Bridge performance analysis method and system based on digital twinning technology

Abstract

The invention provides a bridge performance analysis method and system based on a digital twinning technology, and relates to the field of crossing technologies of civil engineering and information technologies, and the method comprises the steps: obtaining real-time monitoring data of a physical bridge; the method comprises the following steps: constructing a digital twinborn model containing configurable physical parameters on the basis of structural information and a physical mechanism of a physical bridge; the method further comprises the following steps: (a) based on real-time monitoring data and a digital twinborn model, constructing a hybrid prediction model by using a physical information artificial intelligence algorithm; (b) carrying out dynamic and closed-loop self-calibration on physical parameters of the hybrid prediction model through an advanced optimization strategy of Bayesian inference or reinforcement learning so as to minimize deviation between prediction response of the hybrid prediction model and real-time monitoring data; and (c) analyzing the service performance of the bridge based on the calibrated hybrid prediction model in the step (b). The method has high fidelity and strong robustness.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of civil engineering and information technology, in particular to the technical field of bridge structure health monitoring. More specifically, the present application relates to a method and system for high-fidelity analysis of the service performance of a physical bridge using digital twin technology and artificial intelligence. BACKGROUND

[0002] Bridges are critical infrastructure in the transportation network, and their safety and durability are of great importance. Currently, there are still many limitations in the maintenance and health monitoring of bridges. First, the current bridge maintenance work in most cities, such as daily inspection and appearance detection, still heavily relies on manual methods. This manual inspection method has significant drawbacks: first, it is long in cycle and poor in timeliness, for example, the overall inspection cycle of large bridges can be several years, which makes it impossible to track the progressive damage of bridges caused by load, environment, material aging, etc. in a timely manner; second, after severe natural disasters such as earthquakes, floods, or major accidents such as ship collisions, manual inspection cannot immediately provide real-time safety state information of the bridge, affecting emergency decision-making; third, the inspection report is mostly recorded in paper or non-standard electronic format, the data mining foundation is poor, and the professional level of personnel in the basic maintenance department varies greatly, resulting in the quality and consistency of the monitoring data needing to be improved.

[0003] In order to overcome the limitations of manual inspection, the field has begun to use model-based calculation methods for bridge performance analysis, mainly including two categories: The first category is a model based on physical mechanism, such as the Finite Element Method (FEM). This type of model is based on the principles of structural mechanics and material mechanics, and has good interpretability. However, it is heavily dependent on idealized physical parameters such as the elastic modulus of the material, the structural damping coefficient, and the boundary conditions between components. In the actual service process, these parameters of the physical bridge will change dynamically with material aging, fatigue damage, and environmental changes. A static, idealized finite element model cannot capture these real changes, resulting in a distorted model, and the analysis results (such as stress distribution, dynamic response) are far from the true state of the bridge.

[0004] The second category is purely data-driven models, such as traditional neural networks (NNs). These models attempt to learn the behavior of the structure directly from sensor monitoring data. However, they also face two major challenges: First, the problem of data sparsity. In structural health monitoring (SHM) applications, the number of sensors (such as accelerometers) deployed is always limited, typically monitoring only a few key nodes, failing to provide complete information about the entire bridge. Purely data-driven models struggle to train reliable models under sparse data conditions. Second, the black-box and extrapolation problem. The predictive power of neural networks is strictly limited by the range of their training data, lacking physical consistency. When faced with an unprecedented extreme load (such as a super earthquake or a new type of vehicle load), their predictions may completely violate the laws of physics and be unreliable.

[0005] Currently, although some researchers have proposed building information models (BIM) or digital twins for bridges, most of them are still at the level of three-dimensional visualization and data storage, and have not fundamentally solved the distortion problem of the first type of model and the data limitation and black box problem of the second type of model.

[0006] Therefore, there is an urgent need in this field for a new technical solution that must be able to: This approach effectively integrates limited, real-time monitoring data from physical bridges, which may contain noise, with high-dimensional physical mechanism models. It utilizes real-time data to dynamically and in a closed-loop manner calibrate the physical parameters of the digital twin model, ensuring it always aligns with the actual health state of the physical entity. Even with sparse or noisy monitoring data, it can still provide reliable performance predictions that conform to physical logic. It intelligently manages and quantifies the model's inherent uncertainties, optimizing computational resource consumption while maintaining analytical accuracy. Summary of the Invention

[0007] The purpose of this invention is to provide a bridge performance analysis method and system based on digital twin technology to solve the problems mentioned in the background art.

[0008] In a first aspect, embodiments of the present invention provide a bridge performance analysis method based on digital twin technology, comprising: Obtain real-time monitoring data of the physical bridge; Based on the structural information and physical mechanism of the physical bridge, a digital twin model containing configurable physical parameters is constructed. The method is characterized in that it further includes: (a) Based on the real-time monitoring data and the digital twin model, a hybrid prediction model is constructed using a physical information artificial intelligence algorithm; (b) Using advanced optimization strategies such as Bayesian inference or reinforcement learning, the physical parameters of the hybrid prediction model are dynamically and in a closed-loop manner to minimize the deviation between the predicted response of the hybrid prediction model and the real-time monitoring data; and (c) The service performance of the bridge is analyzed based on the calibrated hybrid prediction model in step (b).

[0009] Optionally, the real-time monitoring data originates from a multi-source heterogeneous sensor network, which includes at least one of the following: Strain gauges, accelerometers, displacement sensors, or fiber optic grating sensors.

[0010] Optionally, the hybrid prediction model includes: A physical mechanism model based on the finite element method is used to provide physical constraints for the structural response; and A data-driven model of a neural network is used to learn nonlinear dynamic characteristics not modeled by the physical mechanism model from the real-time monitoring data; The physical information artificial intelligence algorithm integrates the physical constraints into the training process of the data-driven model.

[0011] Optionally, the physical parameters for the dynamic self-calibration include at least one of the following: The elastic modulus of the physical bridge material, the structural damping coefficient, or the boundary conditions between structural components.

[0012] Optionally, step (c) of analyzing the service performance of the bridge includes at least one of the following: Based on the calibrated hybrid prediction model, structural stress analysis under random traffic loads is performed to predict structural fatigue life. Based on the calibrated hybrid prediction model, the structural response of the physical bridge under seismic loads is evaluated to assess its seismic performance and damage; or The ultimate bearing capacity of the physical bridge is analyzed based on the calibrated hybrid prediction model.

[0013] Optionally, the method further includes: Based on the analysis results of the service performance, a comparison is made with a preset safety threshold; and When the analysis results trigger the security threshold, an early warning signal is automatically generated.

[0014] Optionally, the method further includes: Based on the calibrated hybrid prediction model, a What-if scenario analysis is performed to simulate the performance of the physical bridge under assumed extreme loads or maintenance interventions; and Based on the results of the What-if scenario analysis, an optimal maintenance decision recommendation is generated.

