A Kriging-driven system and method for updating bridge digital twin modeling

Abstract

This invention belongs to the field of digital twin technology, specifically providing a bridge digital twin modeling and updating system and method driven by cluster learning and Kriging. The system includes: a data acquisition and processing module for obtaining performance characteristic information of the actual bridge structure; a digital twin model construction module for constructing a bridge digital twin model and processing the performance characteristic information of the simulated bridge structure; an objective function construction module for constructing a residual objective function and obtaining optimization variables and their parameter space based on the residual objective function; an optimization module for guiding the Kriging surrogate model to search for optimal parameters in the parameter space using a cluster optimization mechanism driven by multiple learning functions in parallel; and a model update module for updating the bridge digital twin model using the optimal parameters. This invention effectively improves the accuracy of bridge structure state identification and reduces model update costs, making it suitable for bridge performance prediction and operation and maintenance management under complex working conditions.

Description

Technical Field

[0001] This invention belongs to the field of digital twin technology, specifically relating to a bridge digital twin modeling and updating system and method driven by cluster learning Kriging. Background Technology

[0002] As long-term infrastructure, bridges experience gradual degradation in structural performance and increased uncertainty in their service status under the continuous effects of operational loads and environmental changes. This poses challenges to safety assessment and operation and maintenance management, making accurate simulation and evaluation of the actual condition of bridges increasingly important. Although high-fidelity numerical simulation has been widely used in bridge structural analysis, it is usually based on idealized design parameters and cannot accurately represent the true state of bridges during long-term service.

[0003] To address the aforementioned issues, digital twin technology constructs a virtual mapping model of the physical system and relies on twin updates to achieve accurate dynamic perception and evolution simulation of the structure. However, limited by the scale and structural complexity of bridges, the twin update process often involves large-scale numerical simulations and consumes massive computing resources. To improve update efficiency, existing methods often use surrogate models to replace the original simulation to reduce computational overhead; however, this strategy is highly dependent on the modeling accuracy of the surrogate model and may fail to converge to the true state, thus affecting the quality of twin modeling. Other researchers, based on cluster computing resources, distribute optimization tasks to multiple parallel computing nodes to directly solve the original model. This saves computing resources to some extent while ensuring the accuracy of the results, but the response speed still cannot meet the needs of high-frequency data fusion, and the update efficiency still needs further improvement.

[0004] Kriging, a classic Bayesian surrogate model, has recently been extended into an iterative optimization tool driven by active learning functions. This mechanism enhances the model's ability to identify and approximate key regions through guided sampling, and has achieved good results in some optimization tasks. However, existing methods mostly rely on a single learning function to sequentially perform sampling, generating only one sample point per iteration for surrogate model learning and updating, failing to leverage the potential of parallel learning and multi-point collaborative optimization. While some studies have attempted to introduce multi-objective optimization mechanisms, generating multiple candidate sample points in each iteration to support parallel sampling, these still essentially rely on objective functions constructed under the same learning strategy, failing to break through the single-strategy search paradigm. Furthermore, the sampling distribution remains biased, making it difficult to comprehensively cover complex design spaces, thus limiting its application potential in high-dimensional non-convex, multi-extremum structure optimization problems.

[0005] In practical applications involving the high-dimensional parameter space and complex response characteristics of bridge digital twin modeling and updating, existing optimization strategies still struggle to simultaneously ensure modeling accuracy and update response efficiency. Therefore, there is an urgent need to construct a cluster optimization method with a multi-dimensional parallel collaborative mechanism. This method overcomes the limitations of sampling guidance under a single strategy, collaboratively driving sample point expansion from multiple optimization perspectives. While ensuring global exploration capabilities, it also improves sampling diversity and model convergence efficiency, providing bridge digital twin systems with efficient and stable modeling and updating capabilities. Summary of the Invention

[0006] To address the problems of high-fidelity simulation-dependent updates, low sampling efficiency, and limited modeling accuracy in existing modeling processes, this invention introduces a cluster learning optimization mechanism. This mechanism combines Kriging surrogate model construction with an active learning sample guidance strategy to explore the parameter space in parallel from multiple learning perspectives. Unlike traditional optimization methods that rely solely on a single learning function for sampling, the proposed method integrates multiple learning functions. This multi-strategy collaborative approach enhances sample diversity and accelerates modeling convergence, thereby reducing simulation call frequency and improving the responsiveness of modeling updates.

