Asphalt pavement fatigue damage model calibration method based on AI and digital twin

Abstract

The present invention discloses a method for calibrating asphalt pavement fatigue damage models based on AI and digital twins, relating to the field of damage detection technology. Compared with previous methods, this method solves the problem that traditional asphalt pavement design relies on empirical formulas and static parameters, making it difficult to dynamically reflect material aging, environmental coupling, and load uncertainty, and lacks cross-scale correlation between microscopic interface behavior and macroscopic structural performance. Through multimodal data fusion and dynamic perception systems, multi-source heterogeneous data is collected and preprocessed in real time. Combined with molecular dynamics simulation, the asphalt-aggregate interface binding energy and adhesion work are accurately quantified to construct a cross-scale damage evolution model. Utilizing a physical information neural network embedded in the improved Paris formula, the measured data is deeply integrated with physical laws, effectively correcting the locality assumptions of traditional models.

Description

Technical Field

[0001] The present invention relates to the field of damage detection technology, and in particular to an asphalt pavement fatigue damage model calibration method based on AI and digital twins. Background Art

[0002] Traditional asphalt pavement design has long relied on empirical formulas and laboratory-calibrated static parameters. Asphalt materials undergo irreversible processes such as oxidation aging and aggregate interface adhesion degradation during service, while traditional model parameters use fixed values from short-term laboratory tests and cannot reflect the nonlinear attenuation of material properties over time. Actual pavements are subject to the coupling of cyclic loads and complex environmental fields, while empirical formulas can only independently handle a single factor. There is a lack of physical correlation between macroscopic fatigue damage models and microscopic interface behaviors. Molecular dynamics simulations show that a 10% reduction in aggregate surface texture roughness will reduce interfacial binding energy by 25%, but traditional designs do not incorporate nanoscale interface characteristics into macroscopic models, resulting in a 2-3 times deviation between laboratory small specimen test results and the actual measured life of full-scale pavement.

[0003] With the increasing demand for long-life pavements, the parameter systems established by traditional methods based on indoor accelerated testing are difficult to extrapolate to the long-term performance evolution under the coupling of low-frequency loads and wide temperature ranges in real environments. For example, data from highway research projects show that the fatigue cracking rate of pavements designed using traditional methods within five years of opening to traffic is 65% higher than expected, highlighting the urgent need for cross-scale dynamic modeling and intelligent parameter updating.

[0004] In order to solve the above problems, the present invention proposes an asphalt pavement fatigue damage model calibration method based on AI and digital twins. Summary of the Invention

[0005] The purpose of this invention is to propose an asphalt pavement fatigue damage model calibration method based on AI and digital twins to solve the problems raised in the background technology:

[0006] Traditional asphalt pavement design relies on empirical formulas and static parameters, which makes it difficult to dynamically reflect material aging, environmental coupling, and load uncertainty, and lacks cross-scale correlation between microscopic interface behavior and macroscopic structural performance.

[0007] In order to achieve the above object, the present invention adopts the following technical solutions:

[0008] The asphalt pavement fatigue damage model calibration method based on AI and digital twins includes the following steps:

[0009] Collect road surface condition data and perform preprocessing and spatiotemporal alignment processing in sequence to obtain road surface condition characteristics;

[0010] The energy evolution trajectory of the asphalt-aggregate interface was generated through LAMMPS molecular dynamics simulation, and the interfacial binding energy was calculated based on the energy evolution trajectory. The interfacial binding energy was mapped to adhesion work, and a damage model was constructed based on the adhesion work and the Paris formula.

[0011] The road surface condition characteristics are used as inputs to the physical information neural network to obtain the damage rate and model correction value; the damage model is corrected according to the model correction value to obtain the fatigue damage model;

[0012] Collect test data from pre-set asphalt specimens, including strain, stiffness, and number of load cycles. Use the test data as input to the first channel of a two-channel deep neural network to obtain weights and material constants. Use the pavement state characteristics as input to the second channel of the two-channel deep neural network to obtain aligned baseline features.

[0013] The experimental data and dynamic modulus test data of a specimen with a preset size are obtained and standardized into a dimensionless parameter set. Based on the dimensionless parameter set, damage rate, and aligned benchmark features, the model parameters in the fatigue damage model are optimized using reinforcement learning.

[0014] The asphalt pavement fatigue damage model is calibrated through full life cycle self-evolution through an online-offline dual-loop control strategy.

[0015] Preferably, the road surface condition data is obtained by pre-embedded multi-type sensor arrays in the full-scale accelerated loading test section and vehicle-mounted LiDAR.