[0015] Optionally, the step of constructing a hybrid prediction model and the dynamic, closed-loop self-calibration further includes: Using a probabilistic artificial intelligence framework based on Bayesian physical information neural networks, the aforementioned biases are decoupled and quantified separately: Random uncertainty represents the inherent, non-simplifiable noise in the real-time monitoring data; and Cognitive uncertainty represents the degree of reducible ignorance arising from model defects in the hybrid prediction model; and In the dynamic self-calibration step (b), an uncertainty-based gated data assimilation mechanism is implemented, wherein the dynamic calibration of the physical parameters by the advanced optimization strategy is performed conditionally: the calibration is performed only when the ratio of the quantified random uncertainty to the cognitive uncertainty is lower than a preset assimilation threshold. Conversely, when the random uncertainty is too high, the influence weight of the real-time monitoring data on the physical parameter calibration is rejected or significantly reduced, thereby preventing the physical parameters of the hybrid prediction model from being eroded by data noise.

[0016] Optionally, the method further includes: Based on the quantified cognitive uncertainty, a cognitive uncertainty map representing its spatial distribution on the physical bridge is generated. An adaptive fidelity control loop is established, which takes the cognitive uncertainty map as input and is used for: Identify specific spatial regions where the cognitive uncertainty persists for more than a preset structural reconstruction threshold; When the specific spatial region is identified, the computational structure of the hybrid prediction model is automatically and locally dynamically reconstructed to reduce the cognitive uncertainty of the specific spatial region. The dynamic reconstruction of the computational structure includes at least one of the following: For the physical mechanism model based on the finite element method: driven by the cognitive uncertainty, adaptive mesh refinement is performed in the specific spatial region; or For the data-driven model: more computing resources are dynamically allocated in the specific spatial region.

[0017] Secondly, an embodiment of the present invention provides a bridge performance analysis system based on digital twin technology, comprising: A data acquisition module, configured to acquire real-time monitoring data of the physical bridge; A processor; and A memory that stores computer-executable instructions; The characteristic is that, when the instruction is executed by the processor, it causes the system to function as follows: A hybrid model module configured to construct a hybrid prediction model containing configurable physical parameters based on the real-time monitoring data and physical mechanisms using physical information artificial intelligence algorithms; A dynamic calibration module configured to perform dynamic, closed-loop self-calibration of the physical parameters of the hybrid prediction model using advanced optimization strategies via Bayesian inference or reinforcement learning; and A performance analysis module is configured to analyze the service performance of the bridge based on a calibrated hybrid prediction model.

[0018] The present invention has achieved the following beneficial effects: High fidelity and robustness: This invention constructs a hybrid prediction model using a physical information artificial intelligence algorithm and synchronizes its parameters with physical entities through dynamic self-calibration. More importantly, through an uncertainty-based gating data assimilation mechanism, this invention can intelligently distinguish between inherent noise (random uncertainty) in the data and defects in the model itself (cognitive uncertainty). It can automatically close the calibration gating when sensor data is filled with noise or malfunctions, effectively preventing the physical parameters of the hybrid prediction model from being corrupted (i.e., contaminated) by bad data or noise, ensuring the high fidelity and robustness of the digital twin model during long-term operation.

[0019] Computational Resource Optimization and Adaptive Fidelity: This invention, through an adaptive fidelity control loop, can dynamically generate a cognitive uncertainty map and automatically identify specific spatial regions that the model is unaware of (i.e., those where predictions are inaccurate). Once identified, the system automatically and locally performs dynamic reconstruction in that region, such as performing adaptive mesh refinement (AMR) at the physical model (FEM) level or dynamically allocating more computational resources at the data model (NN) level. This intelligent on-demand allocation strategy solves the problems of high computational cost and inability to run in real time in traditional high-fidelity models, achieving an optimal dynamic balance between computational efficiency and model accuracy.

[0020] Intelligent and forward-looking decision support: This invention is based on a high-fidelity, highly robust, and computationally efficient hybrid prediction model, which transforms bridge health monitoring from a manual, slow-response, and passive mode in the past to an automated, high-fidelity, proactive early warning, and forward-looking decision-making paradigm. It can not only accurately analyze key indicators such as fatigue life and seismic performance, but also provide automatic early warnings of behaviors that trigger safety thresholds. Furthermore, through What-if scenario analysis, it can simulate performance under extreme loads or maintenance interventions for managers and generate optimal maintenance decision recommendations, thus realizing intelligent bridge management and maintenance.

[0021] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0022] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of a bridge performance analysis method based on digital twin technology in an embodiment of the present invention; Figure 2 This is a schematic diagram of a bridge performance analysis system based on digital twin technology in an embodiment of the present invention. Detailed Implementation

[0024] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0025] Example 1: Figure 1 A flowchart of a bridge performance analysis method based on digital twin technology is provided as an embodiment of this application, such as... Figure 1 As shown, the method includes: Obtain real-time monitoring data of the physical bridge; Based on the structural information and physical mechanism of the physical bridge, a digital twin model containing configurable physical parameters is constructed. The method further includes: (a) Based on the real-time monitoring data and the digital twin model, a hybrid prediction model is constructed using a physical information artificial intelligence algorithm; (b) Using advanced optimization strategies such as Bayesian inference or reinforcement learning, the physical parameters of the hybrid prediction model are dynamically and in a closed-loop manner to minimize the deviation between the predicted response of the hybrid prediction model and the real-time monitoring data; and (c) The service performance of the bridge is analyzed based on the calibrated hybrid prediction model in step (b).

[0026] This embodiment provides a basic process for a bridge performance analysis method based on digital twin technology, such as... Figure 1 As shown, its detailed working principle is as follows: First, the system continuously acquires real-time monitoring data from sensors through data acquisition modules deployed on the physical bridge. Simultaneously, based on the structural design information (e.g., drawings, material reports) and physical mechanisms (e.g., structural mechanics principles) of the physical bridge, the system constructs a high-fidelity digital twin model. This model contains a set of configurable physical parameters (e.g., elastic modulus representing material stiffness, damping coefficient representing energy dissipation capacity, etc.), which are initially set to design or empirical values. Next, the system executes step (a), using a physical information artificial intelligence algorithm to construct a hybrid prediction model based on the acquired real-time monitoring data and the digital twin model. This model can both simulate physical mechanisms and learn data characteristics. Subsequently, the system enters the core step (b), namely the dynamic, closed-loop self-calibration process. The system uses the latest real-time monitoring data (e.g., acceleration at a certain moment) to generate a new data set. The system inputs a prediction response (e.g., a certain degree of accuracy) into the hybrid prediction model, which then outputs a predicted response. The system immediately compares this predicted response with real-time monitoring data and calculates the deviation between the two. Once a deviation is detected, the system immediately initiates an advanced optimization strategy (e.g., Bayesian inference or reinforcement learning). The sole objective of this strategy is to automatically and dynamically adjust (i.e., calibrate) the set of configurable physical parameters in the digital twin model until the deviation is minimized. This calibration process is closed-loop (i.e., a cycle of data-model-deviation-calibration parameters-new model) and dynamic (i.e., parameters are constantly updated over time), ensuring that the digital twin model remains highly synchronized with the actual state of the physical bridge. Finally, the system executes step (c), which, based on the high-fidelity hybrid prediction model that has just been dynamically calibrated in step (b), accurately analyzes and predicts the bridge's service performance (e.g., load-bearing capacity, remaining lifespan), thereby providing managers with reliable decision-making support.

