Intelligent agent system for asphalt pavement construction based on neural operator counterfactual world model

Abstract

The application discloses an asphalt pavement construction intelligent agent system based on a neural operator counterfactual world model, which comprises a construction state perception and continuous material field construction, a process semantic understanding and candidate strategy generation, a counterfactual construction scene construction, a neural operator counterfactual world model, a multi-future evaluation and strategy selection, and a deterministic constraint and execution feedback module. The system constructs temperature, skeleton, compaction and equipment action state into a continuous material field, generates candidate strategies such as strong vibration, weak vibration, rubber wheel, static pressure, slow paving and machine group sequence adjustment by a large model, forms multiple counterfactual futures under the same current state, predicts future temperature field, skeleton state field and compaction uniform field by using a neural operator, selects an optimal strategy according to construction quality, efficiency, risk and prediction uncertainty, and solves final equipment control parameters through a deterministic constraint layer and rolling feedback correction, so that predictive intelligent control and machine group cooperation of asphalt pavement construction are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction technology for road engineering, and specifically to an intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model. Background Technology

[0002] Asphalt pavement construction is a dynamic process involving the combined effects of material temperature decay, mixture particle rearrangement, asphalt viscosity changes, compaction work input, equipment trajectory coverage, and environmental heat exchange. The operating sequence, speed, spacing, vibration mode, and number of compaction passes among pavers, double-drum rollers, rubber-tired rollers, and static compaction equipment all affect the final compaction degree, uniformity, skeleton interlocking state, and smoothness of the pavement.

[0003] Current asphalt pavement construction control typically relies on pre-set construction parameters, manual experience, threshold alarms, or statistical models based on single-point data. For example, based on current temperature, number of compaction passes, or compaction degree test results, control suggestions may be given such as "increase vibration intensity," "reduce paving speed," or "adjust roller speed to a certain value." This type of method mainly suffers from the following problems.

[0004] First, existing construction scheduling methods based on large models or general intelligent agents readily convert natural language reasoning results into specific control commands. While large models can understand construction processes, equipment names, and anomaly descriptions, their outputs are essentially probabilistically generated results, and cannot guarantee the physical feasibility of the generated speeds, vibration frequencies, amplitudes, number of passes, or machine group sequences under the current material and equipment conditions. Directly providing and executing specific equipment parameters from large models can easily lead to risks such as exceeding equipment capacity boundaries, violating compaction temperature windows, causing machine group conflicts, or inducing over- or under-pressure.

[0005] Second, existing machine learning prediction models mostly use a limited number of sensor values ​​or statistical features as input, and output single-point compaction degree, final quality grade, or temperature at a certain moment. They are difficult to simultaneously predict the continuous temperature field, skeleton state field, and compaction uniformity field of asphalt mixtures within the construction area. The material state of asphalt mixtures has obvious spatiotemporal continuity and equipment action propagation characteristics. Relying solely on point value predictions is insufficient to describe spatial anomalies such as joints, edges, shutdown areas, localized low-temperature zones, and areas of repeated equipment coverage.

[0006] Third, existing construction optimization methods typically calculate the result of only one preset action or select an immediate action based on the current error, failing to simultaneously construct multiple counterfactual futures under the same initial construction conditions. For example, when localized temperature drops and uneven compaction occur on the pavement, strong vibration, weak vibration, rubber-tired paving, static pressure, slowing down paving, and adjusting the tractor sequence may produce different material field evolution results a few minutes later. Existing technologies struggle to simultaneously generate multiple virtual futures before actual execution and select actions based on future temperature, framework, and compaction status.

[0007] Fourth, traditional digital twin or autonomous driving world models are mainly geared towards predicting discrete states, visual scenes, traffic participant trajectories, or vehicle motion states. Their prediction objects and constraint methods differ significantly from the continuous material field of asphalt mixtures. Asphalt construction world models need to handle the coupling relationships between material thermal state, particle skeleton state, compaction response, and the boundaries of multiple equipment actions, and be able to perform continuous field simulations under different construction lengths, different grid resolutions, and different machine group combinations.

[0008] Fifth, in existing intelligent decision-making systems, safety rules are typically input into the model as prompts or as soft penalty terms in the evaluation function. When the model's return is high enough, soft penalties cannot guarantee, in terms of mechanism, that the output instructions will meet the requirements of equipment limits, temperature windows, equipment spacing, and control rate of change.

[0009] Therefore, a new intelligent agent for asphalt pavement construction needs to be constructed. This agent does not allow large models to directly control construction equipment. Instead, the large model understands the process and generates candidate strategies. A neural operator world model that can predict the evolution of continuous material fields performs counterfactual deductions on multiple candidate strategies. Finally, a deterministic constraint layer completes the solution and execution of the final control quantity, thus forming a closed loop of "process understanding - multiple future prediction - deterministic verification - equipment execution - feedback correction". Summary of the Invention

[0010] Based on the above description, this invention provides an intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model, in order to solve the problems of existing construction intelligent agents that directly generate specific equipment instructions from large models, cannot predict continuous material fields, cannot compare multiple counterfactual futures before execution, and lack deterministic safety constraints.

[0011] The technical solution of the present invention to solve the above-mentioned technical problems is as follows: An intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model, comprising: The construction status perception and continuous material field construction module is used to acquire material data, construction equipment data, quality inspection data and environmental data during the asphalt pavement construction process. It performs time synchronization, coordinate unification, anomaly handling and spatial reconstruction on the acquired data to form a continuous material state including temperature field, skeleton state field, compaction state field, compaction uniformity field, equipment action field and state credibility field. The process semantic understanding and candidate strategy generation module is used to identify the current construction status and abnormal areas based on continuous material status, construction equipment status, construction tasks and construction process knowledge, and to generate multiple structured candidate strategies using the large model. The candidate strategy includes at least action type, acting equipment, acting area, acting time, parameter allowable range, preconditions and prohibition conditions. The large model does not directly issue the final control parameters to the construction equipment. The counterfactual construction scenario construction module is used to freeze the current construction facts that have already occurred. It uses the same current continuous material state, equipment initial state, material properties and environmental conditions as common initial conditions, and converts different candidate strategies into corresponding equipment trajectories, equipment action areas, operation modes and time relationships to form multiple counterfactual construction scenarios. The neural operator counterfactual world model module is used to predict the temperature field, skeleton state field, compaction state field and compaction uniform field at multiple future moments after the application of different candidate strategies, based on the continuous material state, equipment spatiotemporal action conditions, environmental boundary conditions and construction area geometric conditions in each counterfactual construction scenario. The multi-future evaluation and strategy selection module is used to evaluate the construction quality, construction efficiency, equipment operating status, safety risks and prediction uncertainties of candidate strategies based on the future continuous material field of different counterfactual construction scenarios, determine the preferred strategy and retain at least one suboptimal candidate strategy. The deterministic constraint and execution feedback module is used to solve the final control parameters of the construction equipment based on the parameter allowable range of the preferred strategy, equipment capacity, material state, safety distance, construction area boundary and control change requirements. It then sends the final control parameters that meet the deterministic constraints to the corresponding construction equipment and updates the continuous material state and subsequent counterfactual predictions based on the actual construction state after the equipment execution.

[0012] Based on the above technical solution, the present invention can be further improved as follows.

[0013] Furthermore, the construction state perception and continuous material field construction module establishes the longitudinal coordinates of the road based on the paver's forward direction, the lateral coordinates based on the road width direction, and combines them with the construction time to form a construction spatiotemporal coordinate system. This system performs a unified spatiotemporal mapping on at least one of the following: mixture type, gradation, asphalt-aggregate ratio, factory temperature, arrival temperature, paving thickness, paving speed, roller position, travel speed, vibration mode, vibration frequency, amplitude, number of compaction passes, vibration response, infrared temperature, ambient temperature, wind speed, underlying layer temperature, density detection results, and core sampling results. The state credibility field is used to characterize the completeness of observations and the reliability of state estimation at different spatial locations, and it completes the data by combining observations at adjacent times, neighboring spatial states, and material evolution states when data is missing.

