Road structure design method based on large model and reinforcement learning

Abstract

The invention belongs to the crossing field of road engineering technology, artificial intelligence and engineering mechanics, and particularly relates to a road structure design method based on a large model and reinforcement learning. A technical closed loop of'natural language input-parameter automatic mapping-specification standard value acquisition-finite element verification-reinforcement learning optimization 'is constructed: by constructing a load parameter mapping table and a material semantic encoder, the system can automatically convert natural language input into engineering parameters such as load, material, layer thickness and the like, and the problem of adaptability to non-standard working conditions is solved. And a deviation characteristic space is constructed by further combining a mechanical theory solution and finite element simulation, and the reinforcement learning agent is driven to quickly optimize under the condition of meeting theoretical constraints. Compared with a traditional trial calculation method and an existing intelligent optimization scheme, the method has the advantages that the number of design iterations and the material cost are greatly reduced, the safety and compliance of an output scheme are improved, and intelligent spanning from experience driving to theory guiding is achieved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of road engineering technology, artificial intelligence and engineering mechanics. Specifically, it relates to an intelligent design method for road structures that integrates semantic parsing of large language models and reinforcement learning optimization. It is applicable to the intelligent design and optimization of flexible pavement structures such as highways and urban roads. Background Technology

[0002] Traditional road structure design follows the theory of elastic layered systems, employing trial-and-error methods based on empirical formulas to determine the thickness of each structural layer. A typical process involves: initially selecting structural combinations based on traffic load levels, adjusting thickness parameters through mechanical response verification, and iterating repeatedly until design standards are met. This method suffers from two major problems: firstly, low level of automation, as design parameter input relies on manually filling out standardized forms, lacking support for natural language interaction and exhibiting poor adaptability to non-standard working conditions; secondly, insufficient optimization efficiency, as thickness adjustment depends on engineer experience, requiring repeated trial calculations, resulting in high finite element verification computation costs and difficulties in multi-objective collaborative optimization.

[0003] In recent years, the emergence of artificial intelligence technologies, especially deep learning algorithms, has made rapid design and optimization of pavement structures possible. For example, neural network-based asphalt pavement structure optimization methods have improved the efficiency of parameter sensitivity analysis through data-driven strategies, but have not solved the problem of deep integration between natural language interaction and reinforcement learning strategies. While optimization schemes combining genetic algorithms can achieve global optimization, they suffer from slow convergence speed and susceptibility to getting trapped in local optima. Reinforcement learning techniques, especially proximal policy optimization (PPO) algorithms, offer a new paradigm for pavement structure design and optimization due to their stability, efficiency, multi-objective collaboration, and high sample reusability. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention proposes a road structure design method based on large models and reinforcement learning. By constructing a technical closed loop of "natural language input - automated parameter mapping - standard value acquisition - finite element verification - reinforcement learning optimization", the method achieves intelligent upgrading of road structure design and significantly improves design efficiency.

[0005] Technical solution A road structure design method based on large models and reinforcement learning includes the following steps: Step 101: The user inputs natural language design requirements, and the natural language design requirements are converted into prompt words according to the preset prompt word template and sent to the pre-trained large language model; Step 102: The pre-trained large language model performs semantic parsing on the input prompt text to identify key engineering parameters and the relationships between parameters; Step 103, based on the engineering parameter mapping rule library, map the keywords to standard engineering parameters as parameter vectors that can be used for subsequent standard tolerance value extraction and finite element modeling; Step 104, generate a finite element modeling script according to the parameter vectors parsed in steps 102 and 103, drive the finite element software to perform finite element modeling, boundary condition application and response solving of the four-layer road structure (surface layer, base layer, subbase layer, soil base), and extract the finite element analysis result vector . Step 105, query the standard value, according to the key design parameters (especially traffic load grade, pavement structure combination type) obtained in steps 102 and 103, based on JTG D50-2017 "Highway Asphalt Pavement Design Specification", calculate the design traffic volume N e according to the input pavement structure parameters, and according to the design traffic volume N e and the corresponding standard tolerance value vector of the design parameters . Step 106, take the deviation of the finite element analysis result and the standard tolerance value as the input of the reinforcement learning agent, construct the state space of the reinforcement learning agent, and use the reinforcement learning algorithm (PPO) of the proximal policy optimization to iteratively adjust the thickness of the structure layer, realizing the optimization of the pavement structure.

[0006] Step 107, set the reward function for pavement structure design optimization and the training stability guarantee mechanism in the reinforcement learning agent, which is used to train the generated agent action suggestion.

[0007] Step 108, the agent outputs the reinforcement learning action suggestion , combined with the action suggestion given by the large language model , hybrid output pavement structure layer adjustment scheme, forming a new pavement structure layer.

