Intelligent optimization method and system for asphalt pavement structure based on federated generative learning

Abstract

The application discloses an intelligent optimization method and system for asphalt pavement structure based on federated generative learning, and belongs to the field of intelligent road design. The method comprises the following steps: introducing a weighted aggregation based on data value (residual life, scarcity and innovation degree) in federated learning, so that the global model focuses on high-quality experience; adopting a generative model with an engineering constraint layer to ensure that the output scheme naturally conforms to engineering specifications; mixing and screening historical schemes and generated innovative schemes during design, and determining the optimal design through agent evaluation and simulation optimization; and utilizing long-term performance data feedback to drive model closed-loop evolution. The application realizes high-quality knowledge fusion and reliable design innovation under the premise of protecting data information security, and significantly improves the scientificity, innovation and long-term performance guarantee capability of pavement structure design.

Description

Technical Field

[0001] This invention relates to the field of intelligent road design technology, and in particular to an intelligent optimization method and system for asphalt pavement structure based on federated generative learning. Background Technology

[0002] Asphalt pavement structure design is transforming towards intelligentization, but it currently faces several core bottlenecks. First, the contradiction between data silos and knowledge integration is prominent. High-value design, environmental, and long-term performance data are scattered across different regions and institutions, making direct sharing difficult due to data security concerns, creating "data silos" and hindering the construction of a unified knowledge base to support global innovative design. Second, there is insufficient innovation in design paradigms. Traditional methods rely excessively on local experience and standards, and intelligent optimization is often limited to screening and fine-tuning historical solutions, lacking genuine creative design capabilities. Furthermore, when directly applying generative AI models, the generated solutions often fail to meet the hard constraints of discretization and standardization in engineering practice, resulting in low generation efficiency and reliability. Finally, there is a lack of continuous self-optimization capabilities. The data chain throughout the "design-construction-operation" lifecycle is broken, and long-term performance data of completed roads cannot be effectively fed back to the design end, preventing design models and methods from undergoing reverse calibration and iterative evolution based on actual usage effects.

[0003] Federated learning, as a distributed machine learning paradigm, offers the possibility of cross-institutional knowledge fusion while protecting data security. Existing technologies have already developed methods for applying federated learning to specific domains. Existing technology document CN121483464A discloses a smart healthcare information sharing method based on a federated learning framework. This method coordinates through a central server, with each medical institution training models locally using encrypted data and only uploading model parameters for aggregation. This allows for the construction of a global disease prediction or diagnostic model while protecting patient information security. Its aggregation strategy employs a federated averaging algorithm (FedAvg) or a weighted averaging method based on data volume. However, this method has significant limitations: firstly, its aggregation strategy primarily allocates weights based on the number of data samples at each participating node, failing to consider the heterogeneity of the data's inherent value. In the field of road design, data from different regions varies greatly in value in addressing extreme environments, achieving long-life designs, or innovating solutions. Simply averaging by data volume can dilute high-quality design experience in the global model, making it difficult to effectively improve model performance. Secondly, this technical framework aims to use existing data for analysis and prediction, but it does not have the ability to creatively generate new and feasible design solutions and cannot break through the limitations of historical experience.

[0004] Furthermore, existing technical document CN121170604A discloses an AI detection system for the layout of sidewalk paving bricks, which also integrates a federated learning framework for model updates. It collects human feedback cases via mobile devices, and each detection terminal fine-tunes its model using local feedback data. The cloud server uses a federated averaging algorithm (weighted according to the proportion of data volume from each terminal) to aggregate the model parameters of each terminal to achieve collaborative updates. The core objective of this technology is to use federated learning to achieve incremental optimization of the detection model and information security protection. Its federated learning mechanism also focuses on the utilization of existing data and model iteration, without involving the generation of design-oriented solutions, and its aggregation strategy does not differentiate data quality assessment.

[0005] In summary, while existing federated learning methods can solve the problem of collaborative training under data security, their aggregation strategies fail to effectively identify and integrate high-value knowledge. Furthermore, their functionality is limited to the identification and prediction of existing patterns, lacking the ability to generate reliable solutions under engineering constraints, and failing to form a complete design optimization closed loop that integrates feedback throughout the entire lifecycle. Summary of the Invention

[0006] The main objective of this invention is to provide an intelligent optimization method and system for asphalt pavement structure based on federated generative learning, which aims to solve the problems of existing data barriers and contradictions in the quality of knowledge integration, insufficient innovation in design paradigms, and lack of continuous optimization loop.

[0007] To achieve the above objectives, this invention provides an intelligent optimization method for asphalt pavement structure based on federated generative learning, wherein the method includes the following steps:

[0008] Step 1: Multiple participating nodes independently train local models using local historical road data and upload the model parameters to the central server; wherein, the central server performs weighted aggregation of the uploaded model parameters according to the value weight of each node's local data, generates a global model, and distributes it; each node performs local fine-tuning of the global model to obtain a personalized federated model;

[0009] Step 2: Based on the target conditions, filter historical reference solutions and call the personalized federated model to generate innovative solutions that meet engineering constraints. Combine rapid evaluation and fine simulation optimization to determine the final design solution.

[0010] Step 3: Feed back the long-term performance data of the road built according to the final design scheme to the local database of the corresponding node to trigger a new round of federated model training and drive the continuous evolution of the system.

[0011] Optionally, in step 1, the weighted aggregation based on the value weight of the local data of each node specifically includes:

[0012] Obtain the data value weight Vi of each node, and perform weighted aggregation on the received model parameters. The aggregation formula is as follows: ;

[0013] in V is the model parameter uploaded for the i-th node. i The data value weight of the i-th node is... These are the parameters of the aggregated global model, where N is the total number of nodes.

[0014] Optionally, the data value weight V of each node i The calculation formula is: ;

[0015] in, The average remaining lifespan L of the roads in the node database i The normalized value, This represents the scarcity coefficient of the environmental partition to which the node belongs. Structural Innovation Level I for Node Design Scheme i The normalized value of α, β, and γ is a preset weight coefficient, and α+β+γ=1.

[0016] Optionally, step 2 specifically includes:

[0017] Step 21: Based on the target conditions of the road to be built, select historical roads with high comprehensive similarity from the local database to form the first scheme set;

[0018] Step 22: Calculate the remaining lifespan of each road in the first scheme set based on the long-term performance data of each road, and select the scheme with the longest remaining lifespan to form a second scheme set as a reference for historical experience.

[0019] Step 23: Input the target conditions into the personalized federated generative design model. The generator of the model has an engineering constraint layer at the end, which is used to map the generated continuous parameters into discretized and standardized structural parameters that conform to engineering specifications, and output multiple innovative solutions to form a third solution set.