[0014] Furthermore, the skeleton state field is used to characterize the evolution of asphalt mixture particles from loose rearrangement and contact chaining to interlocking stability and skeleton locking. The skeleton state is established based on the roller vibration response, mixture temperature, cumulative compaction effect, field quality test data, and indoor test or core samples. The compaction uniformity field is constructed based on the dispersion, spatial gradient, and compliance probability of the compaction state in the local neighborhood. It is used to identify construction areas where the average compaction state meets the requirements but the local spatial differences exceed the allowable range.

[0015] Furthermore, the process semantic understanding and candidate strategy generation module includes a state semantic compression unit, a construction knowledge retrieval unit, a candidate strategy generation unit, and a strategy verification unit. The state semantic compression unit is used to extract the location, range, trend of change, severity, remaining effective compaction time, and status of nearby construction equipment from the continuous material state of abnormal areas. The construction knowledge retrieval unit is used to obtain process knowledge matching the current construction state from construction specifications, construction organization design, mixture process requirements, equipment capacity descriptions, project-specific plans, historical construction records, and abnormal handling cases. The candidate strategy generation unit is used to generate strong vibration, weak vibration, rubber-tired compaction, static compaction, slowing down paving, adjusting equipment following distance, changing the operation sequence of the machine group, local compaction, and combinations of two or more of these actions. The strategy verification unit is used to eliminate candidate strategies that are unavailable, have ineffective action time, do not match the action object with the action area, or have process conflicts.

[0016] Furthermore, the counterfactual construction scenario construction module includes a fact state freezing unit, a candidate strategy intervention analysis unit, and a multi-branch scenario generation unit. The fact state freezing unit is used to fix the current material state, environmental conditions, mixture properties, equipment location, and geometric conditions of the construction area. The candidate strategy intervention analysis unit is used to convert candidate strategies into equipment trajectories, action widths, action areas, operation modes, action durations, and sequential, following, waiting, and switching relationships between multiple devices that vary with space and time. The multi-branch scenario generation unit is used to establish a baseline branch that maintains the current construction plan and an intervention branch that adopts different candidate strategies, starting from a common fact state, and to maintain consistency of non-intervention conditions between different branches.

[0017] Furthermore, the neural operator counterfactual world model module includes a continuous field input enhancement layer, a spatiotemporal coding layer, a global neural operator layer, a local compaction coupling layer, an equipment control injection layer, and a multi-field output layer. The global neural operator layer is used to establish nonlocal state transfer relationships between different spatial locations. The local compaction coupling layer is used to characterize the nonlinear coupling relationships between material temperature, skeleton rearrangement, compaction growth, and uniformity changes within the roller's operating area. The equipment control injection layer is used to inject construction interventions corresponding to candidate strategies into the continuous material state based on the future trajectory, operation mode, operating time, and equipment sequence of the roller or paver. The multi-field output layer is used to output the temperature field, skeleton state field, compaction state field, and compaction uniformity field at multiple future time points.

[0018] Furthermore, the neural operator counterfactual world model is constructed using Fourier neural operators, graph neural operators, wavelet neural operators, low-rank integral operators, or combinations thereof, and is trained using field construction data, quality inspection data, indoor test data, and numerical simulation data. During the training process, continuous material field prediction constraints, temperature evolution constraints, compaction evolution constraints, construction boundary constraints, and prediction uncertainty constraints are set to ensure that the model prediction results satisfy the engineering laws corresponding to temperature transmission, material compaction response, equipment action area, and construction area boundary, and output at least one of the prediction confidence interval, prediction variance, or state achievement probability.

[0019] Furthermore, the multi-future evaluation and strategy selection module extracts at least one evaluation index from various counterfactual construction scenarios, including the compaction compliance rate, under-compaction area, over-compaction area, compaction uniformity, temperature segregation degree, skeleton stability, equipment waiting time, repetitive work volume, equipment idle distance, construction cycle time, equipment energy consumption, safety risk, and prediction uncertainty. It also sets corresponding weights for different evaluation indicators according to the construction stage. When the construction quality evaluation results of multiple candidate strategies meet the preset requirements, the candidate strategy with fewer equipment action switching times, less disturbance to the original construction organization, more continuous equipment operation, and lower prediction uncertainty is selected first.

[0020] Furthermore, the deterministic constraint and execution feedback module establishes a safe and feasible domain consisting of equipment capability constraints, material state constraints, machine group safety constraints, construction area constraints, and control continuity constraints. Equipment capability constraints limit equipment speed, acceleration, vibration frequency, amplitude, steering, and mode switching conditions. Material state constraints limit compaction actions based on current and predicted temperature, frame, and compaction conditions. Machine group safety constraints limit spatial distance and trajectory conflicts between different construction equipment. Construction area constraints restrict equipment from entering prohibited operating areas. Control continuity constraints limit the magnitude of changes in speed, vibration parameters, and motion state between adjacent control cycles. When some control parameters of the preferred strategy are outside the safe and feasible domain, the corresponding control parameters are corrected to be within the safe and feasible domain. When the preferred strategy cannot obtain feasible control parameters, it switches to the suboptimal candidate strategy.

[0021] Furthermore, the deterministic constraint and execution feedback module adopts a rolling control method, issuing control actions corresponding to the current control cycle only to the construction equipment. After the current control cycle is executed, it reacquires the actual equipment trajectory, material temperature, vibration response, and compaction state, compares the actual construction state with the prediction results of the neural operator counterfactual world model, updates the current continuous material state, equipment response parameters, or model calibration parameters based on the comparison results, and then triggers the generation of candidate strategies, construction of counterfactual scenarios, prediction of future material fields, strategy evaluation, and deterministic control solution for the next control cycle.

[0022] Compared with the prior art, the technical solution of this application has the following beneficial technical effects: 1. This invention constructs continuous material states such as temperature field, skeleton state field, compaction state field and compaction uniform field, realizing the transformation of asphalt pavement construction state from discrete measurement point monitoring to continuous spatiotemporal characterization, and improving the ability to identify low temperature zone, under-compaction zone, over-compaction zone and local non-uniform area; 2. This invention generates multiple candidate strategies, such as strong vibration, weak vibration, rubber tires, static pressure, slowing down paving, and adjusting the sequence of machine groups, by combining a large model with construction specifications, equipment capabilities, and historical cases. It also limits the large model to only output the action type and the allowable range of parameters, avoiding the direct generation of the final control value of the equipment by the probabilistic language model, thereby improving the safety and controllability of intelligent construction decision-making. 3. This invention constructs multiple counterfactual construction branches under the same current construction state, enabling different candidate strategies to be virtually simulated before actual execution. This achieves a shift from "taking actions directly based on current anomalies" to "comparing the future results of different actions before making a decision," thereby improving the foresight and rationality of construction strategy selection. 4. This invention learns the function space mapping between the construction control action function and the future continuous material field through the neural operator world model. It can simultaneously predict the temperature, skeleton evolution, compaction state and uniformity changes at multiple future moments, and can adapt to different construction area lengths, spatial resolutions and equipment trajectory conditions, thereby improving the prediction generalization ability in complex construction scenarios. 5. This invention achieves a synergistic balance between construction quality, efficiency, and risk by comprehensively evaluating various counterfactual future compaction compliance rates, uniformity, temperature segregation, skeleton stability, construction rhythm, equipment energy consumption, and prediction uncertainties, thus avoiding over-compaction, machine group conflicts, or construction rhythm imbalance caused by only pursuing local compaction improvement. 6. This invention uses a deterministic constraint layer to perform hard constraint verification on equipment capacity, effective compaction temperature window, safety distance, construction boundary, and control quantity change rate. Combined with rolling execution and actual feedback to continuously correct the world model, it achieves closed-loop control of "prediction-selection-constraint-execution-feedback", thereby improving the stability, executability, and engineering safety of the coordinated control of asphalt pavement construction machinery groups. Attached Figure Description

