Bridge digital twinning risk assessment method under extreme event
By constructing a bridge digital twin assessment method based on multi-source data repair and causal constraints, the instability and causal interpretability issues caused by sensor failure in the assessment are resolved. This achieves rapid, robust, and interpretable bridge risk assessment, generates direct traffic strategies, and supports rapid emergency management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-03-24
AI Technical Summary
Existing rapid assessment methods for bridge digital twins are susceptible to sensor failure under extreme events, resulting in unstable results, lack of causal interpretability, difficulty in quantifying uncertainty across the entire chain, and difficulty in directly translating assessment results into traffic strategies.
We construct an overall process for anomaly detection and redundancy repair, causal constraint triple calibration, damage dynamics evolution, end-to-end uncertainty propagation, and agent strategy generation. Through multi-source data repair, causal chain constraints, and end-to-end uncertainty propagation, we generate interpretable traffic strategies.
It achieves rapid, robust, and interpretable bridge risk assessment under extreme events, can automatically generate traffic strategies, improves the method's fault tolerance and decision-making transparency, and supports rapid emergency management.
Smart Images
Figure CN121723818A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of civil engineering and traffic engineering, and particularly relates to a bridge digital twin risk assessment method under extreme events. BACKGROUND
[0002] Bridges are often subjected to extreme events such as earthquakes, typhoons, ship collisions, and overloads during service. How to quickly and accurately assess the safety state and traffic capacity of bridges after the occurrence of events is an important problem in structural health monitoring and operation management.
[0003] Existing bridge digital twin rapid assessment methods based on monitoring data usually take finite element or other physical models as priori basis, collect structural response data (such as acceleration, displacement, etc.) during extreme events, and use filtering, variational assimilation or parameter optimization methods for one-way updating of the model. Then, the updated model is used to calculate damage indicators or performance degradation, and the bridge safety level or operation suggestion is output according to the empirical threshold. Such technology has been preliminarily applied in some earthquake emergency and typhoon monitoring scenarios. However, there are still the following shortcomings in practical application:
[0004] Dependence on complete observation: usually assuming that the sensors are intact and the data are sufficient, once the sensors fail or the observation is insufficient, the assessment result is easy to be distorted or unstable.
[0005] Lack of multiple constraints: most methods are one-way data-driven, and multiple priori constraints such as physical feasible region, statistical confidence region and damage sparsity are not introduced simultaneously in the assimilation process, resulting in insufficient physical rationality of the identification result.
[0006] Lack of causal relationship: existing methods are mostly based on correlation matching, without explicitly modeling the causal chain between “load-response-damage-performance”, resulting in limited explainability and generalization ability of the result.
[0007] Insufficient uncertainty propagation: usually only local confidence intervals are provided in the parameter estimation stage, without realizing the whole-link uncertainty propagation from observation error to damage identification to bearing assessment.
[0008] Assessment result and decision-making are disconnected: the output is mostly damage indicators or state levels, which are difficult to directly map to specific traffic strategies such as speed limit, load limit, closure, etc., lacking closed-loop decision support.
[0009] Therefore, the existing technology is still difficult to meet the comprehensive needs of short-time window data adaptability, robustness, causal explainability, whole-link uncertainty quantification and executable strategy output in the rapid assessment of bridges under extreme events. SUMMARY
[0010] To address the problems existing in the prior art, this invention provides a method for risk assessment of bridge digital twins under extreme events. The aim is to achieve rapid, robust, interpretable, and executable post-disaster bridge operation and management by constructing an overall process that includes anomaly perception and redundancy repair, causal constraint triple calibration, damage dynamic evolution, full-link uncertainty propagation, agent strategy generation, and multi-bridge collaborative updates. This solves the problems of insufficient adaptability to short-time window data, unstable results due to sensor failure, lack of causal interpretability, inability to quantify full-link uncertainty, and difficulty in directly converting assessment results into traffic strategies.