[0007] A cluster learning Kriging-driven bridge digital twin modeling and update system includes:

[0008] The data acquisition and processing module is used to collect actual bridge structural response data under preset working conditions using sensors deployed on in-service bridges, and to process the data to obtain the performance characteristic information of the actual bridge structure.

[0009] The digital twin model construction module is used to construct a digital twin model of a bridge, simulate the bridge structure response data under the preset working conditions, and process the data to obtain the performance characteristic information of the simulated bridge structure.

[0010] The objective function construction module is used to construct a residual objective function using the performance characteristics information of the actual bridge structure and the performance characteristics information of the simulated bridge structure, and to obtain the optimization variables and the parameter space of the optimization variables based on the residual objective function;

[0011] The optimization module is used to guide the Kriging surrogate model to search for optimal parameters in the parameter space using a cluster optimization mechanism driven by multiple learning functions in parallel.

[0012] The model update module is used to update the bridge digital twin model with optimal parameters, thereby completing the construction and tracking of the bridge digital twin.

[0013] Preferably, the data acquisition and processing module uses sensors including a vertical laser displacement meter and a lateral vibration pickup, both of which are arranged at the mid-span of the main beam; wherein, the vertical laser displacement meter is used to measure the deflection response and assist in identifying the vertical vibration characteristics; the lateral vibration pickup is used to capture the lateral vibration characteristics of the bridge.

[0014] Preferably, in the objective function construction module, the process of constructing the residual objective function includes:

[0015] The actual bridge natural frequency is extracted from the actual bridge structural response data, and the simulated bridge natural frequency is obtained using the bridge digital twin model;

[0016] Based on the actual bridge natural frequency and the simulated bridge natural frequency, the residual objective function is constructed with the fundamental frequencies of the bridge's horizontal and vertical bends as the target frequencies.

[0017] Preferably, the optimization module includes:

[0018] An initial sample construction unit is used to select initial input sample points in the parameter space and calculate the response output corresponding to the initial input sample points to construct initial modeling samples;

[0019] The proxy model construction unit is used to construct an initial Kriging proxy model based on the initial modeling sample.

[0020] The first optimization unit is used to search for the optimal parameters in the parameter space and evaluate the convergence of the optimal parameters. If the preset convergence condition is met, the current search result is output as the global optimal solution and the iteration is terminated; otherwise, the second optimization unit is executed to enter the next round of optimization.

[0021] The second optimization unit is used to construct an active learning function cluster based on the residual objective function, and to perform sub-optimization processes on the active learning function cluster in parallel based on different computing nodes to obtain several optimal new input sample points.

[0022] The correlation evaluation unit is used to evaluate the spatial correlation between the optimal new input sample points and between the optimal new input sample points and the initial input sample points, and to delete redundant sample points whose spatial correlation does not meet a preset threshold.

[0023] The sample expansion unit is used to compute the response output corresponding to the optimal new input sample point after deleting redundant sample points in parallel on different computing nodes, obtain the newly added sample points, and expand them into the initial modeling sample.

[0024] The proxy model update unit is used to update the initial Kriging proxy model based on the expanded initial modeling samples, and return to the first optimization unit to continue iterative optimization until the preset convergence condition is met.

[0025] Preferably, in the initial sample construction unit, the Latin hypercube experimental design method is used to sample in the parameter space to obtain the initial input sample points.

[0026] Preferably, in the second optimization unit, the active learning function cluster is constructed using the maximum mean square error function, the expected improvement function, the minimum surrogate model prediction function, and the confidence lower bound function.

[0027] This invention also provides a cluster learning Kriging-driven bridge digital twin modeling and update method, which, when applied to the system, includes:

[0028] S1: Using sensors deployed on in-service bridges, collect actual bridge structural response data under preset working conditions, and process the data to obtain performance characteristic information of the actual bridge structure.

