Bridge whole life cycle management system based on digital twinning

Abstract

The application discloses a bridge full life cycle management and maintenance system based on digital twinning, which is used for the full life cycle management and maintenance of a cable-stayed bridge and comprises a data acquisition and fusion module, a digital twinning model construction and dynamic updating module and a structure health assessment and prediction module.The data acquisition and fusion module is used for acquiring monitoring data of the cable-stayed bridge in real time and fusing and processing multi-source data by using an adaptive weighting method.The digital twinning model construction and dynamic updating module is used for constructing and continuously calibrating a parameterized finite element simulation model based on bridge design parameters and fused data output by the data acquisition and fusion module, so as to serve as a digital twinning model corresponding to the physical bridge, and the digital twinning model can simulate material performance degradation.The structure health assessment and prediction module is used for assessing the current health state of bridge components by using residual characteristics of digital twinning model output and measured data based on the updated digital twinning model output by the digital twinning model construction and dynamic updating module.The application realizes intelligent management and maintenance of the full life cycle of the cable-stayed bridge.

Description

Technical Field

[0001] This invention relates to the field of bridge monitoring technology, specifically to a bridge lifecycle management system based on digital twins. Background Technology

[0002] Bridge lifecycle management refers to the systematic management and maintenance of bridges (especially large-scale projects like cable-stayed bridges with complex structures and special stresses) from design, construction, operation, and final decommissioning. The core objective is to ensure the long-term safe and stable operation of bridges by continuously monitoring their structural condition, promptly identifying damage and degradation, and scientifically formulating maintenance strategies, while minimizing overall maintenance costs and extending the structural lifespan. However, cable-stayed bridges are susceptible to multiple factors during long-term operation, including traffic loads, wind loads, and environmental erosion, facing problems such as material performance degradation and cumulative component damage. Traditional maintenance models rely on manual inspection and experience-based judgment, which have limitations such as fragmented monitoring data, delayed damage identification, and one-sided condition assessments, making it difficult to achieve dynamic and precise management throughout the entire lifecycle.

[0003] To overcome the bottlenecks of traditional maintenance, a bridge life-cycle maintenance system has emerged. However, existing technologies still have many shortcomings: data fusion often uses fixed-weight methods, which cannot dynamically adapt to the real-time error differences of multiple sensors, resulting in insufficient data reliability; digital twin models lack a hierarchical optimization and continuous calibration mechanism, have poor synchronization with the physical bridge, and are difficult to accurately simulate material degradation and structural mechanical behavior; health assessment relies on a single characteristic indicator, which cannot quantify the probability distribution and uncertainty of different health states; performance prediction ignores the randomness of the degradation process, and the remaining life estimation results are one-sided; maintenance decisions do not fully integrate the state evolution law and cost-risk balance, and lack a virtual pre-validation link for candidate strategies. At the same time, no effective maintenance experience reuse mechanism has been established, resulting in insufficient scientificity and pertinence of decision-making, making it difficult to meet the complex and dynamic life-cycle maintenance needs of cable-stayed bridges. Therefore, a bridge life-cycle maintenance system based on digital twins is proposed. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a digital twin-based bridge lifecycle management and maintenance system for cable-stayed bridges, comprising:

[0005] The data acquisition and fusion module is used to acquire monitoring data of the cable-stayed bridge in real time and to fuse multi-source data using an adaptive weighting method.

[0006] The digital twin model construction and dynamic update module is used to construct and continuously calibrate a parametric finite element simulation model based on bridge design parameters and fused data output by the data acquisition and fusion module. As a digital twin model corresponding to the physical bridge, the digital twin model can simulate the degradation of material properties.

[0007] The structural health assessment and prediction module is used to assess the current health status of bridge components based on the updated digital twin model output by the digital twin model construction and dynamic update module, by using the residual characteristics between the digital twin model output and the measured data, and to predict its performance evolution based on a stochastic process model.

[0008] The whole life cycle maintenance decision optimization module is used to integrate the current health status and performance evolution prediction results output by the structural health assessment and prediction module, and generate the optimal maintenance strategy for future cycles based on the decision optimization model;

[0009] The modules are connected sequentially to form a closed-loop process from data perception, model synchronization, state assessment to decision optimization.

[0010] Furthermore, the adaptive weighting method executed by the data acquisition and fusion module specifically includes:

[0011] Simultaneously acquire monitoring data sequences characterizing the same physical quantity from multiple sensor nodes distributed at different parts of the cable-stayed bridge;

[0012] For each sampling time, calculate the absolute deviation between the monitoring data of each sensor and the median of the monitoring data of all sensors of the same type at that time;

[0013] The confidence weight of each sensor is dynamically calculated based on its absolute deviation; the smaller the absolute deviation, the higher the confidence weight. The calculation process is as follows: First, the absolute deviation of each sensor is input into a negative exponential function to obtain its unnormalized weight factor; then, the unnormalized weight factors of all sensors are summed, and the unnormalized weight factor of each sensor is divided by the sum to obtain the normalized confidence weight w. i (t), its mathematical expression is:

[0014] ;

[0015] Where, d i (t) represents the data of the i-th sensor at time t, median(D(t)) is the median of the data of all N sensors of the same type at that time, and β is the adjustment coefficient;

[0016] Finally, based on the confidence weights, the data from all sensors of the same type at the same time are weighted and averaged to output the fused data.

[0017] Furthermore, the process by which the structural health assessment and prediction module assesses the current health status of bridge components includes:

[0018] The fused data from the data acquisition and fusion module is input into the digital twin model to calculate the simulation response values ​​of key measurement points;

[0019] Calculate the residual vector between the simulated response value and the corresponding measured response value;

[0020] Extract multidimensional feature vectors, including time-domain statistical features and frequency-domain energy features, from the residual vector;

[0021] The multidimensional feature vector is input into a pre-trained multi-class support vector machine model; during the training phase, the multi-class support vector machine model is trained using samples generated from historical data or simulation data, labeled with different health status types and locations.

[0022] The output of the multi-class support vector machine model is the probability distribution of the current bridge structure being in various preset health states.

[0023] Furthermore, the parameter continuous calibration strategy implemented in the digital twin model construction and dynamic update module adopts a hierarchical optimization mechanism, specifically including:

[0024] The first layer of local parameter fast update: Construct an optimization problem with the objective of minimizing the real-time response residual; the objective function of this problem is defined as the squared L2 of the difference between the measured response vector and the simulated response vector of the digital twin model, i.e. , where θ l y represents a subset of local physical parameters directly associated with the sensor measurement point. m and y s These are the measured and simulated response vectors, respectively.

[0025] The Levenberg-Marquardt optimization algorithm is used to solve the optimization problem online iteratively, achieving the solution of θ. l Rapid tracking and adjustment.

[0026] Second-level global parameter periodic calibration: Global calibration is initiated once every fixed time period;

[0027] Global calibration constructs a global optimization problem with the objective of minimizing the weighted average of substructure interface force balance error and overall response residual. The decision variables are a subset of global parameters θ that affect the overall mechanical behavior of the structure. g A genetic algorithm is used to solve the global optimization problem and update θ. g ;

[0028] The results of the second-level calibration are used to constrain the range of values ​​for local parameters in the first-level calibration.

[0029] Furthermore, the structural health assessment and prediction module predicts the evolution of component performance based on a stochastic process model, specifically including:

[0030] For unhealthy components identified by the multi-class support vector machine model, extract the time series of key indicators reflecting their performance degradation from historical monitoring data;

[0031] Assuming the degradation process follows a Wiener process, i.e. the degradation increment follows a normal distribution with mean and variance proportional to the time interval; using the maximum likelihood estimation method and time series data of key indicators, the drift coefficient λ and diffusion coefficient σ of the Wiener process are estimated.

[0032] Set the performance failure threshold Df for the component; based on the estimated parameter drift coefficient λ and diffusion coefficient σ, use the Monte Carlo simulation method to randomly generate a large number of future degradation paths that conform to the Wiener process starting from the current state;

[0033] The time when all simulated future degradation paths first reach or exceed the failure threshold Df is statistically analyzed to form a probability distribution of the component's remaining useful life; the median of this probability distribution is taken as the component's predicted remaining useful life.

[0034] Furthermore, in the whole life cycle maintenance decision optimization module, the decision optimization model is a partially observable Markov decision process model, and its construction and solution process includes:

[0035] The health status of key bridge components is discretized into a finite number of levels, and the impact of different maintenance actions on state transitions is modeled as a state transition probability matrix.

[0036] The health status probability distribution output by the multi-class support vector machine model is used as incomplete observation information of the true health status of the components to update the belief state in the partially observable Markov decision process model.

[0037] The value function is defined as the long-run discounted expected total cost, which includes the cost of immediate maintenance actions and the risk cost of components being in different health states;

[0038] A point-based iterative algorithm is used to approximate the solution of a partially observable Markov decision process model, yielding the optimal maintenance strategy based on the current belief state.

