Fusion method of measured data and mechanical simulation data

Abstract

The invention relates to a method for fusing measured data and mechanical simulation data. The method comprises the steps of obtaining actual site information and measured data, extracting measured response, calling a large language model to generate a to-be-optimized large-parameter matrix, constructing a finite element simulation model and solving the finite element simulation model, and generating simulation response. Calculating a response difference value between the simulation response and the actual measurement response, and optimizing a parameter value of the large-parameter matrix; judging whether the iteration is terminated or not, if so, judging whether a fusion result output by the current finite element simulation model accords with an actual condition or not, otherwise, updating the finite element simulation model, and carrying out next iteration; and if a fusion result accords with an actual condition, applying an equivalent load to the finite element simulation model obtained by iteration to obtain fusion data for performance evolution analysis of the pavement structure, otherwise, modifying selection of parameters to be optimized, regenerating the finite element simulation model, updating a large parameter matrix, and performing iteration again. Compared with the prior art, the method has the advantages of fully improving the data quality and the like.

Description

Technical Field

[0001] This invention relates to the field of performance evolution analysis of airport pavement structures, and in particular to a method for fusing measured data and mechanical simulation data. Background Technology

[0002] The study, analysis, and prediction of the evolution of airport runway structural performance are crucial foundations for airport runway structure, materials science design, long-term durable operation of pavement structures, and flight area operational safety. Currently, the analysis of pavement structural performance utilizes the following methods: 1. Constructing a pavement structure model using mechanical principles and analyzing the evolution of structural performance by obtaining explicit expressions for different parameters. This approach is based on the laws of mechanics, is theoretically sound, highly interpretable, and less susceptible to randomness. However, due to the complexity of pavement structures, constitutive models cannot always be constructed directly; therefore, this method often relies on numerical simulation models, such as the finite element method, to build the model.

[0003] 2. Utilizing massive amounts of measured data, statistical methods are used to analyze the changing patterns of structural performance and predict the future values ​​of relevant indicators, thereby inferring the structural state of the pavement. This method relies on statistical patterns derived from large amounts of data to reflect the actual conditions of complex systems. The results can also reflect the impact of material heterogeneity, construction quality, and defects, which cannot be simulated in traditional models. Related techniques include linear regression analysis, time series analysis, and stochastic process analysis. With the continuous development of artificial intelligence, various supervised and unsupervised learning methods are also being applied to pavement structural performance analysis.

[0004] However, the above methods also have their drawbacks. For example, during the construction of the simulation model, some assumptions made by humans may be incorrect. In order to improve the efficiency of the solution, the treatment of some mechanical relationships may be simplified, and many situations in real-world scenarios cannot be considered. These factors will affect the results. During the data analysis process, it is easily affected by data noise, abnormal measurement sensors, etc., which will affect the data analysis results.

[0005] Mechanical simulation and data analysis each have their advantages and disadvantages. Combining data analysis with mechanical simulation, and using measured data and structural simulation models to obtain higher quality fused data, can provide a more scientific and reasonable basis for subsequent structural performance evolution analysis, and form more reliable and highly interpretable analysis results.

[0006] Chinese patent CN114117840A discloses a structural performance prediction method based on a hybrid simulation and experimental data-driven approach. In static tests, it combines limited measurement and simulation data to establish a fusion model. A large amount of simulation data and a small amount of experimental data are used to improve the accuracy of the surrogate model constructed from the hybrid data. This method improves model fidelity and reconstructs the structural state field. By utilizing a multi-precision deep neural network model, it adaptively learns the linear and nonlinear relationships between experimental and simulation data, achieving spatial health prediction. In health monitoring tests, optimization algorithms are used to correct the fusion model. The constructed fusion model incorporates health monitoring data and can reflect the true physical state of the structure, achieving temporal health prediction. However, this method requires modeling and solving large-scale structures when generating simulation data, which is time-consuming and has a high barrier to entry, making it unsuitable for practical structural monitoring. Furthermore, the search method used is not conducive to large-scale iterations, leading to difficulties in obtaining high-quality solutions. In addition, the method's quantification of response differences is difficult to comprehensively evaluate structural response differences, and the optimization results cannot reflect the true physical state of the structure.

[0007] Therefore, there is a need for a simulation and experimental data fusion method that is computationally efficient, can effectively obtain high-quality solutions, and whose optimization results can truly reflect the physical state, in order to improve the possibility of practical applications. Summary of the Invention

[0008] The purpose of this invention is to provide a method for fusing measured data and mechanical simulation data, thereby integrating measured data with simulation data to establish a large-scale simulation model with mechanical knowledge supervision and measured data correction, which can significantly improve data quality and provide a foundation for the study of pavement structure performance evolution under the fusion of mechanical simulation and data analysis.