[0023] Figure 1 A flowchart illustrating the overall system architecture provided in this embodiment of the invention; Figure 2 A flowchart for construction status perception and continuous material field construction provided in this embodiment of the invention; Figure 3 A flowchart of process semantic understanding and candidate strategy generation provided for embodiments of the present invention; Figure 4 This is a flowchart of the counterfactual construction scenario construction and multi-branch generation provided in an embodiment of the present invention; Figure 5 A flowchart of deterministic constraint solving and execution feedback closed-loop control provided for embodiments of the present invention; Figure 6 This is a schematic diagram of the on-site deployment structure of the intelligent agent for asphalt pavement construction provided in an embodiment of the present invention; Figure 7 This is a schematic diagram illustrating the unified representation of a continuous material field provided in an embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the transformation from candidate strategies to spatiotemporal control boundaries provided in an embodiment of the present invention. Figure 9 A schematic diagram of the real-time construction status monitoring and construction fleet operation status display interface of the intelligent agent for asphalt pavement construction provided in an embodiment of the present invention; Figure 10 A schematic diagram of the interface for large-scale model process semantic understanding, candidate strategy generation, and structured strategy output provided in this embodiment of the invention; Figure 11This is a comparison diagram of the future temperature field evolution under different candidate strategies provided in the embodiments of the present invention; Figure 12 This is a comparison diagram of the evolution of the skeleton state field and the compaction state field under different candidate strategies provided in the embodiments of the present invention; Figure 13 This is a statistical chart showing the compaction uniformity evaluation results under different candidate strategies provided in the embodiments of the present invention. Figure 14 This is a graph showing the combined ranking results of multiple future evaluation index weights and strategies provided in an embodiment of the present invention. Figure 15 A statistical diagram showing the allocation of computing power for training and online inference of the neural operator counterfactual world model provided in this embodiment of the invention. Detailed Implementation

[0024] To facilitate understanding of this application, a more complete description will be provided below with reference to the accompanying drawings, which illustrate embodiments of the present application. However, the present application can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that the disclosure of this application will be thorough and complete.

[0025] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the application.

[0026] Example 1: refer to Figures 1-15 A neural operator counterfactual world model-based intelligent agent system for asphalt pavement construction includes: a construction state perception and continuous material field construction module, a process semantic understanding and candidate strategy generation module, a counterfactual construction scenario construction module, a neural operator counterfactual world model module, a multi-future evaluation and strategy selection module, and a deterministic constraint and execution feedback module. Specifically: 1. Construction Status Perception and Continuous Material Field Construction Module

[0027] The construction status perception and continuous material field construction module is used to convert discrete data scattered across materials, equipment, environment, and quality inspection stages during asphalt pavement paving, initial compaction, secondary compaction, and final compaction into a continuous spatiotemporal state that can be directly invoked by subsequent counterfactual scenario generation and neural operator world models. The system establishes the longitudinal coordinates of the road based on the paver's forward direction, the lateral coordinates based on the road width, and the time axis based on construction time. The construction area is divided into a rolling computational domain that continuously moves forward with the paving operation, and a unified location identifier and timestamp are assigned to any spatial location within the computational domain. Collected data includes mixture type, gradation, asphalt-aggregate ratio, factory temperature, arrival temperature, paving thickness, paving speed, roller position, travel speed, vibration mode, vibration frequency, amplitude, number of compaction passes, vibration response, infrared temperature, ambient temperature, wind speed, underlying layer temperature, coreless density detection results, and core sampling results. Data generated by different devices and sensors undergoes time synchronization, coordinate transformation, outlier identification, and validity filtering before being mapped to the same construction spatiotemporal grid.

[0028] The current continuous construction state is represented as a multi-field coupled state: ;in, For spatial coordinates, The temperature field reflects the surface temperature, internal equivalent temperature, temperature gradient, and remaining compaction time. It is a skeletal state field that reflects aggregate orientation, particle contact density, interlocking degree, and potential for further rearrangement; This is the compaction state field, reflecting the degree of compaction, compaction increment, and cumulative compaction work; It is a uniform compaction field used to characterize the dispersion of compaction state, spatial gradient, and probability of achieving the standard within a local neighborhood; For the equipment application field, record the coverage trajectory, application mode, application sequence, and cumulative number of passes of different compaction equipment; This is a state confidence field used to describe the completeness of data and the reliability of estimates at different locations.

[0029] During temperature field construction, infrared thermography results, point thermography results, initial mixture temperature, pavement thickness, ambient temperature, wind speed, and underlying layer temperature are jointly input into the thermal state reconstruction unit. This continuously fills in areas obscured by equipment, where temperature measurements are missing, or where sampling is sparse, ensuring that any location within the computational domain has a temperature state corresponding to the construction time. The skeleton state field is not directly equivalent to compaction degree. Instead, it estimates the mesoscopic state of particle evolution from loose rearrangement to contact chaining, interlocking stability, and skeleton locking by utilizing roller vibration acceleration, vibration response changes, mixture temperature, cumulative compaction work, and calibration relationships formed from indoor tests or core samples. The compaction state field is dynamically updated based on the number of passes, vibration parameters, compaction response, and field density detection results. The compaction uniformity field identifies areas where the average compaction degree meets the standard but local differences remain significant through neighborhood statistics and spatial gradient calculations.

[0030] After unified fusion of multi-source data, the continuous field input required by the neural operator is formed: ;in, For material data, For construction equipment and trajectory data, For quality inspection data, For environmental data, This involves a continuous field fusion mapping that includes spatiotemporal alignment, state reconstruction, data assimilation, and confidence assignment. For areas with missing data, the system uses observations from adjacent time points, neighboring spatial states, and material evolution patterns to perform interpolation or prediction completion, while simultaneously reducing the confidence of the corresponding location. When the confidence is below a set threshold, the area is assigned a higher uncertainty weight in subsequent counterfactual inferences to prevent incomplete data from being equated with real observations. Through the above processing, the discrete sensor data in the field is transformed into the state of construction materials with unified coordinates, continuous distribution, dynamic updates, and confidence descriptions, providing unified initial conditions for subsequent candidate strategy interventions, generation of multiple counterfactual futures, and prediction of future temperature fields, skeleton state fields, and compaction uniform fields.

[0031] 2. Process Semantic Understanding and Candidate Strategy Generation Module

[0032] The process semantic understanding and candidate strategy generation module converts temperature anomalies, skeleton evolution, insufficient compaction, decreased uniformity, and fleet operation status in a continuous material field into process semantics that can be understood and reasoned about by the construction agent. Based on the current construction objectives, it generates candidate disposal strategies with significant differences. This module does not handle the final control parameter solution for the equipment; instead, it performs high-level decision-making on "what is happening now—what might be causing it—what process actions can be taken," establishing clear functional boundaries between the large model and the subsequent neural operator world model and deterministic constraint layer. This prevents probabilistic language generation results from directly entering the control interface of the roller or paver.

[0033] The continuous material field first undergoes state semantic compression, transforming large-scale grid data into structured features such as anomalous regions, anomaly types, trends, severity, duration, and reliability. For example, a region might be described as having "increased local temperature decay rate behind the paver, still high skeleton rearrangement potential, lower compaction state than adjacent areas, and a continuously shortening effective compaction window," while simultaneously associating the location, working direction, current task, and estimated arrival time of nearby dual-drum and rubber-tired rollers. State semantic compression preserves spatial extent, temporal trends, and multi-field coupling relationships, rather than directly inputting all grid data into the large model, thereby reducing interference from invalid information on process reasoning.