[0008] Step 109, according to the action suggestion in step 108, modify and construct a new pavement structure parameter vector, re-perform finite element modeling and input into the reinforcement learning agent for training until the convergence requirement is met.

[0009] Wherein, the convergence requirement is that all key mechanical responses (σ , , ) of the current design scheme meet the standard tolerance value (i.e. ≤ ), and in the last N times (N=5) of iteration, the comprehensive reward r is less than the threshold value δ (δ=0.01) or the change of the structure layer parameters is small.

[0010] Step 110, output the optimized road structure design scheme meeting the safety, economy and engineering requirements.

[0011] Advantages: The application breaks through the limitation of traditional methods relying on manual form filling by constructing a semantic analysis engine through pre-training of a large language model, and automatically converting user natural language input into standardized load, material and structure parameters. Combined with an engineering parameter mapping rule library, intelligent analysis of unstructured text into engineering parameters is realized, significantly improving non-standard working condition adaptability and input efficiency; The application innovatively integrates standard allowable standard values and finite element numerical simulation results to construct a state space centered on the deviation between the two, driving the reinforcement learning optimization process. Compared with pure data-driven solutions, it can ensure that the optimization direction meets both specification requirements and engineering practical constraints; The application adopts a multi-objective reward function driven engineering decision mechanism, constructs a reward function integrating safety, economy, dynamic penalty and smoothness, and ensures stable convergence of structure reliability, economy and meeting engineering specification requirements; The intelligent optimization process of the application adopts closed-loop verification and has an interpretable decision support, embeds a parameter verification module in the modeling stage to ensure the compliance of design assumptions. At the same time, the output optimization layer thickness parameters, material configuration and multi-dimensional verification report provide automatic and interpretable decisions that cannot be achieved by traditional methods. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 is a road structure design method flowchart based on the large model and reinforcement learning of the application; Figure 2 is a natural language input and parameter completion logic schematic diagram of an embodiment of the application; Figure 3 is a schematic diagram of constructing a reinforcement learning agent state space of an embodiment of the application; Figure 4 is a schematic diagram of road surface structure layer parameter adjustment and training stability guarantee mechanism of an embodiment of the application. DETAILED DESCRIPTION

[0013] The technical solutions provided by the application will be further described below in combination with specific embodiments and their accompanying drawings. The advantages and features of the application will be clearer in combination with the following description.

[0014] In related technologies, traditional road structure design methods rely on manual experience and trial-and-error algorithms, requiring repeated adjustments to layer thickness parameters and verification through finite element analysis. This results in low levels of intelligence and long optimization cycles. While existing intelligent optimization schemes (such as neural networks and genetic algorithms) can improve efficiency, they face two major bottlenecks: first, parameter input relies on standardized tables, making it unable to parse non-standard working conditions described in natural language; second, the optimization process lacks standardized constraints, easily leading to engineering infeasibility solutions such as unbalanced layer thickness ratios and excessive material modulus, requiring secondary manual verification.

[0015] To address the aforementioned issues, this embodiment proposes a road structure design method integrating a large language model and reinforcement learning. By constructing a load parameter mapping table and a material semantic encoder, the system can automatically convert natural language input into engineering parameters such as load, material, and layer thickness, solving the adaptability problem for non-standard working conditions. Furthermore, by combining mechanical theory solutions with finite element simulations to construct a deviation feature space, the reinforcement learning agent is driven to rapidly find the optimal solution while satisfying theoretical constraints. Compared to traditional trial-and-error methods and existing intelligent optimization schemes, this embodiment significantly reduces the number of design iterations and material costs, improves the safety and compliance of the output solution, and achieves an intelligent leap from experience-driven to theory-guided approaches.

[0016] like Figure 1 As shown, a road structure design method based on large models and reinforcement learning includes the following steps: Step 101: The user inputs natural language design requirements. Based on a preset prompt word template, the natural language design requirements are converted into prompt words and sent to a pre-trained large language model, such as... Figure 2 As shown.

[0017] The prompt word template is as follows: systemPrompt = ['You are a road engineering expert. Parse road structure parameters based on user input and return them strictly in JSON format:' ... '{"traffic_level":"light load / medium load / heavy load / extremely heavy load","road_type":"highway / urban road / rural road / industrial road",' ... '"vehicle_speed_kmh": value, "subgrade_conditions": "soft soil / general soil / hard soil / rock",' ... '"subgrade_treatment":"natural soil foundation / improved soil foundation / replacement soil foundation / composite foundation",' ... '"layers":[{"layer_type":"surface layer","thickness_cm":18,"material":"asphalt concrete","modulus_mpa":1200,"poisson":0.25},' ... '{"layer_type":"base layer","thickness_cm":35,"material":"cement-stabilized crushed stone","modulus_mpa":400,"poisson":0.35},' ... '{"layer_type":"substrate","thickness_cm":25,"material":"graded crushed stone","modulus_mpa":180,"poisson":0.40},' ... '{"layer_type":"soil base","thickness_cm":150,"material":"modulus_mpa":50,"poisson":0.45}]}').