[0009] The objective of this invention can be achieved through the following technical solutions: A method for fusing measured data and mechanical simulation data, wherein the fused data is used for pavement structure performance evolution analysis, the method includes the following steps: S1, obtain actual site information and measured data; S2, Based on the measured data, extract the measured response using cross-power spectrum processing; S3. Based on the actual site information, call the large language model to generate a large parameter matrix to be optimized, and construct a finite element simulation model. S4, Solve the finite element simulation model to generate the simulation response; S5, calculate the response difference between the simulated response and the measured response; S6, Optimize the parameter values ​​of the large parameter matrix based on the response difference; S7, determine whether the preset number of iterations or the response difference has reached the preset threshold. If yes, determine whether the fusion result output by the current finite element simulation model matches the actual situation. Otherwise, return to step S3 to update the finite element simulation model based on the parameter values ​​of the optimized large parameter matrix and proceed to the next iteration. S8. If the fusion result in step S7 matches the actual situation, the finite element simulation model obtained by iteration is used as the fusion model. An equivalent load is applied to the fusion model to obtain fusion data. Otherwise, the parameter selection to be optimized is modified, and the process returns to step S3 to regenerate the finite element simulation model and update the large parameter matrix, and iterates again.

[0010] The specific steps for constructing the finite element simulation model are as follows: Based on computing power, basic site conditions, and target design parameter set; Natural language description is performed based on actual site information to obtain descriptive text; The parameter set and descriptive text are input into a large language model, which then performs the following steps to generate a simulation model: Automatically locate the physical quantities of the model in the descriptive text and extract the parameters; Identify the necessary parameters not mentioned in the description text, and perform reasoning to complete the necessary parameters or call the relevant knowledge base to assign them default values; The extracted parameters and the required parameters for completion are uniformly converted, and their rationality is verified. Based on the verified parameters, a large parameter matrix to be optimized is set, and a finite element simulation model is established by combining the finite element modeling method of the track panel, and working condition definition statements are generated.

[0011] The large parameter matrix to be optimized includes material parameters, geometric parameters, and boundary condition parameters. The material parameters include elastic modulus, Poisson's ratio, and density. The geometric parameters include cross-sectional width, thickness, and length. The boundary condition parameters include foundation support stiffness and joint spring stiffness.

[0012] In constructing the finite element simulation model, the structure was divided into multiple sub-regions based on mechanical principles and the installation of sensing optical fibers. Parameters were set for each region. The structure was divided into two layers: horizontal and vertical. The horizontal layer represents the horizontal direction of the structure, and the vertical layer represents the depth direction. In the vertical layer, the thickness of the layers was adjusted to ensure that the sensing optical fibers passed evenly through the parameter adjustment area. In the horizontal layer, the partitions were densified within a preset range near the load-bearing area. Simultaneously, measurement lines were set in the finite element model according to the location of the on-site optical fiber sensing devices for subsequent data extraction.

[0013] In step S5, before calculating the response difference, the simulation response is aligned using the linear interpolation method of simulation data. Specifically: For the simulation response time series { , }, to the time point set { Data alignment, where, For any time point in the simulation time series, For time points The corresponding simulated response value, For any point in the set of time points of the measured data; Find the corresponding interval index such that: ; By using linear interpolation, the adjustment parameter λ is calculated, and then the interpolated result is calculated using the adjustment parameter λ to achieve data alignment. , , in, This is the interpolated response value of the simulation data at the corresponding time point of the measured data after interpolation.

[0014] The method for calculating the response difference is as follows: using the mode shape as a physical quantity characterizing the response, and quantifying the response difference based on the modal confidence criterion. , , in, Indicates the response difference. For the first The weight of the MAC value corresponding to each test line. Indicates the first The test line is based on the modal shape of the simulated response. a and measured response mode shape b The calculated MAC value, They represent the first Modal vibration modes of the strip measurement line a and b of k A vector of order 1, where * denotes the conjugate transpose. n For order, L This represents the number of survey lines.

[0015] The method for calculating the response difference is as follows: using the mode shape and frequency difference as physical quantities characterizing the response, and after quantizing the mode shape based on the modal confidence criterion, the response difference is obtained by weighting. , , , , in, The response difference is represented by the MAC value and the frequency difference. constitute, , The first The weights of the MAC value and frequency difference corresponding to each test line. L This refers to the number of survey lines; Indicates the first The test line is based on the modal shape of the simulated response. a and measured response mode shape b The calculated MAC value, They represent the first Modal vibration modes of the strip measurement line a and b of k A vector of order 1, where * denotes the conjugate transpose. n It is the order; For the first Frequency difference of the test lines For the first strip survey line k The difference in natural frequencies of the first order. For the first Simulation response of the test line k First natural frequency, For the first Measured response of the test line k The first natural frequency.

[0016] The parameter values ​​for optimizing the large parameter matrix based on the response difference are specifically as follows: a search algorithm is used, with the reciprocal of the response difference as the optimization objective function, the parameters to be optimized are set as a population, and the optimization process of the parameter value range is transformed into a process of reducing the initial values ​​of the parameters to different degrees based on the reduction factor, and the reduction factor is optimized.

