A sparse data-driven global mechanical field reconstruction method and system for asphalt pavement

Abstract

The application discloses a sparse data-driven global mechanical field reconstruction method and system for asphalt pavement, and the method comprises the following steps: acquiring real-time monitoring data of a sparse embedded sensor network; constructing a three-dimensional finite element physical twin benchmark model containing a viscoelastic constitutive relation, and offline training a rapid mechanical calculation model covering temperature-load working conditions; using the monitoring data as a constraint, calling a parameter inversion optimization algorithm driven by the rapid model, online correcting material constitutive parameters in a closed loop, and fitting a digital model with a real pavement state in real time; reconstructing global stress, strain and damage fields based on the corrected model, identifying high-risk areas of hidden diseases, and establishing a long-term performance evolution database to support preventive maintenance decisions. The application realizes fine management and maintenance of the whole life cycle of the asphalt pavement through deep integration of data and mechanism.

Description

Technical Field

[0001] This invention relates to the field of digital twin and health monitoring technology for road engineering, and in particular to a sparse data-driven method and system for reconstructing the global mechanical field of asphalt pavement. Background Technology

[0002] During its service life, the material properties of asphalt pavement continuously evolve under the coupled effects of environmental factors and traffic loads, leading to defects such as rutting and fatigue cracking. Traditional manual inspections and periodic testing methods suffer from large monitoring blind spots, slow response times, and difficulty in detecting early-stage hidden damage. In recent years, digital twin technology has provided a new paradigm for the virtual mapping and prediction of pavement conditions. However, applying digital twins to asphalt pavement health monitoring still faces a series of key technical challenges.

[0003] In existing technologies, some studies have attempted to apply digital and intelligent methods to pavement maintenance. For example, existing technology CN115063053A discloses a digital construction and preventive maintenance decision-making method for asphalt pavement. This method acquires pavement surface technical condition assessment data (such as PCI, RQI, RDI, etc.), combines it with actual detection data to diagnose defects, and predicts and outputs maintenance plans based on evolutionary patterns. The limitations of this method are: First, its decision-making heavily relies on pavement surface condition indicators, which is a "post-diagnosis" and cannot provide early warning and quantitative assessment before internal damage manifests as surface defects. Second, this method lacks the ability to perceive and invert the true mechanical state of the pavement (such as stress, strain, and interlayer bonding state), and fails to establish a real-time interactive closed loop between monitoring data and physical mechanism models. Therefore, it cannot achieve mechanism-based, forward-looking performance degradation prediction and accurate location of hidden defects.

[0004] Another existing technology, CN121458272A, discloses an intelligent maintenance system and method for highways. This method constructs a four-dimensional spatiotemporal digital twin model using multimodal sensing data and utilizes artificial intelligence models such as graph neural networks for health status prediction and maintenance decisions. This technology represents an advanced direction in data-driven predictive maintenance. However, its core is based on historical and real-time monitoring data, directly establishing a mapping relationship from "multi-source data" to "health status" or "maintenance decisions" through machine learning models, belonging to a "pure data-driven" or "black box model" paradigm. For viscoelastic materials like asphalt pavement, which exhibits strong nonlinearity and dependence on temperature and load rate, this method suffers from insufficient physical interpretability, high dependence on the quantity and quality of training data, and difficulty in guaranteeing extrapolation reliability in scenarios with sparse sensor deployment. More importantly, it fails to effectively integrate and utilize mature physical mechanism models such as solid mechanics and finite element analysis, thus failing to achieve the core function of "inverting material constitutive parameters based on sparse monitoring point data, and then deducing the continuous mechanical field across the entire domain," making it difficult to solve the "perception blind spot" problem under the constraint of sensor deployment cost.

[0005] In summary, the existing technical solutions have the following shortcomings: (1) Static model parameters and inaccurate prediction: Most digital twin models based on finite element simulation use fixed material parameters, which cannot reflect the actual deterioration of materials during long-term service, resulting in a decrease in prediction accuracy over time. (2) Blind spots in perception under sparse monitoring: Due to engineering cost constraints, sensors are sparsely deployed, and existing forward simulation or pure data interpolation methods cannot accurately know the mechanical state of non-monitored areas, making it difficult to detect hidden defects in a timely manner. (3) Lack of real-time adaptive and closed-loop correction capabilities: The model does not have the ability to automatically calibrate online based on real-time monitoring data, and gradually becomes disconnected from the real physical state. (4) Contradiction between computational efficiency and engineering real-time performance: The calculation time of full-order finite element models is too long, which cannot meet the stringent real-time requirements of online parameter inversion (which requires hundreds to thousands of iterations). (5) Lack of specialized technologies for asphalt material properties: Existing general-purpose accelerated calculation methods (such as reduced-order models) and parameter inversion techniques do not fully consider the significant viscoelasticity and temperature-frequency dependence of asphalt mixtures, and direct application presents technical barriers.

[0006] Therefore, there is an urgent need to develop a technical solution that can deeply integrate physical mechanisms and real-time monitoring data, specifically targeting the characteristics of asphalt pavement materials, and enabling online adaptive correction and full-domain performance reconstruction of the model within an acceptable engineering timeframe. Summary of the Invention

[0007] The main objective of this invention is to provide a sparse data-driven method and system for reconstructing the global mechanical field of asphalt pavement, aiming to solve the problems of inaccurate predictions by existing static models, prominent blind spots in perception under sparse monitoring, and the inability of full-order models to meet the requirements of online inversion in computation.

[0008] To achieve the above objectives, this invention provides a sparse data-driven method for reconstructing the global mechanical field of asphalt pavement, wherein the method includes the following steps:

[0009] Step 1: Acquire monitoring data from the sensor array deployed inside the asphalt pavement structure. The monitoring data includes mechanical response, ambient temperature, and traffic load data.