[0034] Before generating candidate strategies for the large model, knowledge fragments matching the current state are retrieved from construction specifications, construction organization design, mixture process requirements, roller equipment capacity specifications, project-specific plans, historical high-quality construction records, and anomaly handling cases. These fragments are then combined with the current construction stage to form a local process context. When an abnormal area still has high potential for skeleton rearrangement, candidate actions may include changing the vibration mode, adjusting the entry sequence of the re-compaction equipment, or changing the equipment coverage path. When the effective compaction time is rapidly shortened, strategies such as slowing down paving, shortening the spacing between pavers, and advancing the operation of rubber-tired rollers can be generated. When the skeleton is close to stable but local uniformity is insufficient, weak vibration, static compaction, or local supplementary compaction strategies are prioritized. Candidate actions are jointly determined by the material state and construction stage, rather than being mechanically triggered based on a single temperature threshold.

[0035] Each candidate strategy uses a unified structure encoding: ;in, For the first One candidate strategy, This indicates the type of action, such as strong vibration, weak vibration, rubber tires, static pressure, paving adjustment, track adjustment, or machine group sequence adjustment. Indicates the equipment or group of equipment involved in the execution. Indicates the area of ​​action. Indicates the allowed execution time range. This indicates the range of parameters that can be used in subsequent deterministic control solutions. This indicates the preconditions for the strategy to be valid. This indicates the conditions that prohibit execution. Through this structure, construction suggestions in natural language form are converted into computable, filterable policy objects that can be passed on to the counterfactual scenario building module.

[0036] The candidate strategy generation adopts a "differentiation-first" principle, avoiding the generation of multiple schemes with only slight numerical differences but the same process mechanism. Instead, it prioritizes covering different mechanisms of action. For example, in the same underburden area, multiple strategy branches can be generated, such as "enhancing vibration input," "changing the timing of rubber-tired wheel intervention," "reducing paving speed," "adjusting the sequence of double steel wheels and rubber-tired wheels," and "maintaining the current scheme." This allows the subsequent counterfactual world model to truly compare the future results produced by different construction mechanisms. When necessary, two or more compatible actions can be combined to form composite candidate strategies such as "slowing down paving + weak vibration compensation" and "shortening the spacing between pavers + advancing the rubber-tired wheels."

[0037] After strategy generation, the rule verification unit checks whether the equipment is available, whether the target area still meets the operational conditions, whether there are logical conflicts between strategies, and whether explicit construction prohibitions are violated. Strategies that cannot find execution equipment, whose action time has expired, whose action object does not match the area, or have obvious process conflicts are directly eliminated. For strategies that can be executed but whose specific parameters are not yet determined, only the action category, the allowed range of parameters, and the condition boundaries are retained; the final values ​​of roller speed, vibration frequency, amplitude, etc., are not specified in advance. This ultimately forms a candidate set containing baseline maintenance strategies and several intervention strategies, providing discrete decision-making entry points with process differences for constructing multiple counterfactual construction futures under the same initial conditions in the next stage.

[0038] 3. Counterfactual Construction Scenario Construction Module

[0039] The counterfactual construction scenario construction module transforms candidate strategies into spatiotemporal intervention conditions that can be computed by the neural operator world model, and constructs multiple independent future evolution branches based on a completely consistent current construction state. Instead of simply replicating several construction schemes for conventional simulation, this module first freezes the already occurred construction facts and then only changes the future actions to be executed. This allows the differences in state between different futures to be attributed to specific construction strategies, providing a unified causal starting point for subsequent comparisons of "what would happen if strong vibration were used," "what would happen if rubber-tired paving were used earlier," and "what would happen if paving were slowed down."

[0040] At this moment of decision-making The system extracts factual states that determine subsequent material evolution, including the current continuous material state, future short-term environmental boundaries, inherent properties of the mixture, and the instantaneous operating state of all construction equipment, and forms an initial state set for counterfactual deduction: ;in, Indicates the current material state. This indicates the environmental and heat transfer boundary conditions. This indicates material properties such as mixture type, layer thickness, gradation, and viscosity-temperature characteristics. This indicates the location, orientation, movement status, and current operating mode of each construction device. This represents the geometric state of the construction area, joints, edges, and restricted areas. After entering a counterfactual scenario, these established factual conditions remain consistent and are not randomly generated due to different candidate strategies, thus avoiding the masking of the effects caused by the construction actions themselves due to differences in initial temperature, equipment location, or material batches.

[0041] Before candidate strategies enter the counterfactual scenario, they need to be converted into continuous spatiotemporal control boundaries by the action parser. Strong vibrations, weak vibrations, rubber tires, static pressure, slowing down paving, and changing the machine group sequence cannot simply be used as discrete category inputs to the model; they must be further expanded into corresponding equipment trajectories, action areas, durations, process modes, permissible control intervals, and temporal relationships between equipment. The intervention effect of each candidate strategy can be expressed as: ;in, Indicates the first A strategy in position and future time The control field formed below; Indicates the future operating trajectory of the equipment; This indicates how operating modes such as vibration, static pressure, rubber tires, or paving change over time. Indicates the scope of the strategy; This indicates the sequential, following, waiting, and switching relationships between multiple devices; It is responsible for unfolding structured process strategies into spatiotemporal boundary inputs that can be recognized by neural operators. Through this transformation, the same "weak vibration compensation" strategy can form different control effects in different regions, different equipment trajectories, and different entry times, instead of being compressed into an action label lacking spatial information.

[0042] After the control boundary is constructed, the system uses Establish a baseline future and multiple intervention futures from a common starting point. The baseline future continues to evolve with the current construction organization and equipment actions, answering the question, "What will happen if the current strategy is not changed?" The other branches respectively implement strong vibration, weak vibration, advance tire movement, static pressure, slow down paving, change the sequence of tractors, or a combination of actions. A counterfactual scenario is represented as follows: ;in, This refers to the human intervention in future construction activities, that is, replacing the original planned control boundary with the new one without changing the current state of affairs. The control field corresponds to each candidate strategy. This forms a counterfactual construction tree that branches from the same moment to multiple possible futures.

[0043] To ensure comparability across different branches, all external conditions except the target intervention variable are treated with consistent boundaries in a single comparison. For example, when comparing strong and weak vibrations, both future scenarios use the same initial temperature distribution, the same batch of mixture, the same initial compaction degree, the same equipment starting position, and the same short-term environmental predictions, only changing the vibration action mode and the subsequent trajectory and material response naturally caused by this change. When comparing "maintaining the original machine group sequence" and "advanced rubber-tired tires," the construction area, total task, and equipment set remain unchanged, only changing the order and temporal intervals of equipment entering the target area. By limiting changes in non-target variables, differences in future temperature, framework, and compaction state can be primarily attributed to the candidate actions themselves.

[0044] For anomalies requiring multi-device coordination to resolve, the system allows the construction of combined counterfactual branches. For example, when a local area simultaneously experiences both a shortened temperature window and insufficient compaction, combined actions such as "slowing down paving + shortening the following distance of the double steel drums," "weak vibration compaction + advancing the rubber-tired wheels," and "local static compaction + maintaining the original process in adjacent areas" can be set. The combined strategy first checks whether there are mutually exclusive relationships between actions, and then generates a directed time sequence diagram according to the action dependencies, so that the processes of equipment entry, exit, waiting, and handover can all be transformed into continuous control boundaries, avoiding the neglect of the synergistic effects between paving, initial compaction, intermediate compaction, and final compaction by only comparing isolated actions of a single device.

[0045] Considering that factors such as changes in on-site wind speed, material supply intervals, and short-term equipment speed reduction cannot be fully determined, this module can also add a limited number of disturbance sub-scenarios under the same candidate strategy to test the strategy's sensitivity to external changes. However, the disturbance sub-scenarios and the main strategy branches are managed hierarchically. First, the differences caused by different construction actions are compared, and then the stability of the same action under slight environmental disturbances is analyzed to prevent environmental randomness from undermining the core counterfactual comparison.