[0018] Example of a natural language design requirement: "Design the base course of a heavy-load highway, with the subbase course using graded crushed stone"; Step 102: The pre-trained large language model (such as DeepSeek) receives the above prompt text and uses the natural language understanding capabilities built into the large language model to perform semantic parsing on the input content to identify key engineering parameters and the relationships between parameters.

[0019] The Large Language Model (LLM) receives natural language prompts input by the user, which contain road structure design requirements; the prompts are semantically segmented using a pre-trained large language model (such as DeepSeek) to extract or infer key engineering parameters and the relationships between them; The key engineering parameters include: the thickness of each pavement structural layer. and material parameters; material parameters include: elastic modulus Compared to Poisson ,in Indicates the structural layer number; The specific processing steps of a large language model include: (1) Entity recognition: The large language model recognizes key entities related to road engineering in the text, including but not limited to traffic load level (e.g., "heavy load"), road type (e.g., "expressway"), structural layer material (e.g., "graded crushed stone"), etc. (b) Relationship extraction: Large language model identifies semantic and engineering relationships between entities, such as "bottom base layer adopts graded gravel" is parsed as the material membership relationship between "bottom base layer" and "graded gravel"; (c) Keyword extraction: Based on the above analysis, the large language model extracts a set of keywords that can be used for subsequent parameter mapping, such as "heavy load", "highway", "graded gravel", etc. These keywords will be used as input for engineering parameter mapping.

[0020] Through the above processing, the large language model completes the deep semantic analysis of the user's natural language demand, identifies the key design parameters and their relationship, and provides a structured information basis for the subsequent standard mapping of parameters.

[0021] Step 103, based on the engineering parameter mapping rule library, map the keywords to standard engineering parameters, and return to the system in JSON format as a parameter vector that can be used for subsequent standard allowable value extraction and finite element modeling.

[0022] The engineering parameter mapping rule library includes load parameter mapping library, material semantic encoder and structure parameter reasoning module, which is used to map the key engineering parameters identified by the large language model to the engineering parameter database, and generate a vector of standard loads P, layer thickness , material parameters (elastic modulus , Poisson's ratio ) that can be used for subsequent standard specification allowable value query and finite element modeling.

[0023] For example, the load parameter mapping table matches "heavy load highway" as the axle load P=0.7MPa and the action radius r=15.3cm; The material semantic encoder parses "graded gravel" as modulus =180MPa, Poisson's ratio =0.3; The structure parameter reasoning module completes the missing bottom base layer thickness h=35cm (according to the soil base modulus Es=50MPa).

[0024] Step 104, generate a finite element modeling script according to the parameter vector parsed in step 102 and step 103, drive the finite element software to perform finite element modeling, boundary condition application and response solving of four-layer road structure (surface layer, base layer, bottom base layer, soil base), and extract the finite element analysis result vector : including the maximum tensile stress at the bottom of the surface layer , the maximum tensile strain allowable value of the base layer and the maximum deflection of the roadbed top surface .

[0025] The finite element software is used for finite element modeling. Key engineering parameters are input into the finite element software (such as ABAQUS, MATLAB built-in PDEToolbox, etc.), and an initial road structure finite element model is constructed through an automatic script generation module, and the finite element software is driven to solve, and the finite element solving results are extracted through post-processing files .

[0026] Specifically, it includes: Material property definition: associate modulus , Poisson's ratio to each structural layer section; Load application: according to BZZ-100 standard, uniformly distributed load P = 0.7 MPa is applied in the wheel trace area, and the action radius r = 15.3 cm; Meshing: according to the size and boundary conditions of the current pavement structure, the mesh is divided to generate a refined model; Solving and result extraction: solve through the finite element simulation software and extract the response values of the structural layer including the maximum tensile stress at the bottom of the surface layer , the maximum tensile strain allowable value at the bottom of the base layer and the maximum deflection at the top of the subgrade .