[0029] S2: Construct a digital twin model of the bridge, simulate the bridge structure response data under the preset working conditions, and process it to obtain the performance characteristic information of the simulated bridge structure;

[0030] S3: Using the performance characteristics of the actual bridge structure and the performance characteristics of the simulated bridge structure, construct a residual objective function, and obtain the optimization variables and the parameter space of the optimization variables based on the residual objective function;

[0031] S4: The Kriging surrogate model is guided to search for optimal parameters in the parameter space by using a cluster optimization mechanism driven by multiple learning functions in parallel.

[0032] S5: Update the bridge digital twin model with optimal parameters to complete the construction and tracking of the bridge digital twin.

[0033] Preferably, in step S4, the method for searching for the optimal parameters within the parameter space includes:

[0034] S41: Select initial input sample points in the parameter space and calculate the response output corresponding to the initial input sample points to construct initial modeling samples;

[0035] S42: Based on the initial modeling sample, construct the initial Kriging proxy model;

[0036] S43: Search for the optimal parameters in the parameter space and evaluate the convergence of the optimal parameters. If the preset convergence condition is met, the current search result is output as the global optimal solution and the iteration is terminated; otherwise, proceed to step S44 and enter the next round of optimization.

[0037] S44: Construct an active learning function cluster based on the residual objective function, and perform sub-optimization processes in parallel on the active learning function cluster based on different computing nodes to obtain several optimal new input sample points;

[0038] S45: Evaluate the spatial correlation between the optimal new input sample points and between the optimal new input sample points and the initial input sample points, and delete redundant sample points whose spatial correlation does not meet a preset threshold;

[0039] S46: Calculate the response output corresponding to the optimal new input sample point after deleting redundant sample points in parallel on different computing nodes to obtain the newly added sample points and expand them into the initial modeling sample;

[0040] S47: Update the initial Kriging proxy model based on the expanded initial modeling sample, and return to step S43 to continue iterative optimization until the preset convergence condition is met.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0042] This invention provides a cluster learning Kriging-driven bridge digital twin modeling and updating system and method, aiming to improve the accuracy of bridge structural state identification and the updating efficiency of digital twin models. The system constructs a corresponding virtual twin model based on the response data of actual bridge structures, and achieves dynamic calibration and continuous updating of the model through parameter optimization, thereby realizing accurate modeling and evolution tracking of bridge structural states.

[0043] The KCPL (Multi-Learning Function Parallel Driven Cluster Optimization Mechanism) algorithm proposed in this invention integrates multiple active learning functions, possesses multi-objective orientation and parallel sampling capabilities, significantly improves optimization accuracy and convergence efficiency, and is suitable for parameter optimization in high-dimensional complex problems. The digital twin update method built based on this algorithm can achieve efficient updating and accurate correction of virtual models with lower computational overhead, breaking through the limitation of traditional methods that are difficult to balance response efficiency and modeling quality, improving the real-time performance of bridge structure state identification and parameter identification accuracy, and has broad application scenarios.

[0044] This invention not only improves the mapping accuracy of bridge digital twin models to the actual structural state, but also significantly enhances their performance in high-dimensional parameter identification, nonlinear response modeling, and real-time updates. The constructed system possesses good scalability and engineering applicability, and is particularly suitable for bridge digital twin application scenarios where the dynamic evolution of structural service status is significant, monitoring frequency is high, and update requirements are stringent. Attached Figure Description

[0045] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a schematic diagram of the structure of the bridge digital twin modeling and update system driven by Kriging cluster learning, according to an embodiment of the present invention.

[0047] Figure 2 This is a flowchart illustrating the framework of the cluster optimization mechanism (KCPL) driven by multiple learning functions according to an embodiment of the present invention.

[0048] Figure 3 This is a schematic diagram of the sensor arrangement in the data acquisition and processing module of an embodiment of the present invention;

[0049] Figure 4 The following are spectrum diagrams reflecting the natural vibration characteristics of the bridge structure according to embodiments of the present invention; wherein, (a) is the spectrum diagram of the transverse fundamental frequency; and (b) is the spectrum diagram of the vertical fundamental frequency.