[0011] To achieve the above objectives, the specific solution of the present invention is as follows:
[0012] A bridge risk assessment method based on digital twins under extreme events includes the following steps:
[0013] Step 1, Short-term window multi-source data acquisition: After the extreme event is triggered, multi-source monitoring data such as acceleration, displacement, traffic load, video recognition displacement, and group mobile terminal data are collected within a short time window to form a unified observation vector;
[0014] Step 2, Abnormal Sensor Identification and Redundancy Repair: Based on the unified observation vector obtained in Step 1, calculate the residuals and Mahalanobis distances of the sensor channels used to monitor the bridge status. After identifying abnormal sensor channels, use data from other sensor channels, video monitoring data, group mobile terminal data, or sensor data from nearby bridges to perform multi-source redundancy repair.
[0015] Step 3, Triple calibration of causal constraints: Using the data after multi-source redundancy repair in Step 2, a triple calibration model is constructed, which includes physical feasible domain constraints, causal chain constraints, sparsity and total variation regularization constraints. The model is solved by variational methods or particle filtering combined with projection algorithms to obtain the damage parameter identification results.
[0016] Step 4, Damage dynamic evolution field identification: Based on the damage parameter identification results obtained in Step 3, the evolution law of damage over time is characterized, a load-driven evolution model is introduced, and recursive estimation is achieved by combining ensemble Kalman filtering or particle filtering to obtain the damage evolution results.
[0017] Step 5, Propagation of uncertainty across the entire chain: Based on the damage evolution results obtained in Step 4, a propagation system for uncertainty across the entire chain from observation to risk assessment is constructed. Monte Carlo sampling or first-order sensitivity method is used for propagation, and the propagated uncertainty results with confidence intervals are output.
[0018] Step 6, Remaining Capacity and Risk Quantification: Based on the uncertainty results after propagation in Step 5, calculate the remaining carrying capacity and risk index. When the risk index or its confidence interval covers 1.0, output a conservative traffic strategy.
[0019] Step 7, the policy agent generates a traffic control scheme: combining the risk index and traffic policy obtained in step 6, a state vector containing the risk index and its confidence interval width, traffic demand intensity, and the number of remaining available lanes is constructed. The policy function is obtained through reinforcement learning training and combined with the regularized threshold to generate a traffic control scheme.
[0020] Step 8, Multi-bridge Collaborative Update: For multiple bridges in the area that need to be collaboratively evaluated, based on the traffic control scheme generated in Step 7, a graph regularization collaborative mechanism is introduced to jointly optimize the local objective function of each bridge.
[0021] Furthermore, the formula for forming the unified observation vector described in step 1 is as follows:
[0022] ,
[0023] In the formula: Represents a unified observation vector; Indicates time; The duration of the short time window is indicated, preferably 60-300 seconds; Indicates the first The acceleration at each measuring point; Indicates the first Displacement of each measuring point; Indicates the first The equivalent effect of traffic load over time; This indicates the amount of structural displacement or deformation extracted from video recognition. This indicates weak observational features of acceleration or location acquired by a group of mobile terminals.
[0024] Furthermore, the sensor channels used to monitor the bridge status in step 2 include acceleration sensors, displacement sensors, strain sensors, and traffic load sensors;
[0025] The formula for calculating the residuals of the sensor channels used to monitor the bridge's condition is as follows:
[0026] ,
[0027] In the formula: Indicates the first The residual vector of the channel; Let be the observation value of the i-th channel; These are values predicted based on historical baselines or short-window regression.
[0028] The formula for calculating the Mahalanobis distance is as follows:
[0029] ,
[0030] In the formula: Let represent the Mahalanobis distance of the i-th channel; Represents the residual vector; The duration of the short time window is indicated, preferably 60-300 seconds; Denotes the residual covariance matrix; if If it is, then it is judged as abnormal; among which Let the degrees of freedom be d and the confidence level be d. The critical value of the chi-square distribution;
[0031] The calculation formula for the multi-source redundancy repair is as follows:
[0032] ,
[0033] In the formula: This represents the observed values after the repair. The fusion weights are distributed proportionally to the confidence level and the signal-to-noise ratio. This indicates a redundant data source.
[0034] Furthermore, the dynamic equilibrium equations for the physical feasible region constraints described in step 3 are as follows:
[0035] ,
[0036] In the formula: Represents the mass matrix; Represents the time of each node in the structure The acceleration vector at time; Represents the damping matrix; Represents the time of each node in the structure The velocity vector at time; Represents the stiffness matrix and damage parameters. control; Represents the displacement vector; Represents the external load vector; This represents the damage parameter.