[0017] The parameter values ​​for optimizing the large parameter matrix based on the response difference are specifically determined as follows: A reinforcement learning algorithm is used to process the input state of the agent, which is based on an actor network. ,in, In response to the amount of change in the MAC value within the difference, To respond to the frequency difference in the variation, a corresponding action vector is generated. ,in, The number of parameter types to be estimated. To define the spatial region dimension, the output action vector is split into matrices corresponding to each parameter, and the output values ​​are transformed to obtain specific physical quantity values. The reward function consists of two parts: a differential quantification index and a step size penalty, expressed as: , , in, This is a reward function for quantifying differences, where α is a user-defined coefficient. For the search step size, This is the overall reward function.

[0018] The determination of whether the fusion result output by the current finite element simulation model matches the actual situation specifically includes parameter mutation checking and response mutation checking, wherein, The parameter mutation check is specifically as follows: set a threshold δ, calculate the absolute value of the difference between each parameter and the parameters of its adjacent upper, lower, left and right sub-regions in space, and find the maximum value Δ. If Δ≤δ, it is determined that there is no mutation; otherwise, the parameter mutation check fails. The response mutation check specifically involves: performing a continuity check on the response data, based on... Remove the parabolic trend from the response data, where, For residuals used in continuity checks, In response to the data, Based on location points i Fitted parabolic data; for residuals Use CUSUM detection to check for mutation points. If both the upper and lower CUSUM exceed the threshold, it is determined that a mutation has occurred, and the mutation check fails. If either the parameter mutation check or the response mutation check fails, the fusion result output by the current finite element simulation model is determined to be inconsistent with the actual situation.

[0019] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention introduces an integrated platform based on large language model interaction. The large language model trained on the structural simulation calculation script can quickly respond to the user's natural language commands, complete the generation of finite element calculation scripts, and integrate functions such as finite element calculation software calling and optimal parameter search, reducing the complexity of the fusion process and making it easy to deploy and use quickly on the engineering site.

[0020] 2. This invention introduces cross-power spectrum analysis to process the acquired data, extracts mode shapes and natural frequencies to construct a response difference function, and introduces MAC (Modal Confidence Criterion) to quantify the mode shape differences, thereby accurately capturing the differences in structural vibration response. Introducing this difference calculation method into the reward function of reinforcement learning ensures that the results closely approximate the actual structural condition.

[0021] 3. When using a search algorithm to solve the problem, this invention uses a search scale factor instead of direct search in terms of search strategy, which makes the numerical scale more user-friendly, avoids the optimizer from optimizing for different dimensions / orders of magnitude, and naturally ensures that the parameters meet the boundary conditions.

[0022] 4. This invention uses mechanical constraints as its core and measured data as its basis to ensure that the fusion results simultaneously conform to the actual site structure and basic mechanical laws. Through dual control of "parameter mutation check" and "response mutation check", anomalies in model settings and numerical outputs are identified and located synchronously, ensuring the credibility and robustness of the simulation-data fusion results. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention; Figure 2 A flowchart for constructing a finite element simulation model; Figure 3 This is a schematic diagram of parameter categories and region division during the construction of a finite element simulation model in one embodiment; Figure 4 This is a schematic diagram of the DDPG network operation process; Figure 5 This is a schematic diagram of the solution loop using the search algorithm; Figure 6 A flowchart is generated to integrate the data; Figure 7 This is a schematic diagram of the specimen beam in Example 2; Figure 8 This is a schematic diagram of the measured response of layer 1 extracted using cross-power spectrum in Example 2; Figure 9 This is a schematic diagram of the measured response of layer 2 extracted using cross-power spectrum in Example 2; Figure 10 This is a schematic diagram of the measured response of layer 3 extracted using cross-power spectrum in Example 2; Figure 11 A schematic diagram of the mode shape curves of layer 1 generated using the initial values ​​in Example 2; Figure 12 A schematic diagram of the mode shape curves of layer two generated using the initial values ​​in Example 2; Figure 13A schematic diagram of the mode shape curves of layer 3 generated using the initial values ​​in Example 2; Figure 14 This is a flowchart of the response difference calculation in Example 2; Figure 15 This is a schematic diagram of the optimal parameters determined in Example 2; Figure 16 This is a schematic diagram of the simulation model with assigned optimization parameter values ​​in Example 2; Figure 17 This is a comparison chart of the fused data and the measured data in Example 2. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0025] Example 1 This embodiment provides a method for fusing measured data and mechanical simulation data. The fused data is used for pavement structure performance evolution analysis, such as... Figure 1 As shown, the fusion method includes the following steps: S1, obtain actual site information and measured data; S2, based on measured data, extracts the measured response using cross-power spectrum processing; S3, based on actual site information, calls a large language model to generate a large parameter matrix to be optimized, and constructs a finite element simulation model; S4, Solve the finite element simulation model to generate the simulation response; S5, calculate the response difference between the simulated response and the measured response; S6, Optimize the parameter values ​​of a large parameter matrix based on response difference; S7, determine whether the preset number of iterations or the response difference has reached the preset threshold. If yes, determine whether the fusion result output by the current finite element simulation model matches the actual situation. Otherwise, return to step S3 to update the finite element simulation model based on the parameter values ​​of the optimized large parameter matrix and proceed to the next iteration. S8. If the fusion result in step S7 matches the actual situation, the finite element simulation model obtained by iteration is used as the fusion model. An equivalent load is applied to the fusion model to obtain fusion data. Otherwise, the parameter selection to be optimized is modified, and the process returns to step S3 to regenerate the finite element simulation model and update the large parameter matrix, and iterates again.