[0020] Step 24: Merge the second and third scheme sets, and use the personalized federated agent evaluation model to perform rapid performance prediction and screening to form the fourth scheme set;

[0021] Step 25: Perform finite element simulation and parameter fine-tuning on the schemes in the fourth scheme set, with the goal of maximizing the cumulative equivalent axle number, and output the final optimal pavement structure design scheme.

[0022] Optionally, the engineering constraint layer in step 23 includes a thickness discretization module, a modulus material library mapping module, and an interlayer modulus gradient verification module, which are executed sequentially.

[0023] Optionally, the specific steps for selecting historical roads with high comprehensive similarity in step 21 are as follows: using the analytic hierarchy process (AHP), an evaluation index system with climate zoning and soil resilient modulus as the core is established, the comprehensive similarity score between the road to be built and the historical road is calculated, and historical roads with scores higher than a preset threshold are selected.

[0024] Optionally, in step 22, the specific steps for calculating the remaining life are as follows: based on the decay model of the Road Surface Damage Index (PCI) and the Structural Capacity Index (PSSI), calculate the remaining time required for the road to reach the minimum permissible performance value, and take the smaller value between the remaining time corresponding to PCI and PSSI as the remaining life of the road. The specific formula for the decay model is PCI = 100e (-0.06t) PSSI=100e (-0.04t) .

[0025] Optionally, the personalized federated model includes a personalized federated generative design model and a personalized federated agent evaluation model; wherein the generative design model is a conditional generative adversarial network, and the agent evaluation model is a deep neural network.

[0026] Furthermore, to achieve the above objectives, the present invention also provides an intelligent optimization system for asphalt pavement structure for implementing the above-described method, the system comprising:

[0027] A central coordination server is configured to execute a model parameter weighted aggregation algorithm based on data value weights;

[0028] Multiple regional node servers are deployed at each participating party. Each server includes a local historical road parameter database, a local model training module, and a personalized fine-tuning module.

[0029] The design terminal includes a condition input module, a scheme generation module, a rapid evaluation module, and a simulation optimization module;

[0030] The system is configured to feed back long-term performance data of newly built roads to the local database of the regional node server to trigger a new round of federated model training.

[0031] In addition, to achieve the above objectives, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the intelligent optimization method for asphalt pavement structure based on federated generative learning as described in any of the preceding claims.

[0032] Beneficial effects:

[0033] (1) High-quality knowledge fusion under information security protection is achieved: By introducing a weighted federated aggregation strategy based on data value (remaining lifetime, environmental scarcity, and structural innovation), the central server can guide the global model to prioritize learning and aggregating high-value design experiences (such as long-life roads and solutions for scarce environments) from different regions. This effectively solves the problem of "high-quality experience being diluted by massive amounts of ordinary data" caused by traditional federated averaging (FedAvg) or simple weighting by data volume.

[0034] (2) Breaking through the limitations of experience, reliable design innovation is achieved: By innovatively adding an "engineering constraint layer" (integrating thickness discretization, modulus material library mapping, and interlayer gradient verification module) at the generator end of the generative adversarial network (c-GAN), the system can not only select excellent schemes from historical data, but also creatively generate new pavement structure schemes in batches, and ensure that each generated scheme naturally conforms to all engineering hard constraints (thickness is a discrete standard value, modulus is taken from the standard material library, and interlayer mechanical coordination).

[0035] (3) A data-driven, continuously self-iterable intelligent design ecosystem has been formed: long-term performance data after road construction is used as key feedback, re-injected into the local database, and participates in subsequent federated learning cycles. This enables the system's design knowledge base and model to be continuously reverse-calibrated and optimized based on actual engineering results, driving model performance and design quality to spirally increase with the accumulation of engineering practice. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating an intelligent optimization method for asphalt pavement structure based on federated generative learning, according to the present invention.

[0037] Figure 2 Comparison chart of the effects of federated learning aggregation strategies;

[0038] Figure 3 This chart compares the compliance rates of different generation methods. 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] This invention provides an intelligent optimization method for asphalt pavement structure based on federated generative learning, the core of which lies in the following three stages executed sequentially and iteratively:

[0041] Phase 1: Knowledge Integration and Model Personalization

[0042] This phase is conducted within a federated learning framework. Its core objective is to construct a personalized intelligent model that integrates high-value global knowledge while adapting to specific local conditions, all while protecting the data security of all participating parties. This phase can be broken down into the following four key steps:

[0043] Step 11: Data preparation and local model training. Each participating node independently trains a local generative design model and a local agent evaluation model using its local historical road parameter database. The local generative design model is a conditional generative adversarial network with an engineering constraint layer at the end of its generator.

[0044] Participating nodes: Multiple independent road design or management units (such as design institutes and data centers in various provinces and cities) serve as participating nodes in federated learning.

[0045] Local Database: Each node maintains a local historical road parameter database, storing complete data on roads already built within its jurisdiction, including: design parameters (thickness of each structural layer, material modulus), environmental parameters (climate zones, subgrade resilient modulus, etc.), long-term performance data (Pavement Damage Index (PCI), Structural Bearing Capacity Index (PSSI) and its decay over time), and traffic data (design traffic volume, cumulative equivalent axle loads).

[0046] Model training: Each node independently trains two local models using its own local database.

[0047] Local generative design models are typically Conditional Generative Adversarial Networks (c-GANs). The generator (G) takes the target environment and traffic conditions as input, and the goal is to output road surface structure parameters; the discriminator (D) distinguishes the generated scheme from real historical schemes. The key innovation lies in adding an engineering constraint layer at the end of the generator to ensure the compliance of subsequent outputs.

[0048] This layer consists of a thickness discretization module, a modulus material library mapping module, and an interlayer modulus gradient verification module connected in series. It is used to convert the continuous values ​​output by the generator into structural parameters that conform to engineering specifications.

[0049] The thickness discretization module has a preset set of commonly used engineering thickness discretization values ​​H. set ={4cm, 5cm, 6cm, 8cm, 10cm, 12cm, 15cm, 18cm, 20cm, 25cm, 30cm}, using the nearest neighbor floor function h out =argmin{h∈H set}|h raw -h| Specifies the continuous thickness value h output by the generator. raw The thickness value is mapped to the closest value in the discrete value set.

[0050] The modulus material library mapping module has a pre-set standard material modulus library, including the asphalt mixture modulus set E. AC ={8000,9000,10000,11000,12000} (unit: MPa), base material modulus set E base ={1200,1500,1800,2000,2200} (unit: MPa), Subbase material modulus set E sub ={400,500,600,700,800} (unit: MPa), obtained through the nearest neighbor matching function E out =argmin{E∈E material}|E raw -E| will output the continuous modulus value E from the generator. raw The value is mapped to the corresponding modulus value in the standard material library.