[0010] Step 2: Construct a three-dimensional finite element physical twin reference model of asphalt pavement, wherein the asphalt mixture layer adopts a viscoelastic constitutive relation, the interlayer contact adopts a cohesive contact model with the interlayer bond stiffness parameter as the state variable, and the material constitutive parameters of the physical twin reference model are defined in the form of variable parameters; and generate a mechanical response sample library covering the preset temperature and load parameter space offline based on the physical twin reference model, and construct a fast mechanical calculation model based on the mechanical response sample library, wherein the fast mechanical calculation model is a reduced-order model or a surrogate model;

[0011] Step 3: Using the monitoring data from Step 1 as constraints, the rapid mechanical calculation model from Step 2 is called for iterative simulation to construct an objective function characterizing the deviation between the simulated response value and the measured value of the monitoring point. An optimization algorithm is then used to dynamically adjust the material constitutive parameters of the physical twin benchmark model. The material constitutive parameters include the complex modulus master curve parameters and interlayer bond stiffness parameters of the asphalt mixture, until the objective function converges to a preset threshold, thereby obtaining a digital twin that is in real-time fitted to the current pavement condition.

[0012] Step 4: Based on the real-time digital twin obtained in Step 3, the rapid mechanical calculation model is called to reconstruct the global stress field, modulus field, strain field and damage field. Based on the damage field, high-risk areas with potential hidden defects are identified (the fatigue damage variables in the damage field are compared with preset critical values ​​to locate high-risk areas), and a full-road performance distribution cloud map is generated.

[0013] Optionally, in step 1, the sensor array is deployed at the interface between each structural layer directly below the wheel track on the road surface.

[0014] Optionally, in step 2, the fast mechanical calculation model is a reduced-order model based on intrinsic orthogonal decomposition, or a surrogate model based on physical information neural network or radial basis function neural network.

[0015] Optionally, in step 2, the fast mechanics calculation model dynamically switches between a reduced-order model and different proxy models based on the available memory capacity of the edge computing nodes.

[0016] Optionally, in step 2, the reduced-order model constructs a reduced-order basis matrix by performing eigenorthogonal decomposition on the displacement field of the mechanical response sample library and extracting the first N eigenmodes, where the value of N is used to ensure that the reduced-order basis matrix covers more than 99% of the energy of the original displacement field; during online calculation, the high-dimensional finite element control equations are projected onto the reduced-order basis space for solution.

[0017] Optionally, in step 3, the objective function is the weighted sum of squares of the deviations between the simulated response values ​​and the measured values ​​at each monitoring point, as shown in the following formula:

[0018] ;

[0019] Where J is the objective function. Let i be the simulated response value of the i-th monitoring point. Let i be the measured value of the i-th monitoring point. Here, y represents the weighting coefficient, and y generally refers to the mechanical response value. These are weighting coefficients, and are assigned values ​​based on the sensor type and measurement reliability of each monitoring point.

[0020] Optionally, in step 3, the material constitutive parameters may further include phase angle parameters and elastic modulus of each structural layer.

[0021] Optionally, in step 3, the optimization algorithm includes Bayesian inference, genetic algorithm, or particle swarm optimization.

[0022] Optionally, step 3 also includes an accuracy verification step: after the objective function converges, the inversion parameters are substituted into the fast mechanical calculation model to recalculate the response of all monitoring points. When the root mean square error between the simulation value and the measured value is lower than a preset threshold, the reconstruction is confirmed to be effective; otherwise, the inversion iteration is retried.

[0023] Optionally, in step 4, when the interlaminar shear stress in the high-risk area exceeds the interlaminar bonding material strength threshold, it is determined to be poor interlaminar bonding; when the tensile strain at the bottom of the surface layer in the high-risk area exceeds the fatigue failure limit strain, it is determined to be surface layer fatigue cracking.

[0024] Optionally, the method further includes: step 5, storing the material constitutive parameters obtained from step 3 and the fatigue damage variables from step 4 in a time series manner, and constructing a performance evolution database by synchronously storing historical environmental data and historical traffic load data at the corresponding time; based on the performance evolution database, using time series analysis methods and combining the coupling effects of environment and load, predicting the pavement performance degradation trajectory; when the predicted time when the predicted material modulus is lower than the maintenance intervention critical value is less than a preset time threshold, generating a preventive maintenance decision report containing maintenance intervention timing and measure recommendations.

[0025] Furthermore, to achieve the above objectives, the present invention also provides a sparse data-driven global mechanical field reconstruction system for asphalt pavement, used to implement the above-described method, the system comprising:

[0026] The multi-source sensing module is used to acquire real-time data from sensors, the environment, and the load.

[0027] The digital twin modeling module is used to construct the physical twin benchmark model and the rapid mechanical calculation model.

[0028] The virtual-real interaction and inversion module is used to execute online parameter inversion closed loop and correct the digital twin in real time;

[0029] The global field analysis and diagnosis module is used to reconstruct the global mechanical field, identify hidden defects, and generate a cloud map of the performance distribution of the entire road domain.

[0030] Beneficial effects:

[0031] (1) Real-time dynamic adaptation of the digital twin model is achieved: By integrating a fast mechanical calculation model specifically designed for asphalt viscoelasticity (POD reduced-order model or PINN surrogate model), the time for a single mechanical calculation in parameter inversion is reduced from about 300 seconds for the full-order finite element model to within seconds (e.g., 0.8 seconds). Combined with optimization algorithms, the overall parameter inversion closed-loop response time from data input to obtaining the corrected digital twin can be controlled within 10 seconds under typical working conditions, meeting the stringent real-time requirements of engineering sites and fundamentally solving the long-term prediction inaccuracy problem caused by parameter solidification in traditional static models.