[0016] Preferably, the road surface state data includes three-dimensional strain data, environmental data, acoustic emission signals, and three-dimensional road surface deformation data. The method for obtaining road surface state characteristics includes:

[0017] A median filter is used to eliminate noise in the three-dimensional strain data to obtain a de-noised strain sequence. The environmental data is smoothed based on a Kalman filter to obtain smoothed environmental data. The energy accumulation curve is extracted based on the Hilbert transform. The surface crack length is calculated from the three-dimensional pavement deformation data using the RANSAC algorithm. The de-noised strain sequence, smoothed environmental data, energy accumulation curve, and surface crack length are dynamically fused using a spatiotemporal alignment algorithm to obtain pavement condition characteristics.

[0018] Preferably, the method of generating the energy evolution trajectory of the asphalt-aggregate interface through LAMMPS molecular dynamics simulation includes:

[0019] Combining the preset atomic types and force field parameters, an asphalt-aggregate interface model was constructed through LAMMPS molecular dynamics simulation. The model parameters of the asphalt-aggregate interface model were obtained based on the pavement condition data. Based on the preset running step size, the energy evolution trajectory of the asphalt-aggregate interface was obtained.

[0020] Preferably, the method for calculating the interface binding energy based on the energy evolution trajectory includes:

[0021] The energy mean after equilibrium state is extracted from the energy evolution trajectory, and the interfacial binding energy is calculated based on the energy of the individual asphalt system, the energy of the individual aggregate system and the interfacial contact area.

[0022] Preferably, the method of mapping the interfacial binding energy into adhesion work comprises:

[0023] According to the theory of interface physical chemistry, the adhesion work and the interface binding energy are equal in thermodynamic equilibrium and other energy dissipation is ignored. The calculated interface binding energy is directly mapped to the adhesion work.

[0024] Preferably, the first channel in the dual-channel deep neural network is composed of an input layer, a convolutional layer, an LSTM layer, a fully connected layer and an output layer, and the second channel in the dual-channel deep neural network is composed of an input layer, a contrastive learning module, a feature alignment module, a shared fully connected layer and an output layer; wherein, the output layer of the first channel and the input layer of the second channel are jointly input into the contrastive learning module, and the contrastive learning module calculates the similarity between the feature vector input into the output layer of the first channel and the road surface state feature input into the input layer of the second channel, and amplifies the difference in similarity between different features through logarithmic operation to construct a loss function.

[0025] Preferably, the method for optimizing model parameters in the fatigue damage model based on reinforcement learning includes:

[0026] Define the state space S: pavement state characteristics, initial values of model parameters in the fatigue damage model, aligned benchmark characteristics, and dimensionless parameter sets;

[0027] Define the action space A: adjust the model parameters in the fatigue damage model;

[0028] Design reward function: the inverse of the error between the predicted value and the actual value of the fatigue damage model;

[0029] Update the Q-value function: Use the Q-learning algorithm to learn the optimal strategy, initialize the Q-value function as a table, where the rows correspond to different environmental states and the columns correspond to different actions, set the learning rate, discount factor, and exploration rate; at each time step, detect the current environmental state, and randomly select an action with a probability of ε based on the exploration rate, and select the action with the largest Q value under the current environmental state with a probability of 1-ε; execute the selected action to obtain the new environmental state, and at the same time, calculate the total reward function value based on the reward function; update the Q-value function according to the update formula;

[0030] Repeatedly update the Q-value function until the Q-value converges to obtain the corresponding optimal Q-value function. According to the environmental state, the optimal action is selected from the action space using the optimal Q-value function and executed.

[0031] Preferably, the method for performing full life cycle self-evolution calibration of the asphalt pavement fatigue damage model through the online-offline dual-loop control strategy includes:

[0032] The model parameters in the fatigue damage model are updated in real time based on the extended Kalman filter. The GAN network is used to generate virtual data input into the fatigue damage model, and the virtual data is input into the fatigue damage model for training enhancement.