[0046] When there are many candidate strategies, the system performs pre-screening based on process differences, overlap of action areas, similarity of actions, and obvious infeasibility conditions. Strategies with only slight changes in parameter values ​​and essentially the same construction mechanism are grouped into the same strategy family, leaving parameter refinement for subsequent deterministic solutions; branches that significantly exceed the remaining compaction time, where equipment cannot reach the target area, or that would cause interruption of the established process are deleted in advance. The final result is a set of counterfactual scenarios with controlled scale, diverse process mechanisms, and a unified starting point, providing standardized input for the neural operator world model to parallel extrapolate the evolution of temperature field, skeleton state field, compaction state field, and uniformity over the next few minutes.

[0047] 4. Neural Operator Counterfactual World Model Module

[0048] The neural operator counterfactual world model module receives continuous material states, equipment spatiotemporal boundaries, construction area geometry, and environmental disturbance conditions corresponding to various counterfactual construction scenarios. It then predicts the continuous spatial evolution of asphalt mixture temperature, skeleton, compaction state, and uniformity at several future moments. Instead of outputting a discrete prediction value for a single measuring point, this module learns the function space mapping from the current continuous material field and construction intervention function to the future continuous material field. This allows the same world model to adapt to different construction lengths, road widths, grid densities, and equipment trajectory conditions, providing a unified physical evolution predictor for multiple counterfactual futures.

[0049] The overall mapping relationship of the neural operator world model is expressed as: ;in, For parameters The neural operator; The continuous material state at the moment of decision-making; This is the device spatiotemporal action function formed by candidate strategies within the prediction time range; External boundaries include ambient temperature, wind speed, solar radiation, and heat exchange of the underlying layer; Geometric information of the construction area; For counterfactual prediction time domain; Predicting future states. By changing... While keeping other initial conditions consistent, the same model can calculate the future evolution results corresponding to different candidate strategies.

[0050] The model input is configured with a continuous field lifting and spatiotemporal coding structure. Temperature state, skeleton state, compaction state, local uniformity, equipment operation history, and state reliability are first organized according to construction spatial coordinates, and spatiotemporal features such as the time since paving, the time since the most recent compaction, the longitudinal and transverse positions of the road, and the distance to regional boundaries are added. Equipment actions are not directly input as text labels such as "strong vibration" or "weak vibration," but are encoded as the roller's position trajectory, action width, movement direction, operation mode, and temporal relationship within the prediction time domain. This allows the model to identify the actual differences in the effects of the same process action at different locations, different entry times, and different equipment combinations. Construction edges, longitudinal and transverse joints, areas near bridge structures, and prohibited compaction areas participate in the model calculation through geometric masks to prevent different spatial boundaries from being treated as completely identical material areas.

[0051] Continuous field features are mapped to a high-dimensional latent space by the input lifting layer and then enter the global neural operator layer. The global layer is responsible for establishing non-local dependencies between different spatial locations, allowing heat diffusion in a given area, continuous operation of adjacent equipment, and the preceding compaction state to influence subsequent regional predictions. The computation process of the layer neural operator can be represented as follows: ;in, For position First Hidden state, The local transformation matrix, It is a learnable nonlocal integral kernel. This includes the equipment's operating method, material type, and time conditions. For the current construction calculation domain, This is a nonlinear mapping function. The integral kernel learns the state transfer relationships between different spatial locations, enabling the model to predict without relying on a fixed number of measurement points. In practice, Fourier neural operators, graph neural operators, wavelet neural operators, low-rank integral operators, or combinations of multiple operator structures can be used. For regular construction grids, frequency domain global operations are prioritized to improve computational efficiency, while irregular road widths, joints, and special structural areas can have their boundary adaptability enhanced through graph structures or coordinate condition operators.

[0052] A local compaction coupling layer is set after the global operator layer to focus on describing the highly nonlinear material evolution in and around the roller track. The compaction process of asphalt mixtures is not simply determined by the number of compaction passes: material temperature affects asphalt viscosity and particle mobility; external compaction work promotes aggregate rearrangement and void reduction; as the contact chain gradually stabilizes, the difficulty of further skeleton rearrangement increases. At this point, further increasing vibration may only yield a small compaction increment, or even lead to displacement, aggregate breakage, or local over-compaction risks. Therefore, the model uses a multi-field coupled hidden state to uniformly describe material evolution, and its dynamic relationship can be expressed as: ;in, Represents the thermal state. Represents the evolutionary state of the skeleton. This indicates the compacted state. It represents a uniform spatial state. This represents a continuous evolution operator learned from neural operators. This expression allows temperature decay, skeleton rearrangement, and compaction growth to no longer rely on independent prediction networks, but rather to transmit information within a shared latent space. For example, a rapid temperature drop reduces the subsequent skeleton rearrangement capability, skeleton locking reduces the compaction increment per unit compaction effort, and localized high compaction caused by repeated equipment coverage further affects uniformity evaluation.

[0053] To reflect the localized impact of roller actions, the model generates a dynamic action mask based on the future trajectory of the equipment. When the roller enters a certain spatial location, the current operating mode, action time, and historical accumulated compaction work are injected into the corresponding area. After the equipment leaves, this area no longer receives new mechanical action, but the material temperature and internal state continue to evolve over time. For adjustments to the roller group sequence, the model injects control boundaries sequentially based on the expected arrival times of different equipment in the target area. Therefore, it can distinguish the impact of different processes, such as "double steel drums acting first, then rubber-tired rollers," "rubber-tired rollers intervening early," and "vibration after static compaction," on the future material state.

[0054] The neural operator employs multi-timescale rolling output, not only predicting the next control cycle but also simultaneously calculating the spatial state at multiple key future moments. Based on the needs of asphalt construction control, multiple future moments such as 30 s, 60 s, 120 s, 180 s, and 300 s can be selected to form a continuous prediction sequence. Short-time results are used to determine the upcoming roller actions, medium-time results are used to determine whether the temperature window is about to disappear and whether the machine group scheduling will cause local waiting, and longer-time results are used to determine the continued impact of the current actions on the final uniformity and skeleton stability. Each future moment outputs temperature distribution, low-temperature region range, skeleton stability, potential for further rearrangement, compaction state, compaction increment, uniformity, and corresponding prediction reliability, thus enabling subsequent strategy selection to compare the complete future trajectory, rather than just comparing a single termination moment.

[0055] The world model training utilizes a combination of actual on-site construction data, quality inspection data, laboratory test data, and necessary numerical simulation data to construct a sample. On-site data is used to learn the statistical dynamics relationship between real equipment actions and material responses; laboratory tests supplement material responses under different temperatures, compaction work, and skeleton stages; and rarely encountered on-site conditions such as extreme low temperatures, shutdowns, and temporary equipment withdrawals can be expanded using numerical calculations or controlled experiments to broaden the training scope. During training, continuous field prediction errors and engineering physics constraints are simultaneously incorporated, with the comprehensive loss defined as: ;in, Used to limit the discrepancy between future material fields and actual observations; Used to constrain temperature changes over time to satisfy basic heat conduction, convection, and heat dissipation laws; Used to constrain the relationship between compaction work, skeleton evolution, and compaction increment within the effective temperature range; Used to limit unfounded mechanical compaction responses in areas not covered by equipment and to maintain consistent spatial conditions at road edges and joints; Used to calibrate the prediction confidence of the model output; each This is the loss weight.

[0056] The model calculates confidence intervals, prediction variances, or achievement probabilities along with the output prediction mean to distinguish between "poor prediction performance" and "model indeterminacy." When candidate actions fall within a working area with sufficient historical data coverage, the model can provide a highly reliable future material field; when encountering unprecedented mixture types, extreme environmental conditions, or rare machine cluster combinations, the prediction uncertainty increases accordingly. This information is further passed to the multi-future evaluation module, penalizing strategies with slightly better prediction quality but significantly higher uncertainty.