[0039] Furthermore, the system also includes a virtual simulation verification module, used to pre-evaluate the effectiveness of candidate maintenance strategies generated by the full lifecycle maintenance decision optimization module. The process includes:

[0040] The planned maintenance actions in the candidate maintenance strategy are transformed into modification instructions for the corresponding component parameters in the digital twin model to simulate the maintenance effect;

[0041] On the modified digital twin model, input the typical load sequence for the future planning period predicted based on historical data to perform structural response simulation;

[0042] Based on the simulation results, the structural health assessment and prediction module was called again to re-predict the performance degradation trajectory and remaining life of the relevant components after simulated maintenance.

[0043] Based on the re-predicted component state sequence, the expected total cost of implementing the candidate strategy in subsequent planning periods is estimated using a partially observable Markov decision process model; the selection is made by comparing the long-term expected total costs of different candidate strategies.

[0044] Furthermore, the system also includes an event triggering and data management module; the working mechanism of the event triggering and data management module is as follows:

[0045] Calculate the statistics of specific monitoring indicators in real time within a sliding time window and compare them with the simulation output statistics of the digital twin model under the same time window and the same input conditions;

[0046] When the statistical difference exceeds the preset threshold for multiple consecutive calculation periods, and the current environmental conditions belong to the preset typical conditions library, a significant model deviation event is determined to have occurred.

[0047] The system automatically caches all relevant original monitoring data, fused data, environmental data, and corresponding input and output data of the digital twin model within a preset time period before and after a significant deviation event in the model, forming a dedicated calibration dataset.

[0048] Send the calibration-specific dataset to the digital twin model building and dynamic update module, and trigger it to start or prioritize the execution of the model parameter calibration process.

[0049] Furthermore, the system also includes a real-time load effect assessment and early warning module based on digital twins, used to dynamically assess the impact of actual loads on the safety of the bridge structure. Its workflow specifically includes:

[0050] Based on the real-time traffic load data and wind load data acquired by the data acquisition and fusion module, the actual load spectrum for the current time period is generated.

[0051] Using the actual load spectrum as input, the digital twin model is driven to perform structural response simulation, and the stress time history, displacement response and overall structural response extreme values ​​of key components under the current actual load are calculated.

[0052] The simulated stress time history of the components is compared with the time-varying material strength of the components in real time. The time-varying material strength is determined based on the material property degradation model that can be simulated by the digital twin model. The real-time load effect ratio of each component is calculated. The calculation formula is as follows: ,in f represents the peak stress of the i-th component obtained from the simulation under the current actual load spectrum. i(t) The material strength of the component at the current moment is given by the digital twin model;

[0053] Set multi-level early warning thresholds; when the real-time load effect of any component is higher than... When the first-level warning threshold is exceeded, a notice-level warning message is generated; when the second-level higher warning threshold is exceeded, an action-level warning message is generated and pushed to the full life cycle maintenance decision optimization module as input to trigger immediate assessment or adjustment of maintenance strategies.

[0054] Furthermore, the system also includes a maintenance history digital archive and knowledge accumulation module, used to achieve structured storage and experience mining of data throughout the entire lifecycle. Its specific implementation methods include:

[0055] Establish a database associated with the geometric topology information of the digital twin model to store raw data and fused data from the data acquisition and fusion module, health status assessment records and remaining life prediction records from the structural health assessment and prediction module, decision records from the whole life cycle maintenance decision optimization module and their corresponding verification results output by the virtual simulation verification module, as well as records of actual maintenance actions and post-effect evaluation data, by component and by time slice.

[0056] Design a case-based reasoning-based intelligent recommendation unit for maintenance schemes: When the structural health assessment and prediction module identifies a new damage mode or degradation state, this unit retrieves historical similar cases from the database. The similarity is determined by comparing the Euclidean distance between multiple feature vectors of damage mode, component type, environmental conditions, and current structural state.

[0057] The maintenance plans used in similar historical cases retrieved, the predicted effects verified by virtual simulation, and the long-term effects after actual implementation are comprehensively and weighted to generate a set of recommended maintenance plans for the current new situation and their expected effects analysis, which can be used as a reference for the whole life cycle maintenance decision optimization module.

[0058] The beneficial effects of this invention are reflected in:

[0059] This case provides a precise, intelligent, and efficient closed-loop solution for the full life-cycle management and maintenance of cable-stayed bridges. It comprehensively ensures the structural safety and stability of the bridge while significantly reducing overall maintenance costs, improving maintenance efficiency, and enhancing the scientific basis of decision-making. Specifically, it is reflected in: the adaptive weighting method of the data acquisition and fusion module, which dynamically allocates confidence weights based on the absolute deviation between sensor data and the median, effectively improving the reliability and accuracy of multi-source monitoring data; the hierarchical optimization and calibration mechanism of the digital twin model construction and dynamic update module, which uses the Levenberg-Marquardt algorithm for rapid local parameter updates and a genetic algorithm for periodic global parameter calibration, ensuring that the digital twin model is highly synchronized with the physical bridge and can accurately simulate material performance degradation; and the structural health assessment and prediction module, which extracts residual features, outputs a health status probability distribution using a multi-class support vector machine model, and combines it with a stochastic process model based on Wiener processes and Monte Carlo simulations to realize the current health status of components. The system includes: scientific assessment and reliable prediction of remaining lifespan; a life-cycle maintenance decision optimization module based on a partially observable Markov decision process model, integrating health assessment and lifespan prediction results, generating a maintenance strategy with the optimal long-term discounted expected total cost through a point-base value iterative algorithm, and verifying the maintenance effect and subsequent expected total cost through a virtual simulation module, further ensuring the optimality of the decision; an event triggering and data management module that can monitor model deviation in real time, automatically cache calibration data and trigger calibration processes to ensure the model's continuous accuracy; a digital twin-based real-time load effect assessment and early warning module that calculates the real-time load effect ratio to achieve multi-level early warning and avoid safety risks in advance; and a maintenance history digital archive and knowledge accumulation module that uses structured storage of life-cycle data and an intelligent recommendation unit based on case reasoning to quickly retrieve similar historical cases and comprehensively evaluate the effectiveness of maintenance plans, providing suitable recommendations for new damage or degradation states, efficiently reusing maintenance experience, and forming a virtuous cycle of maintenance knowledge accumulation and continuous optimization. Attached Figure Description

[0060] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0061] Figure 1 This is a system block diagram of the present invention. Detailed Implementation

[0062] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.

[0063] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0064] like Figure 1 As shown, a digital twin-based bridge lifecycle management and maintenance system is used for the full lifecycle management and maintenance of cable-stayed bridges, including:

[0065] The data acquisition and fusion module is used to acquire monitoring data of the cable-stayed bridge in real time and to fuse multi-source data using an adaptive weighting method.

[0066] The digital twin model construction and dynamic update module is used to construct and continuously calibrate a parametric finite element simulation model based on bridge design parameters and fused data output by the data acquisition and fusion module. As a digital twin model corresponding to the physical bridge, the digital twin model can simulate the degradation of material properties.

[0067] The structural health assessment and prediction module is used to assess the current health status of bridge components based on the updated digital twin model output by the digital twin model construction and dynamic update module, by using the residual characteristics between the digital twin model output and the measured data, and to predict its performance evolution based on a stochastic process model.

[0068] The whole life cycle maintenance decision optimization module is used to integrate the current health status and performance evolution prediction results output by the structural health assessment and prediction module, and generate the optimal maintenance strategy for future cycles based on the decision optimization model;

[0069] The modules are connected sequentially to form a closed-loop process from data perception, model synchronization, state assessment to decision optimization.

[0070] Furthermore, the adaptive weighting method executed by the data acquisition and fusion module specifically includes:

[0071] Simultaneously acquire monitoring data sequences characterizing the same physical quantity from multiple sensor nodes distributed at different parts of the cable-stayed bridge;

[0072] For each sampling time, calculate the absolute deviation between the monitoring data of each sensor and the median of the monitoring data of all sensors of the same type at that time;

[0073] The confidence weight of each sensor is dynamically calculated based on its absolute deviation; the smaller the absolute deviation, the higher the confidence weight. The calculation process is as follows: First, the absolute deviation of each sensor is input into a negative exponential function to obtain its unnormalized weight factor; then, the unnormalized weight factors of all sensors are summed, and the unnormalized weight factor of each sensor is divided by the sum to obtain the normalized confidence weight w. i(t), its mathematical expression is:

[0074] ;

[0075] Where, d i (t) represents the data of the i-th sensor at time t, median(D(t)) is the median of the data of all N sensors of the same type at that time, and β is the adjustment coefficient;

[0076] Finally, based on the confidence weights, the data from all sensors of the same type at the same time are weighted and averaged to output the fused data;

[0077] The adaptive weighting method enables precise fusion of monitoring data from multiple sensors, dynamically adapting to the real-time operating status of the sensors, effectively filtering out abnormal data interference, and significantly improving data reliability and accuracy. The weight allocation is based on the characteristics of the data itself rather than a fixed setting, adapting to the dynamic changes in the complex monitoring environment of cable-stayed bridges. This provides high-quality and reliable data support for subsequent modules such as digital twin model construction and structural health assessment, ensuring the accuracy and effectiveness of subsequent processes in the full life cycle maintenance system.