[0051] After the thickness and modulus mapping is completed, the interlayer modulus gradient verification module calculates the modulus ratio r between adjacent structural layers. i =E i / E i-1 If r i Greater than the preset gradient threshold R max Then E i Adjusted to E i =R max ×E i-1 If r i Less than the preset gradient threshold R min Then E i Adjusted to E i =R min ×E i-1 .

[0052] The three modules described above are executed in cascade order to ensure that the final output pavement structure parameter vector simultaneously satisfies the triple constraints of thickness discretization, material library availability, and interlayer mechanical compatibility. A discriminator D is used to distinguish the generated scheme from real historical schemes.

[0053] And local agent evaluation models: typically deep neural networks. Their goal is to establish a rapid mapping relationship from pavement structural parameters to key mechanical responses (such as tensile strain at the bottom of the layer) and fatigue life, for initial screening of schemes.

[0054] Step 12: Model Parameter Upload and Data Value Assessment. Each participating node performs information security protection processing on its local model parameters after training and uploads the processed model parameters to the central server. The central server assesses the data value weight of each participating node's local data.

[0055] Information security protection during upload: After training is completed, each node does not share the original data. Instead, it performs information security protection processing on the trained local model parameters (such as adding differential information security noise) and then uploads the processed model parameters to the central server.

[0056] Data value assessment: Before aggregation, the central server assesses the "data value" uploaded by each node.

[0057] Specifically, for the i-th participating node, the data value weight V of its local database is evaluated. i V i The average remaining lifespan L of the road in this node data i The scarcity coefficient E of the environmental factors among all nodes i And its structural innovation score I i The comprehensive calculations yielded the following results, with the calculation methods for each factor as follows:

[0058] Mean remaining life L i The remaining life is obtained by taking the arithmetic mean of the remaining life of all roads in the local database of node i. The remaining life is calculated based on the corresponding decay model of the pavement damage index (PCI) and the structural bearing capacity index (PSSI). The specific formula is as follows:

[0059] PCI=100e (-0.06t) PSSI=100e (-0.04t) In practical applications, the smaller value between PCI and PSSI is used to determine the remaining service life. If the service life is the same, the remaining service life is determined directly by the remaining service life value. If the service life is different, the remaining service life is predicted to the same baseline time point using a service life prediction model and then compared.

[0060] Environmental scarcity coefficient E i Each node reports its environmental zone identifier (e.g., a combination of climate zone and subgrade grade codes) along with the model parameters. The central server counts the number of nodes n in each environmental zone c. c The total number of nodes is N. For node i, its partition is c. i Then the scarcity coefficient E i =1-(n ci / N), this value is between 0 and 1, and the fewer the number of nodes in a partition, the higher its scarcity.

[0061] Structural Innovation Score I i The calculation is performed independently at node i locally. First, the typical pavement structure reference parameter vector T=[t1,t2,…,t] is determined according to current pavement design specifications. kThis includes recommended values ​​for key indicators such as the thickness of each structural layer and the material modulus, as well as the median of common value ranges. For the j-th design scheme in the local database of node i, its feature vector X is extracted. j (The dimension is the same as T, such as surface layer thickness, base layer thickness, subbase layer thickness, surface layer modulus, base layer modulus, etc.), calculate its Euclidean distance d with the reference vector. j =||X j -T||. Innovation score of node i. M i This represents the total number of design schemes in the database of this node. Each node will calculate the I... i As metadata, it is uploaded to the central server, which then processes the I / O data of all nodes. i Normalization is performed to adjust its value range to 0~1.

[0062] After calculating and normalizing the above three factors, the data value weight V of node i is... i Determined by the following formula: ;

[0063] in, The average remaining lifespan L of the roads in the node database i The normalized value of the remaining lifetime indicates that the design experience is superior and more valuable. This is the scarcity coefficient of the environmental partition to which the node belongs. The calculation formula is E. i =1-n ci / N, where n ci N represents the number of nodes in the same partition as this node, and N is the total number of nodes. The scarcer the partition, the higher its empirical value in solving special environmental problems. Structural Innovation Level I for Node Design Scheme i The normalized score is calculated by comparing the design scheme with the typical scheme in the standard; the greater the difference, the higher the innovation. α, β, and γ are preset weighting coefficients, and satisfy α+β+γ=1.

[0064] Step 13: Value-based weighted aggregation. Based on the data value weights, the received node model parameters are weighted and aggregated to generate a global generative design model and a global proxy evaluation model.

[0065] Aggregation Algorithm: Instead of using a simple parametric average, the central server employs a weighted aggregation based on data value weights. The aggregation formula is as follows: ;

[0066] in V is the model parameter uploaded for the i-th node. i The data value weight of the i-th node is... These are the aggregated global model parameters.

[0067] Weighted processing allows nodes with longer-lasting roads, scarcer environmental solutions, and more innovative designs to have a greater say in the generation of the global model. This ensures that the global model can focus on and integrate truly high-value cross-regional design knowledge, solving the problem of "high-quality experience being diluted by massive amounts of ordinary data" in the traditional federated averaging method.

[0068] Step 14: Model distribution and personalized fine-tuning. The central server distributes the aggregated global model to each participating node. Each participating node uses its local data to fine-tune the received global model to obtain a personalized federated generative design model and a personalized federated agent evaluation model.

[0069] Global model distribution: The central server distributes the aggregated global generative design model and global proxy evaluation model, which incorporates high-value knowledge from all parties, to all participating nodes.

[0070] Local personalization: After receiving the global model, each node does not use it directly, but instead uses its own local historical road parameter database to fine-tune the global model. For example, the model's initial feature extraction layer can be frozen, and only the parameters of subsequent network layers that are strongly correlated with the local environment and materials can be updated.

[0071] Output: After fine-tuning, each node ultimately yields a personalized federated model (including a personalized federated generative design model and a personalized federated agent evaluation model). This model not only possesses advanced knowledge and experience learned globally, but also deeply adapts to the specific conditions and standards of the local area, thus preparing for subsequent intelligent design tasks.

[0072] The first stage – Knowledge Fusion and Model Personalization – is a complete process of distributed training, centralized intelligent aggregation, and then distributed adaptation. Through innovative mechanisms of data value assessment and weighted aggregation, it ensures high-quality and directional knowledge fusion under information security protection. Furthermore, through local fine-tuning, it achieves precise implementation of global intelligence in local scenarios, providing a powerful and practical model foundation for subsequent intelligent design.