[0032] (2) Breakthrough in the perception blind spot of sparse monitoring: Through the technical path of "online parameter inversion calibration mechanism model - reconstruction of the whole field based on calibration model", the limited and discrete point monitoring information is expanded into continuous and high-resolution whole field mechanical field (stress, strain, damage) information under the constraints of physical laws. Virtual imaging of the internal mechanical state of sensorless areas (such as interlayer interfaces and road shoulders) is realized, so that hidden defects such as early accumulation of material damage and interlayer voids can be detected and located in time before they develop into surface defects.

[0033] (3) A new paradigm of dual-driven integration of data and mechanism was established: real-time monitoring data (data-driven) and finite element physical mechanism (mechanism-driven) were deeply integrated. This not only utilized the strong extrapolation capability, interpretability and reliability in sparse data regions brought by the physical model, but also endowed the model with the adaptive ability to dynamically track the actual service state through real-time data closed-loop correction. This overcame the defects of "open-loop" inaccuracy of pure physical simulation model and "black box" and insufficient extrapolation reliability of pure data-driven model.

[0034] (4) Online quantitative identification of key materials and interface conditions was achieved: The complex modulus and phase angle parameters reflecting the viscoelasticity of asphalt mixtures, as well as the interlayer bond stiffness parameters characterizing the overall structure, were incorporated into the online inversion system, enabling real-time tracking of the evolution of these key parameters with service time and environmental effects. This compensates for the deviation between the initial laboratory measurements and the actual field conditions, allowing the system to maintain high-precision condition assessment and prediction capabilities even when the material has aged but the surface has not yet shown significant defects.

[0035] (5) Provides precise decision support for preventive maintenance throughout the entire life cycle: By storing and inverting material state parameters and damage variables in a time series manner, and combining them with environmental and load history, a long-term pavement performance evolution database is constructed. Based on this database, time series prediction models can be used to predict the pavement performance degradation trajectory in a forward-looking manner, and quantitatively provide the best timing, measures, and expected benefits of maintenance intervention (such as extending the pavement service life by more than 5 years), promoting the transformation of the maintenance model from "passive response" to "active prevention" and "precision economy", and significantly reducing the cost of maintenance throughout the entire life cycle. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be 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 the structures shown in these drawings without creative effort.

[0037] Figure 1 This is a flowchart illustrating a sparse data-driven global mechanical field reconstruction method for asphalt pavement according to the present invention.

[0038] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0039] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0040] See Figure 1 This invention provides a flowchart illustrating a sparse data-driven method for reconstructing the global mechanical field of asphalt pavement. The method includes the following steps:

[0041] Step 1 - Data Acquisition: Acquire monitoring data from sensor arrays deployed within the asphalt pavement structure. This data includes mechanical response, ambient temperature, and traffic load data. Mechanical response data includes strain field data and / or compressive stress data; ambient temperature data includes temperature field data and / or humidity data; and traffic load data includes traffic flow data and axle load distribution data. The sensor arrays are deployed at the interfaces between structural layers directly beneath the pavement wheel tracks. Specifically, the sensor arrays include distributed fiber optic strain sensors (deployed at the bottom of the surface layer and the top of the intermediate layer to measure longitudinal and transverse tensile strain), MEMS temperature sensors (deployed in the middle of the surface layer to collect real-time pavement temperature data), and compressive stress sensors (deployed at the top of the base layer to monitor vertical compressive stress). The spacing between adjacent sensor arrays is 50 to 200 meters.

[0042] Preferably, sensor data is transmitted to the roadside edge computing unit via a ZigBee wireless network, with a sampling frequency of 500Hz. The system also connects to a weather station to obtain ambient temperature and rainfall data, and obtains traffic flow and axle load data through the WIM dynamic weighing system's standardized API interface.

[0043] Step 2 - Construction of Physical Twin Baseline Model and Rapid Calculation Model: Construct a three-dimensional finite element physical twin baseline model of asphalt pavement, and construct a rapid mechanical calculation model covering preset temperature and load conditions based on the physical twin baseline model. The rapid mechanical calculation model is a reduced-order model or a proxy model.

[0044] Specifically, the physical twin benchmark model is constructed as follows: Based on actual pavement design parameters, a three-dimensional model of the road section is established using finite element software, comprising a top layer, intermediate layer, bottom layer, and base layer. The top layer, intermediate layer, and bottom layer constitute the asphalt mixture layer, and a viscoelastic constitutive relation is adopted. The dynamic modulus master curve is expressed as a bivariate function of temperature and loading frequency (usually using the Sigmoidal function), and the phase angle parameter characterizes the viscoelastic hysteresis characteristics. The initial material parameters are determined based on laboratory tests. The base layer adopts a linear elastic constitutive relation, and is assigned an initial elastic modulus and Poisson's ratio. The interlayer contact adopts a contact model that reflects adhesion and slippage, such as a cohesive model, and initial bond strength, stiffness, and other parameters are set. All material constitutive parameters in this model (such as dynamic modulus, phase angle, elastic modulus, and bond stiffness) are defined and stored as variable parameters. Preferably, the asphalt mixture layer adopts a viscoelastic constitutive relation with the shape parameter and phase angle parameter of the complex modulus master curve as state variables, and the interlayer contact adopts a cohesive contact model with the interlayer bond stiffness parameter as the state variable.

[0045] The construction of a rapid mechanical calculation model includes three selectable methods.

[0046] (1) POD reduced-order model based on intrinsic orthogonal decomposition: Based on the physical twin benchmark model, a large number of offline simulations are performed in the preset parameter space to generate a mechanical response sample library covering typical working conditions, with no less than 1000 samples. Then, intrinsic orthogonal decomposition is performed on the displacement field (or other key field variables) of the samples to extract the first N eigenmodes that can represent their most important change characteristics, forming a "reduced-order basis matrix". The value of N is from 20 to 100, and it is necessary to ensure that it can cover more than 99% of the energy of the original displacement field. In addition, during online calculation, the high-dimensional finite element control equations are projected onto this low-dimensional reduced-order basis space for solution, thereby reducing the single calculation time from hundreds of seconds for the full-order model to less than seconds.