[0037] The formula for the causal chain constraint, taking a typhoon as an example, is as follows:
[0038] ,
[0039] In the formula: Indicates typhoon wind pressure; Indicates cable tension decay; This indicates the change in the displacement of the main beam; Indicates the remaining bearing capacity;
[0040] The calibration objective function for the sparse and total variation regularization constraints is as follows:
[0041]
[0042] In the formula: J represents the calibration objective function; Indicates the difference between observation and simulation errors; Indicates the penalty term for the physically feasible region; Represents the residual term in the causal chain; This represents sparse + total variation regularization, constraining the damage field distribution; , , This represents the weighting parameter.
[0043] Furthermore, the equations for the load-driven damage dynamic evolution model introduced in step 4 are as follows:
[0044] ,
[0045] In the formula: Indicates time Damage parameters; The damage parameter represents time t; This represents the load-damage evolution gain matrix; This represents the average or envelope of the load over a period of time. Indicates the time step; For evolution power factor; This is the process noise term.
[0046] The formula for the recursive estimation is as follows:
[0047] ,
[0048] In the formula: This represents all observation data up to time t. The obtained damage parameters The posterior distribution of; This represents the sequence of observed data from the initial time to time t; This represents the observation likelihood function, i.e., given the current damage parameters. Below, response data was observed. The probability of; It represents the state transition probability and describes the evolution of damage parameters over time; This indicates a "proportional relationship," meaning that the posterior distribution is equal to the product of the likelihood function and the prior distribution, divided by a normalization constant.
[0049] Furthermore, the construction of the end-to-end uncertainty propagation system from observation to risk assessment described in step 5 is as follows:
[0050] ,
[0051] In the formula: Indicates the standard deviation of observed noise; This indicates the uncertainty of damage parameters; This indicates uncertainty in bearing capacity; This indicates uncertainty in the risk index; This indicates uncertainty regarding the final control strategy;
[0052] The Monte Carlo sampling method is based on the observation error distribution to randomly sample the input, pass it layer by layer, and statistically output the mean and variance.
[0053] The propagation formula for the first-order sensitivity method is as follows:
[0054] ,
[0055] In the formula, The covariance matrix represents the uncertainty of the output. Describe the output variables; Represents the Jacobian matrix; The covariance matrix represents the uncertainty of the input. Indicates input variables; This represents the transpose of the Jacobian matrix.
[0056] Furthermore, the formula for calculating the remaining bearing capacity mentioned in step 6 is as follows:
[0057] ,
[0058] In the formula: Indicates the remaining bearing capacity; Indicates the safety reduction factor; Indicates the equivalent stiffness after identification; Indicates the initial stiffness;
[0059] The formula for calculating the risk index is as follows:
[0060] ,
[0061] In the formula: Indicates a risk index; This indicates the predicted traffic load effect; Indicates the remaining bearing capacity.
[0062] Furthermore, the formula for constructing the state vector containing the risk index and its confidence interval width, traffic demand intensity, and the number of remaining available lanes in step 7 is as follows:
[0063] [RI, △RI, Q, L]
[0064] In the formula: s represents the state vector; RI represents the risk index; △RI represents the confidence interval width of the risk index; Q represents the traffic demand intensity; L represents the number of remaining available lanes;
[0065] The traffic control plan includes a set of strategies and actions such as normal passage, speed limits, load limits, or closures.
[0066] Furthermore, the formula for introducing the graph regularization collaborative mechanism in step 8 is as follows:
[0067] ,
[0068] In the formula: This represents the overall objective function for bridge evaluation; Indicates the number of bridges within the area; Let b represent the local objective function of the b-th bridge; This represents the damage parameters of the b-th bridge. Represents the set of adjacent edges of a bridge; Indicates the collaborative weights between bridges; This represents the regularization coefficient.
[0069] Advantages of the present invention
[0070] (1) This invention introduces an automatic identification and redundancy repair mechanism for abnormal sensors for the first time in extreme event scenarios. It uses Mahalanobis distance and residual covariance matrix to jointly identify anomalies and uses video monitoring, group mobile terminals and sensor data from nearby bridges for redundancy compensation. Even if some sensors fail or data is lost, complete observations can still be recovered, significantly improving the fault tolerance and robustness of the method.