[0027] The present invention provides embodiments of the above-described calibration options for reinforcement learning (RL): The three essential elements of reinforcement learning are state, action, and reward. State is a snapshot of the system at a specific moment, providing the reinforcement learning agent with sufficient information to make decisions. In this invention, state can be defined as a vector containing at least: first, the deviation between the response predicted by the hybrid prediction model at the current moment and the real-time monitoring data; second, the rate of change of this deviation at the previous moment (i.e., the derivative of the deviation); and third, the values ​​of all configurable physical parameters used in the current hybrid prediction model (e.g., elastic modulus, damping coefficient, etc.). Action is an operation that the agent (i.e., the calibration module) can perform to attempt to change the state and obtain a reward. In this invention, the action space can be defined as a set of discrete adjustment instructions for physical parameters. For example, the action space may include: increasing the elastic modulus parameter by one percent, decreasing the elastic modulus parameter by one percent, increasing the damping coefficient by two percent, decreasing the damping coefficient by two percent, or keeping all parameters unchanged. Reward is a scalar signal fed back to the agent by the environment (i.e., the digital twin model) after the agent performs an action, used to evaluate the quality of the action. In this invention, the objective of the reward function must be to minimize the deviation. Therefore, an exemplary reward function can be defined as follows: if, after performing an action, the absolute value of the deviation observed at the next time step is less than the absolute value of the deviation at the current time step, the system awards the agent a positive reward (e.g., a reward of positive one); conversely, if the absolute value of the deviation increases or remains unchanged, the system awards the agent a negative reward or penalty (e.g., a reward of negative one or zero). Through training, the reinforcement learning agent will learn an optimal policy (i.e., what action to perform in what state) to maximize its long-term cumulative reward, thereby achieving dynamic self-calibration of physical parameters and continuously minimizing the deviation.

[0028] The present invention also provides embodiments for the above-described Bayesian inference calibration options: Unlike traditional methods that treat physical parameters (e.g., elastic modulus) as deterministic, unknown scalars, Bayesian inference treats this parameter as a probability distribution. Before acquiring any new monitoring data, the system establishes a prior distribution for the parameter based on engineering experience or design drawings. For example, the system assumes that the prior distribution of the elastic modulus is a Gaussian distribution with a mean of the design value and a variance of 5%, representing our initial belief in the parameter. When new real-time monitoring data is acquired, the system constructs a likelihood function. This function describes the probability of observing the current set of monitoring data given a specific value of the elastic modulus parameter. The core of Bayesian inference is to use Bayes' theorem to mathematically combine the prior distribution and the likelihood function to calculate a new, updated probability distribution, called the posterior distribution. This posterior distribution represents the system's updated belief in the elastic modulus parameter after observing new data. In practice, due to the complexity of calculating the posterior distribution, the system can employ advanced sampling algorithms (such as Markov Chain Monte Carlo (MCMC) or Hamiltonian Monte Carlo (HMC) methods) to perform this inference. These algorithms approximate the shape of the posterior distribution by performing extensive random sampling in the parameter space. The system ultimately uses the statistical characteristics of this posterior distribution (e.g., the peak or mean of the distribution) as the latest self-calibration result for the physical parameter and updates the hybrid prediction model.

[0029] Example 2: In one embodiment, the real-time monitoring data originates from a multi-source heterogeneous sensor network, which includes at least one of the following: Strain gauges, accelerometers, displacement sensors, or fiber optic grating sensors.

[0030] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 1, its detailed working principle is as follows: The real-time monitoring data acquired by this method originates from a multi-source heterogeneous sensor network installed on the physical bridge. This network is called multi-source heterogeneous because it integrates various types of sensors with different operating principles to capture the complex response of the bridge from multiple dimensions. For example, the network includes at least strain gauges, which are attached or embedded in the surface or interior of key load-bearing components of the bridge (such as the upper and lower edges of the main beam, the anchorage zone of the suspension cables, and steel structure welds) to measure with high precision the local tensile or compressive deformation of the structure under loads (such as vehicles, wind, and temperature changes), i.e., strain data. The network may also include accelerometers, which are typically deployed at globally critical locations on the bridge (such as the top of the bridge towers and the mid-span of the main beam) to measure the vibration characteristics of the structure under dynamic loads (such as earthquakes, wind vibrations, vehicle braking, or jumping), i.e., the acceleration response time history. This is crucial for analyzing the structural dynamic characteristics (such as natural frequency and damping). The network includes core data for the bridge; it may also include displacement sensors (such as wire-guided displacement gauges or laser displacement gauges), which are typically deployed at expansion joints, supports, or between two beam segments to measure relative displacement between structural components or absolute displacement of the structure (such as deflection); in addition, the network may include advanced fiber Bragg grating (FBG) sensors, which can be fabricated into long-distance sensor chains and quasi-distributedly laid along the main cable, bridge deck, or steel box girder of the bridge to simultaneously acquire high-density, multi-point strain and temperature field data along its path; the system's data acquisition module is configured to perform time synchronization, data fusion, and preprocessing of these multi-source heterogeneous data from different sources, of different types, and with potentially different sampling frequencies, thereby providing a unified, comprehensive, and highly informative real-time data input for subsequent hybrid prediction models, ensuring the accuracy of model calibration.

[0031] Example 3: In one embodiment, the hybrid prediction model includes: A physical mechanism model based on the finite element method is used to provide physical constraints for the structural response; and A data-driven model of a neural network is used to learn nonlinear dynamic characteristics not modeled by the physical mechanism model from the real-time monitoring data; The physical information artificial intelligence algorithm integrates the physical constraints into the training process of the data-driven model.

[0032] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 1, the detailed working principle of its hybrid prediction model is as follows: The hybrid prediction model creatively integrates the advantages of two models to overcome the shortcomings of purely physical models (too idealistic) and purely data models (too blind) in the background technology. Firstly, the hybrid model includes a physical mechanism model based on the Finite Element Method (FEM). This model is constructed based on classical structural mechanics, materials mechanics, and dynamic governing equations (usually expressed as a set of partial differential equations). It represents the prior knowledge of bridge structural behavior held by human engineers. Its core role in this hybrid model is to provide physical constraints on the structural response; that is, it defines a physically possible boundary for the model's predictions, ensuring that any predictions (e.g., the relationship between stress and strain, energy conservation, etc.) do not violate fundamental physical laws. This is particularly important in regions where monitoring data is sparse. Secondly, the hybrid model includes a data-driven model of a neural network (e.g., a deep neural network or a recurrent neural network). This model acts as a powerful nonlinear function fitter, and its core function is to learn from the real-time monitoring data obtained in Example 2. The complex nonlinear dynamic characteristics that physical mechanism models (FEMs) do not model or are difficult to model, such as the plastic hysteresis of materials under extreme loads, friction and nonlinear contact of complex supports, or local stiffness degradation caused by the accumulation of micro-damage, are the most difficult parts to accurately describe by traditional FEM models. The third part is the hybrid mechanism of physical information artificial intelligence algorithms. When training the neural network data-driven model, the algorithm adopts a unique strategy: its optimization goal is not only to minimize the error between the model prediction and the sensor data (which is the only goal of traditional neural networks), but also to minimize the extent to which the model prediction violates the physical constraints represented by the FEM model (i.e., physical residuals). In other words, the algorithm forces the neural network to find the best balance between fitting the data and conforming to physics. The resulting hybrid prediction model can capture the complex nonlinear dynamics hidden in the data and not modeled, just like a data model, and can also provide physically logical, interpretable, and extrapolable predictive responses in areas with sparse data or when facing unseen load conditions, just like a physical mechanism model.