[0050] Figure 5 This is a schematic diagram of a digital twin model of a bridge structure according to an embodiment of the present invention;

[0051] Figure 6 The following are comparison charts of prediction results from different models and measured data in embodiments of the present invention; wherein, (a) is a comparison chart of response prediction at a vehicle speed of 40km / h; (b) is a comparison chart of response prediction at a vehicle speed of 60km / h; (c) is a comparison chart of response prediction at a vehicle speed of 70km / h; and (d) is a comparison chart of response maximum value prediction. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] Example 1:

[0055] like Figure 1 As shown, a bridge digital twin modeling and updating system driven by cluster learning Kriging includes: a data acquisition and processing module, a digital twin model construction module, an objective function construction module, an optimization module, and a model updating module.

[0056] The data acquisition and processing module is used to collect actual bridge structural response data under preset working conditions using sensors deployed on in-service bridges, and to process the data to obtain performance characteristic information of the actual bridge structure. In a further implementation, the sensors used in the data acquisition and processing module include vertical laser displacement gauges and lateral vibration pickups, both deployed at the mid-span of the main beam. The vertical laser displacement gauges are used to measure deflection response and assist in identifying vertical vibration characteristics; the lateral vibration pickups are used to capture the lateral vibration characteristics of the bridge. Figure 3 As shown.

[0057] The data acquisition and processing module also includes identifying key indicators reflecting structural performance characteristics by processing actual bridge structural response data. These key indicators are the bridge's natural frequencies. For example... Figure 4 As shown, where, Figure 4 (a) is the spectrum of the transverse fundamental frequency. Figure 4 (b) is the spectrum of the vertical fundamental frequency. By performing spectral analysis on the structural residual vibration response after environmental excitation and train passage, the main peak value in its spectral density curve was extracted to identify the dominant vibration frequency of the structure. Among them, the main peak corresponding to the vertical bending of the U-beam is located at 4.22Hz, and the main peak corresponding to the transverse bending of the pier-beam system is located at 1.94Hz. It can be seen that the actual fundamental frequencies of the vertical bending of the bridge U-beam and the transverse bending of the pier-beam system are 4.22Hz and 1.94Hz, respectively.

[0058] A digital twin model construction module is used to build a digital twin model of the bridge, simulate the bridge structural response data under the preset working conditions, and process the data to obtain the performance characteristics of the simulated bridge structure. Specifically, the digital twin model uses four-node solid elements (SOLID45) to model the U-beam and piers based on the design documents, and simultaneously considers the simulation of detailed structures such as the rail abutment and the thickened section of the support. The secondary dead load is applied through mass elements (MASS21). The material parameters used in the modeling are equivalent elastic modulus and mass density. To improve calculation accuracy and reduce the proportion of ill-conditioned elements, the main beam and piers are divided into appropriately sized hexahedral elements using a volume sweep method. Regarding boundary conditions, the pier-beam connection uses constraint coupling to simulate the simply supported constraint effect of the pier on the main beam, while the pier base uses consolidation constraints to reflect the supporting effect of the foundation on the pier. Based on this, the digital twin model is used to calculate and simulate the bridge structural response and structural performance characteristics under the preset working conditions.

[0059] The objective function construction module is used to construct a residual objective function using the performance characteristics of the actual bridge structure and the simulated bridge structure, and to obtain optimization variables and their parameter space based on the residual objective function. A further implementation involves the following steps in constructing the residual objective function: extracting the actual bridge's natural frequencies from the actual bridge structure's response data, and obtaining the simulated bridge's natural frequencies using a bridge digital twin model. Based on the actual and simulated bridge natural frequencies, and using the fundamental frequencies of the bridge's horizontal and vertical bends as target frequencies, a residual objective function is constructed to describe the deviation between the two in their natural vibration characteristics. The residual objective function takes the following form:

[0060]

[0061] In the formula, η is the parameter to be updated; f i (η) and f i,mea These represent the simulated frequency and the corresponding measured result, respectively; n is the number of target frequencies, and in this embodiment, n = 2.

[0062] Taking into account the sensitivity and variability of various structural parameters to the dynamic characteristics of the bridge, the optimization variables and corresponding parameter spaces (i.e., parameter variable design space) in the objective function are determined. Among them, the elastic modulus parameters are allowed to fluctuate by ±30% relative to their design values, while the mass density parameters are allowed to vary by ±5%, as shown in Table 1.