[0078] Assume that four identical stress sensors (N=4) are installed at a critical stress-bearing component of a cable-stayed bridge to synchronously monitor real-time stress data at that component. In engineering practice, an adjustment coefficient β=0.2 is set (the β value can be adjusted according to the monitoring accuracy requirements to control the sensitivity of the weights to deviations). At time t, the stress data synchronously collected by the four sensors are d1(t)=150MPa, d2(t)=152MPa, d3(t)=151MPa, and d4(t)=170MPa (where d4(t) is abnormally large due to temporary interference with the sensor). The specific calculation process is as follows:

[0079] Calculate the median: Arrange the data in ascending order as 150MPa, 151MPa, 152MPa, 170MPa. The median of an even number of data points is the average of the two middle numbers, i.e., median(D(t)) = (151 + 152) / 2 = 151.5MPa.

[0080] Calculate the absolute deviations: |d1(t)-151.5|=1.5MPa, |d2(t)-151.5|=0.5MPa, |d3(t)-151.5|=0.5MPa, |d4(t)-151.5|=18.5MPa;

[0081] Calculate the unnormalized weighting factor (based on the negative exponential function):

[0082] Sensor 1: ;

[0083] Sensor 2: ;

[0084] Sensor 3: ;

[0085] Sensor 4: ;

[0086] Calculate the total weights: 0.7408 + 0.9048 + 0.9048 + 0.0247 ≈ 2.5751;

[0087] Calculate the normalized confidence weights:

[0088] w1(t)=0.7408 / 2.5751≈0.2877;

[0089] W2(t)=0.9048 / 2.5751≈0.3514;

[0090] w3(t)=0.9048 / 2.5751≈0.3514;

[0091] w4(t)=0.0247 / 2.5751≈0.0096;

[0092] The calculated fused data is: 150×0.2877+152×0.3514+151×0.3514+170×0.0096≈151.26MPa.

[0093] The final fusion result effectively mitigated the impact of abnormal data, more closely approximated the actual monitoring values, and verified the anti-interference ability and accuracy of the method.

[0094] The structural health assessment and prediction module assesses the current health status of bridge components through the following processes:

[0095] The fused data from the data acquisition and fusion module is input into the digital twin model to calculate the simulation response values ​​of key measurement points;

[0096] Calculate the residual vector between the simulated response value and the corresponding measured response value;

[0097] Extract multidimensional feature vectors, including time-domain statistical features and frequency-domain energy features, from the residual vector;

[0098] The multidimensional feature vector is input into a pre-trained multi-class support vector machine model; during the training phase, the multi-class support vector machine model is trained using samples generated from historical data or simulation data, labeled with different health status types and locations.

[0099] The output of the multi-class support vector machine model is the probability distribution of the current bridge structure being in various preset health states;

[0100] The process involves analyzing the residuals of simulated responses from digital twin models and measured data, extracting multi-dimensional features in the time and frequency domains, and then training a pre-trained multi-class support vector machine model. This enables accurate identification of the current health status of bridge components, location of damage, and quantification of state probabilities. It avoids the limitations of single-feature assessment and outputs the probability distribution of various preset health states, thus avoiding the limitations of absolute judgments. Furthermore, the model is trained based on labeled samples of historical or simulated data, adapting to the diverse damage modes of different components of cable-stayed bridges. The assessment results are objective, comprehensive, and reliable, providing direct and crucial status information for subsequent component performance evolution prediction and full life-cycle maintenance decisions, ensuring the pertinence and rationality of maintenance strategies.

[0101] Taking a key load-bearing component of a cable-stayed bridge's main girder as the evaluation object, four health states were preset: S1 (normal, no damage), S2 (minor crack, width < 0.1 mm), S3 (moderate crack, 0.1 mm ≤ width < 0.3 mm), and S4 (severe crack, width ≥ 0.3 mm). A multi-class support vector machine model was trained using 800 sets of labeled samples (including time-domain statistical features and frequency-domain energy feature samples for each health state and different crack locations), achieving a cross-validation accuracy of 93%. The specific evaluation process is as follows:

[0102] Input data: The data acquisition and fusion module outputs the fused strain measured response vector ym=[145,147,146]MPa of the three key measuring points of the component at time t; the fused data is input into the updated digital twin model to calculate the simulation response vector ys=[144.6,146.8,145.7]MPa of the corresponding measuring points;

[0103] Calculate the residual vector: The residual vector e = ym - ys, that is, e = [145 - 144.6, 147 - 146.8, 146 - 145.7] = [0.4, 0.2, 0.3] MPa;

[0104] Extracting multidimensional feature vectors:

[0105] Time-domain statistical characteristics: mean ;

[0106] variance ;

[0107] Peak factor ;

[0108] Frequency domain energy characteristics: A Fast Fourier Transform (FFT) is performed on the residual vector to obtain the energy proportions (after normalization) of the key frequency bands at 5Hz, 15Hz, and 25Hz, respectively: E 5Hz =0.22、E 15Hz =0.41, E 25Hz =0.13;

[0109] The final multidimensional feature vector X = [0.3, 0.01, 4.0, 0.22, 0.41, 0.13];

[0110] Model inference and output: Input the multi-dimensional feature vector X into the pre-trained multi-class support vector machine model. The model outputs the probability distribution of each health state as follows: P(S1)=0.07, P(S2)=0.85, P(S3)=0.06, P(S4)=0.02.

[0111] Application of results: The probability that the component is currently in a state of slight crack (S2) is as high as 85%, while the probability of other states is extremely low. Combined with the sample labeling logic during model training, the crack can be further located in the area near the second measurement point. This provides accurate and quantitative state support for the subsequent prediction of crack propagation trend based on stochastic process model and the formulation of targeted strategies for "local crack repair" by the whole life cycle maintenance decision optimization module.

[0112] The parameter continuous calibration strategy implemented in the digital twin model construction and dynamic update module adopts a hierarchical optimization mechanism, specifically including:

[0113] The first layer of local parameter fast update: Construct an optimization problem with the objective of minimizing the real-time response residual; the objective function of this problem is defined as the squared L2 of the difference between the measured response vector and the simulated response vector of the digital twin model, i.e. , where θ l y represents a subset of local physical parameters directly associated with the sensor measurement point. m and y s These are the measured and simulated response vectors, respectively.

[0114] The Levenberg-Marquardt optimization algorithm is used to solve the optimization problem online iteratively, achieving the solution of θ. l Rapid tracking and adjustment.

[0115] Second-level global parameter periodic calibration: Global calibration is initiated once every fixed time period;

[0116] Global calibration constructs a global optimization problem with the objective of minimizing the weighted average of substructure interface force balance error and overall response residual. The decision variables are a subset of global parameters θ that affect the overall mechanical behavior of the structure. g A genetic algorithm is used to solve the global optimization problem and update θ. g ;

[0117] The results of the second-level calibration are used to constrain the range of values ​​for local parameters in the first-level calibration;

[0118] Through a hierarchical optimization mechanism that combines rapid local parameter updates with periodic global parameter calibration, the digital twin model achieves accurate, efficient, and stable parameter calibration. This ensures both rapid response and dynamic adjustment of local parameters to real-time monitoring data, and constrains the range of local parameter values ​​through global parameter calibration, preventing local optimization from falling into partial optima. This ensures that the model can accurately reflect the real-time local state of the sensor measurement points while maintaining the consistency and accuracy of the overall structural mechanical behavior. Furthermore, the online iterative solution of the Levenberg-Marquardt algorithm and the global optimization of the genetic algorithm adapt to different calibration requirements, allowing the digital twin model to maintain a high degree of synchronization with the physical bridge. This provides a high-precision and high-reliability simulation foundation for subsequent health assessments, performance predictions, and other modules, adapting to the dynamic changes in local damage and overall performance evolution of the complex cable-stayed bridge structure during long-term use.