[0033] Furthermore, the training objective (i.e., loss function) of the physical information artificial intelligence algorithm is defined as a mixed, weighted sum that contains at least two components: Part 1: Data Loss. This part measures the accuracy of the model's predictions. It calculates the difference (e.g., the mean square error of the two) between the neural network model's predicted response (e.g., predicted displacement) at the physical location where the sensor is deployed and the data actually measured by the sensor (e.g., measured displacement).

[0034] Part Two: Physical Loss. This part measures how well the model adheres to the laws of physics. It is calculated at numerous other spatial points (i.e., so-called placement points) where no sensors are deployed. The system substitutes the neural network's predictions (e.g., predicted displacement fields) into the known physical governing equations (e.g., partial differential equations of structural dynamics) that describe the behavior of the bridge structure. If the predictions perfectly adhere to the laws of physics, the residuals of the equations should be zero; otherwise, a non-zero residual value is produced. This physical loss term is used to penalize these non-zero residuals.

[0035] During training, the algorithm optimizes the neural network parameters simultaneously by minimizing the weighted sum of the data loss and physical loss. By adjusting the weight coefficients between the two, the algorithm forces the neural network not only to fit sparse sensor data (data loss) but also to obey physical laws in the vast regions where data is missing (physical loss), thereby achieving high-fidelity predictions.

[0036] Example 4: In one embodiment, the physical parameters of the dynamic self-calibration include at least one of the following: The elastic modulus of the physical bridge material, the structural damping coefficient, or the boundary conditions between structural components.

[0037] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 1 or 3, the detailed working principle of its dynamic self-calibration physical parameters is as follows: The configurable physical parameters adjusted in the dynamic, closed-loop self-calibration step (b) described in Embodiment 1 are core parameters that have clear physical meaning in the digital twin model and whose changes can directly reflect the true health status of the physical bridge. These parameters include at least one of the following: First, the elastic modulus of the physical bridge material. The elastic modulus is a direct measure of the stiffness of engineering materials (such as concrete and steel). During the long-term service of the bridge, concrete may experience performance degradation due to carbonation, creep, alkali-aggregate reaction, or freeze-thaw cycles, and steel may develop microcracks due to fatigue accumulation. These damages and aging will macroscopically manifest as a decrease in the effective elastic modulus of the material. Therefore, dynamically and in a closed loop, the elastic modulus parameter is essentially the system tracking the damage accumulation of the bridge structural materials in real time. The first factor is the degree of aging and wear. The second factor is the structural damping coefficient. Damping is the ability of a structure to dissipate energy during vibration. When the connecting parts of a bridge (such as bolts and welds) become loose, new friction occurs between the parts, or micro-cracks invisible to the naked eye appear, the structural damping usually changes (usually increases). Therefore, calibrating the structural damping coefficient is a way for the system to indirectly monitor early damage or changes in connection status that are difficult to detect directly with traditional sensors. The third factor is the boundary conditions between structural components. This is the most difficult parameter to accurately describe in the FEM model, but it has a huge impact on the global response of the structure (especially the seismic response). For example, a support may be designed to slide, but after years of service, it may become semi-consolidated or stuck due to corrosion, dust accumulation, or damage. Dynamic self-calibration of boundary condition parameters (such as using a nonlinear spring stiffness coefficient as an equivalent) corrects the huge difference between the idealized FEM model and the actual structural behavior, ensuring that the model fundamentally reflects reality.

[0038] Example 5: In one embodiment, step (c) of analyzing the service performance of the bridge includes at least one of the following: Based on the calibrated hybrid prediction model, structural stress analysis under random traffic loads is performed to predict structural fatigue life. Based on the calibrated hybrid prediction model, the structural response of the physical bridge under seismic loads is evaluated to assess its seismic performance and damage; or The ultimate bearing capacity of the physical bridge is analyzed based on the calibrated hybrid prediction model.

[0039] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 1, the detailed working principle of step (c) of analyzing the service performance of the bridge is as follows: After the hybrid prediction model is dynamically calibrated through step (b) of Embodiment 1, it becomes a high-fidelity copy of the physical bridge at the current moment. The system then uses this calibrated model to perform at least one of the following high-performance analysis tasks: First, based on the calibrated hybrid prediction model, structural stress analysis under random traffic loads is performed to predict the structural fatigue life. The working principle is that the system will use the real (Or, based on statistically generated) random traffic load flow (i.e., vehicle flow) is applied to this high-fidelity model for high-precision time-history analysis. This calculates the stress amplitude and cycle count of key fatigue points on the bridge (e.g., steel box girder welds, cable anchorage zones). Then, based on material fatigue cumulative damage theory (e.g., Miner's rule), the cumulative damage degree of these key points is calculated and updated in real time, and their remaining fatigue life is predicted, providing managers with a basis for deciding when repair or replacement is necessary. Secondly, based on the calibrated hybrid prediction model, the structural response of the physical bridge under seismic loads is evaluated to assess its seismic performance. The damage assessment works by applying a standard (e.g., a rare or fortification level as defined by regulations) seismic acceleration time history to a calibrated model that accurately reflects the true stiffness, damping, and boundary conditions (according to Example 4). The model then simulates the nonlinear dynamic response of the bridge under earthquake conditions with high fidelity, precisely assessing whether the bridge's displacement, acceleration, and internal forces will exceed safety limits, and automatically identifying weak points where damage may occur (e.g., the formation of plastic hinges), thus completing the seismic performance and damage assessment. Thirdly, based on the calibrated hybrid prediction model, the extreme values ​​of the physical bridge are analyzed. The load-bearing capacity limiter works by using a calibrated model that has captured nonlinear characteristics (according to Example 3) to virtually and gradually increase the load on the model (e.g., simulating increasing vehicle weight or lateral wind force), while considering the nonlinear effects of materials and geometry, and performing pushover analysis or nonlinear dynamic analysis until the model calculates that the structure has reached the limit point of instability or material strength failure. This limit point represents the bridge's current (not the initial design) ultimate load-bearing capacity, providing the most scientific basis for assessing the bridge's load-bearing capacity and granting passage permits for overweight vehicles.

[0040] Example 6: In one embodiment, the method further includes: Based on the analysis results of the service performance, a comparison is made with a preset safety threshold; and When the analysis results trigger the security threshold, an early warning signal is automatically generated.