[0027] Step 105, query the standard value, according to the key design parameters (especially the traffic load grade, pavement structure combination type) obtained in steps 102 and 103, based on JTG D50-2017 “Highway Asphalt Pavement Design Specification”, calculate the design traffic volume N e according to the input pavement structure parameters, according to the design traffic volume N e and the corresponding standard allowable value vector of the design parameters (computing method is prior art): Among them, represents the allowable tensile strain at the bottom of the surface layer (unit: MPa), which is used to control fatigue cracking; represents the allowable tensile stress at the bottom of the base layer (unit: ), which is used to control structural damage; represents the allowable deflection at the top of the subgrade (unit: mm) ， , which is used to control overall deformation.

[0028] Specifically, the specification standard value database is used to query the core design standard data corresponding to the input pavement structure: the Highway Asphalt Pavement Design Specification (JTG D50-2017) is preset therein, and the key design parameters parsed from the large language model (LLM) and the engineering parameter mapping rule library are received, including the traffic load level (such as extremely heavy, heavy, medium, and light traffic), and the pavement structure combination type (such as the typical structure of asphalt pavement). Based on the above parameters, the Highway Asphalt Pavement Design Specification (JTG D50-2017) is queried, and the allowable standard value vector corresponding to the design condition specified in the specification is output , specifically including: the allowable tensile strain at the bottom of the surface layer (for controlling fatigue cracking), the allowable tensile stress at the bottom of the base layer (mainly for controlling structural damage), and the allowable deflection at the top surface of the subgrade (mainly for controlling overall deformation).

[0029] Step 106, taking the deviation of the finite element analysis result from the allowable standard value as the input of the reinforcement learning agent, and constructing the state space s of the reinforcement learning agent, as shown in Figure 3 . The reinforcement learning algorithm (PPO) is used to iteratively adjust the thickness of the structure layer, and the pavement structure optimization is realized.

[0030] The state space s of the reinforcement learning agent is constructed by the deviation of the finite element analysis result from the standard allowable value as follows: wherein, , , are the finite element modeling stress values, strain values, and displacement values obtained by the finite element software; , are the allowable standard stress values, allowable standard strain values, and allowable standard displacement values obtained by querying the specification standard value database; The state space s of the reinforcement learning agent is constructed by the deviation of the finite element analysis result from the standard allowable value obtained by querying the specification as the input of the agent; Step 107, setting a reward function for pavement structure design optimization and a training stability guarantee mechanism in the reinforcement learning agent, for training the generated agent action suggestion.

[0031] Specifically: Taking into account safety, economy, dynamic penalty, and process stability in pavement structure design, a reward function for reinforcement learning is constructed by adopting a weight allocation mechanism that prioritizes safety, followed by economy, and supplemented by penalty and smoothing effects: in , , , These are respectively: safety item, economic item, dynamic penalty item, and smoothing item.

[0032] Safety items The design utilizes the results of quantified finite element analysis. Compared with the standard allowable value The deviation is used to assess the structural safety. The form fully meets the specifications ( ≤ When the score is exceeded, the item receives full marks; if the score is exceeded, the item receives full marks. > Points are deducted proportionally when stress control is activated. The weighting reflects priority: stress control has the highest weight. ), and adaptability is secondary. ), lowest displacement / deflection ( The formula is as follows: Economic items The design uses an exponential function to measure the ratio of actual material cost to theoretical cost; the lower the actual cost, the higher the reward. This design method employs nonlinear incentives to exponentially reinforce the reward difference for low-cost solutions, accelerating convergence; it also comprehensively considers unit price. ,thickness Modulus The product relationship accurately reflects the cost-driving effect, enabling the optimization of material usage and combination in the design and reducing project costs, as follows: in, , The measured material cost (yuan / m2) for the current design scheme. Let be the unit price of the i-th layer material (yuan / MPa·m3). , These are the thickness and modulus of the i-th layer, respectively; Dynamic penalty items Design: When any finite element extract exceeds 1.15 times its specification allowable value, an additional linear penalty is applied. The penalty factor is set to the most severe value for the stress term. ), and adaptability is secondary. ), lowest displacement / deflection ( ),Right now: Smoothing Term Designed to evaluate thickness changes over the last three iterations. The smaller the fluctuation, the higher the reward. This design scheme avoids oscillatory convergence by suppressing parameter mutations, effectively ensuring the stability of reinforcement learning. Simultaneously, this design ensures gradual optimization of the design scheme, avoiding increased construction difficulty or structural performance instability due to drastic changes in layer thickness. For example, if the base layer thickness fluctuates by more than 5% in adjacent iterations, the system will reduce the learning rate to adjust smoothly. To ensure the rationality of action suggestions, a stability guarantee mechanism for reinforcement learning agent training is set up, including: rollback mechanism, adaptive learning rate, and policy entropy constraint.