[0057] Through the above structure, the neural operator counterfactual world model establishes a unified mapping from the construction intervention function to the continuous future evolution of materials. It no longer relies on direct response rules such as "increase vibration intensity if current pressure is low," but can instead answer how the material state will change in the coming minutes after different actions are performed. Multiple counterfactual branches can be input into the same neural operator in batches and in parallel, obtaining future temperature, skeleton, and compaction evolution results corresponding to different construction strategies while maintaining consistency between model parameters and initial facts. This provides a predictive basis for subsequently selecting the optimal construction action from multiple possible futures.

[0058] 5. Multi-future evaluation and strategy selection module

[0059] The multi-future evaluation and strategy selection module is used to uniformly evaluate the future material field predicted by different counterfactual branches, and to select candidate strategies with higher overall benefits, lower construction risks, and better prediction reliability from multiple possible futures. The evaluation object is not the candidate action itself, but the future temperature distribution, skeleton evolution state, compaction state, and spatial uniformity formed after the candidate action is implemented. For example, "enhancing vibration" may produce a large compaction increment in high-temperature areas where the skeleton is not yet stable, while it may increase the risk of displacement, overcompression, or local inhomogeneity in areas where the skeleton is close to locking or the temperature is rapidly decreasing. Therefore, the merits cannot be determined in advance based on the action name, but need to be compared with the future results predicted by the world model.

[0060] The system first extracts four categories of indicators from each counterfactual future: quality, construction efficiency, equipment operation, and risk. Quality evaluation focuses on statistically analyzing the percentage of area achieving the compaction target, the area of ​​under-compacted areas, the area of ​​over-compacted areas, compaction uniformity, temperature segregation, skeleton locking stability, and the condition of joints and edge areas within the predicted time domain. For the same candidate strategy, it evaluates not only the predicted endpoint state but also analyzes the evolution process throughout the entire predicted time domain, avoiding situations where a scheme achieves the target compaction degree but exhibits significant low temperatures, over-compaction, or skeleton anomalies in the intermediate stages. For areas where the temperature window is about to close, it also calculates the remaining space and time for effective compaction, ensuring that strategy evaluation reflects whether construction actions truly keep pace with changes in material state.

[0061] Construction efficiency evaluation measures the impact of action adjustments on overall construction continuity, including paving progress, roller waiting time, local repetitive work, equipment idle distance, machine group connection time, and cycle time changes caused by handling abnormal areas. Equipment operation evaluation further considers the vibration duration, movement distance, mode switching frequency, start-stop frequency, and energy consumption changes resulting from different candidate strategies, preventing the system from frequently altering the entire machine group's operating state to improve only a small local area. For machine group sequence adjustment strategies, it also evaluates whether the new sequence will transfer the original abnormality from one area to another. For example, while advancing rubber-tired rollers may improve the current low-temperature area, it may cause a delay in the entry time of double-drum rollers in other areas. Therefore, the evaluation scope covers the entire effective construction domain, rather than just the abnormal location that triggered the decision.

[0062] No. The combined cost of a counterfactual future is expressed as: ;in, As a trade-off for quality, it comprehensively reflects undervoltage, overvoltage, uniformity, temperature segregation, and skeleton stability. This is due to the impact on construction pace and schedule; The cost of equipment operation and energy consumption; For safety and abnormal risks; , , , The evaluation weights are set according to the construction stage. In the initial compaction stage, the weights of indicators related to temperature window and skeleton rearrangement can be increased; in the intermediate compaction stage, the weights of compaction increment and uniformity can be increased; and in the final compaction stage, the weights of indicators related to surface stability, overpressure risk, and equipment disturbance can be increased, so that the same evaluation framework can be adapted to different construction stages.

[0063] Beyond predictive performance, strategy selection introduces inherent uncertainty into the model itself. While world models achieve high reliability with action combinations adequately covered by historical data, they cannot be equivalent to high-reliability models when facing extreme environments, rare equipment combinations, or material conditions exceeding the training distribution. The system generates uncertainty indicators based on prediction variance, confidence interval width, state reliability, and deviation from the training distribution over multiple future time points, and applies risk corrections to the overall cost. ;in, This is the risk-adjusted strategy evaluation value. For the first The overall uncertainty of a counterfactual future prediction This represents the uncertainty penalty coefficient. When predicted benefits are similar, the option with lower uncertainty is prioritized; if an option predicts the best compaction effect but its credibility is significantly insufficient, its ranking will be automatically reduced, thereby reducing the decision-making risk caused by the world model being overly optimistic in unfamiliar working conditions.

[0064] After calculating the indicators, the system first performs a hard elimination process, deleting branches that show obvious failure in future temperature windows, exceed the expected overpressure risk limit, have serious conflicts in the operation of the equipment fleet, or fail to meet the minimum construction quality requirements. Then, it uses weighted ranking or Pareto comparison to evaluate the remaining solutions. When the quality improvement effects of multiple strategies are similar, the system prioritizes retaining the strategy with fewer action changes, less disturbance to the original construction organization, more continuous equipment operation, and higher prediction reliability. Simultaneously, it retains the second-best candidate strategies and their corresponding evaluation results, allowing the subsequent deterministic constraint layer to quickly switch when the preferred strategy cannot form a feasible control solution. Through this process, the construction agent transforms from "generating multiple possible actions" to "comparing multiple future results," providing a strategy priority validated by the future material field for the final solution of equipment control parameters.

[0065] 6. Deterministic Constraints and Execution Feedback Module

[0066] The deterministic constraint and execution feedback module converts the optimal strategy obtained from multi-future evaluation into control commands that can be actually executed by pavers, twin-drum rollers, rubber-tired rollers, and other construction equipment. It also verifies process boundaries, mechanical capacity, safety distances, construction areas, and control continuity before issuing equipment actions. This module sits between the intelligent decision-making results and the equipment control interface, forming the final safety barrier for the construction agent. The large model can only determine candidate actions and their allowable parameter ranges, and the neural operator can only predict the future material fields that different actions may generate. Neither can bypass this module to directly change the equipment state. Ultimately, the speed, vibration mode, amplitude, frequency, operating position, entry time, and duration are all generated by the control solution process under deterministic constraints.

[0067] After the selected strategy is entered into the control solver, a set of controllable variables is first established based on the strategy's area of ​​effect and the equipment. For a double-drum roller, controllable variables include travel speed, vibration mode, allowable vibration frequency, amplitude level, travel direction, start and end positions, and reversal time; for a rubber-tired roller, these may include travel speed, area of ​​effect, tire working status, entry order, and number of repeated coverages; for a paver, these may include variations in advance speed within the allowable range, short waiting time, and the target distance to the compaction equipment. The strategy generation stage only provides feasible intervals for these parameters; the control solver further determines a single executable solution by combining real-time construction conditions.

[0068] Equipment capacity constraints first limit the control variables to within the limits of the equipment's physical performance and the controller's allowable range, including minimum and maximum travel speeds, allowable acceleration and deceleration, vibration frequency and amplitude range, minimum turning radius, reversal time, and mode switching conditions. Even if counterfactual predictions show that further increasing mechanical action can improve local compaction, the corresponding control solution is deemed infeasible as long as the target parameters exceed the equipment's rated capacity.

[0069] Material condition constraints limit the applicability of different compaction methods based on the current and short-term predicted material conditions. When the temperature in a region has already left the effective vibration compaction window, the material skeleton has stabilized, and further input of compaction work yields minimal benefit, or the predicted probability of overcompression exceeds the allowable level, it is prohibited to continue increasing high-intensity mechanical action. Only when the region still exhibits significant undercompression and has sufficient rearrangement potential is it permissible to search for further compaction parameters within the equipment's capabilities. This avoids continuously increasing vibration action simply based on insufficient current compaction.