[0041] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 5, its detailed working principle is as follows: Based on the performance analysis described in Embodiment 5, this method further includes an automated early warning mechanism. First, during system deployment, a safety threshold database is built-in. This database is based on current national bridge design specifications, highway bridge maintenance specifications, and pre-set thresholds for the specific bridge's design characteristics and importance. These thresholds are hierarchical, for example, they may include a Level 1 (attention level) threshold (e.g., a certain indicator slightly exceeds the normal fluctuation range, indicating the need for attention), a Level 2 (early warning level) threshold (e.g., the stress of a key section is close to the design limit, or the natural frequency of the structure undergoes a slight but clear change), and a Level 3 (alarm level) threshold (e.g., the measured or predicted displacement or strain exceeds the limit value of the specification, or the fatigue life is lower than the preset number of years). Subsequently, the performance analysis module in this method will, based on a calibrated hybrid prediction model, perform real-time or near-real-time analysis... (For example, every five minutes or whenever a heavy vehicle passes) calculates various key performance indicators of the bridge (e.g., mid-span deflection, main cable stress, bridge tower vibration acceleration, fatigue cumulative damage of key parts, etc.); the system continuously and automatically compares these real-time calculated performance indicators with a preset safety threshold database; once the system detects that the analysis results of any one or more indicators trigger (i.e. reach or exceed) a preset safety threshold, the system will automatically and immediately generate an early warning signal. This signal will clearly indicate the trigger time of the warning, the physical location of the warning (e.g., the top of the second pier or the third suspender), the specific indicator that triggered the warning (e.g., lateral displacement or fatigue damage), the current indicator value, and the threshold level triggered. This warning signal will be automatically pushed to the bridge maintenance department through the system interface, SMS, email, or dedicated maintenance software, thereby minimizing the delay from risk occurrence or risk prediction to manual response and completely changing the passive inspection mode with poor timeliness in the background technology.

[0042] Example 7: In one embodiment, the method further includes: Based on the calibrated hybrid prediction model, a What-if scenario analysis is performed to simulate the performance of the physical bridge under assumed extreme loads or maintenance interventions; and Based on the results of the What-if scenario analysis, an optimal maintenance decision recommendation is generated.

[0043] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 1, its detailed working principle is as follows: Based on the calibrated hybrid prediction model described in Embodiment 1, this method further provides a powerful What-if scenario analysis (i.e., hypothetical scenario analysis) function. This is a digital sandbox providing forward-looking decision support for bridge managers. Its working principle is that the system allows bridge engineers or maintenance personnel to input a hypothetical scenario into this calibrated hybrid prediction model, which is highly synchronized with the physical bridge, through an interactive interface. This scenario analysis includes at least two types: The first type is hypothetical extreme load simulation. For example, an engineer can ask what would happen to the bridge if it encountered a once-in-a-century typhoon, and then apply the corresponding hypothetical typhoon wind field data to the model; or if an overweight convoy (e.g., with a total weight of 200 tons) illegally crosses the bridge, is the bridge safe? Then, this virtual overweight convoy load is applied to the model. The system will use the calibrated model to run these simulations with high fidelity and output the bridge's performance under these conditions. The system simulates the performance under hypothetical extreme loads (e.g., which components will reach their stress limits first, and what the maximum displacement will be). The second category involves hypothetical maintenance intervention simulations. For example, the maintenance department plans to choose between replacing the dampers on the No. 3 main cable and performing preventative reinforcement of the bridge deck. Engineers can input these maintenance interventions (represented in the model as parameter modifications, such as increasing the damping coefficient of the No. 3 main cable or increasing the stiffness of the bridge deck) into the model. The system will simulate and predict the bridge's performance after implementing these interventions (e.g., how much vibration will be reduced under standard wind loads). Finally, based on the simulation results of these What-if scenario analyses (e.g., showing that replacing the dampers has a much better vibration reduction effect than reinforcing the bridge deck), the system will automatically or assist engineers in generating an optimal maintenance decision recommendation. This recommendation clearly indicates which maintenance option is technically most effective and economically most cost-effective, thus elevating bridge maintenance decision-making from a traditional model relying on experience and trial and error to a new intelligent paradigm relying on scientific deduction and data-driven approaches.

[0044] Example 8: In one embodiment, the step of constructing a hybrid prediction model and the dynamic, closed-loop self-calibration further includes: Using a probabilistic artificial intelligence framework based on Bayesian physical information neural networks, the aforementioned biases are decoupled and quantified separately: Random uncertainty represents the inherent, non-simplifiable noise in the real-time monitoring data; and Cognitive uncertainty represents the degree of reducible ignorance arising from model defects in the hybrid prediction model; and In the dynamic self-calibration step (b), an uncertainty-based gated data assimilation mechanism is implemented, wherein the dynamic calibration of the physical parameters by the advanced optimization strategy is performed conditionally: the calibration is performed only when the ratio of the quantified random uncertainty to the cognitive uncertainty is lower than a preset assimilation threshold. Conversely, when the random uncertainty is too high, the influence weight of the real-time monitoring data on the physical parameter calibration is rejected or significantly reduced, thereby preventing the physical parameters of the hybrid prediction model from being eroded by data noise.

[0045] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 1, its detailed working principle in constructing the hybrid prediction model (step a) and dynamic self-calibration (step b) is as follows: To address the core challenge of data inevitably containing noise in real-world monitoring environments, and the inherent imperfections of models, this embodiment employs an advanced probabilistic artificial intelligence framework: Bayesian Physics-Informed Neural Networks (B-PINN). This framework enables deep insight and intelligent processing of the total deviation between model predictions and monitoring data. Its complete working principle is as follows: First, unlike the standard Physical Information Neural Network (PINN) described in Embodiment 3, where the internal parameters (e.g., weights and biases) are treated as fixed, deterministic values ​​after training, the standard PINN outputs only a definite predicted response (e.g., a deflection of 10.1 mm) for a given input (e.g., load location). The B-PINN used in this embodiment, however, follows the Bayesian framework, treating these internal parameters not as fixed values ​​but as probability distributions. This means the model has some uncertainty regarding each of its parameters. When B-PINN makes a prediction, it does not use a fixed set of... Instead of using parameters, B-PINN samples from its internal parameter probability distribution (e.g., through advanced probabilistic programming techniques such as variational inference (VI) or Hamiltonian Monte Carlo (HMC)) to produce a prediction. By repeating this sampling and prediction process over a very short period of time (e.g., five hundred times), B-PINN will output five hundred slightly different predictions for the same input. These five hundred results together constitute a probability distribution of the predicted response (e.g., the system may report a mean deflection prediction of 10.1 mm, but with 95% confidence that it falls within the range of 9.8 mm to 10.4 mm). The width or variance of this probability distribution intuitively quantifies the model's total uncertainty or confidence in its own predictions.