[0047] (2) PINN surrogate model based on physical information neural network: A large number of offline simulation samples (no less than 5000 sets) are used as training data. The unique feature of PINN is that its loss function not only includes the fitting error of the monitoring point response data, but also forcibly incorporates physical law constraints. Specifically, its loss function is a weighted sum of the following three parts:

[0048] Data residual term: The error between the monitoring point response predicted by the neural network and the simulation value of the full-order model.

[0049] Mechanical control equation residuals: At the sampling points within the network domain, the prediction results must satisfy the residuals of the mechanical equilibrium equation and the viscoelastic constitutive equation of asphalt pavement.

[0050] Boundary condition residuals: The prediction results must satisfy the residuals of boundary conditions such as stress and displacement.

[0051] The sum of the three weights is 1. By training, the total loss is minimized, resulting in a surrogate model that both fits the data and obeys the laws of physics. After training, the model weight file is small (approximately 50MB), and the inference speed is extremely fast (approximately 0.08 seconds).

[0052] (3) RBF surrogate model based on radial basis function neural network: This is a more lightweight data-driven surrogate model. It uses a small number of offline samples (e.g., 1000 groups) as the training set, uses the Gaussian kernel function as the radial basis function, and the number of hidden layer nodes is usually 20% to 30% of the number of training samples. The output layer weights are solved directly by the least squares method, the training speed is fast (about 0.5 hours), the resulting model file is very small (about 5MB), and the single inference time is about 0.05 seconds, which is suitable for edge devices with extremely limited computing resources.

[0053] After the rapid mechanical calculation model is built, a certain proportion (e.g., 10%) of the offline samples needs to be randomly selected as a test set to verify its prediction accuracy. The maximum absolute error between the predicted value and the full-order finite element simulation value should not exceed 2% of the measured value; otherwise, the sample size needs to be increased or the model parameters adjusted and rebuilt.

[0054] To adapt to the resource conditions of different edge computing devices, the system supports online dynamic switching of the fast mechanics calculation model. For example, when the available memory of the device is greater than 500MB, the more accurate POD reduced-order model is used first; when the memory is between 50-100MB, it automatically switches to the PINN proxy model; and when the memory is less than 50MB, it switches to the lighter RBF proxy model. The switching process must be seamless, and the difference in prediction results for the same working condition should not exceed 3%.

[0055] Step 3 - Parameter Inversion and Real-time Model Correction: Using the monitoring data from Step 1 as constraints, the rapid mechanical calculation model from Step 2 is called for iterative simulation, and the material constitutive parameters of the physical twin benchmark model are dynamically adjusted using an optimization algorithm until the error between the simulation response and the measured data converges, thus obtaining a digital twin that fits the current road surface condition in real time.

[0056] Specifically, using the real-time monitoring data obtained in step 1 as constraints, the rapid mechanical calculation model constructed in step 2 is called to perform mechanical response simulation, and an objective function J characterizing the deviation between the measured value and the simulated value is constructed. The specific formula is as follows:

[0057] ;

[0058] in Let i be the simulated response value of the i-th monitoring point. Let i be the measured value of the i-th monitoring point. The weighting coefficient, y, generally refers to the mechanical response value, which in a specific embodiment is expressed as the strain value ε or the compressive stress value; the weighting coefficient Based on the sensor type and measurement reliability assignment for each monitoring point, the monitoring points for fiber optic strain sensors are selected as follows: For MEMS strain sensor monitoring points, The weight is 0.8 to 1.0, and for monitoring points with abnormal fluctuations, the weight is reduced to 0.1 to 0.5 to reduce interference.

[0059] Furthermore, the material constitutive parameters of the physical twin benchmark model are iteratively adjusted using an optimization algorithm. Each iteration calls the fast mechanical calculation model to update the simulation values, forming a real-time closed loop of sensing and monitoring, fast simulation, and parameter update. The optimization algorithm includes Bayesian inference, genetic algorithm, particle swarm optimization, or Kalman filter algorithm.

[0060] Key parameters include core parameters reflecting the viscoelasticity of asphalt mixtures, such as the shape and phase angle parameters of the complex modulus master curve, the elastic / dynamic moduli of each structural layer (e.g., surface layer modulus E1, intermediate layer modulus E2, base layer modulus E3), and the bond stiffness parameters reflecting the interlayer interface state. These parameters are set based on prior engineering constraints, and multiple random initialization inversions are performed to verify the stability and identifiability of the results.

[0061] Iterative process: forming a real-time closed loop of "sensing and monitoring - rapid simulation - parameter update":

[0062] (1) The optimization algorithm proposes a new set of candidate material parameters → (2) The system substitutes these new parameters into the fast mechanical calculation model (instead of the full-order model), and recalculates the simulation response of the monitoring point in combination with the current working conditions → (3) Calculate the objective function value J based on the new simulation response → (4) The algorithm judges the quality of the parameters based on the objective function value and generates the next set of better parameter candidate values. Repeat this process, calling the fast model in each iteration, thereby reducing the time of a single mechanical calculation from about 300 seconds for the full-order finite element method to less than a second (such as 0.8 seconds for POD order reduction), making the entire inversion process feasible in engineering.

[0063] When the objective function converges to a preset threshold, a real-time digital twin that fits the current real road surface condition is obtained; the preset threshold ranges from 5% to 20% of the initial objective function value.

[0064] Furthermore, to ensure a reliable solution is obtained even with a limited number of sensors (sparse), the following measures were taken:

[0065] Bounded constraints based on engineering priors are imposed on the inversion parameters (e.g., the search range for the surface modulus E1 is set to 50% to 100% of the initial laboratory value).