[0071] (2) This invention introduces a causal inference framework on the basis of traditional "physics-data" assimilation, and constructs a triple calibration model that includes physical feasible region, causal chain constraints and sparse / total variation regularization. Compared with the shortcomings of existing methods that are based solely on correlation fitting, this invention can explicitly express the causal logic of "extreme load - cable force attenuation - main beam displacement - bearing capacity", ensuring that the identification results are not only numerically reasonable, but also physically clear, avoiding the problem of "correct fitting but physical error".
[0072] (3) This invention proposes a load-driven damage evolution equation, which, combined with filtering recursion, enables dynamic updating and can track the time-varying evolution of damage in real time. Unlike the existing "static snapshot" identification, this method can capture potential failure trends in advance, support rolling prediction and early warning, and provide a forward-looking basis for rapid emergency management after extreme events.
[0073] (4) This invention establishes a full-link uncertainty propagation system covering "observation error - parameter identification - carrying capacity - risk index - traffic strategy". It uses Monte Carlo sampling and first-order sensitivity method to quantify error propagation and ensures that each output is accompanied by a confidence interval. This mechanism not only improves the credibility of the results, but also provides risk boundaries for managers, making decision-making more scientific and transparent.
[0074] (5) By introducing a strategy agent trained through reinforcement learning, this invention achieves the automatic generation of traffic strategy schemes from risk index to speed limit, load limit, and closure, forming a "assessment-control" closed loop. Compared with the existing methods that can only output damage indicators and still require human experience for interpretation, this method achieves immediate assessment and use, greatly shortening emergency response time and improving post-disaster disposal efficiency.
[0075] (6) Based on single-bridge digital twins, this invention proposes a multi-bridge collaborative update mechanism, employing Laplace regularization to jointly optimize bridge damage parameters within the region, fully utilizing monitoring information from neighboring bridges, and improving the consistency and robustness of overall assessment under large-scale disasters (such as earthquakes and typhoons). This breakthrough enables the method to possess network-level disaster recovery capabilities for the first time, supporting global scheduling of the transportation network.
[0076] (7) Each step of this invention is equipped with a mathematical model and parameter description, which can directly connect to existing monitoring methods such as accelerometers, displacement gauges, traffic load monitoring, video recognition, and mobile terminal data, avoiding major modifications to the hardware system. Its algorithm module can be quickly deployed on existing bridge health monitoring platforms and has good prospects for promotion and application.
[0077] (8) Compared with the existing rapid evaluation method of bridge digital twins that only relies on one-way data calibration, assumes complete sensing, and has a theoretical output, this invention organically combines six levels: fault tolerance robustness, causal interpretability, dynamic prediction capability, end-to-end credibility, strategy closure, and network-level collaboration to form a systematic and practical overall solution, which is significantly superior to the existing technology in terms of academic advancement and engineering practicality. Attached Figure Description
[0078] Figure 1 This is a flowchart of a bridge risk assessment method based on digital twins under extreme events according to the present invention. Detailed Implementation
[0079] The following explanation and description are provided in conjunction with the accompanying drawings and specific embodiments. However, it should be noted that these specific embodiments are not intended to limit the scope of the invention.
[0080] like Figure 1 As shown in the illustration, this specific embodiment provides a bridge risk assessment method based on digital twins under extreme events, which includes the following steps:
[0081] Step 1, Short-term multi-source data acquisition: After the extreme event is triggered, multi-source monitoring data such as acceleration, displacement, traffic load, video recognition displacement, and group mobile terminal data are collected within a short time window to form a unified observation vector; this unified observation vector provides the basic input for subsequent anomaly detection, model calibration, and damage evolution identification.
[0082] The formula for forming the unified observation vector mentioned in step 1 is as follows:
[0083] ,
[0084] In the formula: Represents a unified observation vector; Indicates time; The duration of the short time window is indicated, preferably 60-300 seconds; Indicates the first The acceleration at each measuring point; Indicates the first Displacement of each measuring point; Indicates the first The equivalent effect of traffic load over time; This indicates the amount of structural displacement or deformation extracted from video recognition. This indicates weak observational features of acceleration or location acquired by a group of mobile terminals.