[0119] Taking a certain span of the main girder of a cable-stayed bridge as the object, the subset of local parameters in the digital twin model that are directly related to the measuring points of three key sensors in that span is... (These are local stiffness coefficients near three measuring points, in GN / m), a subset of global parameters affecting the overall mechanical behavior of the structure. (E is the elastic modulus of the material, in GPa;) (Poisson's ratio, unitless) The preset global calibration cycle is 30 days. The specific calibration process is as follows:

[0120] Global parameter periodic calibration (second layer):

[0121] Before initiating global calibration, the initial global parameters of the model are E0 = 300 GPa. Construct a global optimization objective function: ,in For substructure interface force balance error , The weighting coefficient is ym, which is the measured response vector of the most recent 30 days (taking the strain mean sequence of 3 measuring points [142,145,143,...,144] MPa). This is the simulation response vector under the corresponding global parameters;

[0122] A genetic algorithm was used to solve the global optimization problem. The population size was set to 50 and the number of iterations was set to 100. The updated global parameters were obtained after solving the problem. At the same time, determine the range of local parameter constraints: ;

[0123] Fast local parameter update (first layer):

[0124] The measured response vector y at a certain time t m (t)=[146,148,147]MPa, initial local parameters Substituting the values ​​into the digital twin model yields the initial simulation response vector. Initial residual L2 norm squared ;

[0125] Construct a local optimization objective function (minimize the squared L2 norm of the response residual): , where y s1 ,y s2 ,y s3 The simulated strain values ​​at three measuring points are given under the corresponding local parameters;

[0126] The Levenberg-Marquardt algorithm is used for online iterative solution. An initial damping coefficient (iteration step size factor) α = 0.01 is set (used to adjust the iteration step size; decreasing α accelerates convergence when the residual is large, and increasing α ensures stability when the residual is small), and a convergence threshold are also set. The iterative process is as follows:

[0127] First iteration: Calculate the parameter Jacobian matrix J (reflecting the impact of parameter changes on the simulation response), based on... Solving for parameter update amount I represents the identity matrix, and we obtain Update parameters Substitute the values ​​into the model to obtain the simulation response. residual norm square Since the residual has decreased significantly, α = 0.005 is adjusted to increase the step size of subsequent iterations.

[0128] Second iteration: based on Calculate the new Jacobian matrix Solve for update quantity ,get Update parameters ;

[0129] Simulation Response Residual L2 norm squared = The residual continues to decrease, so adjust α = 0.002.

[0130] Third iteration: Calculate the Jacobian matrix Solve for update quantity ,get Update parameters ;

[0131] Simulation Response residual norm square The convergence threshold is satisfied.

[0132] Final updated local parameters All are within the constraints set by the global calibration;

[0133] Calibration effect verification: After calibration, the residual between the simulated response and the measured response of the model changes from the initial value. , down to This significantly improves the simulation accuracy of the model. Furthermore, the iteration step size factor α is dynamically adjusted, which not only ensures the iteration convergence speed but also avoids excessive oscillations in parameter updates. Global parameter constraints ensure that local stiffness adjustments do not disrupt the overall mechanical equilibrium of the main beam. The model can accurately reflect the local state of the measuring points while maintaining the reliability of the overall structural simulation.

[0134] The structural health assessment and prediction module predicts the evolution of component performance, based on a stochastic process model, and specifically includes:

[0135] For unhealthy components identified by the multi-class support vector machine model, extract the time series of key indicators reflecting their performance degradation from historical monitoring data;

[0136] Assuming the degradation process follows a Wiener process, i.e. the degradation increment follows a normal distribution with mean and variance proportional to the time interval; using the maximum likelihood estimation method and time series data of key indicators, the drift coefficient λ and diffusion coefficient σ of the Wiener process are estimated.

[0137] Set the performance failure threshold Df for the component; based on the estimated parameter drift coefficient λ and diffusion coefficient σ, use the Monte Carlo simulation method to randomly generate a large number of future degradation paths that conform to the Wiener process starting from the current state;

[0138] The time when all simulated future degradation paths first reach or exceed the failure threshold Df is statistically analyzed to form a probability distribution of the component's remaining useful life; the median of this probability distribution is taken as the component's predicted remaining useful life.

[0139] By targeting unhealthy components identified by multi-class support vector machines, and employing a scientific process from extracting time series of key degradation indicators to Wiener process modeling, maximum likelihood estimation of parameters, and Monte Carlo simulation of degradation paths, this approach achieves accurate quantitative prediction of component performance evolution trajectory and remaining life. It fully considers the randomness and uncertainty of the degradation process, outputting a probability distribution of remaining life rather than a fixed value, which better reflects the actual degradation patterns of cable-stayed bridge components and avoids the one-sidedness of traditional fixed life prediction. The Wiener process effectively fits the progressive degradation characteristics of components, the maximum likelihood estimation method ensures the accuracy of model parameter (drift coefficient λ, diffusion coefficient σ) estimation, and numerous Monte Carlo simulation paths guarantee the reliability of remaining life prediction. This provides the core life-cycle maintenance decision optimization module with the criteria for "when to maintain and whether to replace," and helps avoid cost waste caused by over-maintenance or safety risks caused by under-maintenance, thus improving the foresight and scientific nature of maintenance strategies.

[0140] For example, for a critical component of the main girder of a cable-stayed bridge (identified as having a minor crack state S2, with an initial crack width of 0.08 mm), "crack width" is selected as a key indicator reflecting its performance degradation. The specific prediction process is as follows:

[0141] Extracting key indicator time series: Crack width data for the component over the past 3 years (monitored once every 6 months, for a total of 6 time points) were obtained from historical monitoring data. The time series is t=[0,6,12,18,24,30] (unit: month), and the corresponding crack width w(t)=[0.08,0.10,0.13,0.15,0.18,0.21] (unit: mm).

[0142] Wiener process modeling and parameter estimation: Assuming that the crack width degradation process follows the Wiener process. Where w0 = 0.08 mm (initial crack width), λ is the drift coefficient (characterizing the average degradation rate), σ is the diffusion coefficient (characterizing the randomness of degradation), B(t) is the standard Brownian motion, satisfying B(t) ~ N(0,t) (N represents a normal distribution, i.e., B(t) follows a normal distribution with a mean of 0 and a variance of t), and the degradation increment is... ( (months, i=1,2,...,5).

[0143] Calculate the degradation increment: , , , , ;

[0144] Maximum likelihood estimation parameters:

[0145] Estimated value of drift coefficient λ Substituting n=5, , ,have to mm / month;

[0146] Estimated value of diffusion coefficient σ: First calculate

[0147] , then calculate Ultimately (Unit: millimeters per unit per month);

[0148] Setting the failure threshold and Monte Carlo simulation:

[0149] Based on the component design standards, the crack width failure threshold Df is set at 0.3mm (corresponding to the critical value of healthy state S4, i.e., crack width ≥ 0.3mm is judged as failure).

[0150] Use Monte Carlo simulation to generate 10,000 future degradation paths, and the degradation process of each path is (At the current time t = 30 months, the current crack width is 0.21 mm). The simulation time interval for each step is 1 month until the crack width is first ≥ 0.3 mm, and record this time as the failure time Ti of each path;

[0151] Remaining life statistics and output:

[0152] Sort and statistically analyze the failure times Ti of 10,000 paths to obtain the remaining life probability distribution: P(T ≤ 20 months) = 12%, P(20 < T ≤ 25 months) = 35%, P(25 < T ≤ 30 months) = 40%, P(T > 30 months) = 13%;

[0153] Take the median of the probability distribution as the predicted remaining life, that is months (about 2 years), indicating that there is a 50% probability that the component will reach the failure state within the next 24 months, providing an accurate quantitative life basis for the targeted maintenance strategy of "conduct a special inspection again after 18 months and implement crack grouting repair before 22 months" formulated by the full life cycle maintenance decision optimization module.

[0154] In the full life cycle maintenance decision optimization module, the decision optimization model is a partially observable Markov decision process model, and its construction and solution process include:

[0155] Discretize the health state of the key components of the bridge into a finite number of levels, and model the impact of different maintenance actions on state transition as a state transition probability matrix;

[0156] Use the health state probability distribution output by the multi-class support vector machine model as the incomplete observation information of the true health state of the component to update the belief state in the partially observable Markov decision process model;

[0157] Define the value function as the long-term discounted expected total cost, which includes the immediate maintenance action cost and the risk cost brought by the component being in different health states;

[0158] Use the point-based value iteration algorithm to approximately solve the partially observable Markov decision process model to obtain the optimal maintenance strategy based on the current belief state;

[0159] A full life-cycle maintenance decision optimization system is constructed using a partially observable Markov decision process (POMDP) ​​model. This system seamlessly integrates the outputs of preceding modules (health state probability distribution of multi-class support vector machines and performance evolution prediction of stochastic process models). The health status of key bridge components is discretized and the impact of maintenance actions on state transitions is quantified. The belief state is updated with incomplete observation information. A value function is defined by combining the long-term discounted expected total cost (including immediate maintenance costs and risk costs of each health state). The optimal strategy is efficiently solved using a point-based value iteration algorithm. This system fully considers the uncertainty of health status and the dynamic impact of maintenance actions, while achieving a global balance between maintenance costs and safety risks. It avoids blind or insufficient maintenance and outputs an optimal strategy that is forward-looking, economical, and targeted. This provides a scientific basis for decision-making in the full life-cycle maintenance of cable-stayed bridges, significantly improving the efficiency of maintenance resource utilization and the level of structural safety assurance.