[0033] The rollback mechanism design, through finite element analysis, detects when plastic strain εp ≥ 2% or layer thickness ratio > 2.0, terminating the current training cycle and restoring the optimal state of the five most recent feasible solutions from the historical cache. This mechanism can promptly terminate parameter adjustments that may lead to structural failure (such as plastic deformation caused by excessively thin base layers), thereby achieving risk avoidance. Simultaneously, it avoids wasting computational resources on invalid paths, improving training efficiency and achieving resource optimization. In actual road structure design, if the algorithm excessively reduces the base layer thickness (e.g., from 30cm to 25cm) in pursuit of economy, it may lead to an imbalance in the layer thickness ratio. This mechanism will automatically revert to the historical feasible state, ensuring that the design meets the interlayer compatibility requirements. Adaptive learning rate design: In the PPO algorithm, a linearly decaying learning rate is designed as follows: in The preset maximum number of training steps, t is the current number of training steps ( ).

[0034] This design method enables rapid exploration of the parameter space in the early stages of design and fine-tuning in the later stages, while avoiding oscillations or premature convergence caused by a fixed learning rate. Policy entropy constraint: Set the target KL divergence threshold to 0.02, that is, when the actual KL divergence... Automatically reduce the learning rate using the following formula: This design method can limit the magnitude of policy updates and prevent abrupt changes. At the same time, it can automatically suppress the learning rate when the KL divergence exceeds the limit, maintain stable training, and achieve adaptive adjustment.

[0035] The objective function of the PPO algorithm is: wherein, is the objective function of the PPO, is the optimization objective of the policy parameter , and therefore it is desired to maximize this value in the operation; is the importance sampling ratio, wherein represents the probability of the new policy selecting action a t in state s t , and represents the probability of the old policy selecting action a t in state s t ; is the expectation of the time step t, which is usually the average value in a batch of sampling data; is the advantage function estimate value, which represents the good or bad of taking action a t in state s t relative to the average performance; is the clipping function, which limits to the interval[ , and this operation is the core of the PPO policy, which is used to prevent the policy update from being too aggressive, is the clipping parameter.

[0036] Step 108, the agent outputs the reinforcement learning action suggestion , combines the action suggestion given by the large language model, and outputs the road structure layer adjustment scheme to form a new road structure layer. Among them: The reinforcement learning agent action suggestion is as follows: According to the constructed state space and the environment interaction, the reinforcement learning agent outputs the action on each layer structure, and the original adjustment value is as follows: wherein, ; are the original adjustment values of the surface layer, the base layer, and the subbase layer thickness, respectively; are the original adjustment values of the surface layer, the base layer, and the subbase layer modulus, respectively; It should be noted that in the process of road structure design, the soil foundation is often regarded as a semi-infinite body and only used for checking calculation without participating in actual design, so the reinforcement learning agent only iteratively optimizes the surface layer, the base layer, and the subbase layer in the road structure.

[0037] The layer thickness adjustment value calculation formula is: wherein, respectively represent the surface layer, the base layer, and the subbase layer; The modulus adjustment calculation formula is: wherein, respectively represent the surface layer, the base layer, and the subbase layer; The reinforcement learning agent action suggestion is composed of the calculated layer thickness adjustment value and modulus adjustment value: Large language model action suggestion As follows: Convert the reinforcement learning state space s into a natural language description, construct a prompt word, and send it to the large language model: Wherein, the large language model action suggestion prompt word template is as follows: "You are a senior road structure engineer. Focus on structural performance, load distribution, and JTG D50-2017 specification compliance.

[0038] The current structure state is: Stress utilization rate: {σ_FEA / σ_std × 100:.1f}% Strain utilization rate: {ε_FEA / ε_std × 100:.1f}% Deflection utilization rate: {D_FEA / D_std × 100:.1f}%,, Current: Surface layer: {surface layer thickness:.1f} cm, {surface layer modulus:d} MPa Base layer: {base layer thickness:.1f} cm, {base layer modulus:d} MPa Subbase layer: {subbase layer thickness:.1f} cm, {subbase layer modulus:d} MPa, Current cost: {current total cost:.0f} yuan / m², please provide structure optimization suggestions focusing on JTG D50-2017 specification compliance. Please return in JSON format: {"thickness_adjustments": [t1,t2,t3],"modulus_adjustments": [m1,m2,m3],"engineering_analysis": {"design_recommendation": "Specific recommendations", "priority_layer": "Priority layer", "structural_concern": "Main problems", "compliance_status": "Compliance status with specifications"}}. Adjustment values range between -0.15 and 0.15.