[0033] Compared with the existing technology, the present invention provides an asphalt pavement fatigue damage model calibration method based on AI and digital twins, which has the following beneficial effects:

[0034] The present invention uses a multimodal data fusion and dynamic perception system to collect and preprocess multi-source heterogeneous data in real time, combines molecular dynamics simulation to accurately quantify the asphalt-aggregate interface binding energy and adhesion work, and constructs a cross-scale damage evolution model; uses a physical information neural network to embed an improved Paris formula to deeply integrate measured data with physical laws, effectively correcting the locality assumptions of traditional models; realizes cross-domain migration of laboratory data and full-scale tests through dual-channel deep neural networks and reinforcement learning, dynamically optimizes model parameters, and reduces prediction errors; drives the adaptive evolution of the model through an online-offline dual-loop control strategy, significantly improving the robustness in complex environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flow chart of the method mentioned in Example 1 of the present invention;

[0036] Figure 2 Schematic diagram of the dual-channel deep neural network structure mentioned in Example 1 of the present invention. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0038] The present invention uses a multimodal data fusion and dynamic perception system to collect and pre-process multi-source heterogeneous data in real time, combines molecular dynamics simulation to accurately quantify the asphalt-aggregate interface binding energy and adhesion work, and constructs a cross-scale damage evolution model. It uses a physical information neural network embedded in the improved Paris formula to deeply integrate measured data with physical laws, effectively correcting the locality assumptions of traditional models. It uses a dual-channel deep neural network and reinforcement learning to achieve cross-domain migration of laboratory data and full-scale tests, dynamically optimize model parameters, and reduce prediction errors. It drives the adaptive evolution of the model through an online-offline dual-loop control strategy, significantly improving the robustness in complex environments. Specifically, it includes the following contents.

[0039] Example 1:

[0040] See also Figure 1-2 The present invention provides an asphalt pavement fatigue damage model calibration method based on AI and digital twins, comprising the following steps:

[0041] Collect 3D strain data, environmental data, acoustic emission signals, and 3D road surface deformation data, and perform pre-processing and spatiotemporal alignment in sequence to obtain road surface state characteristics;

[0042] A multi-type sensor array is embedded in the full-scale accelerated loading test section to collect strain data, environmental data, and acoustic emission signals. Combined with the on-board LiDAR, the three-dimensional deformation data of the road surface is collected in real time. The multi-type sensor array consists of FBG fiber optic strain sensors, MEMS temperature and humidity sensors, and acoustic emission sensors. The FBG fiber optic strain sensor is used to monitor the strain data of the asphalt layer, the MEMS temperature and humidity sensor is used to capture temperature data and moisture data, and the acoustic emission sensor is used to locate the sound source of crack propagation in real time. The test vehicle combined with the on-board LiDAR can synchronously obtain three-dimensional deformation data of the road surface.

[0043] A median filter is used to eliminate noise in the three-dimensional strain data to obtain a de-noised strain sequence. The environmental data is smoothed based on a Kalman filter to obtain smoothed environmental data. The energy accumulation curve is extracted based on the Hilbert transform. The surface crack length is calculated from the three-dimensional pavement deformation data using the RANSAC algorithm. The de-noised strain sequence, smoothed environmental data, energy accumulation curve, and surface crack length are dynamically fused using a spatiotemporal alignment algorithm to obtain pavement condition characteristics.

[0044] After collecting pavement condition data, preprocessing removes noise interference to improve data validity. Spatiotemporal alignment then unifies the temporal and spatial benchmarks of multi-source data, establishing logical connections between heterogeneous data. The resulting pavement condition characteristics provide precise, standardized input variables for the asphalt pavement fatigue damage model, supporting dynamic tuning of model parameters during intelligent calibration. This ensures the model more accurately maps pavement fatigue damage mechanisms, thereby improving the accuracy and reliability of damage evolution predictions and enabling intelligent, adaptive iteration of the calibration process.

[0045] The energy evolution trajectory of the asphalt-aggregate interface is generated through LAMMPS molecular dynamics simulation, and the interface binding energy is calculated based on the energy evolution trajectory; the interface binding energy is mapped to adhesion work, and the damage model is constructed based on the adhesion work and the Paris formula; ;in, is the crack growth rate; is the material constant; is the gradient term coefficient; is the strain gradient; is the exponential parameter of the strain gradient term; is the stress intensity factor range; is the exponential parameter of the stress intensity factor range.

[0046] By using LAMMPS molecular dynamics simulations to generate the energy evolution trajectory of the asphalt-aggregate interface, the interfacial binding energy is calculated and mapped to adhesion work, providing microscopic physical mechanism support for the intelligent calibration of asphalt pavement fatigue damage models. Adhesion work, as a cross-scale transfer parameter, directly links microscopic interfacial behavior with macroscopic damage evolution. The damage model constructed in conjunction with the Paris formula can quantify the impact of interfacial bonding on crack propagation. In the subsequent intelligent calibration process, dynamic optimization of model parameters through reinforcement learning or data-driven methods enables the model to simultaneously reflect the combined effects of the material's intrinsic properties and the service environment, improving the physical rationality and engineering adaptability of fatigue damage prediction, and ultimately achieving cross-scale, precise calibration from microscopic interfacial behavior to macroscopic fatigue life.