[0073] Multiple participating nodes (such as design institutes in various provinces and cities) independently train local models (including generative design models and proxy evaluation models) using local historical road databases. The key innovation lies in the fact that when the central server aggregates model parameters, it does not simply average them, but rather uses the data value weights (Vi) of each node's local data. i The data value weights are dynamically evaluated based on factors such as the average remaining lifespan of the road in the node data, the scarcity of the environmental zone to which it belongs, and the structural innovation of the design scheme.

[0074] The aggregation formula ensures that nodes with better, newer, and scarcer design experience contribute more to the global model, thereby achieving high-quality knowledge fusion under information security protection. The aggregated global model is then fine-tuned using local data from each node to obtain a personalized federated model adapted to specific local conditions.

[0075] Phase Two: Hybrid-Driven Intelligent Design

[0076] This stage aims to generate and determine the optimal pavement structure design scheme for a specific road construction project. Its "hybrid-driven" approach combines historical experience mining with artificial intelligence generation, and rapid intelligent evaluation with detailed physical simulation for verification. This stage can be broken down into the following five core sub-steps executed sequentially:

[0077] Step 21: Initial screening of historical schemes based on environmental similarity.

[0078] Input: Target design conditions for the road to be built, including geographical environmental parameters (such as climate zone, subgrade resilient modulus, groundwater level) and traffic parameters (such as design cumulative equivalent axle loads).

[0079] Processing: The system employs the analytic hierarchy process (AHP) to quantify the similarity between the target conditions and each historical road in the local database. First, an evaluation system is established: this involves determining the core comparison indicators (such as climate zoning, subgrade modulus, and traffic grade) and their weights. Specifically, pairwise comparison judgment matrices for the evaluation indicators are constructed, and the weight coefficients of each indicator are calculated.

[0080] Then, individual similarity is calculated. For qualitative indicators (such as climate zones): 1 is assigned to identical items, and 0 to different items. For quantitative indicators: an adaptive function (such as a Gaussian kernel function) is used, taking into account the overall distribution of the indicator in historical data. The individual similarity score is calculated using the following normalization function: ;

[0081] in, The similarity score for the i-th parameter; This parameter value represents the road to be constructed. This parameter value represents a historical road case. Let be the standard deviation of all values ​​of the i-th parameter in the historical road parameter database.

[0082] And calculate the overall similarity score: sum the similarity scores of each indicator according to their weights to get a score between 0 and 1.

[0083] Finally, all historical roads with a comprehensive similarity score higher than a preset threshold (e.g., >0.85) were selected to form Scheme Set 1. This step ensures that subsequent reference schemes are in a similar environment and load context to the current project.

[0084] Step 22: Optimize high-value historical solutions based on long-term performance.

[0085] Input: Solution set 1 obtained in step S21.

[0086] Processing: From "Solution Set 1," the best-performing schemes with long-term validation were selected as benchmarks. Specifically, for each road in "Solution Set 1," its current Pavement Damage Index (PCI), Structural Bearing Capacity Index (PSSI), and service time were obtained. These values ​​were then substituted into preset PCI and PSSI decay models to calculate the remaining time required for the pavement structure to reach the minimum allowable values ​​specified in the code. The smaller of the two values ​​was taken as the remaining life of the road.

[0087] Then, all the schemes in "Scheme Set 1" are sorted from longest to shortest remaining lifespan. Preferably, the top 1 to 3 schemes (i.e., the best performing and longest-lived designs) are directly selected and output to form Scheme Set 2. This set represents the most durable design schemes that have been tested in practice under similar environments.

[0088] Step 23: Generate innovative solutions based on generative models.

[0089] Input: Target conditions for the road to be built (same as step 21).

[0090] Processing: The personalized federated generative design model (c-GAN) trained in the first stage is invoked. Its core process is as follows: The target conditions are input as a condition vector into the generator. The generator then outputs a continuous vector of original road structure parameters (such as layer thickness and modulus). The original vector is then immediately fed into the cascaded engineering constraint layer for forced normalization. This includes a thickness discretization module, which maps continuous thickness values ​​to the nearest value in a preset set of commonly used engineering discrete thicknesses (e.g., {4, 5, 6, 8, 10, 12, 15, 18, 20, 25, 30} cm; a modulus material library mapping module, which maps continuous modulus values ​​to the nearest value in a standard material modulus library (e.g., asphalt mixture: {8000, 9000, 10000, 11000, 12000} MPa); and an interlayer modulus gradient verification module, which calculates and adjusts the modulus ratio of adjacent structural layers to ensure it falls within a reasonable range (e.g., 0.18~3.0) to guarantee structural stress coordination.

[0091] Finally, output N (N≥10) innovative pavement structure schemes that fully comply with engineering specifications, forming scheme set three.

[0092] Step 24: Rapid evaluation and screening of the hybrid scheme set.

[0093] Input: Merge "Solution Set 2" (historical experience solutions) and "Solution Set 3" (innovative generated solutions).

[0094] Processing: Proxy Model Rapid Prediction: The structural and environmental parameters of each scheme in "Scheme Set Two" and "Scheme Set Three" are input into the personalized federated proxy evaluation model (deep neural network) trained in the first stage. This model predicts the key mechanical responses (such as tensile strain at the bottom of the asphalt layer) and fatigue life of each scheme within milliseconds. All schemes are then ranked according to the fatigue life predicted by the proxy model.

[0095] Finally, the 5 to 8 schemes with the best prediction performance are output, forming scheme set four (potential scheme set). This step utilizes the efficiency of the surrogate model to quickly identify the few candidate schemes with the greatest potential from a large number of (historical + generated) schemes, greatly narrowing the scope of subsequent detailed simulations.

[0096] Step 25: Fine-scale simulation optimization and final scheme determination.

[0097] Input: Scheme set four, which includes structural layer parameters and target environment parameters.

[0098] Processing: For each scheme in "Scheme Set 4", a refined mechanical response model is established using finite element software. Preferably, a three-dimensional refined model (e.g., 5m × 5m × 3m) is established, and the mesh is refined to more realistically simulate the structural response. This also requires consideration of complex mechanical states such as the nonlinear behavior of the material and the contact conditions between structural layers.

[0099] Apply a standard axle load (such as a double-circle uniformly distributed load with a pressure of 0.7 MPa and a radius of 10.65 cm) to the critical load position (usually the center of the wheel gap), and calculate the accurate mechanical response of the scheme under the standard load. The optimal response is the maximum tensile strain (ε) at the bottom of the asphalt layer. a This indicator is a key driver of road fatigue cracking.