[0039] The large language model will return a JSON response, for example: { "thickness_adjustments": [-0.05, 0.08, 0.03], "modulus_adjustments": [-0.10, 0.05, 0.00], "engineering_analysis": { "design_recommendation": "Recommend reducing surface layer thickness and increasing base layer thickness to improve stress distribution", "priority_layer": "Base layer", "structural_concern": "Surface layer stress utilization is too high", "compliance_status": "Exceeds JTG D50-2017 stress limit"}}.

[0040] After system analysis, the large language model action vector is formed: Reinforcement learning and large language model mixed action recommendations: According to the mixed strategy defined in the invention content, the reinforcement learning action recommendation and the large language model action recommendation are mixed as follows: where, is the mixed recommendation given by the agent for the pavement structure layer adjustment; β=0.3 is the preset influence weight of the large language model recommendation; The mixed action recommendation is decomposed into layer thickness adjustment component and modulus adjustment component : where, Structure layer parameter update rule: Based on hybrid action suggestion , each structure layer parameter is updated as follows: Layer thickness update: Wherein and are the layer thickness before and after the update, respectively; Modulus update: Wherein and are the modulus before and after the update, respectively; In the reinforcement learning agent, a progressive action constraint method is set. When the adjacent layer thickness ratio exceeds the preset buffer threshold, that is, when the ratio of the base layer thickness to the bottom base layer thickness satisfies the following formula , the action of increasing the base layer thickness or reducing the bottom base layer thickness is prohibited. In the formula, represents the base layer thickness, represents the bottom base layer thickness, and the constraint process is shown in Figure 4 ; Step 109, according to the action suggestion of step 108, modify and construct a new pavement structure parameter vector, re-perform finite element modeling and input into the reinforcement learning agent training until the convergence requirement is met. Among them, the convergence requirement is that all key mechanical responses (σ FEA, ε FEA , D FEA ) of the current design scheme after finite element modeling meet the specification allowable standard value (i.e. ≤ ), and in the last N (N=5) iterations, the comprehensive reward r is less than the threshold value δ (δ=0.01) or the structure layer parameter changes.

[0041] Step 110, output the optimized road structure design scheme that meets the safety, economy and engineering requirements, including: Optimized thickness of each structure layer (such as “base layer from 32 cm to 28 cm”), material parameters and safety factor; Visual comparison: comparison of theoretical solution and finite element stress nephogram; Economic analysis: generate material consumption comparison table, for example, “save cement 15.2 tons / km, cost reduction 14.7%”.

[0042] The application breaks through the limitation of traditional methods relying on manual form filling by pre-training a large language model to construct a semantic analysis engine, which automatically converts user natural language input into standardized loads, materials and structure parameters. The large language model has high accuracy in parameter analysis. As shown in Table 1, the introduction of the large language model can greatly improve efficiency.

[0043] Table 1 Comparison table of parameter analysis accuracy of different large models The above description is only a description of the preferred embodiments of the application, and is not any limitation on the scope of the application. Any modification or modification made by any person skilled in the art according to the above disclosed technical content should be regarded as an equivalent effective embodiment, and belongs to the protection scope of the technical scheme of the application.