[0047] Methods for generating energy evolution trajectories of the asphalt-aggregate interface through LAMMPS molecular dynamics simulations include:

[0048] Combining the preset atomic types and force field parameters, an asphalt-aggregate interface model was constructed through LAMMPS molecular dynamics simulation. The model parameters of the asphalt-aggregate interface model were obtained based on the pavement condition data. Based on the preset running step size, the energy evolution trajectory of the asphalt-aggregate interface was obtained.

[0049] Methods for calculating interfacial binding energy based on energy evolution trajectories include:

[0050] The energy mean after equilibrium state is extracted from the energy evolution trajectory, and the interfacial binding energy is calculated based on the energy of the individual asphalt system, the energy of the individual aggregate system and the interfacial contact area.

[0051] Methods for mapping interfacial binding energy to work of adhesion include:

[0052] According to the theory of interface physical chemistry, the adhesion work and the interface binding energy are equal in thermodynamic equilibrium and other energy dissipation is ignored. The calculated interface binding energy is directly mapped to the adhesion work.

[0053] The road surface state characteristics are used as the input of the physical information neural network to obtain the damage rate and model correction value; the damage model is corrected according to the model correction value to obtain the fatigue damage model; ;in, is the crack growth rate; is the material constant; is the gradient term coefficient; is the strain gradient; is the exponential parameter of the strain gradient term; is the stress intensity factor range; is the exponential parameter of the stress intensity factor range; It is a data-driven correction item that is dynamically updated through real-time monitoring data; This is the fatigue damage model correction value output by the physical information neural network. The physical information neural network is a novel neural network model that combines physical laws with deep learning. During training, the physical information neural network incorporates known physical laws of pavement damage (such as differential equations and conservation laws) as constraints, enabling the model to not only fit the data but also strictly adhere to physical laws.

[0054] Pavement condition characteristics (multimodal data such as strain energy density, temperature gradient, and crack length) are input into a physical information neural network (PINN). By integrating physical equations (such as the Paris formula) with a data-driven mechanism, this method achieves the dual functions of damage rate prediction and model parameter correction. The damage rate output by the PINN provides a benchmark for model validation, while the model correction value dynamically calibrates the physical parameters of the damage model to better reflect actual operating conditions. This approach optimizes model parameters through real-time data feedback, enhancing the physical interpretability and engineering adaptability of fatigue damage prediction. It also forms a closed loop of "data acquisition-PINN correction-model update," promoting the intelligent and adaptive calibration of asphalt pavement fatigue damage models and ultimately improving the accuracy of long-term service performance assessments.

[0055] Test data from pre-set asphalt specimens, including strain, stiffness, and number of load cycles, is collected. This data is used as input to the first channel of a two-channel deep neural network to obtain weights and material constants. The pavement condition characteristics are used as input to the second channel of the two-channel deep neural network to obtain aligned benchmark features. This two-channel deep neural network integrates laboratory and field data to achieve intelligent calibration of asphalt pavement fatigue damage models. The first channel inputs the strain, stiffness, and number of load cycles of the asphalt specimens to learn intrinsic material properties (such as Paris formula parameters) and outputs weights and material constants. The second channel uses the spatiotemporally aligned pavement condition characteristics as input to extract benchmark features under service conditions. The two-channel features are then fused through a network mechanism to generate comprehensive parameters, which are then used to dynamically modify the physical equation parameters and reinforcement learning parameters in the damage model. This method combines material properties from precisely controlled laboratory conditions with complex field data, ensuring the physical rationality of the model while improving its adaptability to real-world conditions. Ultimately, this method enables intelligent, multi-source data-driven calibration of fatigue damage model parameters, enhancing the accuracy and robustness of long-term performance predictions.

[0056] Reference Figure 2 The first channel in the dual-channel deep neural network consists of an input layer, a convolutional layer, an LSTM layer, a fully connected layer, and an output layer. The second channel in the dual-channel deep neural network consists of an input layer, a contrastive learning module, a feature alignment module, a shared fully connected layer, and an output layer. The output layer of the first channel and the input layer of the second channel are input into the contrastive learning module together. The contrastive learning module calculates the similarity between the feature vector input into the output layer of the first channel and the road surface state feature input into the input layer of the second channel, and amplifies the difference in similarity between different features through logarithmic operations to construct a loss function, thereby driving the dual-channel deep neural network to learn more discriminative feature representations.