[0046] The next crucial step is to utilize this probabilistic AI framework to decouple the total bias and quantify its sources: The first type of uncertainty to be quantified is aleatoric uncertainty, which represents the inherent, non-reducible noise in the real-time monitoring data. Its physical sources include sensor measurement noise (e.g., current fluctuations), unpredictable minor vibrations in the environment (e.g., background vibrations caused by distant traffic), or random interference introduced during data transmission. This uncertainty is non-reducible, meaning that even a perfect, omniscient hybrid prediction model cannot eliminate it, as it is a property of the data itself, not the model. In the B-PINN framework, this uncertainty is typically quantified by explicitly learning a noise variance associated with the input data at the model's final output layer. The second type of uncertainty to be quantified is epistemological uncertainty. micUncertainty represents a simplifiable level of ignorance inherent in the hybrid prediction model itself due to model defects. Its physical sources include oversimplification of the physical mechanism model (FEM) (e.g., overly coarse meshing, inappropriate boundary condition assumptions), undertraining of the data-driven model (NN) (e.g., in a specific area of ​​a bridge, sensor data is very sparse, causing the model to be unfamiliar with that area and hesitant to make a guess), or limitations in the model structure itself (e.g., failure to capture a complex nonlinear effect). This uncertainty is simplifiable, meaning it can be reduced by improving the model (e.g., using a denser mesh, a more complex network) or increasing training data. In the B-PINN framework, this uncertainty is captured precisely through the probability distribution of the aforementioned model parameters (weights and biases). If the model is ignorant of a certain area, the probability distribution of its internal parameters will be wide, resulting in a wide final prediction distribution, thus quantifying a high level of cognitive uncertainty.

[0047] After successfully quantifying and decoupling these two fundamentally different uncertainties, the core mechanism of this embodiment, the uncertainty-based gated data assimilation mechanism, is activated and implemented in the dynamic self-calibration step (b). Data assimilation refers to the process of integrating new real-time monitoring data into the existing model to calibrate the model's physical parameters (such as the elastic modulus, damping coefficient, etc. described in Embodiment 4). The gated mechanism in this invention is an intelligent, conditional filter that uses the aforementioned quantified uncertainty ratio to determine whether and how to assimilate (i.e., use) the newly arrived data point. The system first presets an assimilation threshold, which represents the system's tolerance boundary for data noise and model ignorance. For each newly acquired real-time monitoring data point, the system executes the following gated judgment logic: The system calculates the ratio of the random uncertainty (AU) to the cognitive uncertainty (EU) corresponding to the data point (or some preset relationship between the two). In the first scenario (i.e., the gate is open), when the system determines that the random uncertainty caused by the data point is much lower than the cognitive uncertainty (e.g., the ratio of AU to EU is lower than the preset assimilation threshold), the system interprets it as follows: I (the model) am aware of the uncertainty caused by the data point. In the previous scenario, the prediction was highly uncertain (high EU), but I judged that the new data point itself was clean and reliable (low AU), which is a valuable new knowledge. At this time, the gating was turned on, and the system sent this reliable data point completely (or with high weight) to the dynamic calibration module. Advanced optimization strategies (such as Bayesian inference) will use the deviation between this data point and the model prediction to perform meaningful dynamic calibration of the physical parameters (such as elastic modulus) of the hybrid prediction model. In the second scenario (i.e., the gating was turned off), conversely, when the system judges that the random uncertainty caused by the data point is too high (for example, the ratio of AU to EU is higher than the preset assimilation threshold), the system will interpret it as follows: the new data point itself is full of noise (high AU), and its noise level even exceeds the ignorance of the model itself. This is very likely to mean that the sensor has failed, there is a momentary error in the data transmission, or the bridge has suffered a sudden local impact that the model cannot understand (such as construction misoperation). At this time, the system judges this to be bad data or contaminated data, the gating is turned off, and the system will refuse to use this data point for parameter calibration or significantly reduce the influence weight of the data point in the calibration process.

[0048] Through the aforementioned gating data assimilation mechanism, this invention solves a crucial yet often overlooked model corruption problem in long-running digital twin systems: In traditional models without this mechanism, if a sensor malfunctions and begins sending continuous, erroneous data, traditional data assimilation algorithms blindly accept this bad data and incorrectly calibrate the model's physical parameters (e.g., incorrectly judging the stiffness of a bridge as a sharp decline). This leads to the corruption or contamination of the entire digital twin model, decoupling it from the real world and resulting in catastrophic performance analyses and warnings based on this flawed model. The gating mechanism of this invention acts as an intelligent guardian. Through the probabilistic reasoning capabilities of the Bayesian physical information neural network, it self-perceives when data is unreliable (high AU) and when the model needs learning (high EU). It only opens the calibration gate when the data is reliable and the model needs learning, thus preventing the physical parameters of the hybrid prediction model from being corrupted by data noise and ensuring the long-term robustness, reliability, and high fidelity of the digital twin model throughout its entire lifecycle.

[0049] Example 9: In one embodiment, the method further includes: Based on the quantified cognitive uncertainty, a cognitive uncertainty map representing its spatial distribution on the physical bridge is generated. An adaptive fidelity control loop is established, which takes the cognitive uncertainty map as input and is used for: Identify specific spatial regions where the cognitive uncertainty persists for more than a preset structural reconstruction threshold; When the specific spatial region is identified, the computational structure of the hybrid prediction model is automatically and locally dynamically reconstructed to reduce the cognitive uncertainty of the specific spatial region. The dynamic reconstruction of the computational structure includes at least one of the following: For the physical mechanism model based on the finite element method: driven by the cognitive uncertainty, adaptive mesh refinement is performed in the specific spatial region; or For the data-driven model: more computing resources are dynamically allocated in the specific spatial region.

[0050] This embodiment provides a bridge performance analysis method based on digital twin technology. As a further limitation of Embodiment 8, its detailed working principle is as follows: This embodiment builds upon the successful quantification of cognitive uncertainty in Embodiment 8, aiming to resolve the inherent contradiction between computational efficiency and model accuracy in digital twin models. Its core is to establish an adaptive fidelity control loop that enables the model to achieve self-optimization and intelligent evolution. Its complete working principle is as follows: First, based on the probabilistic evaluation of the full-bridge model using a Bayesian Physical Information Neural Network (B-PINN) as described in Example 8, the system obtains quantitative values ​​of epistemic uncertainty (EU) at various spatial locations of the bridge. Epistemic uncertainty (EU) represents the degree of ignorance or lack of confidence in the model's predictive ability at that location. The system then overlays these EU values ​​distributed across the bridge structure onto the bridge's digital twin model (e.g., a visualization model of BIM or FEM) to generate an epistemic uncertainty map (EU). The pistemicUncertaintyMap is dynamically updated and visually displays to the system, in the form of heat maps (similar to infrared thermography), which parts of the bridge the model predicts with high confidence (low EU values, displayed as cool blue areas on the map) and which parts (such as a complex tower-beam connection node, the vicinity of a cable whose parameters have just been changed, or an area never covered by sensor data) are highly uncertain or unknown (high EU values, displayed as warm red hot spots on the map).