[0100] Based on the cumulative equivalent number of axes (N) f With maximizing as the clear objective, individual parameters of the preferred scheme are fine-tuned in the finite element model to output the final optimal pavement structure design scheme.

[0101] Preferably, small-scale iterative simulations are performed by adjusting the structural layer thickness (e.g., ±1 cm) and material modulus parameters in the model (switching between adjacent values ​​in the standard material library or adjusting within a range of ±500 MPa). The optimized parameter range remains consistent with the engineering constraint layer in step 23 to ensure that the optimization result is still a compliant engineering solution.

[0102] The optimized maximum tensile strain (ε) at the bottom of the asphalt layer a ) and other parameters (such as surface modulus E)a Substituting into the fatigue equation in the specification, calculate the final predicted cumulative equivalent shaft number (N) for each scheme. f The fatigue equation is: ;

[0103] in, The fatigue cracking life of the asphalt mixture layer is [number] times. As a reliable indicator of the target; For seasonally frozen soil areas, an adjustment factor is used. For fatigue loading mode coefficients; VFA is the temperature adjustment factor; VFA is the bitumen saturation. The dynamic compression modulus of asphalt mixture at 20℃, in MPa; For the tensile strain at the bottom of the asphalt mixture layer, 10 -6 .

[0104] The second stage is a funnel-shaped decision-making process that moves from "broad screening" to "focused optimization." By creating a candidate solution pool through a "hybrid" approach of historical data mining and AI generation, and then using a "hybrid" evaluation strategy of rapid evaluation by surrogate models and fine simulation by finite elements, the design is efficiently and reliably driven to converge toward the optimal solution, achieving a balance between experience and innovation, efficiency and accuracy.

[0105] Phase Three: Continuous Learning and Closed-Loop Evolution

[0106] The goal of this stage is to transform the long-term performance of the design results in the real world into the fuel driving model evolution, forming a complete intelligent closed loop of "design-verification-feedback-optimization". This stage can be broken down into the following five specific steps:

[0107] Step 31: Long-term performance data collection and feedback. After the road is built and put into use according to the final design scheme, long-term performance monitoring data of the road is collected regularly. The long-term performance data is associated with its corresponding design parameters and environmental parameters and stored in the local historical road parameter database of the participating node.

[0108] The road is constructed and put into use based on the final design scheme output from the second phase. Long-term performance monitoring data of the road is collected regularly (e.g., annually), and key indicators include: Pavement Damage Index (PCI, reflecting the health of the pavement surface), Structural Bearing Capacity Index (PSSI, reflecting the overall strength of the pavement structure), rutting depth, and smoothness (according to database requirements).

[0109] These performance data that change over time are then correlated with the road’s original design parameters (from the final design), target environmental parameters, actual traffic volume, and other information to form a complete “design-performance” record.

[0110] Feedback to Database: This newly added complete record is stored in the local historical road parameter database of the participating node corresponding to the area to which this road belongs. At this point, the node's local database is expanded to include the latest practical results.

[0111] Step 32: Incremental learning of the local model. Each participating node uses the newly added long-term performance data in its database to incrementally train the local model.

[0112] Triggering condition: When a node's local database accumulates a certain amount of new long-term performance data (for example, reaching a preset batch size or time period), the node initiates an update of the local model.

[0113] Learning process: Nodes use new data (or combine some historical data) to incrementally train or fine-tune the existing personalized federated generative design model and personalized federated agent evaluation model.

[0114] For proxy evaluation models, the new performance data can serve as new "labels" to calibrate and improve the accuracy of their long-term performance predictions.

[0115] For generative design models, new cases that perform well (good long-term performance) can reinforce the model's tendency to generate similar successful solutions.

[0116] Furthermore, the incrementally trained local model will participate in a new round of federated model training and aggregation. In this new round of training, the central server will re-evaluate the data value weights of each node based on the updated local database containing new feedback data.

[0117] Step 33: Trigger a new round of federal training.

[0118] Coordinated initiation: The central server initiates the process periodically, or coordinates the start of a new round of federated training after a significant number of nodes have completed local incremental learning.

[0119] Process restart: Each node restarts training its local model based on the updated local database (containing the new feedback data) (i.e., returns to step 11 of "Phase One"). After training is complete, the updated model parameters are uploaded to the central server after information security processing.

[0120] Step 34: Reassessment and aggregation of data value weights.

[0121] Weight recalculation (key evolutionary step): Before aggregation, the central server re-evaluates the data value weights (V) of each node. i At this point, the node with a newly constructed road exhibiting excellent long-term performance has a mean remaining life (L). iThe E factor will be enhanced if the road is in a scarce environment or has an innovative design. i Or I i Factors may also be affected.

[0122] Evolutionary aggregation: Due to changes in the value weights of node data, a new round of weighted aggregation occurs. The focus will shift towards nodes that produce more and more reliable successful case studies. This means that proven and effective design practices will be further reinforced and disseminated throughout the global model.

[0123] Step 35: Model update and distribution.

[0124] Global model update: The central server performs weighted aggregation to generate a new generation of global model. This model incorporates knowledge learned from the latest engineering practices (both successful experiences and lessons learned from failures).

[0125] Model redistribution and personalization: The new generation global model is distributed to all nodes. Each node then combines its own complete local data (including both old and new data) for personalized fine-tuning, resulting in an evolved personalized federated model.

[0126] The formation of a closed loop: The evolved model will be used in subsequent new design projects (i.e., entering the "second phase"), thereby generating new design solutions. The performance data of these new solutions will then be fed back, driving the next round of evolution. This cycle repeats itself, and the system's design knowledge base, generation capabilities, and evaluation accuracy can continuously improve and enhance themselves over time and with the accumulation of engineering cases, forming a truly evolving intelligent design ecosystem.

[0127] Long-term performance data of roads built according to the optimal solution (such as the Pavement Damage Index (PCI) and the Structural Bearing Capacity Index (PSSI)) will be used as new samples and fed back to the local database of the corresponding nodes. This new data will be used for incremental learning of the node's local model and will increase the data value weight of the node in subsequent federated training. This will continuously integrate the successful (or unsuccessful) experiences represented by the node into the global knowledge base, driving a spiral improvement in model performance and design quality, and forming a complete intelligent closed loop of "design-verification-feedback-optimization".