[0063] Table 1

[0064]

[0065] The optimization module guides the Kriging surrogate model to search for optimal parameters in the parameter space using a cluster optimization mechanism driven by multiple learning functions in parallel. For example... Figure 2As shown.

[0066] A further implementation method is that the optimization module includes:

[0067] An initial sample construction unit is used to select initial input sample points in the parameter space and calculate the corresponding response outputs to construct initial modeling samples. A further implementation involves using a Latin hypercube experimental design method to sample in the parameter space to obtain initial input sample points. Specifically, the Latin hypercube experimental design method divides each parameter dimension into 50 equidistant intervals, and random sampling is performed within each interval. This sampling is then randomly recombined in the multidimensional space to generate 50 sets of well-covered initial sample points, ensuring uniform exploration of the parameter space. Based on the established bridge digital twin model, the response outputs corresponding to the input sample points are numerically calculated to form the initial modeling samples. During the numerical simulation, the computational tasks are distributed across four different computational cores to accelerate the response output.

[0068] The surrogate model building unit is used to build an initial Kriging surrogate model based on the initial modeling samples.

[0069] The first optimization unit searches for the optimal parameters in the parameter space and evaluates their convergence. If the preset convergence condition is met, the current search result is output as the global optimal solution, and the iteration terminates; otherwise, the second optimization unit is executed, and the next round of optimization begins. Specifically, a particle swarm optimization algorithm is used. By initializing the particle swarm, the particle positions are updated in each iteration based on the individual historical best value and the swarm's best value to search for the optimal parameter value and perform convergence evaluation. The convergence criteria include the following three conditions, and termination occurs if any one of them is met: the objective function value is lower than 1e-4, the number of iterations reaches 100, or the objective function has continuously improved by less than 1e-6 for 30 iterations. If the evaluation result meets the set convergence condition, the current result can be output as the global optimal solution, and the iteration terminates; otherwise, the second optimization unit is executed, and the next round of optimization begins.

[0070] The second optimization unit is used to construct an active learning function cluster based on the residual objective function, and to perform sub-optimization processes on the active learning function cluster in parallel based on different computing nodes to obtain several optimal new input sample points. The active learning function cluster can initially obtain sample points for expansion, and the spatial correlation analysis screens the sample points expanded in the previous step by evaluating the correlation between samples and deleting redundant samples with excessive correlation.

[0071] A further implementation involves constructing an active learning function cluster (i.e., the multiple learning functions mentioned earlier) in the second optimization unit using the maximum mean square error (MSE), expected improvement function (EI), minimum surrogate model prediction function (MSP), and confidence lower bound function (LCB). Sub-optimization processes are then executed in parallel on different computing nodes to obtain multiple optimal new input sample points. The functions and sampling criteria are as follows:

[0072] The expression for MSE is MSE(x) = σ(x), the feature description is the model mean squared error estimate, and the sampling criterion is x. new =argmax[MSE(x)].

[0073] The expression for EI is The feature description is the balance trend between the current optimal value (minimum) and the global search, and the sampling criterion is x. new =argmax[EI(x)].

[0074] The expression for MSP is The feature description is the model's predicted mean, and the sampling criterion is x. new =argmin[MSP(x)].

[0075] The expression for LCB is The feature description is based on the uncertainty penalty of the model prediction, and the sampling criterion is x. new =argmin[LCB(x)].

[0076] Note: y min Φ(·) represents the current optimal value (minimum value) of the model; Φ(·) and φ(·) are the standard normal cumulative distribution function and probability density function, respectively.

[0077] The correlation evaluation unit assesses the spatial correlation between optimal new input sample points and between the optimal new input sample point and the initial input sample points, and removes redundant sample points whose spatial correlation does not meet a preset threshold. When the correlation is too high, the spatial positions of two input points are very close or even coincident (when the spatial correlation equals 1, the two points completely coincide). Since the two points are very close, their contributions to subsequent optimization are almost identical. Therefore, for several points with very close spatial positions, retaining only one point can achieve almost the same optimization effect as not deleting any points. Thus, when the spatial correlation between different input samples is too high, it can be considered that there are unnecessary redundant points to retain, and deleting them can save subsequent optimization costs. Furthermore, Kriging modeling does not allow two completely identical input samples to exist (spatial correlation equals 1).