[0070] The safety constraints for the machine group calculate the minimum spatial distance and timing conflicts between the equipment based on their current position, direction of movement, and predicted trajectory. When two machines are expected to intersect, rear-end, approach each other from the same work zone, or enter a narrow construction area simultaneously, the control solver automatically adjusts the entry time, speed, or work sequence. Roadsides, bridge structures, joint protection zones, personnel work areas, and temporarily prohibited construction areas are all considered as insurmountable spatial boundaries in the control solution, ensuring that the optimized equipment trajectory always remains within the permissible work domain.

[0071] Control continuity constraints are used to limit abrupt changes in velocity, vibration parameters, and motion direction between adjacent control cycles. The system does not allow frequent switching between strong vibration, weak vibration, and static pressure modes due to slight changes in the world model's predictions with each roll, nor does it allow repeated large increases and decreases in equipment speed within a short period. For multiple parameter solutions that can achieve similar future quality effects, the solution with the smallest difference from the equipment state in the previous cycle is preferred to reduce mechanical shock, operating frequency, and fluctuations in cluster operation.

[0072] Considering the above constraints, the final control quantity is determined through constraint optimization: ;in, For the final device control vector, To optimize the search range of parameters allowed by the preferred strategy, Taking into account the predicted quality deviation, changes in motion, equipment waiting time, and control costs, This involves considering equipment capacity, temperature conditions, safety distances, construction boundaries, control rates of change, and other hard constraints that must be met. The solution process further transforms high-rise process strategies into defined equipment control values, clearly separating "selecting weak vibrations" from "specific speeds, application locations, and durations."

[0073] When the overall optimal strategy is reasonable, but some candidate parameters fall into the infeasible region, the strategy is not discarded directly. Instead, a safe control solution that is closest to the original objective is searched within the control domain that satisfies all hard constraints. The safe projection relationship is expressed as: ;in, This is a reference control quantity formed based on the strategic objectives. The safe and feasible region is formed by all deterministic constraints. This represents the actual control value that can be issued. If no valid solution exists in the safe and feasible domain, the current preferred strategy is abandoned and the suboptimal solution reserved by the previous module is invoked; if none of the candidate strategies can form a feasible solution, the current safe operating conditions are maintained, the workload is reduced, equipment is restricted from entering abnormal areas, or manual takeover is triggered.

[0074] After the final control command is issued through the edge controller, vehicle terminal, or equipment control interface, the system executes only a small segment of action corresponding to the current rolling cycle, rather than executing the entire predicted time domain at once. During execution, the system continuously collects data on the actual equipment trajectory, temperature changes, vibration response, and mass status, and compares the actual results with the future field previously predicted by the world model. When a persistent deviation occurs between the prediction and the actual situation, the system corrects the current material state, equipment response parameters, and model calibration coefficients, and then regenerates the next control cycle with the new field conditions. This forms a closed loop of "strategy selection—deterministic solution—short-term execution—re-sensing—prediction correction—re-decision," enabling the construction agent to continuously adjust under conditions of environmental changes, material supply fluctuations, equipment delays, and material batch changes, rather than executing a fixed control scheme generated only once for a long period of time.

[0075] The working principle of this application is as follows: This invention achieves intelligent decision-making and collaborative control of the asphalt pavement paving and compaction process based on a closed-loop mechanism of "continuous characterization of construction state - large model generation of candidate strategies - counterfactual multi-future construction - neural operator continuous field prediction - multi-future evaluation and optimization - deterministic constraint execution - feedback correction". Its core lies in not directly generating the final control parameters of the equipment from the large model, but first predicting the future material states that may be formed by different construction actions, and then determining the actual execution actions based on the future results.

[0076] During construction, the system continuously acquires data on asphalt mixture properties, paver operating status, the operating trajectories of dual-drum and rubber-tired rollers, vibration modes, number of compaction passes, infrared temperature, vibration response, ambient temperature, wind speed, and quality inspection data. Through time synchronization, coordinate unification, anomaly data processing, and spatial reconstruction, discrete observations are mapped to the current construction area. This forms a continuous material state that rolls forward with the paving operation, mainly including a temperature field, skeleton state field, compaction state field, compaction uniformity field, equipment action field, and state reliability field. Among these, the skeleton state characterizes the evolution of aggregates from loose rearrangement, contact chaining, to interlocking stability and skeleton locking, enabling the agent to determine whether the material still has the potential for further compaction and rearrangement.

[0077] After establishing the continuous material state, semantic extraction is performed on the location, range, trend of change, remaining effective compaction time, and status of nearby machinery in abnormal areas. The large-scale model, combining construction specifications, construction organization design, equipment capacity descriptions, and historical construction cases, understands the current working conditions and generates several candidate strategies with different process mechanisms, including strong vibration, weak vibration, advance tire compaction, static compaction, slowing down paving, shortening equipment following distance, changing machinery sequence, and combinations of multiple actions. The large-scale model only determines the action type, acting equipment, acting area, allowable time, and allowable parameter range; it does not directly output specific speed, vibration frequency, amplitude, or other final control values ​​to the equipment.

[0078] Subsequently, the current construction facts are frozen, using identical temperature conditions, framework conditions, compaction conditions, initial equipment positions, material properties, and environmental conditions as a common starting point. Different candidate strategies are then transformed into spatiotemporal control boundaries composed of equipment trajectories, areas of action, operation modes, and time sequences, forming multiple counterfactual construction branches. For example, future scenarios such as "maintaining the original process," "strong vibration compaction," "weak vibration compaction," "advanced tire installation," "slowing down paving," and "adjusting the machine group sequence" are simultaneously constructed, ensuring that the main differences between the branches arise only from the construction intervention actions to be compared.

[0079] After each counterfactual branch enters the neural operator world model, the neural operator learns the function space mapping from the current continuous material field and control function to the future continuous material field, and predicts in parallel the changes in temperature, skeleton, compaction state, and uniformity at multiple future moments. The model simultaneously considers the effects of temperature decay on asphalt viscosity and particle rearrangement capacity, the effect of mechanical compaction work on skeleton evolution, and the decreasing benefits of continued compaction after skeleton locking, thus obtaining the continuous spatial evolution of the material state over the next few minutes after different strategies are implemented, rather than predicting only a single result at a certain measuring point.

[0080] The multi-future evaluation module further compares the compaction compliance rate, uniformity, temperature segregation, skeleton stability, under- and over-compaction risks, construction cycle time, equipment waiting time, energy consumption, and prediction uncertainty of each counterfactual branch. Strategies with good prediction quality but low model confidence are penalized, while the strategy with the best overall performance and several suboptimal strategies are retained.

[0081] The optimal strategy ultimately enters the deterministic constraint layer, where specific execution parameters are solved based on equipment speed and vibration capacity, effective compaction temperature window, equipment safety distance, construction area boundaries, control quantity change rate, and manual intervention conditions. Any action that violates hard constraints cannot be directly issued; for strategies with infeasible parameters, modifications are made within the safe feasible region, and if no feasible solution can be obtained, the system automatically switches to a suboptimal strategy or a conservative construction state.

[0082] During operation, the equipment performs only one segment of control action within the current rolling cycle. Subsequently, it re-collects actual temperature, trajectory, vibration response, and compaction results, compares the real material field with the prediction results of the world model, corrects the current state and model parameters, and then regenerates the next round of counterfactual future. Through a continuously rolling process of "perception-generation-deduction-selection-constraint-execution-correction," the asphalt pavement construction is transformed from experience-based response based on the current state to predictive intelligent control based on the future evolution of the material field.