[0128] The optimization method of this invention is based on a federated learning framework. The central server evaluates the value weight of each node's data based on its remaining lifetime, environmental scarcity, and structural innovation, and performs weighted aggregation accordingly. This allows the global model to focus on integrating high-quality design experience across regions. A generative adversarial network with engineering constraint layers is used to ensure that the generated schemes automatically meet hard engineering constraints such as thickness discretization, material standardization, and interlayer mechanical coordination. During the design phase, the system mixes and filters high-value historical schemes with generated innovative schemes. After rapid evaluation by a proxy model and fine optimization through finite element simulation, the optimal scheme is output. Finally, the long-term performance data of the newly built roads is used to drive the continuous evolution of the model, forming a complete closed loop of "design-verification-feedback-optimization".

[0129] Furthermore, to better illustrate the beneficial effects of the present invention, specific embodiments are described below.

[0130] Example 1

[0131] This embodiment takes a section of a riverside expressway in a certain province as an example to illustrate the practical application effect of the method of the present invention.

[0132] 1. Engineering Background and Data Foundation

[0133] This section of road is located in a hot summer and mild winter zone (Zone IV1), with a subgrade composed of low liquid limit clay (dry state). The design traffic volume level is heavy traffic, with a cumulative equivalent axle load of N. e =25 million times. There are 5 nodes participating in federated learning: Node A (eastern part of the province), Node B (western part of the province), Node C (southern part of the neighboring province), Node D (northern part of the neighboring province), and Node E (research institute experimental road database). The data of each node is shown in Table 1.

[0134]

[0135] 2. Federated Model Training and Value-Weighted Aggregation

[0136] Calculate the environmental scarcity coefficient: Count the number of nodes in each partition: Partition IV1 has 2 nodes (A and E combined), Partition IV2 has 1 node (B combined), Partition III2 has 1 node (C combined), and Partition III1 has 1 node (D combined). Therefore, E... A =1-2 / 5=0.6, E B =1 - 1 / 5 = 0.8, E C =0.8, E D =0.8, E E =0.6.

[0137] Normalization: for L i and I i After normalization, we get =[0.58,0.25,0.88,0,1], =[0.14,0,0.69,0.44,1].

[0138] Given the weights α=0.4, β=0.3, and γ=0.3, calculate V. i :

[0139] V A =0.4×0.58+0.3×0.6+0.3×0.14=0.232+0.18+0.042=0.454

[0140] V B =0.4×0.25+0.3×0.8+0.3×0=0.1+0.24+0=0.34

[0141] V C =0.4×0.88+0.3×0.8+0.3×0.69=0.352+0.24+0.207=0.799

[0142] V D =0.4×0+0.3×0.8+0.3×0.44=0+0.24+0.132=0.372

[0143] V E =0.4×1+0.3×0.6+0.3×1=0.4+0.18+0.3=0.88

[0144] Aggregation: The central server aggregates the c-GAN and proxy model parameters uploaded by each node according to their weights to obtain a global model, which is then distributed to each node. After receiving the model, node A uses 150 local data points to fine-tune the global model (freezing the first 3 layers and updating the last 2 layers) to obtain a personalized federated model.

[0145] 3. Local intelligent design

[0146] Node A designers input target parameters through the design terminal:

[0147] Climate zone: Zone IV1, Soil subgrade: Dry, Traffic volume: 25 million trips

[0148] Step S1: Calculate similarity using the analytic hierarchy process. Using climate zone (weight 0.5), subgrade grade (0.3), and traffic volume (0.2) as indicators, select 25 historical roads with similarity > 0.9 from the local database of node A to form scheme set one.

[0149] Step S2: Calculate the remaining lifetime of each scheme based on the PCI and PSSI decay models. For example: Scheme H-23 (a highway completed in 2008) currently has PCI=82, PSSI=85, has been in service for 12 years, and its decay model is PCI=100e. (-0.06t)PSSI=100e (-0.04t) Assuming a lower limit of 75 for PCI and 70 for PSSI, the remaining lifetimes are 1.5 years and 3.8 years respectively. Taking the smaller value of 1.5 years, the overall service life is 13.5 years. After sorting, the two schemes with the longest overall service life are selected: Scheme H-05 and Scheme H-12, forming Scheme Set Two.

[0150] Step S3: The structural parameters of the optimized alternative scheme set (H-05: 16cm AC-16 surface layer, 24cm cement-stabilized crushed stone base layer, 20cm cement-stabilized crushed stone subbase layer; H-12: 18cm AC-20 surface layer, 20cm cement-stabilized crushed stone base layer, 22cm graded crushed stone subbase layer) are used as empirical references. Simultaneously, the target conditions are input into the personalized federated generative model, setting the generation quantity N=13. Example of the generator's internal processing:

[0151] Original output: Surface layer thickness h1=16.8cm, base layer thickness h2=22.3cm, subbase layer h3=19.1cm; Surface layer modulus E1=11200MPa, base layer E2=1900MPa, subbase layer E3=650MPa.

[0152] Thickness discretization module: Discrete set H set ={4,5,6,8,10,12,15,18,20,22,25,30}, the nearest neighbor mapping yields h1=18cm, h2=22cm, h3=20cm.

[0153] Modulus Library Mapping Module: Material Library E AC ={8000,9000,10000,11000,12000}, E1=11200 maps to 11000MPa; E base ={1200,1500,1800,2000,2200}, E2=1900 maps to 2000MPa; E sub ={400,500,600,700,800}, E3=650 is mapped to 700MPa.

[0154] Gradient verification module: Preset R max =3, R min =0.18. The surface layer is 12000, the base layer is 2000, and the modulus ratio r = 2000 / 12000 = 0.17 < 0.18. Therefore, the base layer modulus is adjusted to R. min ×12000=2160, so we take 2200. 2200 / 12000=0.183>0.18, which satisfies the condition. In practical applications, gradient constraints may be more complex. This embodiment assumes that the system outputs feasible values ​​after two iterations: the base layer modulus is 2200, and the surface layer modulus is 1200. This is just an illustration of the process.

[0155] After constraint layer processing, 13 schemes were finally generated. The thickness of each scheme is a discrete value, the modulus is taken from the standard library, and the interlayer modulus ratio is in the range of 0.18 to 3. Scheme set three is formed.

[0156] Step S4: Merge scheme set two (2 schemes) with scheme set three (13 schemes) to obtain 15 candidate schemes. Input the personalized federated agent evaluation model (based on a deep neural network, with 6 neurons in the input layer corresponding to surface layer thickness, base layer thickness, subbase layer thickness, surface layer modulus, base layer modulus, and subbase layer modulus, and 2 neurons in the output layer corresponding to tensile strain and fatigue life) to predict the tensile strain and fatigue life of the asphalt layer bottom for each scheme. Sort by fatigue life and select 5 schemes as the potential scheme set, which includes 2 historical optimized schemes and 3 innovative generated schemes.