[0160] For example, the key components of the main girder of a cable-stayed bridge (current health status probability distribution: P(S1)=0.07, P(S2)=0.85, P(S3)=0.06, P(S4)=0.02, predicted remaining life of 24 months, set decision cycle of 6 months, for a total of 4 planning cycles), the specific decision-making process is as follows:

[0161] Basic settings:

[0162] Discrete health status: S1 (normal and undamaged, crack width < 0.1 mm), S2 (minor crack, 0.1 mm ≤ width < 0.3 mm), S3 (moderate crack, 0.3 mm ≤ width < 0.5 mm), S4 (severe crack, width ≥ 0.5 mm, failure status).

[0163] Maintenance actions: A1 (regular inspection, no repair effect, only updates status information), A2 (grouting repair of cracks, can delay / reverse degradation), A3 (complete component replacement, restore to S1);

[0164] Cost parameters: Maintenance costs C(A1) = 10,000 yuan, C(A2) = 200,000 yuan, C(A3) = 1,000,000 yuan; Risk costs (period losses of components in this state) R(S1) = 0 yuan, R(S2) = 50,000 yuan, R(S3) = 500,000 yuan, R(S4) = 2,000,000 yuan; Long-term discount factor γ = 0.9 (characterizing the discount weight of future costs);

[0165] State transition probability matrix (determined based on historical data and simulation):

[0166] When executing A1 (detection only): (Rows represent the current state, and columns represent the state of the next cycle);

[0167] When performing A2 (crack repair): ;

[0168] When executing A3 (component replacement): (All will be restored to S1 after replacement);

[0169] Initial belief state (incomplete observation of the actual health state): b0=[P(S1),P(S2),P(S3),P(S4)]=[0.07,0.85,0.06,0.02].

[0170] Belief state update and value function definition:

[0171] Belief State Update: After performing maintenance actions, beliefs are updated based on new state observations (such as detection results). The formula is as follows: Assuming no observational error, it simplifies to Value function (expected total cost with long-run discount):

[0172] ,in For the updated belief state, a smaller V(b) indicates a better strategy.

[0173] Point base value iterative solution process:

[0174] Step 1: Initialize the value function V0(b) = 0 (assuming no initial cost);

[0175] Step 2: Iteratively calculate the value (taking 2 iterations as an example):

[0176] First iteration (V1(b)):

[0177] If A1 is selected: maintenance cost is 10,000 yuan, current risk cost is 0.07×0+0.85×5+0.06×50+0.02×200=4.25+3+4=11.25 million yuan;

[0178] The updated belief b1 = [0.07×0.95+0.85×0.02+0.06×0+0.02×0,0.07×0.05+0.85×0.7+0.06×0.05+0.02×0,0.07×0+0.85×0.28+0.06×0.6+0.02×0,0.07×0+0.85×0+0.06×0.35+0.02×1.0] = [0.0845,0.6095,0.258,0.048]; Expected future cost ;

[0179] Total value Ten thousand yuan;

[0180] If A2 is selected: maintenance cost is 200,000 yuan, current risk cost is 112,500 yuan; after updating, belief b2 = [0.07×0.95+0.85×0.8+0.06×0.1+0.02×0,0.07×0.05+0.85×0.18+0.06×0.6+0.02×0,0.07×0+0.85×0.02+0.06×0.25+0.02×0,0.07×0+0.85×0+0.06×0.05+0.02×1.0] = [0.7325,0.2145,0.032,0.021];

[0181] Total value Ten thousand yuan;

[0182] If A3 is selected: maintenance cost is 1 million yuan, current risk cost is 112,500 yuan; after updating, belief b3 = [1.0,0,0,0];

[0183] Total value Ten thousand yuan;

[0184] The optimal action in the first iteration is A1. Ten thousand yuan;

[0185] Second iteration ( ):

[0186] Choose A1: Total Value ,in Ten thousand yuan, so Ten thousand yuan;

[0187] Choose A2: Total Value ,in Ten thousand yuan, so Ten thousand yuan;

[0188] Choose A3: Total Value

[0189] Ten thousand yuan;

[0190] Third iteration ( As iterations deepen, the future cost advantage of A2 gradually becomes apparent (the risk cost of b2 is significantly reduced), and finally, after iteration convergence, Ten thousand yuan (A1), 10,000 yuan (A2) The optimal action is A2 (crack grouting repair), which costs 10,000 yuan (A3).

[0191] Application of Results: Under the current belief state, choosing A2 (crack repair) has the lowest expected total cost of long-term discount (321,000 yuan). It avoids the high subsequent risk costs of crack expansion caused by A1, and does not have to bear the high replacement cost of A3. It provides accurate decision-making for full life cycle maintenance, matches the previously predicted 24-month remaining life, and can be re-evaluated 18 months after repair, forming a closed-loop maintenance.

[0192] The system also includes a virtual simulation verification module, used to pre-evaluate the effects of candidate maintenance strategies generated by the full lifecycle maintenance decision optimization module. Its process includes:

[0193] The planned maintenance actions in the candidate maintenance strategy are transformed into modification instructions for the corresponding component parameters in the digital twin model to simulate the maintenance effect;

[0194] On the modified digital twin model, input the typical load sequence for the future planning period predicted based on historical data to perform structural response simulation;

[0195] Based on the simulation results, the structural health assessment and prediction module was called again to re-predict the performance degradation trajectory and remaining life of the relevant components after simulated maintenance.

[0196] Based on the re-predicted component state sequence, the expected total cost of implementing the candidate strategy in subsequent planning periods is estimated using a partially observable Markov decision process model; the selection is made by comparing the long-term expected total costs of different candidate strategies.

[0197] A virtual simulation verification module constructs a "pre-evaluation closed loop" for candidate maintenance strategies, quantifying their long-term effects without actual maintenance actions. This connects the candidate strategies in the preceding full life-cycle maintenance decision optimization module with the performance evolution model in the structural health assessment and prediction module. Furthermore, it simulates maintenance effects by modifying parameters in a digital twin model, simulates structural responses by inputting future load sequences, and re-predicts degradation trajectories and remaining lifespan. Finally, it estimates the long-term expected total cost based on a partially observable Markov decision process model, enabling horizontal comparison of different candidate strategies. This effectively avoids safety risks and cost waste caused by improper actual strategy implementation, ensuring that the output optimal maintenance strategy is feasible, economical, and long-lasting. Simultaneously, it reduces resource consumption from physical experiments, shortens the decision-making cycle, provides dual guarantees for full life-cycle maintenance decisions, and further enhances the scientific rigor and reliability of maintenance strategies.

[0198] For example, for key components of the main girder of a cable-stayed bridge (current health status P(S1)=0.07, P(S2)=0.85, P(S3)=0.06, P(S4)=0.02, predicted remaining life of 24 months, decision cycle of 6 months, planning period of 2 years, totaling 4 cycles), the candidate maintenance strategies are A1 (periodic inspection), A2 (crack grouting repair), and A3 (component replacement). The specific verification process is as follows:

[0199] Candidate strategies are transformed into model parameter modifications:

[0200] Strategy A1 (Periodic Inspection): No maintenance or repair is required; the parameters of the digital twin model remain unchanged, i.e., the local stiffness remains constant. GN / m, material elastic modulus E=298.5GPa, time-varying material strength f(t)=298.5GPa;

[0201] Strategy A2 (Crack Grouting Repair): Simulates crack closure and stiffness increase after repair, adjusting the local stiffness of the model to... GN / m, material strength increased to GPa (corresponding to a decrease in degradation rate);

[0202] Strategy A3 (Component Replacement): Simulate the replacement process and restore the system to a damage-free state; reset the model parameters to their initial design values. GN / m, GPa, GPa;

[0203] Input future typical load sequences: Based on historical traffic load and wind load data, predict the typical load spectrum for four decision cycles in the next two years. The peak stress sequences for each cycle are L1=[145,146,144]MPa, L2=[148,147,149]MPa, L3=[146,145,147]MPa, and L4=[147,148,146]MPa. Input the corresponding modified digital twin model to perform structural response simulation.

[0204] Re-predicting performance degradation trajectory and remaining lifetime:

[0205] The structural health assessment and prediction module is invoked, and the performance evolution prediction (Wiener process model) is re-executed based on the simulation response:

[0206] A1: Degradation parameters remain unchanged ( ), and re-simulate 10,000 degradation paths. The median remaining lifespan is still 24 months. The probability of health status in each cycle is: S2→S3 probability 28% / cycle;

[0207] A2: The degradation rate decreases after repair. mm / month, The remaining life expectancy remains unchanged, increasing to 60 months. The probability of health status in each cycle is as follows: 2% probability of S2→S3 per cycle, and 80% probability of S2→S1.