[0085] Step 2, Abnormal Sensor Identification and Redundancy Repair: Under extreme events, some sensors may fail or generate abnormal data. To ensure the robustness of the assessment, and based on the unified observation vector obtained in Step 1, the residuals and Mahalanobis distances of the sensor channels used to monitor the bridge status are first calculated to identify abnormal sensor channels. For the identified abnormal sensor channels, multi-source redundancy repair is performed using data from other sensor channels, video monitoring data, group mobile terminal data, or sensor data from nearby bridges.
[0086] The sensor channels used to monitor the bridge status mentioned in step 2 include acceleration sensors, displacement sensors, strain sensors, and traffic load sensors;
[0087] The formula for calculating the residuals of the sensor channels used to monitor the bridge's condition is as follows:
[0088] ,
[0089] In the formula: Indicates the first The residual vector of the channel; Let be the observation value of the i-th channel; These are values predicted based on historical baselines or short-window regression.
[0090] The formula for calculating the Mahalanobis distance is as follows:
[0091] ,
[0092] In the formula: Let represent the Mahalanobis distance of the i-th channel; Represents the residual vector; The duration of the short time window is indicated, preferably 60-300 seconds; Denotes the residual covariance matrix; if If it is, then it is judged as abnormal; among which Let the degrees of freedom be d and the confidence level be d. The critical value of the chi-square distribution;
[0093] The calculation formula for the multi-source redundancy repair is as follows:
[0094] ,
[0095] In the formula: This represents the observed values after the repair. The fusion weights are distributed proportionally to the confidence level and the signal-to-noise ratio. This indicates a redundant data source.
[0096] Step 3, Triple calibration of causal constraints: Using the data after multi-source redundancy repair in Step 2, a triple calibration model is constructed, which includes physical feasible domain constraints, causal chain constraints, sparsity and total variation regularization constraints. The model is solved by variational methods or particle filtering combined with projection algorithms to obtain the damage parameter identification results.
[0097] Physically feasible region constraint: Ensures that the model calibration results are physically feasible;
[0098] Causal chain constraints include "extreme loads" — Cable tension attenuation —Main beam displacement —Load capacity The relationship is used to explicitly introduce causal relationships in physical processes, such as wind pressure causing cable force attenuation in a typhoon scenario, which in turn affects the main beam displacement and remaining bearing capacity.
[0099] Sparse regularization constraint: The distribution of damage parameters is constrained by sparse regularization methods (such as L1 regularization) so that the solution of damage parameters is sparse, that is, damage exists at only a few locations.
[0100] Total variation regularization constraint: The distribution of damage parameters is constrained by the total variation regularization method, so that the solution of damage parameters has spatial smoothness, that is, the change of damage parameters will not be too drastic.
[0101] The dynamic equilibrium equations for the physical feasible region constraints mentioned in step 3 are as follows:
[0102] ,
[0103] In the formula: Represents the mass matrix; Represents the time of each node in the structure The acceleration vector at time; Represents the damping matrix; Represents the time of each node in the structure The velocity vector at time; Represents the stiffness matrix and damage parameters. control; Represents the displacement vector; Represents the external load vector; Indicates damage parameters;
[0104] The formula for the causal chain constraint, taking a typhoon as an example, is as follows:
[0105] ,
[0106] In the formula: Indicates typhoon wind pressure; Indicates cable tension decay; This indicates the change in the displacement of the main beam; Indicates the remaining bearing capacity;
[0107] The calibration objective function for the sparse and total variation regularization constraints is as follows:
[0108]
[0109] In the formula: J represents the calibration objective function; Indicates the difference between observation and simulation errors; Indicates the penalty term for the physically feasible region; Represents the residual term in the causal chain; This represents sparse + total variation regularization, constraining the damage field distribution; , , This represents the weighting parameter.
[0110] Step 4, Damage dynamic evolution field identification: Based on the damage parameter identification results obtained in Step 3, the evolution law of damage over time is characterized, and a load-driven evolution model is introduced to obtain the damage evolution results.
[0111] The equations for the load-driven damage dynamic evolution model are as follows:
[0112] ,
[0113] In the formula: Indicates time Damage parameters; The damage parameter represents time t; This represents the load-damage evolution gain matrix; This represents the average or envelope of the load over a period of time. Indicates the time step; For evolution power factor; This is the process noise term.