[0157] Step S5: Optimize the five potential schemes using finite element simulation. A three-dimensional model (5m × 5m × 3m) was built using ABAQUS with refined mesh, and a double-circular uniformly distributed load (pressure 0.7MPa, radius 10.65cm) was applied. The tensile strain ε at the bottom of the asphalt layer under standard axle load was calculated for each scheme. a Substitute into the fatigue equation to calculate the cumulative equivalent number of shafts: ;

[0158] Where β is 1.65, k a =1.1, k T1 =1.0, VFA=75%, E a The modulus of the surface layer is used. A small-range parameter scan is performed by adjusting the thickness of each layer (within ±1 cm) and the modulus (within ±500 MPa) to find the optimal N. f The optimal combination was determined as follows: 18cm AC-20 surface layer (modulus 10000MPa), 22cm cement-stabilized crushed stone base layer (modulus 1800MPa), and 20cm cement-stabilized crushed stone subbase layer (modulus 600MPa). The predicted cumulative equivalent axle cycles is 31.2 million, which is higher than the design requirement of 25 million cycles and represents an 11.4% improvement over the initial optimized scheme H-05 (28 million cycles).

[0159] 4. Implementation Results

[0160] This section of road was completed in 2024 and has been open to traffic for two years. Initial performance monitoring data shows good agreement with the surrogate model predictions, validating the reliability of this method. Data will continue to be collected and fed back into the federated learning system for model evolution.

[0161] 5. Comparative Verification

[0162] To verify the advantages of the present invention, under the same design conditions, the traditional standard method (looking up tables according to JTG D50-2017) and the historical scheme screening method (steps S1+S2+S5) were compared respectively, and the results are shown in Table 2.

[0163]

[0164] The results show that the present invention significantly improves the performance of the solution while shortening the design cycle, and achieves a truly innovative design.

[0165] Furthermore, in step S1, the core is to assess environmental similarity through an objective and quantitative method, specifically implemented as follows:

[0166] (1) Establish a geographical environment evaluation index system

[0167] First, the system defines a set of geographic environmental parameters for similarity comparison. This set preferably includes, but is not limited to, the following: climate zones (such as severe cold, cold, hot summer and cold winter, warm and temperate), annual average precipitation (mm, average over 30 consecutive years), soil resilient modulus (MPa, reflecting soil strength), and groundwater depth (m, reflecting hydrogeological conditions).

[0168] (2) Determining indicator weights based on the analytic hierarchy process

[0169] Because the aforementioned indicators have varying degrees of influence on pavement structural performance, this invention employs the analytic hierarchy process (AHP) to scientifically determine their weights. The system invites domain experts to conduct pairwise importance comparisons of the indicators based on the 1-9 scale, constructing a judgment matrix. For example, it determines the degree of influence of "climate zone" relative to "average annual precipitation" on pavement design. Subsequently, the built-in algorithm calculates the eigenvectors of this judgment matrix and performs a consistency check. After passing the check, the resulting eigenvectors represent the weights of each indicator. The final calculated weight allocation is: climate zone (0.4), average annual precipitation (0.3), subgrade resilient modulus (0.2), and groundwater depth (0.1). This process ensures the objectivity and scientific rigor of the weight allocation, avoiding the subjectivity of direct human assignment.

[0170] (3) Calculate the overall similarity score

[0171] The system compares the target parameters of the road to be built with the corresponding parameters of each historical road in the database. For each indicator:

[0172] If it is a qualitative indicator (such as climate zone), the score is 1 if they are the same and 0 if they are different.

[0173] For quantitative indicators (such as precipitation or resilient modulus), the individual similarity score is calculated using the following normalization function: ;

[0174] in, The similarity score for the i-th parameter; This parameter value represents the road to be constructed. This parameter value represents a historical road case. Let be the standard deviation of all values ​​of the i-th parameter in the historical road parameter database.

[0175] Suppose we are comparing the parameter of "average annual precipitation," and the historical database shows the standard deviation of the average annual precipitation for all historical roads. =200 mm, target value for the road to be built =1000mm, value of historical road A =1100mm, value of historical road B =1400mm,

[0176] Calculate the precipitation similarity score with historical road A: ;

[0177] Calculate the precipitation similarity score with historical road B: .

[0178] The system will calculate the similarity scores between the proposed road and historical roads A and B in terms of quantitative indicators such as subgrade resilient modulus and groundwater depth, using this method. Then, it will combine the weights of each indicator determined by the analytic hierarchy process (AHP) and perform a weighted summation to obtain a final comprehensive similarity score. The higher the score, the more similar the environments are.

[0179] This function can adaptively quantify the degree of similarity based on the data distribution characteristics of the parameters themselves, making the scoring results more scientific and reasonable.

[0180] (4) Final screening based on threshold

[0181] The system presets a similarity threshold (e.g., 0.85, which can be adjusted according to the stringency of design requirements). All historical roads with a comprehensive similarity score greater than or equal to this threshold are automatically selected, forming Solution Set 1. This step enables the rapid and accurate identification of candidate solutions that best match the environmental conditions from massive amounts of data.

[0182] Furthermore, to verify the actual effectiveness of the weighted aggregation strategy based on data value assessment and the generative design model with engineering constraint layer proposed in this invention, this embodiment designed a comparative experiment to quantitatively evaluate the aggregation accuracy of federated learning and the compliance rate of the generated scheme.

[0183] 1. Experimental setup

[0184] (1) Data preparation

[0185] Five participating nodes were selected (denoted as A, B, C, D, and E respectively), and the basic information of the local historical road parameter database of each node is shown in Table 3.

[0186]

[0187] The remaining lifespan is calculated according to the method in step S2, the environmental zoning is based on the current specifications, and the structural innovation score is determined according to the method of this invention (calculating the Euclidean distance based on the typical pavement structure of JTG D50-2017 as the reference vector).

[0188] (2) Setting up the comparison method

[0189] The method of this invention employs a weighted aggregation strategy based on data value assessment to calculate the weight V of each node. i The aggregation formula is The generator includes an engineering constraint layer (thickness discretization module, modulus material library mapping module, and interlayer modulus gradient verification module).

[0190] Traditional Federated Average: This method uses a weighted average based on data volume, and the aggregation formula is as follows: , where n i Let represent the amount of data at node i. The generator is a standard c-GAN without an engineering constraint layer.

[0191] Ordinary GAN method: used only for comparing the compliance rate of generated solutions, using a standard c-GAN structure without constraint layers.

[0192] (3) Evaluation indicators

[0193] The effectiveness of the aggregation strategy is evaluated using the "accuracy of high-value road plan predictions." 20% of the data from each node's database is reserved as a test set (120 entries in total), containing road plans of different value levels. After each round of federated communication, the global model is used to predict the test set, and the prediction accuracy is calculated.