[0026] In step S3, the large language model is used to generate simulation files according to the specified prompt word template. The prompt word template generates simulation files through multi-stage prompt chains, intermediate workpiece constraints and automatic calibrators. The site information described in natural language is extracted, inferred and integrated to generate program statements that ABAQUS can understand, thereby better adjusting the response of the simulation model. This provides support for the subsequent rapid generation of finite element simulation models for simulation response data generation and for modifying parameters under the action of optimization methods.

[0027] like Figure 2 As shown, constructing a finite element simulation model includes the following steps: S31, based on computing power, basic site conditions and target design parameter set; S32, Based on actual site information, perform natural language description to obtain descriptive text; S33, Input the parameter set and descriptive text into the large language model, and the large language model performs the following steps to generate a simulation model: S331 automatically locates physical quantities of the model in the description text, such as the thickness, modulus, Poisson's ratio and density information of the pavement, the ground response modulus under healthy conditions, the normal joint spring stiffness, the aircraft load speed, etc., and extracts the values ​​to obtain the extracted parameters. S332, Identify necessary parameters not mentioned in the description text, such as grid type and grid size, and perform reasoning to complete the necessary parameters or call the pavement engineering industry specifications and standard knowledge base to assign them default values; S333 performs a unified conversion on the extracted parameters and the required parameters for completion, and verifies their rationality (value range, relationship between variables); S334, based on the verified parameters, sets the large parameter matrix to be optimized, combines the finite element modeling method of the track panel, establishes the finite element simulation model, and generates the working condition definition statement.

[0028] In this embodiment, the large parameter matrix to be optimized includes material parameters, geometric parameters, and boundary condition parameters, wherein the material parameters include the elastic modulus. Poisson's ratio and density Geometric parameters include cross-sectional width ,thickness and length Boundary condition parameters include foundation support stiffness. and joint spring stiffness There are a total of 8 parameters, as shown in equations (1)-(4).

[0029] (1) (2) (3) (4) in, For material parameters, For geometric parameters, These are the boundary condition parameters.

[0030] In constructing the finite element simulation model, the structure is divided into multiple sub-regions based on mechanical laws and the installation of sensing optical fibers, such as... Figure 3 As shown, the structure is divided into two layers: horizontal and vertical. The horizontal layer represents the horizontal direction of the structure, and the vertical layer represents the depth direction of the structure. Parameters are set for each region. In the vertical structural layer, since the optical fibers are buried at different depths, the thickness of the layers is adjusted to ensure that the sensing optical fibers pass evenly through the parameter adjustment area (0.05m, 0.5m, and 0.8m). In the horizontal structural layer, considering that the structural performance is prone to change at the load loading point, the partitions are densely spaced within a preset range near the load loading area to better reflect changes in structural performance.

[0031] Based on the location of the fiber optic sensing device buried on site, measurement lines are simultaneously set in the finite element model for subsequent data extraction.

[0032] Combining the parameters obtained from the analysis and the settings of the parameters to be optimized, a finite element simulation model is established using the simulation model generation module and the finite element modeling method for track panels. This module generates working condition definition statements that can be understood by commercial finite element software. The module ultimately generates a .py file, which contains a description of the model and ABAQUS subroutine call statements. Running this file allows ABAQUS to calculate the global response results and extract the simulation response at the set survey line location. The code for the track panel model generation component is as follows: s=mdb.models["Model-1"].constrainedsketch(name="profile",sheetsize=200.0) g,v,d,c=s.geometry,s.vertices,s.dimensions,s.constraints s.setPrimaryObject(option=STANDALONE) Subsequently, in step S4, the commercial finite element software ABAQUS is called to perform response analysis. Modal analysis is used, with a maximum of 10 mode shapes, to solve for the mode response.

[0033] The data point spacing between the simulated response and the measured response is often difficult to keep consistent in practical applications. Therefore, in step S5, before calculating the response difference, the simulated response is aligned using the linear interpolation method. Specifically: For the simulation response time series { , }, to the time point set { Data alignment, where, For any time point in the simulation time series, For time points The corresponding simulated response value, For any point in the set of time points of the measured data; Find the corresponding interval index such that: ; By using linear interpolation, the adjustment parameter λ is calculated, and then the interpolated result is calculated using the adjustment parameter λ to achieve data alignment. (5) (6) in, This is the interpolated response value of the simulation data at the corresponding time point of the measured data after interpolation.