[0057] Experimental data and dynamic modulus test data for specimens of preset dimensions are obtained and standardized into a dimensionless parameter set. Based on this dimensionless parameter set, reinforcement learning is used to optimize the model parameters in the fatigue damage model. The dimensionless parameter set conveys the intrinsic material properties under precise laboratory control, while the baseline characteristics reflect the heterogeneity of the complex field environment. These two are integrated in the state space of reinforcement learning to drive adaptive adjustment of model parameters. With the damage rate as the optimization objective, reinforcement learning dynamically balances the contribution weights of laboratory parameters and field characteristics through a policy gradient algorithm, ensuring that the fatigue damage model conforms to both the physical laws of the material and actual service behavior. Dimensional interference is eliminated by standardizing the data, and the model parameters are updated in real time using online iterations of reinforcement learning. This forms a closed-loop system of "data input-strategy optimization-model modification-verification feedback," which improves the accuracy and robustness of fatigue damage prediction.

[0058] Methods for optimizing model parameters in fatigue damage models based on reinforcement learning include:

[0059] Define the state space S: pavement state characteristics, initial values of model parameters in the fatigue damage model, aligned benchmark characteristics, and dimensionless parameter sets;

[0060] Define the action space A: adjust the model parameters in the fatigue damage model;

[0061] Design a reward function: the inverse of the error between the fatigue damage model’s predicted value and the actual value; the smaller the error, the higher the reward function value, which can drive the agent to optimize parameters.

[0062] Update the Q-value function: Use the Q-learning algorithm to learn the optimal strategy, initialize the Q-value function as a table, where the rows correspond to different environmental states and the columns correspond to different actions, set the learning rate, discount factor, and exploration rate; at each time step, detect the current environmental state, and randomly select an action with a probability of ε based on the exploration rate, and select the action with the largest Q value under the current environmental state with a probability of 1-ε; execute the selected action to obtain the new environmental state, and at the same time, calculate the total reward function value based on the reward function; update the Q-value function according to the update formula;

[0063] Repeatedly update the Q-value function until the Q-value converges to obtain the corresponding optimal Q-value function. According to the environmental state, the optimal action is selected from the action space using the optimal Q-value function and executed.

[0064] The asphalt pavement fatigue damage model is calibrated through full life cycle self-evolution through an online-offline dual-loop control strategy.

[0065] The model parameters in the fatigue damage model are updated in real time based on the extended Kalman filter. The GAN network is used to generate virtual data input into the fatigue damage model, and the virtual data is input into the fatigue damage model for training enhancement.

[0066] An online-offline dual-loop control strategy enables full-lifecycle self-evolutionary calibration of asphalt pavement fatigue damage models. Its core functions are as follows: the online loop processes minute-level strain and temperature data in real time through edge computing, dynamically updating model parameters using the Extended Kalman Filter (EKF) to ensure rapid response to changes in pavement conditions; the offline loop deeply pre-trains the fatigue damage model based on historical data and extreme operating condition data generated by a Generative Adversarial Network (GAN), enhancing its adaptability to unusual environments. These two loops work together to form a closed-loop mechanism of "real-time feedback correction - extreme operating condition enhancement - model generalization improvement": online parameter updates ensure short-term prediction accuracy, while offline enhancements optimize long-term robustness. This strategy enables the fatigue damage model to continuously evolve throughout its service life, calibrating short-term damage predictions with real-time data and enhancing long-term durability assessment capabilities through virtual data pre-training. Ultimately, this strategy enables intelligent management of the fatigue damage model throughout its service life, from data acquisition to fatigue damage model optimization.