[0194] Compliance rate of generated schemes: 1000 pavement structure schemes were generated using each method, and the percentage of schemes that conformed to engineering specifications was calculated. Compliance judgment criteria included: the thickness of each structural layer was a commonly used discrete value in engineering (e.g., 4, 5, 6, 8, 10, 12, 15, 18, 20, 22, 25, 30 cm), the material modulus was within the range of the standard material library, and the interlayer modulus ratio was between 0.18 and 3.0.

[0195] 2. Experimental Results

[0196] (1) Comparison of the effects of federated learning aggregation strategies

[0197] Perform 40 rounds of federated communication as described above, and record the prediction accuracy of the global model on the high-value solution test set in each round. The results are shown in Table 4 and... Figure 2 As shown.

[0198]

[0199] Experimental results show that the weighted aggregation strategy based on data value assessment adopted in this invention enables the global model to achieve an accuracy of over 85% after 30 rounds of communication, while the traditional federated averaging method still falls short of 80% even after more than 30 rounds. Ultimately, the accuracy of this invention's method reaches 92.4%, an improvement of 14.2 percentage points compared to the traditional method (78.2%). This demonstrates that by prioritizing the aggregation of node parameters with high remaining lifetime, high environmental scarcity, and high structural innovation, the global model can learn high-value design experience faster and more accurately.

[0200] (2) Comparison of compliance rates of generated solutions

[0201] Each of the three methods generated 1000 pavement structure schemes, and the compliance status was statistically analyzed. The results are shown in Table 5. Figure 3 As shown.

[0202]

[0203] Experimental results show that only 31.6% of the solutions generated by ordinary GANs can be used directly. The main problems are that the output thickness is a non-discrete value and the modulus is not in the standard material library. After adding post-processing correction, the compliance rate is improved to 78.2%, but manual intervention is still required and some corrected solutions have mechanical inconsistencies. This invention, by embedding a thickness discretization module, a modulus material library mapping module, and an interlayer modulus gradient verification module in the generator, enables all 1,000 generated solutions to meet engineering specifications, achieving a compliance rate of 100%.

[0204] 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 method for intelligent optimization of asphalt pavement structure based on federated generative learning, characterized in that, The method includes the following steps: Step 1: Multiple participating nodes independently train local models using local historical road data and upload the model parameters to the central server; wherein, the central server performs weighted aggregation of the uploaded model parameters according to the value weight of each node's local data, generates a global model, and distributes it; each node performs local fine-tuning of the global model to obtain a personalized federated model; Step 2: Based on the target conditions, filter historical reference solutions and call the personalized federated model to generate innovative solutions that meet engineering constraints. Combine rapid evaluation and fine simulation optimization to determine the final design solution. Step 3: Feed back the long-term performance data of the road built according to the final design scheme to the local database of the corresponding node to trigger a new round of federated model training and drive the continuous evolution of the system.

2. The method according to claim 1, characterized in that, In step 1, the weighted aggregation based on the value weight of the local data of each node specifically includes: Obtain the data value weight Vi of each node, and perform weighted aggregation on the received model parameters. The aggregation formula is as follows: ； in V is the model parameter uploaded for the i-th node. i The data value weight of the i-th node is... These are the parameters of the aggregated global model, where N is the total number of nodes.

3. The method according to claim 2, characterized in that, Data value weight V of each node i The calculation formula is: ； in, The average remaining lifespan L of the roads in the node database i The normalized value, This represents the scarcity coefficient of the environmental partition to which the node belongs. Structural Innovation Level I for Node Design Scheme i The normalized value of α, β, and γ is a preset weight coefficient, and α+β+γ=1.

4. The method according to claim 1, characterized in that, Step 2 specifically includes: Step 21: Based on the target conditions of the road to be built, select historical roads with high comprehensive similarity from the local database to form the first scheme set; Step 22: Calculate the remaining lifespan of each road in the first scheme set based on the long-term performance data of each road, and select the scheme with the longest remaining lifespan to form a second scheme set as a reference for historical experience. Step 23: Input the target conditions into the personalized federated generative design model. The generator of the model has an engineering constraint layer at the end, which is used to map the generated continuous parameters into discretized and standardized structural parameters that conform to engineering specifications, and output multiple innovative solutions to form a third solution set. Step 24: Merge the second and third scheme sets, and use the personalized federated agent evaluation model to perform rapid performance prediction and screening to form the fourth scheme set; Step 25: Perform finite element simulation and parameter fine-tuning on the schemes in the fourth scheme set, with the goal of maximizing the cumulative equivalent axle number, and output the final optimal pavement structure design scheme.

5. The method according to claim 4, characterized in that, The engineering constraint layer in step 23 includes a thickness discretization module, a modulus material library mapping module, and an interlayer modulus gradient verification module, which are executed sequentially.

6. The method according to claim 4, characterized in that, The specific steps for selecting historical roads with high comprehensive similarity in step 21 are as follows: using the analytic hierarchy process (AHP), an evaluation index system with climate zoning and soil resilient modulus as the core is established, the comprehensive similarity score between the road to be built and the historical road is calculated, and historical roads with scores higher than the preset threshold are selected.

7. The method according to claim 4, characterized in that, In step 22, the specific steps for calculating the remaining life are as follows: Based on the decay model of the Road Surface Damage Index (PCI) and the Structural Capacity Index (PSSI), calculate the remaining time required for the road to reach the minimum permissible performance value, and take the smaller value between the remaining time corresponding to PCI and PSSI as the remaining life of the road. The specific formula for the decay model is PCI = 100e (-0.06t) PSSI=100e (-0.04t) .

8. The method according to claim 1, characterized in that, The personalized federated model includes a personalized federated generative design model and a personalized federated agent evaluation model; wherein the generative design model is a conditional generative adversarial network and the agent evaluation model is a deep neural network.

9. An intelligent optimization system for asphalt pavement structure for implementing the method of any one of claims 1-8, characterized in that, The system includes: A central coordination server is configured to execute a model parameter weighted aggregation algorithm based on data value weights; Multiple regional node servers are deployed at each participating party. Each server includes a local historical road parameter database, a local model training module, and a personalized fine-tuning module. The design terminal includes a condition input module, a scheme generation module, a rapid evaluation module, and a simulation optimization module; The system is configured to feed back long-term performance data of newly built roads to the local database of the regional node server to trigger a new round of federated model training.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent optimization method for asphalt pavement structure based on federated generative learning as described in any one of claims 1 to 8.

Images