[0208] A3: After replacement, restore to S1, degraded parameters. mm / month, median remaining life expectancy increased to 120 months, probability of health status in each cycle: S1→S2 probability 5% / cycle;

[0209] Estimate the long-term expected total cost:

[0210] Based on the partially observable Markov decision process model, the cost parameters are used (maintenance cost C(A1) = 10,000, C(A2) = 200,000, C(A3) = 1,000,000);

[0211] Given risk costs R(S1) = 0,000, R(S2) = 50,000, R(S3) = 500,000, and R(S4) = 2,000,000; and a discount factor γ = 0.9, calculate the expected total cost of long-term discounts for each strategy:

[0212] A1: Initial belief b0 = [0.07, 0.85, 0.06, 0.02];

[0213] ;

[0214] Update belief b1 = [0.0845, 0.6095, 0.258, 0.048];

[0215]

[0216] Update belief b2 = [0.0913, 0.439, 0.3328, 0.1369];

[0217] ;

[0218] Update belief b3=[0.098,0.307,0.385,0.21];

[0219] ;

[0220] Iterative convergence (cumulative discount term):

[0221] Substitute these values ​​sequentially into the calculations for subsequent belief states, and accumulate them. After discounting terms, the final convergence is:

[0222] Ten thousand yuan;

[0223] A2: ;

[0224] in, ;

[0225] Initial belief b0 = [0.07, 0.85, 0.06, 0.02];

[0226] ;

[0227] Update belief b1 = [0.7525, 0.1925, 0.032, 0.0225];

[0228] ;

[0229] Update belief b2 = [0.8721, 0.0914, 0.0119, 0.0241];

[0230] ;

[0231] Update belief b3=[0.92,0.065,0.008,0.007];

[0232] ;

[0233] Iterative convergence (cumulative discount term):

[0234] Grand total As subsequent risk costs continue to decrease, the iteration converges to:

[0235] Ten thousand yuan;

[0236] A3: ;

[0237] in, (After replacement) (The rest are 0)

[0238] Initial belief b0 = [0.07, 0.85, 0.06, 0.02];

[0239] ;

[0240] Update belief b1=[1.0,0,0,0];

[0241] ;

[0242] Update belief b2=[0.95,0.05,0,0]:

[0243] ;

[0244] Update belief b3=[0.9025,0.0925,0.005,0]:

[0245] ;

[0246] Iterative convergence (cumulative discount term):

[0247] Grand total After iteration, due to the high initial cost, the following convergence was obtained:

[0248] Ten thousand yuan;

[0249] Strategy Selection: Through virtual simulation verification, A2 (crack grouting repair) has the lowest long-term expected total cost (567,000 yuan), which is better than both the high subsequent risk cost of A1 and the high initial cost of A3. Furthermore, the simulation predicts that it can extend the remaining life of the component from 24 months to 60 months, verifying the long-term effectiveness and economy of the strategy. Finally, A2 was determined to be the optimal maintenance strategy, avoiding the resource waste or safety hazards that might have been caused by the actual implementation of A1 or A3.

[0250] The system also includes an event triggering and data management module; the working mechanism of the event triggering and data management module is as follows:

[0251] Calculate the statistics of specific monitoring indicators in real time within a sliding time window and compare them with the simulation output statistics of the digital twin model under the same time window and the same input conditions;

[0252] When the statistical difference exceeds the preset threshold for multiple consecutive calculation periods, and the current environmental conditions belong to the preset typical conditions library, a significant model deviation event is determined to have occurred.

[0253] The system automatically caches all relevant original monitoring data, fused data, environmental data, and corresponding input and output data of the digital twin model within a preset time period before and after a significant deviation event in the model, forming a dedicated calibration dataset.

[0254] Send the calibration-specific dataset to the digital twin model building and dynamic update module, and trigger it to start or prioritize the execution of the model parameter calibration process once;

[0255] Through event-triggered mechanisms and intelligent data management functions, the system monitors the real-time matching degree between the digital twin model and the physical bridge, accurately identifies significant model deviation events, and avoids systemic deviations in subsequent health assessments, life predictions, and maintenance decisions due to model inaccuracies. Simultaneously, it automatically caches all relevant data before and after deviation events (raw monitoring data, fused data, environmental data, and model input / output data), forming a standardized calibration dataset. This significantly reduces the cost of manual data screening and processing and triggers a priority model calibration process, ensuring the digital twin model quickly returns to an accurate state. It seamlessly connects the digital twin model construction and dynamic update modules, strengthening the closed-loop collaboration of "data monitoring - deviation identification - model calibration," improving the system's responsiveness to structural state changes under complex working conditions of cable-stayed bridges, and ensuring the reliability and accuracy of all aspects of the entire lifecycle maintenance process.

[0256] For example, key components of the main girder of a cable-stayed bridge (local stiffness) GN / m, elastic modulus E=298.5GPa, current health status is slight crack S2), the specific implementation process is as follows:

[0257] Basic parameter settings:

[0258] The sliding time window length T = 10 minutes (600 seconds), the sampling frequency is 1 time / second, and each window contains 600 data points; the monitoring index is the strain value of the key measuring points of the component;

[0259] The statistic is the mean strain within the window. and variance ;

[0260] Preset threshold: Absolute deviation of mean threshold MPa, relative deviation threshold of variance ;

[0261] The typical working condition library includes 5 common working conditions: normal traffic load (50-100 vehicles / hour) + light wind (wind speed v≤8m / s), medium traffic load (100-150 vehicles / hour) + light wind, etc. The current environmental working condition is "traffic load 85 vehicles / hour + wind speed 6.2m / s", which belongs to the typical working condition library;

[0262] Real-time monitoring and difference calculation:

[0263] The system calculates the monitoring statistics and simulation statistics for each sliding window in real time. The formula for calculating the difference is as follows:

[0264] Absolute deviation of the mean: ( To monitor the average, (The average value is the simulation result).

[0265] Variance relative bias: ( To monitor variance, (for simulation variance).

[0266] Calculation results for four consecutive sliding windows:

[0267] Window 1

[0268] (Exceeding the threshold) (Not exceeding the threshold);

[0269] Window 2:

[0270] (Exceeding the threshold) (Exceeding the threshold);

[0271] Window 3:

[0272] (Exceeding the threshold) (Exceeding the threshold);

[0273] Window 4:

[0274] (Exceeding the threshold) (Exceeding the threshold);

[0275] Event triggering and data management:

[0276] Because both types of statistics exceeded the preset threshold for three consecutive windows (windows 2-4), and the current working condition belongs to the typical working condition library, a significant model deviation event was determined to have occurred.

[0277] The system automatically caches the full data from the two windows before the event (windows 2-3), the window during the event (window 4), and the window after the event (window 5, for subsequent data collection): the original monitoring strain sequence, data fusion results, environmental data (traffic volume, wind speed), and digital twin model input parameters. E) and simulated strain sequences form a dedicated calibration dataset. ;

[0278] Immediately The data is sent to the digital twin model building and dynamic update module, triggering a priority calibration process. The module pauses regular periodic calibrations, based on... Enable rapid local parameter updates;

[0279] Calibration effect verification:

[0280] Local stiffness is updated after calibration using the Levenberg-Marquardt algorithm. GN / m, statistics for window 6 after calibration:

[0281] ; (Not exceeding the threshold) (If the threshold is not exceeded), model bias is eliminated, ensuring the accuracy of subsequent health assessments and maintenance decisions.

[0282] The system also includes a digital twin-based real-time load effect assessment and early warning module, used to dynamically assess the impact of actual loads on the safety of bridge structures. Its workflow specifically includes:

[0283] Based on the real-time traffic load data and wind load data acquired by the data acquisition and fusion module, the actual load spectrum for the current time period is generated.

[0284] Using the actual load spectrum as input, the digital twin model is driven to perform structural response simulation, and the stress time history, displacement response and overall structural response extreme values ​​of key components under the current actual load are calculated.

[0285] The simulated stress time history of the components is compared with the time-varying material strength of the components in real time. The time-varying material strength is determined based on the material property degradation model that can be simulated by the digital twin model. The real-time load effect ratio of each component is calculated. The calculation formula is as follows: ,in f represents the peak stress of the i-th component obtained from the simulation under the current actual load spectrum. i(t) The material strength of the component at the current moment is given by the digital twin model;

[0286] Set multi-level early warning thresholds; when the real-time load effect of any component is higher than... When the first-level warning threshold is exceeded, attention-level warning information is generated; when the second-level higher warning threshold is exceeded, action-level warning information is generated and pushed to the full life cycle maintenance decision optimization module as input to trigger immediate assessment or adjustment of maintenance strategies.