[0114] The equations of this evolutionary model are combined with the calibration results from step three, and recursive estimation is achieved using ensemble Kalman filtering or particle filtering. The state transition equation is given by the damage evolution equation, the observation equation is given by the kinetic equation obtained from the calibration in step three, and the update form of the filtered recursive estimation is:
[0115] ,
[0116] In the formula: This represents all observation data up to time t. The obtained damage parameters The posterior distribution of; This represents the sequence of observed data from the initial time to time t; This represents the observation likelihood function, i.e., given the current damage parameters. Below, response data was observed. The probability of; It represents the state transition probability and describes the evolution of damage parameters over time; This indicates a "proportional relationship," meaning the posterior distribution equals the product of the likelihood function and the prior distribution, divided by a normalization constant. Determined by the kinetic equations in step three, The damage evolution equation is given by step four.
[0117] Step 5, Propagation of uncertainty across the entire chain: Based on the damage evolution results obtained in Step 4, a propagation system for uncertainty across the entire chain from observation to risk assessment is constructed. Monte Carlo sampling or first-order sensitivity method is used for propagation, and the propagated uncertainty results with confidence intervals are output.
[0118] The proposed end-to-end uncertainty propagation system, from observation to risk assessment, is as follows:
[0119] ,
[0120] In the formula: Indicates the standard deviation of observed noise; This indicates the uncertainty of damage parameters; This indicates uncertainty in bearing capacity; This indicates uncertainty in the risk index; This indicates uncertainty regarding the final control strategy;
[0121] The Monte Carlo sampling method is based on the observation error distribution to randomly sample the input, pass it layer by layer, and statistically output the mean and variance.
[0122] The propagation formula for the first-order sensitivity method is as follows:
[0123] ,
[0124] In the formula, The covariance matrix represents the uncertainty of the output. Describe the output variables; Represents the Jacobian matrix; The covariance matrix represents the uncertainty of the input. Indicates input variables; This represents the transpose of the Jacobian matrix.
[0125] Step 6, Remaining Capacity and Risk Quantification: Based on the uncertainty results after propagation in Step 5, calculate the remaining carrying capacity and risk index. When the risk index or its confidence interval covers 1.0, output a conservative traffic strategy.
[0126] The formula for calculating the remaining bearing capacity is as follows:
[0127] ,
[0128] In the formula: Indicates the remaining bearing capacity; Indicates the safety reduction factor; Indicates the equivalent stiffness after identification; Indicates the initial stiffness;
[0129] The formula for calculating the risk index is as follows:
[0130] ,
[0131] In the formula: Indicates a risk index; This indicates the predicted traffic load effect; Indicates the remaining bearing capacity.
[0132] when or[ , When covering version 1.0, a conservative traffic strategy is output.
[0133] Step 7: The policy agent generates a traffic control plan: Combining the risk index and traffic policy obtained in Step 6, a state vector is constructed that includes the risk index and its confidence interval width, traffic demand intensity, and the number of remaining available lanes. The policy function is obtained through reinforcement learning training and combined with the regularized threshold to generate a traffic control plan, thus realizing the evaluation and decision-making closed loop.
[0134] The formula for constructing a state vector comprising the risk index RI, the cross confidence interval width ΔRI, the traffic demand Q, and the remaining lane number L is as follows:
[0135] [RI, △RI, Q, L]
[0136] In the formula: s represents the state vector; RI represents the risk index; △RI represents the confidence interval width of the risk index; Q represents the traffic demand intensity; L represents the number of remaining available lanes;
[0137] The traffic control plan includes a set of strategies and actions such as normal passage, speed limits, load limits, or closures.
[0138] Strategy Action Set {Normal, Speed Limit, Load Limit, Closed}
[0139] The policy function is obtained through reinforcement learning training. And by combining regularized thresholds to improve interpretability:
[0140] RI < 0.60: Normal;
[0141] 0.60 ≤ RI < 0.80: Speed limit;
[0142] 0.80≦RI<1.00: Load limit;
[0143] RI≧1.00: Closed.
[0144] Step 8, Multi-bridge Collaborative Update: For multiple bridges in the region that require collaborative evaluation, based on the traffic control scheme generated in Step 7, a graph regularization collaborative mechanism is introduced to jointly optimize the local objective function of each bridge, thereby achieving disaster recovery collaborative evaluation under regional extreme events and ensuring the consistency and robustness of the evaluation results of multiple bridges in extreme events (such as earthquakes and typhoons) within the region.
