, that is, the real response data of the bridge at the sth position; Represents the model prediction value, that is, the bridge response data predicted by the model; To adjust the parameters; , , , , is the physical parameter corresponding to sample s; is the standardized parameter; is the regularization coefficient; For the model Layer weight parameter; Q is the number of network layers.

9. The bridge early warning method based on physical information neural network and machine vision as claimed in claim 2, characterized in that: Step S7 includes: S71, acquiring dynamic response data generated by the bridge during operation in real time through a visual data acquisition system, and performing time synchronization and preliminary preprocessing on the acquired dynamic response data; S72, converting the preprocessed dynamic response data into the input format required by the physical information neural network model, and inputting the data into the trained physical information neural network model, calculating the current operating state parameters of the bridge through the forward propagation process, and the model outputs the health assessment results of the bridge according to the embedded physical constraints and causal association database, including the current deflection, vibration mode and potential structural abnormality indicators; S73. Based on the health assessment results output by the model, the preset safety standards and thresholds are compared to determine whether the bridge operation status is normal, and a multi-level assessment mechanism is used to distinguish between slight abnormalities, obvious abnormalities and serious abnormalities; S74. Generate warning signals of corresponding levels based on the abnormality assessment results, including level 1 warning, level 2 warning and level 3 warning, which correspond to slight abnormalities, obvious abnormalities and severe abnormalities respectively. Each level of warning signal contains specific abnormal parameter values, abnormality categories and recommended response measures. The warning signal is transmitted in real time to the bridge management and operation and maintenance system through wireless communication or wired network.

10. The bridge warning system based on physical information neural network and machine vision is characterized by: The system is used to perform the method according to any one of claims 1 to 9, and the system comprises: The data acquisition module integrates bridge dynamic displacement acquisition equipment and trajectory visual acquisition equipment to collect dynamic response data of the bridge and vehicle trajectory data in real time; The dual calibration module performs at least two known load calibration tests and natural traffic flow collection, evaluates the difference of random load data and calls the error correction algorithm for dynamic correction; The causal relationship database module stores, builds causal relationships, and iteratively updates multi-source heterogeneous data based on a graph database structure containing vehicle nodes, bridge response nodes, and environment nodes; Physical information neural network module, including a physical information neural network model with multi-branch hybrid structure, to evaluate the bridge operation status; The real-time evaluation and early warning module inputs dynamic response data into the trained neural network model in real time; generates multi-level early warning signals according to the model output, and transmits and feedbacks the signals through the early warning setting unit; Communication and storage module, used to transmit warning signals to the bridge management and operation and maintenance system; The user interface and monitoring module provides a visual bridge operation status monitoring interface; displays real-time data, warning information and historical records.

Images