[0067] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. The asphalt pavement fatigue damage model calibration method based on AI and digital twin is characterized by: The steps include: Collect road surface condition data and perform preprocessing and spatiotemporal alignment processing in sequence to obtain road surface condition characteristics; The energy evolution trajectory of the asphalt-aggregate interface was generated through LAMMPS molecular dynamics simulation, and the interfacial binding energy was calculated based on the energy evolution trajectory. The interfacial binding energy was mapped to adhesion work, and a damage model was constructed based on the adhesion work and the Paris formula. The road surface condition characteristics are used as inputs to the physical information neural network to obtain the damage rate and model correction value; the damage model is corrected according to the model correction value to obtain the fatigue damage model; Collect test data from pre-set asphalt specimens, including strain, stiffness, and number of load cycles. Use the test data as input to the first channel of a two-channel deep neural network to obtain weights and material constants. Use the pavement state characteristics as input to the second channel of the two-channel deep neural network to obtain aligned baseline features. The experimental data and dynamic modulus test data of the preset specimen are obtained and standardized into a dimensionless parameter set. Based on the dimensionless parameter set, damage rate, and aligned benchmark features, the model parameters in the fatigue damage model are optimized based on reinforcement learning, including: Define the state space S: pavement state characteristics, initial values of model parameters in the fatigue damage model, aligned benchmark characteristics, and dimensionless parameter sets; Define the action space A: adjust the model parameters in the fatigue damage model; Design reward function: the inverse of the error between the predicted value and the actual value of the fatigue damage model; Update the Q-value function: Use the Q-learning algorithm to learn the optimal strategy, initialize the Q-value function as a table, where the rows correspond to different environmental states and the columns correspond to different actions, set the learning rate, discount factor, and exploration rate; at each time step, detect the current environmental state, and randomly select an action with a probability of ε based on the exploration rate, and select the action with the largest Q value under the current environmental state with a probability of 1-ε; execute the selected action to obtain the new environmental state, and at the same time, calculate the total reward function value based on the reward function; update the Q-value function according to the update formula; Repeatedly update the Q-value function until the Q-value converges to obtain the corresponding optimal Q-value function. According to the environment state, use the optimal Q-value function to select the optimal action from the action space and execute it; The asphalt pavement fatigue damage model is calibrated through full life cycle self-evolution through an online-offline dual-loop control strategy.

2. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The road surface condition data is obtained through a multi-type sensor array pre-buried in a full-scale accelerated loading test section and vehicle-mounted LiDAR.

3. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The road surface state data includes three-dimensional strain data, environmental data, acoustic emission signals, and three-dimensional road surface deformation data. The method for obtaining road surface state characteristics includes: A median filter is used to eliminate noise in the three-dimensional strain data to obtain a de-noised strain sequence. The environmental data is smoothed based on a Kalman filter to obtain smoothed environmental data. The energy accumulation curve is extracted based on the Hilbert transform. The surface crack length is calculated from the three-dimensional pavement deformation data using the RANSAC algorithm. The de-noised strain sequence, smoothed environmental data, energy accumulation curve, and surface crack length are dynamically fused using a spatiotemporal alignment algorithm to obtain pavement condition characteristics.

4. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The method for generating the energy evolution trajectory of the asphalt-aggregate interface through LAMMPS molecular dynamics simulation includes: Combining the preset atomic types and force field parameters, an asphalt-aggregate interface model was constructed through LAMMPS molecular dynamics simulation. The model parameters of the asphalt-aggregate interface model were obtained based on the pavement condition data. Based on the preset running step size, the energy evolution trajectory of the asphalt-aggregate interface was obtained.

5. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The method for calculating the interface binding energy based on the energy evolution trajectory includes: The energy mean after equilibrium state is extracted from the energy evolution trajectory, and the interfacial binding energy is calculated based on the energy of the individual asphalt system, the energy of the individual aggregate system and the interfacial contact area.

6. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The method for mapping the interfacial binding energy into adhesion work comprises: According to the theory of interface physical chemistry, the adhesion work and the interface binding energy are equal in thermodynamic equilibrium and other energy dissipation is ignored. The calculated interface binding energy is directly mapped to the adhesion work.

7. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The first channel in the dual-channel deep neural network consists of an input layer, a convolutional layer, an LSTM layer, a fully connected layer and an output layer, and the second channel in the dual-channel deep neural network consists of an input layer, a contrastive learning module, a feature alignment module, a shared fully connected layer and an output layer; wherein the output layer of the first channel and the input layer of the second channel are jointly input into the contrastive learning module, and the contrastive learning module calculates the similarity between the feature vector input into the output layer of the first channel and the road surface state feature input into the input layer of the second channel, and amplifies the difference in similarity between different features through logarithmic operation to construct a loss function.

8. The asphalt pavement fatigue damage model calibration method based on AI and digital twin according to claim 1 is characterized in that: The method for performing full life cycle self-evolution calibration of the asphalt pavement fatigue damage model through the online-offline dual-loop control strategy includes: The model parameters in the fatigue damage model are updated in real time based on the extended Kalman filter. The GAN network is used to generate virtual data input into the fatigue damage model, and the virtual data is input into the fatigue damage model for training enhancement.

Images