[0287] By generating an actual load spectrum through real-time fusion of traffic and wind load data, a digital twin model is driven to dynamically simulate the stress response of key components. Combined with time-varying material strength (adapting to material degradation), the real-time load effect ratio is calculated, and a multi-level early warning mechanism is established. This mechanism can accurately quantify the safety impact of the current load on the bridge structure, avoiding the one-sidedness of fixed strength assessment. It can also promptly avoid safety risks caused by overloading or degradation through attention-level and action-level early warnings. At the same time, the early warning information is pushed to the full life cycle maintenance decision optimization module, triggering the immediate adjustment of maintenance strategies. This strengthens the linkage between "load monitoring-effect assessment-early warning-decision", improving the timeliness of safety assurance and the dynamic adaptability of maintenance strategies for cable-stayed bridges under complex load conditions.

[0288] For example, the specific implementation process for a key component of the main girder of a cable-stayed bridge (numbered i=1, currently in a state of minor crack S2, with a crack width of 0.21mm at t=30 months) is as follows:

[0289] Basic parameter settings:

[0290] Time-varying material strength: The digital twin model, based on a material property degradation model, provides the time-varying material strength of the component at the current moment. MPa (Initial design strength 300MPa, degraded by 5MPa due to minor cracks);

[0291] Actual load spectrum generation: The data acquisition and fusion module acquires real-time traffic load data (average 120 vehicles / hour, including 3 heavy-duty vehicles) and wind load data (average wind speed 7.5 m / s) for the current hour, generates the actual load spectrum for the current period, inputs it into the updated digital twin model for structural response simulation, obtains the stress-time history curve of the component, and extracts the peak stress. MPa;

[0292] Multi-level warning thresholds: Preset first-level (attention level) warning threshold Level 2 (Action-level) warning threshold ;

[0293] Real-time load effect ratio calculation:

[0294] According to the formula Substituting the data, we get: ;

[0295] Early warning judgment and information application:

[0296] because The system generates a warning message at the attention level, indicating that "the current load effect of the No. 1 main beam component is close to the safety threshold, and it is recommended to increase the monitoring frequency from once every 6 hours to once every 2 hours."

[0297] Over the next two hours, traffic load increased to 150 vehicles / hour (including 5 heavy-duty vehicles), and wind load speed rose to 9 m / s. The actual load spectrum was regenerated and simulated. MPa, recalculate the load effect ratio: It has reached the action-level early warning threshold;

[0298] The system immediately generated an action-level early warning message, which stated: "The load effect of the No. 1 main beam component has reached the safety threshold, and there is a risk of crack propagation. It is recommended to implement temporary traffic control (restricting the passage of heavy-duty vehicles) and start crack grouting repair within 24 hours."

[0299] The action-level early warning information was pushed to the full life cycle maintenance decision optimization module, triggering an immediate adjustment to the original optimal strategy A2 (crack grouting repair), bringing forward the repair action originally planned to be carried out 22 months ago to within 1 month, avoiding safety accidents caused by the superposition of load and degradation.

[0300] The system also includes a maintenance history digital archive and knowledge accumulation module, used to achieve structured storage and experience mining of data throughout the entire lifecycle. Its specific implementation methods include:

[0301] Establish a database associated with the geometric topology information of the digital twin model to store raw data and fused data from the data acquisition and fusion module, health status assessment records and remaining life prediction records from the structural health assessment and prediction module, decision records from the whole life cycle maintenance decision optimization module and their corresponding verification results output by the virtual simulation verification module, as well as records of actual maintenance actions and post-effect evaluation data, by component and by time slice.

[0302] Design a case-based reasoning-based intelligent recommendation unit for maintenance schemes: When the structural health assessment and prediction module identifies a new damage mode or degradation state, this unit retrieves historical similar cases from the database. The similarity is determined by comparing the Euclidean distance between multiple feature vectors of damage mode, component type, environmental conditions, and current structural state.

[0303] The maintenance plans adopted in similar historical cases, the predicted effects verified by virtual simulation, and the long-term effects after actual implementation are comprehensively and weighted to generate a set of recommended maintenance plans and their expected effects analysis for the current new situation, which can be used as a reference for the whole life cycle maintenance decision optimization module.

[0304] By establishing a structured and traceable storage system for maintenance data throughout its entire lifecycle through a digital archive of maintenance history, the system connects the entire data chain from data collection, health assessment, decision execution, and effect verification. Simultaneously, an intelligent recommendation unit based on case-based reasoning accurately matches similar historical cases using Euclidean distance. This comprehensively evaluates the virtual simulation prediction effects and actual long-term effects of maintenance plans, rapidly generating a set of recommended plans adapted to new damage patterns or degradation states. This avoids repeated trial and error and resource waste, while also enabling efficient reuse of maintenance experience and knowledge accumulation, forming a virtuous cycle of "data storage - experience mining - intelligent recommendation - effect feedback." This provides diverse references for the lifecycle maintenance decision optimization module, significantly improving the adaptability of maintenance plans, decision-making efficiency, and overall maintenance level.

[0305] For example, the key components of the main girder of a cable-stayed bridge (component type: main girder load-bearing component; current condition: slight crack S2, crack width 0.21mm, damage mode "linear crack in tension zone"; environmental conditions: annual average humidity 60%, annual average temperature 25℃; current structural condition: average strain 146MPa, local stiffness 31.5-32.6GN / m), the specific implementation process is as follows:

[0306] Fundamentals of structured data storage:

[0307] The database stores historical data in “component-time slices”, containing 300 maintenance cases. The core storage dimensions of each case are: damage mode, component type, environmental conditions (humidity / temperature), current structural state (mean strain / local stiffness), maintenance plan, virtual simulation verification effect (remaining life extension), and long-term effect after actual implementation (crack width change after 1 year).

[0308] Feature vector extraction and similar case retrieval:

[0309] Define the feature vector dimensions: X=[x1,x2,x3,x4,x5,x6], where x1 (damage mode: no damage=1, minor crack=2, moderate crack=3, severe crack=4), x2 (component type: main beam=1, cable-stayed bridge=2, pier=3), x3 (humidity normalized value=humidity / 100), x4 (temperature normalized value=temperature / 50), x5 (mean strain normalized value=strain / 200), and x6 (local stiffness normalized value=average stiffness / 40).

[0310] New situation feature vector: (Average stiffness = (31.5 + 32.6 + 30.5) / 3 ≈ 31.53 GN / m, normalized value = 31.53 / 40 ≈ 0.8025);

[0311] Similarity determination: using the Euclidean distance formula The database retrieved the top 3 similar cases:

[0312] Case 1: Euclidean distance

[0313] Case 2: Euclidean distance ;

[0314] Case 3: Euclidean distance ;

[0315] Comprehensive weighted evaluation of maintenance plan:

[0316] Weighting: Weights for virtual simulation prediction results Weight of actual long-term effects Overall score (The higher the performance score, the better; the maximum score is 10 points).

[0317] Case 1: Maintenance Plan A2 (Crack Grouting Repair), Simulation Effect Score (Remaining lifespan extended to 58 months), actual effect score (The crack width decreased to 0.09 mm after 1 year), overall score ;

[0318] Case 2: Maintenance Plan (Crack grouting + anti-corrosion coating), simulation effect score (Remaining lifespan extended to 62 months), actual effect score (After one year, the crack width decreased to 0.08mm, but the coating cost increased), Overall Score ;

[0319] Case 3: Maintenance Plan A1 (Regular Inspection + Crack Observation), Simulation Effect Score (Remaining lifespan maintained at 22 months), actual performance score (The crack width increased to 0.28 mm after 1 year), overall score ;

[0320] Recommendation set generation and application:

[0321] The system outputs a set of recommended solutions: Top 1 recommendation A2 (overall score 8.92), with expected results of "crack width reduced to 0.09-0.10mm after 1 year, remaining lifespan extended to 55-60 months, and long-term expected total cost of 560,000-580,000 yuan";

[0322] Top 2 Recommendation (Overall score 8.77) Additional explanation: "Anti-corrosion coating can improve durability, but requires an additional cost of 30,000 yuan."

[0323] The recommended solution set was pushed to the full life cycle maintenance decision optimization module, and it was highly consistent with the original optimal strategy A2, further verifying the rationality of the strategy. At the same time, it provided a reference for adjusting the solution under special environments (such as future increases in humidity).

[0324] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A bridge lifecycle management and maintenance system based on digital twins, used for the lifecycle management and maintenance of cable-stayed bridges, characterized in that, include: The data acquisition and fusion module is used to acquire monitoring data of the cable-stayed bridge in real time and to fuse multi-source data using an adaptive weighting method. The digital twin model construction and dynamic update module is used to construct and continuously calibrate a parametric finite element simulation model based on bridge design parameters and fused data output by the data acquisition and fusion module. As a digital twin model corresponding to the physical bridge, the digital twin model can simulate material performance degradation. The structural health assessment and prediction module is used to assess the current health status of bridge components based on the updated digital twin model output by the digital twin model construction and dynamic update module, by using the residual characteristics between the digital twin model output and the measured data, and to predict its performance evolution based on a stochastic process model. The whole life cycle maintenance decision optimization module is used to integrate the current health status and performance evolution prediction results output by the structural health assessment and prediction module, and generate the optimal maintenance strategy for future cycles based on the decision optimization model.