[0066] Perform independent inversions with multiple random initializations (no less than 3 times). If the standard deviation of multiple results exceeds 10% of the mean, the data is considered to be abnormal. The system will automatically reduce the weight of the abnormal monitoring points and perform the inversion again to ensure that the results converge to a physically reasonable solution domain.

[0067] After the inversion convergence, a verification procedure is executed: the optimal parameter set obtained from the inversion is substituted into the fast model, and the simulated values ​​of all monitoring points are recalculated, requiring that the root mean square error between the simulated values ​​and the measured values ​​does not exceed 5% relative to the average level of the measured values. If this condition is met, the accuracy is confirmed to be up to standard; otherwise, the inversion iteration needs to be retried. When the convergence condition is met and the accuracy verification is passed, a real-time digital twin that fits the current actual service state of the road surface is obtained. At this point, the material parameters in the model (such as modulus and bond stiffness) have been corrected online, which can truly reflect the current (potentially deteriorated) mechanical properties of the road surface, providing an accurate basis for subsequent full-domain reconstruction.

[0068] Step 4 - Global Mechanical Field Reconstruction and Disease Identification: Based on the real-time digital twin obtained in Step 3, the rapid mechanical calculation model is called to reconstruct the mechanical response field distribution. The mechanical response field includes stress field, modulus field, strain field and damage field, and high-risk areas of potential hidden diseases are identified based on the damage field.

[0069] Specifically, under the same current environment and load conditions, the mechanical response field of the entire area covered by the digital twin model (the entire road domain) is calculated, not just the monitoring point locations. The field variables that need to be extracted include at least: stress field (including stress components in all directions), strain field (including strain components in all directions), and damage field (field variables calculated based on the selected material damage model, the core of which is the fatigue damage factor (D), which quantifies the degree of fatigue damage accumulated by the material due to cyclic loading, and its value range is usually from 0 (no damage) to 1 (complete failure)).

[0070] Furthermore, the specific steps for identifying high-risk areas of potential hidden defects based on the damage field are as follows: identify high-risk areas of potential hidden defects based on the grid-by-grid comparison results between the damage field and the preset critical value, and generate a cloud map of the performance distribution of the entire road domain.

[0071] Specifically, the fatigue damage factor D of each computational unit (or grid) in the global damage field is compared grid-by-grid with a preset critical value. This critical value is determined based on pavement material fatigue test data or current design specifications, for example, ranging from 0.6 to 0.9, preferably 0.8. All areas where the D value exceeds the aforementioned critical value are automatically marked by the system as "high-risk areas for potential hidden defects." Furthermore, in addition to locating high-risk areas, the system automatically identifies suspected defect types based on the combined characteristics of the mechanical field.

[0072] The specific criteria are as follows: when the fatigue damage factor in a certain area exceeds the limit and the interlaminar shear stress in that area exceeds the threshold of the interlaminar bonding material strength, it is judged as poor interlaminar bonding.

[0073] When the fatigue damage factor in a certain area exceeds the limit, and the tensile strain at the bottom of the surface layer in that area exceeds the limit strain for fatigue failure of asphalt mixture, it is determined to be surface layer fatigue cracking.

[0074] Furthermore, based on the calculated full-field data, a full-area performance distribution cloud map with a spatial resolution of no less than 0.1 m × 0.1 m is automatically generated. These cloud maps may include: a full-area stress distribution cloud map, an interlayer shear stress distribution cloud map, a fatigue damage factor distribution cloud map, and a permanent deformation cumulative distribution cloud map. High-risk areas identified are then visually marked on the cloud maps or associated electronic maps. The markings must include at least: the suspected defect type (e.g., "poor interlayer bonding"), the specific location and station number (e.g., "K100+240"), the quantitative value of the damage degree (e.g., "fatigue damage factor D=0.85", "estimated bond strength 0.15 MPa"), and the recommended treatment priority (e.g., "Level 1, immediate core drilling verification recommended").

[0075] When a high-risk area is identified, the system can automatically output a warning message and trigger subsequent maintenance procedures.

[0076] Step 5 - Long-term performance evolution and maintenance decision-making: Store the parameters obtained from the inversion in Step 3 and the damage data in Step 4 in time series, construct a performance evolution database, predict the pavement performance degradation trajectory based on the database, and generate a maintenance decision report.

[0077] Specifically, the material constitutive parameters obtained from the inversion in step 3 and the damage data from step 4 are stored in time series, and the historical environmental data and traffic load data corresponding to each inversion time are stored synchronously. These data are then used to form a related dataset with the above parameters, thereby constructing a long-term pavement performance evolution database. The storage period of this time series database is designed to be no less than 10 years to cover the medium- and long-term performance evolution process of the pavement.

[0078] Based on the database, the future performance degradation trajectory of the pavement is predicted, and a preventive maintenance decision report is generated. Specifically, time series analysis methods are used to predict the time when the dynamic modulus of the surface layer decays to the critical value for maintenance intervention, and a decision report including maintenance timing and measures is generated accordingly. The time series analysis methods include at least one of the following: autoregressive integral moving average model, long short-term memory network, and state-space model. These methods can capture the trend and periodicity of parameter changes over time. The prediction incorporates the coupled effects of environment and load, that is, temperature, rainfall, axle load, etc., are included as covariates or influencing factors in the prediction model, rather than simply extrapolating material parameters. The core output of the prediction is the future performance degradation trajectory of the pavement, one of the key judgments of which is "the predicted time when the dynamic modulus of the surface layer decays to the critical value for maintenance intervention." This critical value for maintenance intervention needs to be determined according to current design specifications such as the "Specifications for Design of Highway Asphalt Pavement" (JTG D50-2017). That is, the predicted degradation trajectory of the key performance indicators of the pavement specifically predicts the time when the dynamic modulus of the surface layer decays to the critical value for maintenance intervention determined according to the design specifications.