Claims

1. A road structure design method based on a large model and reinforcement learning, characterized by, Includes the following steps: Step 101: The user inputs natural language design requirements, and the natural language design requirements are converted into prompt words according to the preset prompt word template and sent to the pre-trained large language model; Step 102: The pre-trained large language model performs semantic parsing on the input prompt text to identify key engineering parameters and the relationships between parameters; Step 103: Based on the engineering parameter mapping rule library, map keywords to standard engineering parameters, which will serve as parameter vectors for subsequent standard allowable value extraction and finite element modeling. Step 104, generating a finite element modeling script according to the parameter vector parsed in step 102 and step 103, driving the finite element software to perform finite element modeling of the four-layer road structure, boundary condition application and response solving, and extracting a finite element analysis result vector , the four-layer road structure comprising a surface layer, a base layer, a bottom base layer, and a soil base; Step 105, query the specification standard value, according to the key design parameters obtained in steps 102 and 103, based on JTGD50-2017 "Highway Asphalt Pavement Design Specification", calculate the design traffic volume N according to the input pavement structure parameters e , according to the design traffic volume N e , the corresponding specification allowable standard value vector of the design parameter ; Step 106, finite element analysis results deviation from the standard allowable standard value The deviation from the standard allowable standard value is used as the input of the reinforcement learning agent, the state space of the reinforcement learning agent is constructed, the reinforcement learning algorithm PPO of the proximal policy optimization is used to iteratively adjust the thickness of the structure layer, and the pavement structure optimization is realized. Step 107: Set up a reward function and a training stability guarantee mechanism in the reinforcement learning agent to optimize the road structure design, and use them to train and generate action suggestions for the agent. Step 108, the agent outputs a reinforcement learning action suggestion , combined with the action suggestion given by the large language model , the mixed output road structure layer adjustment scheme forms a new road structure layer; Step 109: Based on the action suggestions in Step 108, modify and construct a new road structure parameter vector, re-model the finite element model, and input it into the reinforcement learning agent for training until the convergence requirement is met. Step 110: Output an optimized road structure design scheme that meets safety, economy and engineering requirements. 2.The road structure design method based on large model and reinforcement learning according to claim 1, wherein, In step 101, the prompt word template is: systemPrompt = ['You are a road engineering expert.'] Based on user input, road structure parameters are parsed and returned strictly in JSON format: '...' '{"traffic_level":"light load / medium load / heavy load / extremely heavy load","road_type":"highway / urban road / rural road / industrial road",' ... '"vehicle_speed_kmh": numerical value, "subgrade_conditions": "soft soil / general soil / hard soil / rock",'... '"subgrade_treatment":"natural soil foundation / improved soil foundation / replacement soil foundation / composite foundation",' ... '"layers":[{"layer_type":"surface layer","thickness_cm":18,"material":"asphalt concrete","modulus_mpa":1200,"poisson":0.25},' ... '{"layer_type":"base layer","thickness_cm":35,"material":"cement-stabilized crushed stone","modulus_mpa":400,"poisson":0.35},' ... '{"layer_type":"substrate","thickness_cm":25,"material":"graded crushed stone","modulus_mpa":180,"poisson":0.40},' ... '{"layer_type":"soil base","thickness_cm":150,"material":"modulus_mpa":50,"poisson":0.45}]}'). 3.The road structure design method based on large model and reinforcement learning according to claim 1, characterized in that, In step 102, the large language model receives a user input natural language prompt word containing road structure design requirements; the pre-trained large language model performs semantic segmentation on the prompt word, extracts or infers key engineering parameters and the relationship between parameters; The key engineering parameters include: thickness of each pavement structure layer and material parameters, the material parameters include: elastic modulus and Poisson's ratio wherein denotes the structure layer number; In step 103, the engineering parameter mapping rule library includes a load parameter mapping library, a material semantic encoder, and a structure parameter reasoning module, which is used to map the key engineering parameters identified by the large language model to the engineering parameter database, to generate standard loads P for subsequent specification standard allowable value query, finite element modeling, and thicknesses of each pavement structure layer , material parameters of the vector; In step 104, the finite element analysis result vector is: Maximum tensile stress at the bottom of the surface layer Maximum tensile strain at the bottom of the base layer Maximum deflection at the top of the subgrade . 4.The road structure design method based on large model and reinforcement learning according to claim 3, characterized in that, In step 105, the specification allows the standard value vector to be: Wherein, represents the allowable tensile strain at the bottom of the surface layer for controlling fatigue cracking, in MPa; Representative of the base of the bottom of the allowable tensile stress, for the control of structural damage, units ; representative of the subgrade surface deflection ， for controlling the overall deformation, in mm; The specific process is as follows: Using the specification standard value database for querying the core design standard data corresponding to the input pavement structure: In which the "Highway Asphalt Pavement Design Specification" JTG D50-2017 is preset, receiving the key design parameters parsed from the large language model and the engineering parameter mapping rule library, including the traffic load grade, the pavement structure combination type; Based on the above parameters, query the "Highway Asphalt Pavement Design Specification" JTG D50-2017 and output the allowable standard value vector specified in the specification corresponding to the design condition , Specifically including: the allowable tensile strain at the bottom of the surface layer , the allowable tensile stress at the bottom of the base layer , the allowable deflection of the subgrade top surface . 