[0078] Specifically, the spatial correlation between new input sample points and between new input sample points and initial input sample points is evaluated based on the Gaussian function, and redundant samples with a correlation greater than 0.95 are deleted. The expression of the Gaussian function is shown in equation (2):

[0079]

[0080] In the formula, θ l The relevant function parameters can be solved using maximum likelihood estimation; For different sample points x i x j Components in the l-dimensional dimension.

[0081] The sample expansion unit is used to compute the response output corresponding to the optimal new input sample point after deleting redundant sample points in parallel on different computing nodes, obtain the new sample points, and expand them into the initial modeling sample.

[0082] The surrogate model update unit is used to update the initial Kriging surrogate model based on the expanded initial modeling samples, and then return to the first optimization unit to continue iterative optimization until the preset convergence condition is met.

[0083] To deeply analyze the accuracy and efficiency of the proposed KCPL (Multi-Learning Function Parallel Driven Cluster Optimization Mechanism) algorithm in actual bridge twin updates, the classic standard Kriging and PSO algorithms are introduced for comparison, with the relevant optimization parameter settings being the same as the KCPL algorithm. Table 2 shows the structural transverse and vertical bending vibration frequencies before and after each method update, as well as their errors relative to the measured values. Table 3 summarizes the core computational indicators such as the objective function value, number of iterations, number of modal analyses, and optimization time when each algorithm completes the update, used to quantitatively analyze the performance of the proposed methods.

[0084] Table 2

[0085]

[0086] Table 3

[0087] Calculation indicators KCPL Kriging PSO objective function value 7.90e-5 6.31e-4 9.61e-5 Total running time (s) 551.48 273.13 1312.37 Total number of modal analyses 85 79 250

[0088] It can be seen that although the standard Kriging method has the highest computational efficiency, its lack of an active optimization mechanism makes it difficult to fine-tune key parameters, resulting in a significantly higher frequency fitting error than KCPL and PSO, thus failing to meet the set convergence criterion. In contrast, the PSO method has higher fitting accuracy, but it is highly dependent on large-scale numerical simulations, resulting in significant computational overhead. KCPL, on the other hand, maintains fitting accuracy comparable to or even better than PSO while significantly reducing computational costs, fully demonstrating its excellent ability to coordinate the contradiction between "optimization accuracy and computational cost" in bridge digital twin modeling and updating.

[0089] The model update module is used to update the bridge's digital twin model with optimal parameters, completing the construction and tracking of the bridge's digital twin. For example... Figure 5 As shown. Specifically, after the digital twin is updated, it is applied to the simulation and prediction of the vertical deflection of the bottom slab at the mid-span of the bridge when the train passes through at different speeds and the structural response under the condition of the train being stationary under loading.

[0090] Figure 6 The prediction results of the digital twin and the original design model under different operating conditions are shown, along with their comparison with measured data. Figure 6 (a) Comparison chart of response prediction at a vehicle speed of 40km / h; Figure 6 (b) is a comparison chart of response predictions at a vehicle speed of 60km / h; Figure 6 (c) is a comparison chart of response predictions at a vehicle speed of 70km / h; Figure 6 (d) is a comparison chart of predicted maximum response values. For example... Figure 6 While both digital twins and virtual models based on design documents possess reasonable mechanical expression capabilities, the original design models failed to reflect the actual service state and performance changes of the structure, leading to a general overestimation of the bridge's true deformation under train loads. In contrast, the prediction results of the digital twin model built based on the technical solution of this invention highly match the measured data, more accurately capturing response trends, and significantly improving the peak fitting accuracy, demonstrating strong state mapping and prediction capabilities. This result further verifies the accuracy of the technical solution of this invention in identifying the current state of the structure and demonstrates its practicality and potential for widespread application in bridge structures under complex operating conditions.

[0091] Example 2:

[0092] This invention also provides a cluster learning Kriging-driven bridge digital twin modeling and update method, which, when applied to the system, includes:

[0093] S1: Using sensors deployed on in-service bridges, collect actual bridge structural response data under preset working conditions, and process the data to obtain performance characteristic information of the actual bridge structure.

[0094] S2: Construct a digital twin model of the bridge, simulate the bridge structure response data under the preset working conditions, and process it to obtain the performance characteristic information of the simulated bridge structure.