[0051] Next, the system establishes and initiates an adaptive fidelity control loop, a continuously running, fully automated background monitoring process. It is unconcerned with the bridge's stress or displacement (that's the task of Example 6), focusing solely on the model's own health status; that is, it uses this dynamic cognitive uncertainty map as its sole control input. The first step of this control loop is identification: the system presets a structural reconstruction threshold, representing the upper limit of the system's tolerance for model ignorance (e.g., EU value exceeding a certain percentage). Below this threshold, it means that although the model has some uncertainty, its fidelity is sufficient to meet the needs of performance analysis. The control loop continuously and automatically scans the cognitive uncertainty map and performs identification tasks: it searches for specific spatial regions on the map where the quantified cognitive uncertainty has consistently (e.g., within the past hour or under multiple load conditions) exceeded the preset structural reconstruction threshold. This sustained exceedance is a key judgment, ruling out misjudgments caused by instantaneous data fluctuations. Once such a region is identified (e.g., the system identifies the tower-beam connection area of ​​Bridge Tower 3 as an EU hotspot), the system marks it as an area with severely insufficient model fidelity. After identification, the control loop immediately triggers the second step, which is to automatically and locally perform dynamic reconstruction of the computational structure of the hybrid prediction model. This reconstruction is automatic (without manual intervention) and local (only targeting the identified hotspot regions, not the entire large model). The goal of dynamic reconstruction is to precisely and on-demand allocate computational resources to this high-ignorance region to reduce the cognitive uncertainty in that region, i.e., to improve the model fidelity in that region. Since the hybrid prediction model of this invention (as described in Embodiment 3) comprises two parts: a physical mechanism model based on the finite element method (FEM part) and a data-driven model based on a neural network (NN part), this dynamic reconstruction also includes two targeted actions accordingly: The first type of reconstruction action is Adaptive Mesh Refinement (AMR) for the physical mechanism model. Its trigger condition is that if the system, through diagnostic analysis, determines that the high cognitive uncertainty in the hotspot region is mainly due to the physical mechanism model (FEM) describing that region being too coarse (e.g., the region exhibits drastic stress gradient changes and is a stress concentration zone, but the FEM mesh is too coarse, resulting in inaccurate physical constraints). In this case, the control loop uses the cognitive uncertainty (EU value) as a driving signal to automatically execute AMR on the finite element (FEM) model portion of that specific spatial region. Its working principle is that AMR... The technology automatically subdivides the finite element mesh elements (such as triangular or tetrahedral elements) in the identified high EU regions (e.g., a large element is subdivided into four or eight smaller elements) without stopping the model (or in the next iteration), while the other large blue regions of the bridge model (with low EU values) retain their original coarse mesh. As a result, the resolution of the FEM model in the hot spot region is significantly improved, enabling it to capture the physical behavior (such as stress concentration) of the region more accurately. This leads to improved physical constraint quality in the region, thereby effectively reducing cognitive uncertainty in the region (within the B-PINN framework) while avoiding the huge computational burden brought about by global mesh refinement.

[0052] The second type of reconfiguration action involves dynamically allocating more computational resources to the data-driven model. The trigger condition is that if the system diagnostic analysis determines that the high cognitive uncertainty in the hotspot region is primarily due to the data-driven model (NN) failing to learn the complex nonlinear dynamics of that region (e.g., the nonlinear frictional behavior of a support, the plastic development of a component's material, which are difficult for FEM to model and should be learned by the NN), the control loop will trigger dynamic resource allocation to the data-driven model (NN). Its working principle is that the system automatically and locally adjusts the computational structure of the neural network to increase its learning and expressive capabilities in that specific region. Specific allocation methods may include: in the neural network, the negative... The system dynamically increases the number of neurons on those paths that process inputs from (or influence) a specific spatial region; or adds additional hidden layers to that specific path to enable it to learn deeper, more complex abstract features; or, in subsequent dynamic self-calibration (step b), when data points come from this high EU region, the system allocates more training iterations or higher learning rate weights to it (e.g., by optimizing resource allocation through reinforcement learning strategies), making it work harder to learn the characteristics of that region. The effect is that by dynamically and locally increasing the complexity and computational resources of the NN, the data-driven model can more fully learn the complex nonlinear dynamics of the region from the data, thereby effectively reducing the cognitive uncertainty of the region.

[0053] Finally, this adaptive fidelity control loop is closed-loop: after performing one or two of the above dynamic reconstructions, the computational structure of the hybrid prediction model changes (e.g., the FEM mesh becomes denser, or the NN becomes more complex); the system recalculates the cognitive uncertainty map based on the new model structure; the control loop continues to monitor the map, and if the EU value of the hotspot region has been successfully reduced to below the structural reconstruction threshold (i.e., the hotspot has cooled down), the dynamic reconstruction task for that region is completed and stopped, and computational resources are released; if it is still above the threshold, the next round of reconstruction may be initiated.

[0054] The fundamental value of this embodiment lies in resolving the efficiency-accuracy paradox of digital twins: traditional models either choose low fidelity (fast computation, but unreliable results) or high fidelity (global high-precision FEM / NN, but with huge computational load and unable to operate in real time); this invention creates an intelligent, self-optimizing model through this adaptive fidelity control loop. It only uses the necessary resources (AMR or dynamic NN) in the necessary locations (high EU regions) to achieve on-demand fidelity allocation, enabling the system to dynamically and continuously maintain its highest prediction accuracy in all key regions with the lowest overall computational cost. This is the core key to achieving scalable, maintainable, and cost-effective bridge digital twins.

[0055] Example 10: Figure 2 This application provides a schematic diagram of a bridge performance analysis system based on digital twin technology, as shown in the embodiment of the present application. Figure 2 As shown, the system includes: A data acquisition module, configured to acquire real-time monitoring data of the physical bridge; A processor; and A memory that stores computer-executable instructions; The characteristic is that, when the instruction is executed by the processor, it causes the system to function as follows: A hybrid model module configured to construct a hybrid prediction model containing configurable physical parameters based on the real-time monitoring data and physical mechanisms using physical information artificial intelligence algorithms; A dynamic calibration module configured to perform dynamic, closed-loop self-calibration of the physical parameters of the hybrid prediction model using advanced optimization strategies via Bayesian inference or reinforcement learning; and A performance analysis module is configured to analyze the service performance of the bridge based on a calibrated hybrid prediction model.