[0079] Furthermore, when the prediction results meet the warning conditions, the system automatically generates decision support information. When the predicted time for the material modulus to fall below the maintenance critical value is less than 12 months from now, a preventive maintenance warning is automatically triggered. The system generates a structured preventive maintenance decision report, which is the final output of step S5 and includes at least: maintenance intervention timing - the specific time point recommended for implementing maintenance (e.g., "recommended to implement in the 7th year"), maintenance measure type - the specific plan recommended from the measure library based on the predicted type and degree of damage (e.g., "implement a 3cm thick SMA-13 ​​thin overlay"), and expected effects - a quantitative estimate of the degree of pavement performance recovery and service life extension after maintenance (e.g., "expected to restore the modulus to approximately 5500MPa", "effectively extend the pavement service life by more than 5 years").

[0080] Furthermore, to better illustrate the method of the present invention, a sparse data-driven global mechanical field reconstruction system for asphalt pavement is also provided, the system comprising:

[0081] The multi-source sensing module is used to acquire real-time data from embedded sensors, environmental meteorology, and traffic loads. Specifically, it includes an embedded sensor submodule, an environmental meteorology access submodule, and a traffic load access submodule. The embedded sensor submodule includes distributed fiber optic grating sensors and / or MEMS strain sensors, sparsely distributed at the interfaces between various structural layers of the road surface. The traffic load access submodule connects to the dynamic weighing WIM system through a standardized API interface to acquire traffic flow and axle load distribution data.

[0082] The digital twin benchmark modeling module is used to construct a three-dimensional finite element physical twin benchmark model and a rapid mechanical calculation model. Specifically, it includes a finite element modeling unit and a rapid calculation model construction unit. The finite element modeling unit constructs a three-dimensional finite element physical twin benchmark model based on pavement design parameters, including the geometry of each structural layer, interlayer cohesive contact, and boundary loads. The asphalt mixture layer adopts a viscoelastic constitutive relation with the shape parameters and phase angle parameters of the complex modulus principal curve as the core, and the material constitutive parameter fields are marked as dynamically correctable. The rapid calculation model construction unit, considering the viscoelastic characteristics of the asphalt mixture's complex modulus changing nonlinearly with temperature and frequency, pre-generates simulation samples covering the temperature-load parameter space offline, extracts the reduced-order basis matrix based on the intrinsic orthogonal decomposition (POD) method, or trains a radial basis function (RBF) / physical information neural network (PINN) surrogate model for online inference.

[0083] The virtual-real real-time interaction module is used to invoke a rapid mechanical calculation model, perform parameter inversion optimization, and determine convergence and verify accuracy. Specifically, it includes a rapid model inference unit, a parameter optimization solution unit, a convergence judgment unit, and an accuracy verification unit. The rapid model inference unit stores offline pre-constructed reduced-order basis matrices or surrogate model weights, receives temperature-load condition inputs online, and outputs the simulation response values ​​of monitoring points through matrix projection or neural network forward inference. The parameter optimization solution unit uses the weighted sum of squared residuals between monitoring data and simulation response values ​​as the objective function to iteratively update the shape parameters, phase angle parameters, and interlayer bond stiffness parameters of the complex modulus master curve of asphalt mixture. The convergence judgment unit stops iterating when the objective function value drops below a preset threshold. The accuracy verification unit performs root mean square error verification on the converged parameters; if the standard is not met, parameter optimization is retried.

[0084] The global information extraction module is used to calculate the global mechanical field, identify areas of damage exceeding limits, and generate performance cloud maps. Specifically, it includes a global field calculation unit, a damage exceeding limit judgment unit, and a cloud map rendering unit. The global field calculation unit calls a fast mechanical calculation model based on the corrected digital twin to calculate the stress field, strain field, interlayer shear stress field, and fatigue damage field of the entire road area. The damage exceeding limit judgment unit compares the fatigue damage field with the preset critical value grid by grid, and marks the location station number, suspected defect type, and treatment priority of the exceeding limit area. The cloud map rendering unit generates a performance distribution cloud map with a spatial resolution of not less than 0.1 m × 0.1 m.

[0085] A long-term evolution storage module is used to store time-series performance data and predict future degradation trajectories. Specifically, it includes a time-series database unit and a performance prediction unit. The time-series database unit stores the material constitutive parameters and fatigue damage variables obtained from each parameter inversion according to timestamps, and synchronously stores the corresponding historical environmental data and historical traffic load data. The database storage period is no less than 10 years. The performance prediction unit uses at least one of ARIMA, Long Short-Term Memory Network (LSTM), or state-space model based on the time-series data, combined with the environmental-load coupling effect, to predict the future performance degradation trajectory of the pavement, and generates a preventive maintenance decision report containing maintenance intervention timing and measure recommendations.

[0086] Furthermore, the virtual-real real-time interaction module also includes an edge computing unit, which is deployed in a roadside cabinet for localized execution parameter inversion calculation and supports automatic switching to offline operation mode when the network is interrupted; the edge computing unit uploads the calculation results to the cloud platform via 5G or Ethernet communication.

[0087] Furthermore, the global information extraction module also includes an augmented reality (AR) visualization unit. The AR visualization unit automatically matches the digital twin data of the current location with the positioning information of the mobile terminal, and then superimposes and renders the damage distribution information with the real road surface image after spatial registration, so as to realize the perspective view of hidden defects inside the road surface and the real-time annotation of defect types and treatment suggestions.

[0088] Furthermore, the present invention also provides a specific application embodiment, the details of which are as follows:

[0089] Example 1: Real-time reconstruction of global mechanical behavior based on a reduced-order POD model

[0090] This embodiment takes a test section of an asphalt concrete pavement on a highway that has been open to traffic for 5 years as the subject.