[0095] S3: Using the performance characteristics of the actual bridge structure and the performance characteristics of the simulated bridge structure, construct a residual objective function, and obtain the optimization variables and the parameter space of the optimization variables based on the residual objective function.

[0096] S4: The Kriging surrogate model is guided to search for optimal parameters in the parameter space by a cluster optimization mechanism driven by multiple learning functions in parallel.

[0097] S5: Update the bridge digital twin model with optimal parameters to complete the construction and tracking of the bridge digital twin.

[0098] A further implementation method is that, in step S4, the method for searching for optimal parameters in the parameter space includes:

[0099] S41: Select initial input sample points in the parameter space and calculate the response output corresponding to the initial input sample points to construct the initial modeling samples.

[0100] S42: Construct the initial Kriging proxy model based on the initial modeling sample.

[0101] S43: Search for the optimal parameters in the parameter space and evaluate the convergence of the optimal parameters. If the preset convergence condition is met, the current search result is output as the global optimal solution and the iteration is terminated; otherwise, proceed to step S44 and enter the next round of optimization.

[0102] S44: Construct an active learning function cluster based on the residual objective function, and perform sub-optimization processes in parallel on the active learning function cluster based on different computing nodes to obtain several optimal new input sample points.

[0103] S45: Evaluate the spatial correlation between the optimal new input sample points and between the optimal new input sample points and the initial input sample points, and delete redundant sample points whose spatial correlation does not meet the preset threshold.

[0104] S46: Calculate the response output corresponding to the optimal new input sample point after deleting redundant sample points in parallel on different computing nodes, obtain the newly added sample points, and expand them into the initial modeling sample.

[0105] S47: Update the initial Kriging proxy model based on the expanded initial modeling samples, and return to step S43 to continue iterative optimization until the preset convergence condition is met.

[0106] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A bridge digital twin modeling and update system driven by cluster learning Kriging, characterized in that, include: The data acquisition and processing module is used to collect actual bridge structural response data under preset working conditions using sensors deployed on in-service bridges, and to process the data to obtain performance characteristic information of the actual bridge structure. The digital twin model construction module is used to construct a digital twin model of a bridge, simulate the bridge structure response data under the preset working conditions, and process the data to obtain the performance characteristic information of the simulated bridge structure. The objective function construction module is used to construct a residual objective function using the performance characteristics information of the actual bridge structure and the performance characteristics information of the simulated bridge structure, and to obtain the optimization variables and the parameter space of the optimization variables based on the residual objective function; The optimization module is used to guide the Kriging surrogate model to search for optimal parameters in the parameter space using a cluster optimization mechanism driven by multiple learning functions in parallel. The model update module is used to update the bridge digital twin model with optimal parameters to complete the construction and tracking of the bridge digital twin; The process of constructing the residual objective function in the objective function construction module includes: The actual bridge natural frequency is extracted from the actual bridge structural response data, and the simulated bridge natural frequency is obtained using the bridge digital twin model; Based on the actual natural frequency of the bridge and the simulated natural frequency of the bridge, the residual objective function is constructed with the fundamental frequencies of the bridge's horizontal and vertical bends as the target frequencies. The residual objective function takes the following form: ， In the formula, Parameters to be updated; and These are the simulated frequencies and their corresponding measured results; The target frequency number; The optimization module includes: An initial sample construction unit is used to select initial input sample points in the parameter space and calculate the response output corresponding to the initial input sample points to construct initial modeling samples; The proxy model construction unit is used to construct an initial Kriging proxy model based on the initial modeling sample. The first optimization unit is used to search for the optimal parameters in the parameter space and evaluate the convergence of the optimal parameters. If the preset convergence condition is met, the current search result is output as the global optimal solution and the iteration is terminated; otherwise, the second optimization unit is executed to enter the next round of optimization. The second optimization unit is used to construct an active learning function cluster based on the residual objective function, and to perform sub-optimization processes on the active learning function cluster in parallel on different computing nodes to obtain several optimal new input sample points. In the second optimization unit, the active learning function cluster (i.e., multiple learning functions) is constructed using the maximum mean square error function (MSE), the expected improvement function (EI), the minimum surrogate model prediction function (MSP), and the confidence lower bound function (LCB). Sub-optimization processes are then performed on the learning function cluster in parallel on different computing nodes to obtain multiple optimal new input sample points. The functions and sampling criteria are as follows: The expression for MSE is The feature description is the model mean squared error estimate, and the sampling criterion is... ; The expression for EI is The feature description is the balance trend between the current optimal value and the global search, and the sampling criterion is... ; The expression for MSP is The feature description is the model's predicted mean, and the sampling criterion is... ; The expression for LCB is The feature description is based on the uncertainty penalty of the model prediction, and the sampling criterion is... ; This is the optimal value for the current model; and These are the standard normal cumulative distribution function and the probability density function, respectively. The correlation evaluation unit is used to evaluate the spatial correlation between the optimal new input sample points and between the optimal new input sample points and the initial input sample points, and to delete redundant sample points whose spatial correlation does not meet a preset threshold. The sample expansion unit is used to compute the response output corresponding to the optimal new input sample point after deleting redundant sample points in parallel on different computing nodes, obtain the newly added sample points, and expand them into the initial modeling sample. The proxy model update unit is used to update the initial Kriging proxy model based on the expanded initial modeling samples, and return to the first optimization unit to continue iterative optimization until the preset convergence condition is met.