[0034] After data alignment, the response difference is calculated.

[0035] In one embodiment, mode shapes are used as physical quantities characterizing the response, and response differences are quantified based on the Modal Confidence Criterion (MAC). (7) (8) in, Indicates the response difference. For the first The weight of the MAC value corresponding to each test line. Indicates the first The test line is based on the modal shape of the simulated response. a and measured response mode shape b The calculated MAC value, They represent the first Modal vibration modes of the strip measurement line a and b of k A vector of order 1, where * denotes the conjugate transpose. n For order, L This represents the number of survey lines.

[0036] In another embodiment, the mode shape and frequency difference are used as physical quantities characterizing the response. After mode shape quantization based on the modal confidence criterion, the response difference is obtained by weighting. (9) (10) (11) (12) in, The response difference is represented by the MAC value and the frequency difference. constitute, , The first The weights of the MAC value and frequency difference corresponding to each test line. L This refers to the number of survey lines; Indicates the first The test line is based on the modal shape of the simulated response. a and measured response mode shape b The calculated MAC value, They represent the first Modal vibration modes of the strip measurement line a and b of k A vector of order 1, where * denotes the conjugate transpose. n It is the order; For the first Frequency difference of the test lines For the first strip survey line k The difference in natural frequencies of the first order. For the first Simulation response of the test line k First natural frequency, For the first Measured response of the test line k The first natural frequency.

[0037] In one preferred embodiment, when calculating the modal confidence criterion (MAC) to quantify the response difference, the first-order mode shape that best approximates the measured response is selected for calculating the MAC value, i.e., k is set to 1. The MAC value ranges from [0,1]. When the MAC is close to 1, it indicates that the two modes are very similar, while when it is close to 0, it indicates that the difference is large.

[0038] The MAC value alone can only measure the relative similarity of the responses, while the difference calculated by Equation (9) comprehensively quantifies the difference between the measured response and the simulated response, providing a reference value for finding the balance point between the two.

[0039] After calculating the response difference, the parameter values ​​of the large parameter matrix are optimized based on the response difference.

[0040] In one embodiment, a search algorithm such as a genetic algorithm can be used, with the reciprocal of the response difference as the optimization objective function. The parameters to be optimized are set as a population, and the optimization process of the parameter value range is transformed into a process of reducing the initial parameter values ​​to different degrees based on a reduction factor, and the reduction factor is optimized. The reduction factor is set to (0,1). The algorithm uses a method of randomly generating the parameter matrix as the starting point for the search, and performs 50 iterations, with the population size in each search being 50. The entire optimization solution loop is as follows: Figure 5 As shown.

[0041] In another embodiment, a reinforcement learning algorithm is employed, specifically the DDPG algorithm within the actor-critic framework. This framework consists of a storage pool, a training and target actor network, and a training and target critic network. The critic network learns from Q-values ​​to gradually provide Q-values ​​that approximate the true value. Guided by these Q-values, the actor network gradually updates its strategy and generates the optimal action based on the state.

[0042] Algorithm execution framework such as Figure 4 As shown, for an agent with an actor network at its core, the input state... ,in, In response to the amount of change in the MAC value within the difference, To respond to the frequency difference in the variation, a corresponding action vector is generated. ,in, The number of parameter types to be estimated. To define the spatial region dimension, the output motion vector is split into matrices corresponding to each parameter, as shown in equation (13), and the output values ​​are transformed to obtain specific physical quantity values, where the elastic modulus is... Poisson's ratio ,density The calculations are shown in equations (14)-(16). To improve the general applicability of this method, a parameter conversion method is proposed as shown in equations (17)-(18). For parameters in the range of (0, 1), the conversion method shown in equation (17) is used. For parameters with a larger range of values, the conversion method shown in equation (18) is used. The calculation of physical quantities shown in equations (14)-(16) falls within this category.

[0043] , , (13) (14) (15) (16) (17) (18) in, , , These are the output value vectors corresponding to the modulus, Poisson's ratio, and density, respectively. for , The output value corresponding to each sub-region in a given vector. , These are the maximum and minimum values ​​of the elastic modulus, respectively. , These are the maximum and minimum values ​​of Poisson's ratio, respectively. , These are the maximum and minimum density values, respectively. , , These are the converted elastic modulus, Poisson's ratio, and density, respectively. The converted physical quantity , These represent the maximum and minimum values ​​of the corresponding physical quantities.

[0044] The reward function consists of two parts: a differential quantification index and a step size penalty, expressed as: (19) (20) in, This is a reward function for quantifying differences, where α is a user-defined coefficient. For the search step size, This is the overall reward function.

[0045] During the optimization process, the parameter values ​​obtained from each solution are fed back to the finite element simulation model. The simulation response of the updated model is calculated, and a new response difference is calculated based on the newly calculated simulation response, completing one iteration. Depending on actual needs and experience, the number of iterations or a response difference threshold is set as a termination condition. The loop ends when the set termination condition is met.