[0145] The formula for introducing the graph regularization collaboration mechanism is as follows:
[0146] ,
[0147] In the formula: This represents the overall objective function for bridge evaluation; Indicates the number of bridges within the area; Let b represent the local objective function of the b-th bridge; This represents the damage parameters of the b-th bridge. Represents the set of adjacent edges of a bridge; Indicates the collaborative weights between bridges; This represents the regularization coefficient.
Claims
1. A bridge risk assessment method based on digital twins under extreme events, characterized in that, Includes the following steps: Step 1, Short-term window multi-source data acquisition: After the extreme event is triggered, multi-source monitoring data such as acceleration, displacement, traffic load, video recognition displacement, and group mobile terminal data are collected within a short time window to form a unified observation vector; Step 2, Abnormal Sensor Identification and Redundancy Repair: Based on the unified observation vector obtained in Step 1, calculate the residuals and Mahalanobis distances of the sensor channels used to monitor the bridge status. After identifying abnormal sensor channels, use data from other sensor channels, video monitoring data, group mobile terminal data, or sensor data from nearby bridges to perform multi-source redundancy repair. Step 3, Triple calibration of causal constraints: Using the data after multi-source redundancy repair in Step 2, a triple calibration model is constructed, which includes physical feasible domain constraints, causal chain constraints, sparsity and total variation regularization constraints. The model is solved by variational methods or particle filtering combined with projection algorithms to obtain the damage parameter identification results. Step 4, Damage dynamic evolution field identification: Based on the damage parameter identification results obtained in Step 3, the evolution law of damage over time is characterized. A load-driven damage dynamic evolution model is introduced, and recursive estimation is achieved by combining ensemble Kalman filtering or particle filtering to obtain the damage evolution results. Step 5, Propagation of uncertainty across the entire chain: Based on the damage evolution results in Step 4, a propagation system for uncertainty across the entire chain from observation to risk assessment is constructed. Monte Carlo sampling or first-order sensitivity method is used for propagation, and the propagated uncertainty results with confidence intervals are output. Step 6, Remaining Capacity and Risk Quantification: Based on the uncertainty results after propagation in Step 5, calculate the remaining carrying capacity and risk index. When the risk index or its confidence interval covers 1.0, output a conservative traffic strategy. Step 7, the policy agent generates a traffic control scheme: combining the risk index and traffic policy obtained in step 6, a state vector containing the risk index and its confidence interval width, traffic demand intensity, and the number of remaining available lanes is constructed. The policy function is obtained through reinforcement learning training and combined with the regularized threshold to generate a traffic control scheme. Step 8, Multi-bridge Collaborative Update: For multiple bridges in the area that need to be collaboratively evaluated, based on the traffic control scheme generated in Step 7, a graph regularization collaborative mechanism is introduced to jointly optimize the local objective function of each bridge.
2. The bridge risk assessment method according to claim 1, characterized in that, The formula for forming the unified observation vector mentioned in step 1 is as follows: ， In the formula: Represents a unified observation vector; Indicates time; The duration of the short time window is indicated, preferably 60-300 seconds; Indicates the first The acceleration at each measuring point; Indicates the first Displacement of each measuring point; Indicates the first The equivalent effect of traffic load over time; This indicates the amount of structural displacement or deformation extracted from video recognition. This indicates weak observational features of acceleration or location acquired by a group of mobile terminals.
3. The bridge risk assessment method according to claim 1, characterized in that, The sensor channels used to monitor the bridge status mentioned in step 2 include acceleration sensors, displacement sensors, strain sensors, and traffic load sensors; The formula for calculating the residuals of the sensor channels used to monitor the bridge's condition is as follows: ， In the formula: Indicates the first The residual vector of the channel; Let be the observation value of the i-th channel; These are values predicted based on historical baselines or short-window regression. The formula for calculating the Mahalanobis distance is as follows: ， In the formula: Let represent the Mahalanobis distance of the i-th channel; Represents the residual vector; The duration of the short time window is indicated, preferably 60-300 seconds; Denotes the residual covariance matrix; if If it is, then it is judged as abnormal; among which This represents a system with d degrees of freedom and a confidence level of 1. The critical value of the chi-square distribution; The calculation formula for the multi-source redundancy repair is as follows: ， In the formula: This represents the observed values after the repair; The fusion weights are distributed proportionally to the confidence level and the signal-to-noise ratio. This indicates a redundant data source.