5.The road structure design method based on large model and reinforcement learning according to claim 4, characterized in that, In step 106, the finite element analysis result deviation from the standard allowable value The reinforcement learning agent state space s is constructed as follows: Wherein, , , respectively the stress value, the strain value and the displacement value obtained by finite element modeling of the finite element software; , are the allowable standard stress value, the allowable standard strain value and the allowable standard displacement value respectively obtained by querying the norm standard value database; with the finite element analysis results deviation from the standard tolerance value obtained by querying the specification The reinforcement learning state space is constructed as the agent input. 6.The road structure design method based on large model and reinforcement learning according to claim 4, characterized in that, In step 107, the reward function of reinforcement learning is as follows: wherein , , , are a safety term, an economic term, a dynamic penalty term and a smoothing term, respectively; Safety items The formula is as follows: Economic item As follows: Wherein, , is the measured material cost of the current design scheme, unit: yuan / m2; is the unit price of the i-th layer material, unit: yuan / MPa·m3, , are the thickness and modulus of the i-th layer, respectively; Dynamic penalty term The formula is as follows: smooth term Designed to evaluate the degree of fluctuation in thickness change in the last 3 iterations, the smaller the fluctuation, the higher the reward: ​ In order to ensure the rationality of the action suggestion, a reinforcement learning agent training stability guarantee mechanism is set, including: rollback mechanism, adaptive learning rate and policy entropy constraint; Wherein, the rollback mechanism is designed to detect plastic strain εp≥2% or layer thickness ratio >2.0 by finite element analysis, terminate the current training period and restore the optimal state of the last 5 feasible schemes from the history cache; Adaptive learning rate design: in the PPO algorithm, the learning rate is designed as follows: wherein is a preset maximum training step number, t is a current training step number, ; Policy entropy constraint: Set the target KL divergence threshold to 0.02, i.e., when the actual KL divergence is automatically reduced by the learning rate as follows: ； The objective function of PPO algorithm is: where is the importance sampling ratio, where is the probability of the new policy to select action a t in state s t , is the probability of the old policy to select action a t in state s t ; is the advantage function estimate, is the clipping parameter; is the clipping function, which limits to the interval . 7.The road structure design method based on large model and reinforcement learning according to claim 4, characterized in that, In step 108, the reinforcement learning agent action suggestion As follows: According to the constructed state space and environmental interaction, the reinforcement learning agent outputs the action for each layer structure, and the original adjustment value is as follows: wherein ; original adjustment values of the thickness of the surface layer, base layer, and sub-base layer, respectively; original adjustment values of the modulus of the surface layer, base layer, and sub-base layer, respectively; The layer thickness adjustment value calculation formula is: wherein , respectively, represent the surface layer, the base layer, the subbase layer; The modulus adjustment calculation formula is: wherein , respectively, represent the surface layer, the base layer, the subbase layer; To reinforce the learning agent action suggestion, the calculated layer thickness adjustment value and the modulus adjustment value are composed: ； Large language model action suggestions As follows: Convert the reinforcement learning state space s into natural language description, construct the prompt word and send it to the large language model; The large language model will return a JSON response, which is parsed by the system to form a large language model action vector: 。 8.The road structure design method based on large model and reinforcement learning according to claim 7, characterized in that, Reinforcement learning action suggestions Large language model action suggestions Mixing by the following formula: Wherein, a hybrid suggestion for the agent to adjust to the road structure layer; β=0.3 is the preset influence weight of large language model suggestion; wherein the mixing action recommendation is decomposed into a layer thickness adjustment component and a modulus adjustment component : Wherein, Structure layer parameter update rule: Based on hybrid action suggestion each structure layer parameter is updated as follows: Layer thickness update: wherein and respectively the layer thickness before and after the update; Modulus update: wherein and Epre and Epost are the modulus before and after updating, respectively; A method of setting a design progressive action constraint in a reinforcement learning agent is provided, wherein when the thickness ratio of adjacent layers exceeds a pre-set buffer threshold, i.e. when the action of increasing the thickness of the base layer or decreasing the thickness of the sub-base layer is prohibited from being performed. 9.The road structure design method based on large model and reinforcement learning according to claim 7, characterized in that, In step 108, the template of the prompt word is as follows: "You are a senior road structure engineer. Focus on structural performance, load distribution and JTG D50-2017 specification compliance. The current structure state is: Stress utilization rate: {σ_FEA / σ_std × 100:.1f}% Strain utilization rate: {ε_FEA / ε_std × 100:.1f}% Deflection utilization rate: {D_FEA / D_std × 100:.1f}%, current: Surface layer: {surface layer thickness:.1f}cm, {surface layer modulus:d}MPa Base layer: {base layer thickness:.1f}cm, {base layer modulus:d}MPa Subbase layer: {subbase layer thickness:.1f}cm, {subbase layer modulus:d}MPa, Current cost: {current total cost:.0f} yuan / m², please provide structure optimization suggestions focusing on JTG D50-2017 specification compliance. Please return in JSON format: ​ {"thickness_adjustments": [t1, t2, t3], "modulus_adjustments": [m1, m2, m3], "engineering_analysis": {"design_recommendation": "specific recommendations", "priority_layer": "priority layer", "structural_concern": "main issues", "compliance_status": "compliance status"}}. Adjustment values range between -0.15 and 0.

15. 10.The road structure design method based on large model and reinforcement learning according to claim 1, wherein, In step 109, the convergence requirement is that all key mechanical responses of the current design scheme, after finite element modeling, meet the allowable standard values ​​specified in the code. Furthermore, in five consecutive iterations, the increase in the overall reward r is less than the threshold δ or the change in the structural layer parameters is small.

Images