[0046] After the loop ends, the parameters of the finite element simulation model have changed significantly from their initial settings. These local changes in structure, geometry, and boundary conditions are considered deviations from the ideal situation relative to the actual site. Since the response calculation of the finite element simulation model conforms to mechanical constraints, the response data generated by the simulation model should simultaneously reflect the structural conditions and mechanical laws of the actual site. However, the initial parameter settings included certain assumptions about the defined region, namely that the material properties of adjacent regions might differ significantly. In reality, the changes should be continuous rather than abrupt. Therefore, it is necessary to perform abrupt change checks on the parameters and generated data to determine whether the fused results output by the current finite element simulation model match the actual situation. This includes parameter abrupt change checks and response abrupt change checks. 1. Parameter mutation check The parameters are in the form of an m×n matrix. Therefore, the mutation check for parameters mainly focuses on ensuring that the differences between adjacent elements in the matrix are not too large. A threshold δ is set, and the absolute value of the difference between each parameter and the parameters of its adjacent sub-regions (up, down, left, and right) in space is calculated. The maximum value Δ is found. If Δ ≤ δ, no mutation is considered; otherwise, a large local mutation is considered, and the parameter mutation check fails. The parameter assignment region is then adjusted, using methods such as local encryption or local fusion. The fusion process is repeated until the parameters no longer mutate.

[0047] 2. Response mutation check Perform a continuity check on the response data, and remove the parabolic trend of the response data according to equation (21): (twenty one) in, For residuals used in continuity checks, In response to the data, Based on location points i Fitted parabolic data.

[0048] For residuals Use CUSUM to check for mutation points and calculate the mean. The upper CUSUM and lower CUSUM are calculated as shown in equations (22) and (23) with respect to the standard deviation σ. If both the upper CUSUM and lower CUSUM exceed the threshold (take 5σ), it is determined that a mutation has occurred. The mutation check fails, the simulation model design is modified, and the above steps are repeated.

[0049] (twenty two) (twenty three) in, The allowable offset is set to 0.5σ. This represents the data value at the current detection time point. The cumulative statistic of the upper CUSUM is taken at time n. When it exceeds the threshold (taken as 5σ), it is determined to be a sudden change. Let be the value of the cumulative statistic of the lower CUSUM at time n.

[0050] If either the parameter mutation check or the response mutation check fails, the fusion result output by the current finite element simulation model is determined to be inconsistent with the actual situation.

[0051] Applying specified loads to the final fused model and obtaining response data, this response data can be considered a fusion of simulation and measured data. This data conforms to the actual site conditions and mechanical constraints, enabling better evaluation and prediction of pavement structure performance. The fused data generation process is as follows: Figure 6 As shown.

[0052] In summary, the core of the data fusion in this invention is the process of fusing simulation data generation with actual measured physical quantities. Simulation data is generated and then integrated with the actual measured data in real time. The measured response and simulation response respectively embody the uncertainties of the actual structural service process and the mechanical constraints implemented through the finite element method. Through repeated iterations and optimizations, the fused data gradually approaches the equilibrium point between the measured and simulated data. By employing optimization algorithms to continuously reduce the differences generated by the response difference calculation module, the discrepancy between the measured and simulated response data is significantly reduced, ensuring that the final fused data accurately reflects the combined characteristics of both. Simultaneously, this mechanism supports the adaptive evolution of the model throughout its service life, ensuring that the fusion always accurately reflects the actual operational status of the structure.

[0053] Example 2 This embodiment uses the method described in Embodiment 1 to fuse measured data of asphalt beam optical fiber with mechanical simulation laws as an example for detailed explanation.

[0054] like Figure 7 As shown, the test specimen was a 0.8×0.1×0.3m simply supported asphalt beam. Local heating was used to alter the local stiffness of the material, simulating changes in the structural performance during actual use. Vibration strain data under hammer impact were acquired using optical fibers embedded within the specimen. To better obtain the structural vibration state, the optical fibers were embedded in three layers (top, middle, and bottom) within the beam, providing better coverage of all heights of the structure. The initial material parameters, geometric parameters, and boundary conditions of the specimen were known.

[0055] I. Establishment of the initial finite element simulation model Based on the actual fiber optic cable installation, the entire simply supported beam area is divided into 3×12 regions. Specifically, it is divided into 3 layers along the height and 12 regions along the length of the beam. The region in the middle of the beam is further densified to accurately reflect changes in structural performance. Regarding parameter selection, except for the material modulus, other parameters such as Poisson's ratio, density, bottom support, and thickness are set as constants to save computational resources and speed up calculations. In this embodiment, the shape and size of the large model parameters involved in the optimization are 3×12.

[0056] This approach combines a large language model-based simulation platform with reinforcement learning, embedding the large language model into the reinforcement learning environment and integrating the large language model, agent, and environment into the platform to achieve integration. The integration process and results are then visualized.