[0091] Step 1: System Deployment and Data Acquisition

[0092] Directly beneath the wheel tracks of the test section, a sensor cluster (6 clusters in total) was deployed every 100 meters. Each cluster included a fiber optic strain sensor embedded at the bottom of the surface layer, a temperature sensor in the middle of the surface layer, and a compressive stress sensor at the top of the base layer. The sensors transmitted data via a wireless network. Simultaneously, the system accessed data from meteorological stations along the route and obtained real-time traffic axle load and traffic flow data from a dynamic weighing system through a standardized interface.

[0093] Step 2, Model Building

[0094] Physical twin baseline model: A three-dimensional model of the road section was established using finite element software, including the top layer, intermediate layer, bottom layer, and base layer. The asphalt layer adopted a viscoelastic constitutive model based on the principal curve of the Sigmoidal function, the base layer adopted a linear elastic model, and the interlayer was simulated using a cohesive model. The initial material parameters of the model were determined based on laboratory tests.

[0095] Fast Mechanical Calculation Model (POD Reduced-Order Model): Based on the aforementioned full-order model, the mechanical response of 1000 typical working conditions is calculated offline within the parameter spaces of temperature (-30℃~60℃) and axle load (50kN~200kN). By performing eigenorthogonal decomposition on the displacement field, the first 50 modes are extracted to construct a reduced-order basis matrix, which covers 99.2% of the energy of the original displacement field. In online applications, the full-order equations are projected onto this reduced-order basis space for rapid solution.

[0096] Step 3: Parameter Inversion and Real-time Correction

[0097] Taking the monitoring data at 14:00 on a certain day as an example, the system reads the measured strain, current road surface temperature, and axle load information from each sensor. In the initial state, the simulated strain of the monitoring points calculated by calling the POD reduced-order model deviates from the measured value (for example, the simulated value of a certain point is 120με, while the measured value is 150με).

[0098] The system constructs the objective function using weighted least squares and employs Bayesian inference (MCMC sampling) as the optimization algorithm. The dynamic modulus of the surface layer and intermediate layer, as well as the interlayer bond stiffness, are used as parameters to be inverted, with reasonable physical constraints set (e.g., the surface layer modulus search range is 50%-100% of the initial laboratory value). During the inversion iteration, the POD reduced-order model (approximately 0.8 seconds per calculation) is used for each calculation of the simulated response at the monitoring points, instead of the full-order model (approximately 300 seconds). After iteration, the objective function converges below the threshold. The inversion results show that the surface layer dynamic modulus decays from the initial 6000 MPa to 4750 MPa, and the interlayer bond stiffness decreases to 0.41 MPa, reflecting material damage. Substituting the inversion parameters back into the model, the simulated values ​​at each monitoring point agree well with the measured values, with a root mean square error of less than 2%, verifying the inversion accuracy.

[0099] Step 4: Global Reconstruction and Disease Identification

[0100] Based on the revised digital twin (parameters updated), the POD reduced-order model was called again to calculate the stress, strain, and fatigue damage field distribution at all locations along the entire 500-meter section. The calculations revealed that at the bottom of the mid-layer in the section from K100+200 to K100+260, the fatigue damage factor exceeded the preset critical value of 0.8. Simultaneously, the interlayer shear stress in this area was also abnormally high. Based on this, the system determined that there was a high risk of "poor interlayer bonding" at this location and highlighted it in the generated global damage cloud map, providing the location's station number and suggested treatment priority. Subsequent on-site core drilling verified that the measured interlayer bonding strength at this location was only 0.13 MPa, highly consistent with the system's assessment, achieving an accuracy rate of over 95%.

[0101] Step 5: Long-term performance prediction and maintenance decision-making

[0102] The system continuously retrieves parameters such as material modulus and damage factor, along with corresponding temperature and axle load data, and stores them in a database in time series. Analysis of data from the past year reveals a regular decline in the surface layer modulus. Using a time series model and considering the environmental-load coupling effect, the system predicts that the surface layer modulus will drop to a critical value requiring maintenance in the 8th year after the road opens to traffic. Based on this, the system generates a decision report recommending the implementation of a thin overlay preventative maintenance in the 7th year, which is expected to extend the pavement's service life by more than 5 years.

[0103] Example 2 - A Lightweight Alternative Based on the PINN Proxy Model

[0104] This embodiment demonstrates an alternative solution for scenarios where edge computing devices have limited memory resources. The road surface structure and monitoring layout are the same as in Embodiment 1.

[0105] In the model building phase, a Physical Information Neural Network (PINN) is used to construct the surrogate model. Using 5000 sets of offline simulation samples as the training set, the PINN loss function is a weighted sum of data residuals, mechanical control equation residuals (stress equilibrium equations, viscoelastic constitutive equations), and boundary condition residuals. After training, the model weight file is approximately 50MB.

[0106] In the parameter inversion stage, the particle swarm optimization algorithm was used as the optimizer, and the PINN model was invoked for forward inference in each evaluation (approximately 0.08 seconds per iteration). The final inverted material parameters differed from those in Example 1 by less than 3%, and the complete inversion took approximately 8.5 seconds, also meeting the real-time requirements. The global damage field reconstructed based on the PINN surrogate model successfully identified the same high-risk region, verifying the effectiveness of this lightweight solution.

[0107] Comparison with existing technologies

[0108] To objectively verify the effectiveness of this invention, a comparative test was conducted on the same test section, comparing this invention (using a POD reduced-order model) with two representative existing technologies: (A) traditional fixed-parameter finite element forward simulation; (B) pure kriging space interpolation based on sensor data (without a physical model). Through parallel comparative experiments conducted on the same test section (a section of a highway from K100+000 to K100+500, open to traffic for 5 years), the three methods used the same pavement structure parameters, sensor layout scheme, and evaluation period. The experimental conditions were completely consistent with Example 1 of this specification, and the test results were reproducible. The results are shown in Table 1 below.