4. The bridge risk assessment method according to claim 1, characterized in that, The dynamic equilibrium equations for the physical feasible region constraints mentioned in step 3 are as follows: ， In the formula: Represents the mass matrix; Represents the time of each node in the structure The acceleration vector at time; Represents the damping matrix; Represents the time of each node in the structure The velocity vector at time; Represents the stiffness matrix and damage parameters. control; Represents the displacement vector; Represents the external load vector; Indicates damage parameters; The formula for the causal chain constraint, taking a typhoon as an example, is as follows: ， In the formula: Indicates typhoon wind pressure; Indicates cable tension decay; This indicates the change in the displacement of the main beam; Indicates the remaining bearing capacity; The calibration objective function for the sparse and total variation regularization constraints is as follows: ， In the formula: J represents the calibration objective function; Indicates the difference between observation and simulation errors; Indicates the penalty term for the physically feasible region; Represents the residual term in the causal chain; This represents sparse + total variation regularization, constraining the damage field distribution; , , This represents the weighting parameter.
5. The bridge risk assessment method according to claim 1, characterized in that, The equations for the load-driven damage dynamic evolution model introduced in step 4 are as follows: ， In the formula: Indicates time Damage parameters; The damage parameter represents time t; This represents the load-damage evolution gain matrix; This represents the average or envelope of the load over a period of time. Indicates the time step; For evolution power factor; This refers to the process noise term; The formula for the recursive estimation is as follows: ， In the formula: This represents all observation data up to time t. The obtained damage parameters The posterior distribution of; This represents the sequence of observed data from the initial time to time t; This represents the observation likelihood function, i.e., given the current damage parameters. Below, response data was observed. The probability of; It represents the state transition probability and describes the evolution of damage parameters over time; This indicates a "proportional relationship," meaning that the posterior distribution is equal to the product of the likelihood function and the prior distribution, divided by a normalization constant.
6. The bridge risk assessment method according to claim 1, characterized in that, The uncertainty propagation system from observation to risk assessment described in step 5 is as follows: ， In the formula: Indicates the standard deviation of observed noise; This indicates the uncertainty of damage parameters; This indicates uncertainty in bearing capacity; This indicates uncertainty in the risk index; This indicates uncertainty regarding the final control strategy; The Monte Carlo sampling method is based on the observation error distribution to randomly sample the input, pass it layer by layer, and statistically output the mean and variance. The propagation formula for the first-order sensitivity method is as follows: ， In the formula, The covariance matrix represents the uncertainty of the output. Describe the output variables; Represents the Jacobian matrix; The covariance matrix represents the uncertainty of the input. Indicates input variables; This represents the transpose of the Jacobian matrix.
7. The bridge risk assessment method according to claim 1, characterized in that, The formula for calculating the remaining bearing capacity mentioned in step 6 is as follows: ， In the formula: Indicates the remaining bearing capacity; Indicates the safety reduction factor; Indicates the equivalent stiffness after identification; Indicates the initial stiffness; The formula for calculating the risk index is as follows: ， In the formula: Indicates a risk index; This indicates the predicted traffic load effect; Indicates the remaining bearing capacity.
8. The bridge risk assessment method according to claim 1, characterized in that, The formula for constructing the state vector containing the risk index and its confidence interval width, traffic demand intensity, and the number of remaining available lanes in step 7 is as follows: [RI,△RI,Q,L], In the formula: s represents the state vector; RI represents the risk index; △RI represents the confidence interval width of the risk index; Q represents the traffic demand intensity; L represents the number of remaining available lanes; The traffic control plan includes a set of strategies and actions such as normal passage, speed limits, load limits, or closures.
9. The bridge risk assessment method according to claim 1, characterized in that, The formula for introducing the graph regularization collaboration mechanism in step 8 is as follows: ， In the formula: This represents the overall objective function for bridge evaluation; Indicates the number of bridges within the area; Let represent the local objective function of the b-th bridge; This represents the damage parameters of the b-th bridge. Represents the set of adjacent edges of a bridge; Indicates the collaborative weights between bridges; This represents the regularization coefficient.