[0057] The platform parses the input, and the obtained finite element model parameters are as follows: density is set to 2200 kg / m³. 3 The initial modulus E is 4.5e. 8 Pa, Poisson's ratio is 0.35, the boundary condition is simply supported, and other material parameters are inferred or default values. The parameter to be optimized is set to I. E ×E,I E The parameter matrix to be optimized has a value range limited to (0,1], which represents the reduction of the initial modulus.

[0058] II. Introduction of Measured Response and Calculation of Difference The measured data from the optical fiber were processed using a signal processing method that solves the cross-power spectrum (CPSD) to extract the structural mode shapes. Three mode shape vectors were extracted, with each mode shape having a magnitude of [missing value]. .like Figure 8 , Figure 9 and Figure 10 As shown.

[0059] Run the script file from the previous step to call the ABAQUS software to perform a finite element analysis, solve for and extract the mode shape curves at the survey line location. The survey line settings are consistent with the measured data, i.e., the burial depths are both 0.0725, 0.03, and 0, and the number of data points is both 240. The specific mode shape curves are as follows: Figure 11 , Figure 12 and Figure 13 As shown.

[0060] Rewards are calculated only for ΔMAC, based on, for example Figure 14 The process calculates the response difference.

[0061] III. Data Fusion Cycle The parameter values ​​were optimized using the reinforcement learning DDPG algorithm. The number of episodes was set to 500, with 100 steps within each episode, and the actor network was trained repeatedly. The final motion values ​​were output, and the beam internal modulus distribution displayed by the final motion is shown below. Figure 15 As shown, a decrease in modulus occurs in the beam region, which can counteract the decrease in modulus caused by heating the asphalt beam, qualitatively verifying the effectiveness of the fusion frame operation.

[0062] IV. Data Fusion Generation The optimized parameters are then assigned to the simulation model, completing the construction of the large-scale simulation model. No abrupt changes occurred in the optimal parameters established by the algorithm framework. After verification, the parameters are substituted into the initial finite element simulation model. The simulation model used for verification in ABAQUS software is as follows: Figure 16 As shown, the input condition is set to static load using the loading head to calculate the fused data, as follows. Figure 17 As shown, the data is relatively continuous without any abrupt changes. Upon verification, this data is indeed high-quality fused data.

[0063] The reconstructed fused data was compared with the strain data of the asphalt beam measured by optical fiber under static load. The results showed that the fused data was in good agreement with the measured data and had a low noise level, which further verified the effectiveness of the fusion framework.

[0064] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. A method for fusing measured data and mechanical simulation data, characterized in that, The fused data is used for pavement structure performance evolution analysis. The fusion method includes the following steps: S1, obtain actual site information and measured data; S2, Based on the measured data, extract the measured response using cross-power spectrum processing; S3, based on the actual site information, call the large language model to generate the large parameter matrix to be optimized, and construct the finite element simulation model: Based on computing power, basic site conditions, and target design parameter set; Natural language description is performed based on actual site information to obtain descriptive text; The parameter set and descriptive text are input into a large language model, which then performs the following steps to generate a simulation model: Automatically locate the physical quantities of the model in the descriptive text and extract the parameters; Identify the necessary parameters not mentioned in the description text, and perform reasoning to complete the necessary parameters or call the relevant knowledge base to assign them default values; The extracted parameters and the required parameters for completion are uniformly converted, and their rationality is verified. Based on the verified parameters, a large parameter matrix to be optimized is set, and a finite element simulation model is established by combining the finite element modeling method of the track panel, and working condition definition statements are generated. S4, Solve the finite element simulation model to generate the simulation response; S5, calculate the response difference between the simulated response and the measured response; S6, Optimize the parameter values ​​of the large parameter matrix based on the response difference; S7, determine whether the preset number of iterations or the response difference has reached the preset threshold. If yes, determine whether the fusion result output by the current finite element simulation model matches the actual situation. Otherwise, return to step S3 to update the finite element simulation model based on the parameter values ​​of the optimized large parameter matrix and proceed to the next iteration. S8. If the fusion result in step S7 matches the actual situation, the finite element simulation model obtained by iteration is used as the fusion model. An equivalent load is applied to the fusion model to obtain fusion data. Otherwise, the parameter selection to be optimized is modified, and the process returns to step S3 to regenerate the finite element simulation model and update the large parameter matrix, and iterates again.

2. The method for fusing measured data and mechanical simulation data according to claim 1, characterized in that, The large parameter matrix to be optimized includes material parameters, geometric parameters, and boundary condition parameters. The material parameters include elastic modulus, Poisson's ratio, and density. The geometric parameters include cross-sectional width, thickness, and length. The boundary condition parameters include foundation support stiffness and joint spring stiffness.