2. The system according to claim 1, characterized in that, The data acquisition and processing module uses sensors including a vertical laser displacement meter and a lateral vibration pickup, both of which are deployed at the mid-span of the main beam. The vertical laser displacement meter is used to measure the deflection response and assist in identifying vertical vibration characteristics. The lateral vibration pickup is used to capture the lateral vibration characteristics of the bridge.

3. The system according to claim 1, characterized in that, In the initial sample construction unit, the Latin hypercube experimental design method is used to sample in the parameter space to obtain the initial input sample points.

4. The system according to claim 1, characterized in that, In the second optimization unit, the active learning function cluster is constructed using the maximum mean square error function, the expected improvement function, the minimum surrogate model prediction function, and the confidence lower bound function.

5. A cluster learning Kriging-driven bridge digital twin modeling and update method, applying the system described in any one of claims 1-4, characterized in that, include: S1: Using sensors deployed on in-service bridges, collect actual bridge structural response data under preset working conditions, and process the data to obtain performance characteristic information of the actual bridge structure. S2: Construct a digital twin model of the bridge, simulate the bridge structure response data under the preset working conditions, and process it to obtain the performance characteristic information of the simulated bridge structure; S3: Using the performance characteristics of the actual bridge structure and the performance characteristics of the simulated bridge structure, construct a residual objective function, and obtain the optimization variables and the parameter space of the optimization variables based on the residual objective function; S4: The Kriging surrogate model is guided to search for optimal parameters in the parameter space by using a cluster optimization mechanism driven by multiple learning functions in parallel. S5: Update the bridge digital twin model with optimal parameters to complete the construction and tracking of the bridge digital twin.

6. The method according to claim 5, characterized in that, In step S4, the method for searching for optimal parameters within the parameter space includes: S41: Select initial input sample points in the parameter space and calculate the response output corresponding to the initial input sample points to construct initial modeling samples; S42: Based on the initial modeling sample, construct the initial Kriging proxy model; S43: Search for the optimal parameters in the parameter space and evaluate the convergence of the optimal parameters. If the preset convergence condition is met, the current search result is output as the global optimal solution and the iteration is terminated; otherwise, proceed to step S44 and enter the next round of optimization. S44: Construct an active learning function cluster based on the residual objective function, and perform sub-optimization processes in parallel on the active learning function cluster based on different computing nodes to obtain several optimal new input sample points; S45: Evaluate the spatial correlation between the optimal new input sample points and between the optimal new input sample points and the initial input sample points, and delete redundant sample points whose spatial correlation does not meet a preset threshold; S46: Calculate the response output corresponding to the optimal new input sample point after deleting redundant sample points in parallel on different computing nodes to obtain the newly added sample points and expand them into the initial modeling sample; S47: Update the initial Kriging proxy model based on the expanded initial modeling sample, and return to step S43 to continue iterative optimization until the preset convergence condition is met.

Images