2. The bridge lifecycle management and maintenance system based on digital twins according to claim 1, characterized in that: The adaptive weighting method executed by the data acquisition and fusion module specifically includes: Simultaneously acquire monitoring data sequences characterizing the same physical quantity from multiple sensor nodes distributed at different parts of the cable-stayed bridge; For each sampling time, calculate the absolute deviation between the monitoring data of each sensor and the median of the monitoring data of all sensors of the same type at that time; The confidence weight of each sensor is dynamically calculated based on the absolute deviation. The smaller the absolute deviation of a sensor, the higher its confidence weight. The calculation process of the confidence weight is as follows: First, the absolute deviation of each sensor is input into a negative exponential function to obtain its unnormalized weight factor. Then, the unnormalized weight factors of all sensors are summed, and the unnormalized weight factor of each sensor is divided by the sum to obtain the normalized confidence weight. Finally, based on the confidence weights, the data from all sensors of the same type at the same time are weighted and averaged to output the fused data.

3. The bridge lifecycle management and maintenance system based on digital twins according to claim 1, characterized in that: The structural health assessment and prediction module assesses the current health status of bridge components through the following processes: The fused data from the data acquisition and fusion module is input into the digital twin model to calculate the simulation response values ​​of key measurement points; Calculate the residual vector between the simulated response value and the corresponding measured response value; Extract multidimensional feature vectors, including time-domain statistical features and frequency-domain energy features, from the residual vector; The multidimensional feature vectors are input into the pre-trained multi-class support vector machine model; during the training phase, the multi-class support vector machine model is trained using samples generated from historical data or simulation data, which are labeled with different health status types and locations. The output of the multi-class support vector machine model is the probability distribution of the current bridge structure being in various preset health states.

4. The bridge lifecycle management and maintenance system based on digital twins according to claim 3, characterized in that: The parameter continuous calibration strategy implemented in the digital twin model construction and dynamic update module adopts a hierarchical optimization mechanism, specifically including: The first layer of local parameter fast update: construct an optimization problem with the goal of minimizing the real-time response residual; the objective function of this problem is defined as the squared L2 of the difference between the measured response vector and the simulated response vector of the digital twin model; Second-level global parameter periodic calibration: Global calibration is initiated once every fixed time period; Global calibration constructs a global optimization problem with the goal of minimizing the weighted average of substructure interface force balance error and overall response residual. The decision variables are a subset of global parameters that affect the overall mechanical behavior of the structure. A genetic algorithm is used to solve the global optimization problem and update the global parameter subset. The results of the second-level calibration are used to constrain the range of values ​​for local parameters in the first-level calibration.

5. The bridge lifecycle management and maintenance system based on digital twins according to claim 1, characterized in that: The structural health assessment and prediction module predicts the evolution of component performance, based on a stochastic process model, and specifically includes: For unhealthy components identified by the multi-class support vector machine model, extract the time series of key indicators reflecting their performance degradation from historical monitoring data; Assuming the degradation process follows a Wiener process, i.e. the degradation increment follows a normal distribution with mean and variance proportional to the time interval; using the maximum likelihood estimation method and time series data of key indicators, the drift coefficient and diffusion coefficient of the Wiener process are estimated. Set a performance failure threshold for the component; based on the estimated parameter drift coefficient and diffusion coefficient, use the Monte Carlo simulation method to randomly generate a large number of future degradation paths that conform to the Wiener process starting from the current state; The time when all simulated future degradation paths first reach or exceed the failure threshold is statistically analyzed to form a probability distribution of the component's remaining useful life; the median of this probability distribution is taken as the component's predicted remaining useful life.

6. The bridge lifecycle management and maintenance system based on digital twins according to claim 5, characterized in that: In the whole life cycle maintenance decision optimization module, the decision optimization model is a partially observable Markov decision process model, and its construction and solution process includes: The health status of key bridge components is discretized into a finite number of levels, and the impact of different maintenance actions on state transitions is modeled as a state transition probability matrix. The health status probability distribution output by the multi-class support vector machine model is used as incomplete observation information of the true health status of the components to update the belief state in the partially observable Markov decision process model. The value function is defined as the long-run discounted expected total cost, which includes the cost of immediate maintenance actions and the risk cost of components being in different health states; A point-based iterative algorithm is used to approximate the solution of a partially observable Markov decision process model, yielding the optimal maintenance strategy based on the current belief state.

7. The bridge lifecycle management and maintenance system based on digital twins according to claim 6, characterized in that: The system also includes a virtual simulation verification module, used to pre-evaluate the effectiveness of candidate maintenance strategies generated by the full lifecycle maintenance decision optimization module. Its process includes: The planned maintenance actions in the candidate maintenance strategy are transformed into modification instructions for the corresponding component parameters in the digital twin model to simulate the maintenance effect; On the modified digital twin model, input the typical load sequence for the future planning period predicted based on historical data to perform structural response simulation; Based on the simulation results, the structural health assessment and prediction module was called again to re-predict the performance degradation trajectory and remaining life of the relevant components after simulated maintenance. Based on the re-predicted component state sequence, the expected total cost of implementing the candidate strategy in subsequent planning periods is estimated using a partially observable Markov decision process model; the selection is made by comparing the long-term expected total costs of different candidate strategies.

8. The bridge lifecycle management and maintenance system based on digital twins according to claim 7, characterized in that: The system also includes an event triggering and data management module; the working mechanism of the event triggering and data management module is as follows: Calculate the statistics of specific monitoring indicators in real time within a sliding time window and compare them with the simulation output statistics of the digital twin model under the same time window and the same input conditions; When the statistical difference exceeds the preset threshold for multiple consecutive calculation periods, and the current environmental conditions belong to the preset typical conditions library, a significant model deviation event is determined to have occurred. The system automatically caches all relevant original monitoring data, fused data, environmental data, and corresponding input and output data of the digital twin model within a preset time period before and after a significant deviation event in the model, forming a dedicated calibration dataset. Send the calibration-specific dataset to the digital twin model building and dynamic update module, and trigger it to start or prioritize the execution of the model parameter calibration process.

9. The bridge lifecycle management and maintenance system based on digital twins according to claim 8, characterized in that: The system also includes a digital twin-based real-time load effect assessment and early warning module, used to dynamically assess the impact of actual loads on the safety of bridge structures. Its workflow includes: Based on the real-time traffic load data and wind load data acquired by the data acquisition and fusion module, the actual load spectrum for the current time period is generated. Using the actual load spectrum as input, the digital twin model is driven to perform structural response simulation, and the stress time history, displacement response and overall structural response extreme values ​​of key components under the current actual load are calculated. The stress time history of the component obtained from the simulation is compared with the time-varying material strength of the component in real time. The time-varying material strength is determined based on the material performance degradation model that can be simulated by the digital twin model; the real-time load effect ratio of each component is calculated. Set multi-level early warning thresholds; when the real-time load effect ratio of any component exceeds the first-level early warning threshold, an attention-level early warning message is generated; when it exceeds the second-level higher early warning threshold, an action-level early warning message is generated and pushed to the full life cycle maintenance decision optimization module as input to trigger immediate evaluation or adjustment of maintenance strategies.

10. The bridge lifecycle management system based on digital twins according to claim 9, characterized in that: The system also includes a maintenance history digital archive and knowledge accumulation module, used to achieve structured storage and experience mining of data throughout the entire lifecycle. Its specific implementation methods include: Establish a database associated with the geometric topology information of the digital twin model to store raw data and fused data from the data acquisition and fusion module, health status assessment records and remaining life prediction records from the structural health assessment and prediction module, decision records from the whole life cycle maintenance decision optimization module and their corresponding verification results output by the virtual simulation verification module, as well as records of actual maintenance actions and post-effect evaluation data, by component and by time slice. It also includes a case-based reasoning-based intelligent recommendation unit for maintenance plans: when the structural health assessment and prediction module identifies a new damage mode or degradation state, this unit retrieves historical similar cases from the database. The similarity is determined by comparing the Euclidean distance between multiple feature vectors of damage mode, component type, environmental conditions and current structural state. The maintenance plans used in similar historical cases retrieved, the predicted effects verified by virtual simulation, and the long-term effects after actual implementation are comprehensively and weighted to generate a set of recommended maintenance plans for the current new situation and their expected effects analysis, which can be used as a reference for the whole life cycle maintenance decision optimization module.

Images