3. The method for fusing measured data and mechanical simulation data according to claim 1, characterized in that, In constructing the finite element simulation model, the structure was divided into multiple sub-regions based on mechanical principles and the installation of sensing optical fibers. Parameters were set for each region. The structure was divided into two layers: horizontal and vertical. The horizontal layer represents the horizontal direction of the structure, and the vertical layer represents the depth direction. In the vertical layer, the thickness of the layers was adjusted to ensure that the sensing optical fibers passed evenly through the parameter adjustment area. In the horizontal layer, the partitions were densified within a preset range near the load-bearing area. Simultaneously, measurement lines were set in the finite element model according to the location of the on-site optical fiber sensing devices for subsequent data extraction.

4. The method for fusing measured data and mechanical simulation data according to claim 1, characterized in that, In step S5, before calculating the response difference, the simulation response is aligned using the linear interpolation method of simulation data. Specifically: For the simulation response time series { , }, to the time point set { Data alignment, where, For any time point in the simulation time series, For time points The corresponding simulated response value, For any point in the set of time points of the measured data; Find the corresponding interval index such that: ; By using linear interpolation, the adjustment parameter λ is calculated, and then the interpolated result is calculated using the adjustment parameter λ to achieve data alignment. ， ， in, This is the interpolated response value of the simulation data at the corresponding time point of the measured data after interpolation.

5. The method for fusing measured data and mechanical simulation data according to claim 1, characterized in that, The method for calculating the response difference is as follows: using the mode shape as a physical quantity characterizing the response, and quantifying the response difference based on the modal confidence criterion. ， ， in, Indicates the response difference. For the first The weight of the MAC value corresponding to each test line. Indicates the first The test line is based on the modal shape of the simulated response. a and measured response mode shape b The calculated MAC value, They represent the first Modal vibration modes of the strip measurement line a and b of k A vector of order 1, where * denotes the conjugate transpose. n For order, L This represents the number of survey lines.

6. The method for fusing measured data and mechanical simulation data according to claim 1, characterized in that, The method for calculating the response difference is as follows: using the mode shape and frequency difference as physical quantities characterizing the response, and after quantizing the mode shape based on the modal confidence criterion, the response difference is obtained by weighting. ， ， ， ， in, The response difference is represented by the MAC value and the frequency difference. constitute, , The first The weights of the MAC value and frequency difference corresponding to each test line. L This refers to the number of survey lines; Indicates the first The test line is based on the modal shape of the simulated response. a and measured response mode shape b The calculated MAC value, They represent the first Modal vibration modes of the strip measurement line a and b of k A vector of order 1, where * denotes the conjugate transpose. n It is the order; For the first Frequency difference of the test lines For the first strip survey line k The difference in natural frequencies of the first order. For the first Simulation response of the test line k First natural frequency, For the first Measured response of the test line k The first natural frequency.

7. A method for fusing measured data and mechanical simulation data according to claim 5 or 6, characterized in that, The parameter values ​​for optimizing the large parameter matrix based on the response difference are specifically as follows: a search algorithm is used, with the reciprocal of the response difference as the optimization objective function, the parameters to be optimized are set as a population, and the optimization process of the parameter value range is transformed into a process of reducing the initial values ​​of the parameters to different degrees based on the reduction factor, and the reduction factor is optimized.

8. The method for fusing measured data and mechanical simulation data according to claim 6, characterized in that, The parameter values ​​for optimizing the large parameter matrix based on the response difference are specifically determined as follows: A reinforcement learning algorithm is used to process the input state of the agent, which is based on an actor network. ,in, In response to the amount of change in the MAC value within the difference, To respond to the frequency difference in the variation, a corresponding action vector is generated. ,in, The number of parameter types to be estimated. To define the spatial region dimension, the output action vector is split into matrices corresponding to each parameter, and the output values ​​are transformed to obtain specific physical quantity values. The reward function consists of two parts: a differential quantification index and a step size penalty, expressed as: ， ， in, This is a reward function for quantifying differences, where α is a user-defined coefficient. For the search step size, This is the overall reward function.

9. The method for fusing measured data and mechanical simulation data according to claim 1, characterized in that, The determination of whether the fusion result output by the current finite element simulation model matches the actual situation specifically includes parameter mutation checking and response mutation checking, wherein, The parameter mutation check is specifically as follows: set a threshold δ, calculate the absolute value of the difference between each parameter and the parameters of its adjacent upper, lower, left and right sub-regions in space, and find the maximum value Δ. If Δ≤δ, it is determined that there is no mutation; otherwise, the parameter mutation check fails. The response mutation check specifically involves: performing a continuity check on the response data, based on... Remove the parabolic trend from the response data, where, For residuals used in continuity checks, In response to the data, Based on location points i Fitted parabolic data; for residuals Use CUSUM detection to check for mutation points. If both the upper and lower CUSUM exceed the threshold, it is determined that a mutation has occurred, and the mutation check fails. If either the parameter mutation check or the response mutation check fails, the fusion result output by the current finite element simulation model is determined to be inconsistent with the actual situation.

Images