[0109]

[0110] As can be seen from the table above, the method of the present invention is significantly superior to methods A and B in four dimensions: global reconstruction accuracy, blind zone coverage capability, accuracy of hidden defects identification, and adaptive correction of material parameters. Moreover, it achieves a real-time response speed that is usable in engineering while ensuring physical interpretability. The above comprehensive effect cannot be achieved by using forward simulation or pure data-driven methods alone, which reflects the creative contribution of the data-mechanism dual-driven fusion scheme of the present invention.

[0111] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A sparse data-driven method for reconstructing the global mechanical field of asphalt pavement, characterized in that, The method includes the following steps: Step 1: Acquire monitoring data from the sensor array deployed inside the asphalt pavement structure. The monitoring data includes mechanical response, ambient temperature, and traffic load data. Step 2: Construct a three-dimensional finite element physical twin reference model of asphalt pavement, wherein the asphalt mixture layer adopts a viscoelastic constitutive relation, the interlayer contact adopts a cohesive contact model with the interlayer bond stiffness parameter as the state variable, and the material constitutive parameters of the physical twin reference model are defined in the form of variable parameters; and generate a mechanical response sample library covering the preset temperature and load parameter space offline based on the physical twin reference model, and construct a fast mechanical calculation model based on the mechanical response sample library, wherein the fast mechanical calculation model is a reduced-order model or a surrogate model; Step 3: Using the monitoring data from Step 1 as constraints, the rapid mechanical calculation model from Step 2 is called for iterative simulation to construct an objective function characterizing the deviation between the simulated response value and the measured value of the monitoring point. An optimization algorithm is then used to dynamically adjust the material constitutive parameters of the physical twin benchmark model. The material constitutive parameters include the complex modulus master curve parameters and interlayer bond stiffness parameters of the asphalt mixture, until the objective function converges to a preset threshold, thereby obtaining a digital twin that is in real-time fitted to the current pavement condition. Step 4: Based on the real-time digital twin obtained in Step 3, the rapid mechanical calculation model is called to reconstruct the global strain field, modulus field, stress field and damage field. Based on the comparison between the damage field and the preset critical value, potential hidden defects high-risk areas are identified, and a full-domain performance distribution cloud map is generated.

2. The method according to claim 1, characterized in that, In step 1, the sensor array is deployed at the interface between each structural layer directly below the wheel track on the road surface.

3. The method according to claim 1, characterized in that, In step 2, the fast mechanical calculation model is at least one of the following: a reduced-order model based on intrinsic orthogonal decomposition, a surrogate model based on physical information neural network, and a radial basis function neural network.

4. The method according to claim 3, characterized in that, In step 2, the fast mechanics calculation model dynamically switches between the reduced-order model and different proxy models based on the available memory capacity of the edge computing nodes.

5. The method according to claim 1, characterized in that, In step 2, the reduced-order model performs eigenorthogonal decomposition on the displacement field of the mechanical response sample library and extracts the first N eigenmodes to construct a reduced-order basis matrix, where the value of N is used to ensure that the reduced-order basis matrix covers more than 99% of the energy of the original displacement field; during online calculation, the high-dimensional finite element control equations are projected onto the reduced-order basis space for solution.

6. The method according to claim 1, characterized in that, In step 3, the objective function is the weighted sum of squares of the deviations between the simulated and measured values ​​at each monitoring point, and the specific formula is as follows: ； Where J is the objective function. Let i be the simulated response value of the i-th monitoring point. Let i be the measured value of the i-th monitoring point. These are the weighting coefficients.

7. The method according to claim 1, characterized in that, In step 3, the material constitutive parameters also include phase angle parameters and elastic modulus of each structural layer.

8. The method according to claim 1, characterized in that, In step 3, the optimization algorithm includes Bayesian inference, genetic algorithm, or particle swarm optimization.

9. The method according to claim 1, characterized in that, Step 3 also includes an accuracy verification step: after the objective function converges, the inversion parameters are substituted into the fast mechanical calculation model to recalculate the response of all monitoring points. When the root mean square error between the simulated value and the measured value is less than 5% of the average level of the measured value, the reconstruction is confirmed to be effective; otherwise, the inversion iteration is retried.

10. The method according to claim 1, characterized in that, In step 4, when the interlaminar shear stress in the high-risk area exceeds the interlaminar bonding material strength threshold, it is determined to be poor interlaminar bonding; when the tensile strain at the bottom of the surface layer in the high-risk area exceeds the fatigue failure limit strain, it is determined to be surface layer fatigue cracking.

11. The method according to any one of claims 1 to 10, characterized in that, The method further includes: Step 5, storing the material constitutive parameters obtained from the inversion in Step 3 and the fatigue damage variables from Step 4 in a time series manner, and constructing a performance evolution database by synchronously storing historical environmental data and historical traffic load data at the corresponding time; based on the performance evolution database, using time series analysis methods and combining the coupling effects of environment and load, predicting the pavement performance degradation trajectory; when the predicted time when the predicted material modulus is lower than the maintenance intervention critical value is less than a preset time threshold, generating a preventive maintenance decision report containing maintenance intervention timing and measure recommendations.

12. A sparse data-driven global mechanical field reconstruction system for asphalt pavement, used to implement the method according to any one of claims 1-11, characterized in that, The system includes: The multi-source sensing module is used to acquire real-time data from sensors, the environment, and the load. The digital twin modeling module is used to construct the physical twin benchmark model and the rapid mechanical calculation model. The virtual-real interaction and inversion module is used to execute online parameter inversion closed loop and correct the digital twin in real time; The global field analysis and diagnosis module is used to reconstruct the global mechanical field, identify hidden defects, and generate a cloud map of the performance distribution of the entire road domain.

Images