[0056] This embodiment provides a bridge performance analysis system based on digital twin technology, serving as the hardware and software entity for implementing the methods described in embodiments one through nine above. Its detailed working principle is as follows: The system can be deployed on a local server at the bridge management center or on a cloud platform. The system includes hardware components: a data acquisition module configured to communicate with a multi-source heterogeneous sensor network (as described in Embodiment 2) installed on the physical bridge via wired or wireless communication to acquire real-time monitoring data such as strain, acceleration, displacement, and fiber optic grating readings. This module is also responsible for data collection, preliminary cleaning, and timestamp synchronization; a processor (e.g., a central processing unit (CPU) or a graphics processing unit (GPU), or a combination of both); and a memory (e.g., a hard disk, solid-state drive, or RAM) storing computer-executable instructions. The feature of this embodiment is that, when the instructions are executed by the processor, the system is implemented as a series of collaborative functional modules: First, it is implemented as a hybrid model module, which is configured to execute the model building in Embodiment 1 and the hybrid process in Embodiment 3. It utilizes the physical bridge structure information and physical mechanism (FEM model) stored in the memory, and based on the physical information artificial intelligence algorithm, it fuses the physical mechanism with a neural network (NN) model to construct a hybrid prediction model (e.g., the Bayesian physical information neural network described in Embodiment 8) containing configurable physical parameters (as described in Embodiment 4); Second, it is implemented as a dynamic calibration module, which is configured to execute the dynamic self-calibration in Embodiment 1. It performs dynamic, closed-loop self-calibration of the physical parameters of the hybrid prediction model constructed by the hybrid model module through advanced optimization strategies such as Bayesian inference or reinforcement learning. The dynamic calibration module integrates a hybrid model with ... The uncertainty-based gating data assimilation mechanism described in Example 8 can quantify and decouple accidental uncertainty and cognitive uncertainty using the B-PINN framework, and conditionally (i.e., gating) perform calibration based on the comparison results of the two (e.g., their ratio) with the assimilation threshold to prevent noise corrosion. Finally, it is implemented as a performance analysis module configured to analyze the service performance of the bridge based on a calibrated hybrid prediction model (updated in real time by a dynamic calibration module). This performance analysis module may further include: a specific analysis submodule for performing the fatigue life analysis, seismic performance assessment, and ultimate bearing capacity analysis described in Example 5; an early warning submodule for performing the analysis results compared with a preset safety threshold and automatically generating early warning signals as described in Example 6; and a What-if decision support submodule for performing the user input of hypothetical scenarios and generating optimal maintenance decision recommendations as described in Example 7.In a preferred embodiment, the instructions further enable the system to be implemented as an adaptive fidelity module configured to execute the adaptive fidelity control loop described in Embodiment 9. This module obtains the cognitive uncertainty quantification results from the dynamic calibration module, generates a cognitive uncertainty map, and automatically issues instructions to the hybrid model module when it identifies a specific spatial region where the EU value continuously exceeds the structural reconstruction threshold. This instruction performs automatic, local dynamic reconstruction of the computational structure of the hybrid model module (e.g., the grid of the FEM model or the resources of the NN model).

[0057] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A bridge performance analysis method based on digital twin technology, the method comprising: Obtain real-time monitoring data of physical bridges; Based on the structural information and physical mechanism of the physical bridge, a digital twin model containing configurable physical parameters is constructed. The method is characterized in that it further includes: (a) Based on the real-time monitoring data and the digital twin model, a hybrid prediction model is constructed using a physical information artificial intelligence algorithm; (b) Using advanced optimization strategies such as Bayesian inference or reinforcement learning, the physical parameters of the hybrid prediction model are dynamically and in a closed-loop manner to minimize the deviation between the predicted response of the hybrid prediction model and the real-time monitoring data; and (c) The service performance of the bridge is analyzed based on the calibrated hybrid prediction model in step (b).

2. The method according to claim 1, characterized in that, The real-time monitoring data originates from a multi-source heterogeneous sensor network, which includes at least one of the following: Strain gauges, accelerometers, displacement sensors, or fiber optic grating sensors.

3. The method according to claim 1, characterized in that, The hybrid prediction model includes: A physical mechanism model based on the finite element method is used to provide physical constraints for the structural response; and A data-driven model of a neural network is used to learn nonlinear dynamic characteristics not modeled by the physical mechanism model from the real-time monitoring data; The physical information artificial intelligence algorithm integrates the physical constraints into the training process of the data-driven model.

4. The method according to claim 1 or 3, characterized in that, The physical parameters for dynamic self-calibration include at least one of the following: The elastic modulus of the physical bridge material, the structural damping coefficient, or the boundary conditions between structural components.

5. The method according to claim 1, characterized in that, Step (c) of analyzing the service performance of the bridge includes at least one of the following: Based on the calibrated hybrid prediction model, structural stress analysis under random traffic loads is performed to predict structural fatigue life. Based on the calibrated hybrid prediction model, the structural response of the physical bridge under seismic load is evaluated to assess its seismic performance and damage. or The ultimate bearing capacity of the physical bridge is analyzed based on the calibrated hybrid prediction model.

6. The method according to claim 5, characterized in that, The method further includes: Based on the analysis results of the service performance, a comparison is made with a preset safety threshold; and When the analysis results trigger the security threshold, an early warning signal is automatically generated.

7. The method according to claim 1, characterized in that, The method further includes: Based on the calibrated hybrid prediction model, a What-if scenario analysis is performed to simulate the performance of the physical bridge under assumed extreme loads or maintenance interventions; and Based on the results of the What-if scenario analysis, an optimal maintenance decision recommendation is generated.

8. The method according to claim 1, characterized in that, The steps of constructing a hybrid prediction model and the dynamic, closed-loop self-calibration further include: Using a probabilistic artificial intelligence framework based on Bayesian physical information neural networks, the aforementioned biases are decoupled and quantified separately: Random uncertainty represents the inherent, non-simplifiable noise in the real-time monitoring data; and Cognitive uncertainty represents the degree of reducible ignorance arising from model defects in the hybrid prediction model; and In the dynamic self-calibration step (b), an uncertainty-based gated data assimilation mechanism is implemented, wherein the dynamic calibration of the physical parameters by the advanced optimization strategy is performed conditionally: the calibration is performed only when the ratio of the quantified random uncertainty to the cognitive uncertainty is lower than a preset assimilation threshold. Conversely, when the random uncertainty is too high, the influence weight of the real-time monitoring data on the physical parameter calibration is rejected or significantly reduced, thereby preventing the physical parameters of the hybrid prediction model from being eroded by data noise.

9. The method according to claim 8, characterized in that, The method further includes: Based on the quantified cognitive uncertainty, a cognitive uncertainty map representing its spatial distribution on the physical bridge is generated. An adaptive fidelity control loop is established, which takes the cognitive uncertainty map as input and is used for: Identify specific spatial regions where the cognitive uncertainty persists for more than a preset structural reconstruction threshold; When the specific spatial region is identified, the computational structure of the hybrid prediction model is automatically and locally dynamically reconstructed to reduce the cognitive uncertainty of the specific spatial region. The dynamic reconstruction of the computational structure includes at least one of the following: For the physical mechanism model based on the finite element method: driven by the cognitive uncertainty, adaptive mesh refinement is performed in the specific spatial region; or For the data-driven model: more computing resources are dynamically allocated in the specific spatial region.

10. A bridge performance analysis system based on digital twin technology, the system comprising: A data acquisition module, configured to acquire real-time monitoring data of the physical bridge; A processor; as well as A memory that stores computer-executable instructions; The characteristic is that, when the instruction is executed by the processor, it causes the system to function as follows: A hybrid model module configured to construct a hybrid prediction model containing configurable physical parameters based on the real-time monitoring data and physical mechanisms using physical information artificial intelligence algorithms; A dynamic calibration module configured to perform dynamic, closed-loop self-calibration of the physical parameters of the hybrid prediction model using advanced optimization strategies via Bayesian inference or reinforcement learning; and A performance analysis module is configured to analyze the service performance of the bridge based on a calibrated hybrid prediction model.

Images