[0083] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model, characterized in that, It includes: The construction status perception and continuous material field construction module is used to acquire material data, construction equipment data, quality inspection data and environmental data during the asphalt pavement construction process. It performs time synchronization, coordinate unification, anomaly handling and spatial reconstruction on the acquired data to form a continuous material state including temperature field, skeleton state field, compaction state field, compaction uniformity field, equipment action field and state credibility field. The process semantic understanding and candidate strategy generation module is used to identify the current construction status and abnormal areas based on continuous material status, construction equipment status, construction tasks and construction process knowledge, and to generate multiple structured candidate strategies using the large model. The candidate strategy includes at least action type, acting equipment, acting area, acting time, parameter allowable range, preconditions and prohibition conditions. The large model does not directly issue the final control parameters to the construction equipment. The counterfactual construction scenario construction module is used to freeze the current construction facts that have already occurred. It uses the same current continuous material state, equipment initial state, material properties and environmental conditions as common initial conditions, and converts different candidate strategies into corresponding equipment trajectories, equipment action areas, operation modes and time relationships to form multiple counterfactual construction scenarios. The neural operator counterfactual world model module is used to predict the temperature field, skeleton state field, compaction state field and compaction uniform field at multiple future moments after the application of different candidate strategies, based on the continuous material state, equipment spatiotemporal action conditions, environmental boundary conditions and construction area geometric conditions in each counterfactual construction scenario. The multi-future evaluation and strategy selection module is used to evaluate the construction quality, construction efficiency, equipment operating status, safety risks and prediction uncertainties of candidate strategies based on the future continuous material field of different counterfactual construction scenarios, determine the preferred strategy and retain at least one suboptimal candidate strategy. The deterministic constraint and execution feedback module is used to solve the final control parameters of the construction equipment based on the parameter allowable range of the preferred strategy, equipment capacity, material state, safety distance, construction area boundary and control change requirements. It then sends the final control parameters that meet the deterministic constraints to the corresponding construction equipment and updates the continuous material state and subsequent counterfactual predictions based on the actual construction state after the equipment execution.

2. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The construction state perception and continuous material field construction module establishes the longitudinal coordinates of the road based on the paver's forward direction and the lateral coordinates based on the road width direction. It also combines the construction time to form a construction spatiotemporal coordinate system. This system performs a unified spatiotemporal mapping on at least one of the following: mixture type, gradation, asphalt-aggregate ratio, factory temperature, arrival temperature, paving thickness, paving speed, roller position, travel speed, vibration mode, vibration frequency, amplitude, number of compaction passes, vibration response, infrared temperature, ambient temperature, wind speed, underlying layer temperature, density detection results, and core sampling results. The state credibility field is used to characterize the completeness of observations and the reliability of state estimation at different spatial locations. When data is missing, it combines observations from adjacent time points, neighboring spatial states, and material evolution states to complete the data.

3. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The skeleton state field is used to characterize the evolution of asphalt mixture particles from loose rearrangement, contact chaining to interlocking stability and skeleton locking. The skeleton state is established based on the roller vibration response, mixture temperature, cumulative compaction effect, field quality test data, and indoor test or core samples. The compaction uniformity field is constructed based on the dispersion of compaction state in the local neighborhood, spatial gradient, and compliance probability. It is used to identify construction areas where the average compaction state meets the requirements but local spatial differences exceed the allowable range.

4. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The process semantic understanding and candidate strategy generation module includes a state semantic compression unit, a construction knowledge retrieval unit, a candidate strategy generation unit, and a strategy verification unit. The state semantic compression unit is used to extract the location, range, trend of change, severity, remaining effective compaction time, and status of nearby construction equipment from the continuous material state of abnormal areas. The construction knowledge retrieval unit is used to obtain process knowledge that matches the current construction state from construction specifications, construction organization design, mixture process requirements, equipment capacity descriptions, project-specific plans, historical construction records, and abnormal handling cases. The candidate strategy generation unit is used to generate strategies such as strong vibration, weak vibration, rubber-tired compaction, static compaction, slowing down paving, adjusting equipment following distance, changing the operation sequence of the machine group, local compaction, and combinations of two or more of these actions. The strategy verification unit is used to eliminate candidate strategies that are unavailable, have ineffective action time, do not match the action object and the action area, or have process conflicts.

5. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The counterfactual construction scenario construction module includes a fact state freezing unit, a candidate strategy intervention and analysis unit, and a multi-branch scenario generation unit; the fact state freezing unit is used to fix the current material state, environmental conditions, mixture properties, equipment location, and geometric conditions of the construction area; The candidate strategy intervention and analysis unit is used to convert candidate strategies into device trajectories, action width, action area, operation mode, action duration, and sequential, following, waiting, and switching relationships between multiple devices that vary with space and time. The multi-branch scenario generation unit is used to establish a baseline branch that maintains the current construction plan and an intervention branch that adopts different candidate strategies, starting from a common fact state, and to maintain consistency of non-intervention conditions between different branches.

6. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The neural operator counterfactual world model module includes a continuous field input enhancement layer, a spatiotemporal coding layer, a global neural operator layer, a local compaction coupling layer, an equipment control injection layer, and a multi-field output layer. The global neural operator layer is used to establish nonlocal state transmission relationships between different spatial locations. The local compaction coupling layer is used to characterize the nonlinear coupling relationships between material temperature, skeleton rearrangement, compaction growth, and uniformity changes within the roller's operating area. The equipment control injection layer is used to inject construction interventions corresponding to candidate strategies into the continuous material state based on the future trajectory, operation mode, operating time, and equipment sequence of the roller or paver. The multi-field output layer is used to output the temperature field, skeleton state field, compaction state field, and compaction uniformity field at multiple future time points.

7. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 6, characterized in that: The neural operator counterfactual world model is constructed using Fourier neural operators, graph neural operators, wavelet neural operators, low-rank integral operators, or combinations thereof, and is trained using field construction data, quality inspection data, indoor test data, and numerical simulation data. During the training process, continuous material field prediction constraints, temperature evolution constraints, compaction evolution constraints, construction boundary constraints, and prediction uncertainty constraints are set to ensure that the model prediction results satisfy the engineering laws corresponding to temperature transfer, material compaction response, equipment action area, and construction area boundary, and output at least one of the following: prediction confidence interval, prediction variance, or state achievement probability.

8. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The multi-future evaluation and strategy selection module extracts at least one evaluation index from various counterfactual construction scenarios, including the compaction compliance rate, under-compaction area, over-compaction area, compaction uniformity, temperature segregation degree, skeleton stability, equipment waiting time, repetitive work volume, equipment idle distance, construction cycle time, equipment energy consumption, safety risk, and prediction uncertainty. It also sets corresponding weights for different evaluation indices according to the construction stage. When the construction quality evaluation results of multiple candidate strategies meet the preset requirements, the candidate strategy with fewer equipment action switching times, less disturbance to the original construction organization, more continuous equipment operation, and lower prediction uncertainty is selected first.

9. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The deterministic constraint and execution feedback module establishes a safe and feasible domain consisting of equipment capability constraints, material state constraints, machine group safety constraints, construction area constraints, and control continuity constraints. Equipment capability constraints limit equipment speed, acceleration, vibration frequency, amplitude, steering, and mode switching conditions. Material state constraints limit compaction actions based on current and predicted temperature, frame, and compaction conditions. Machine group safety constraints limit spatial distance and trajectory conflicts between different construction equipment. Construction area constraints restrict equipment from entering prohibited operating areas. Control continuity constraints limit the magnitude of changes in speed, vibration parameters, and motion state between adjacent control cycles. When some control parameters of the preferred strategy are outside the safe and feasible domain, the corresponding control parameters are corrected to be within the safe and feasible domain. When the preferred strategy cannot obtain feasible control parameters, it switches to the suboptimal candidate strategy.

10. The intelligent agent system for asphalt pavement construction based on a neural operator counterfactual world model according to claim 1, characterized in that: The deterministic constraint and execution feedback module adopts a rolling control method, issuing control actions corresponding to the current control cycle only to the construction equipment. After the current control cycle is executed, it reacquires the actual equipment trajectory, material temperature, vibration response, and compaction state, compares the actual construction state with the prediction results of the neural operator counterfactual world model, updates the current continuous material state, equipment response parameters, or model calibration parameters based on the comparison results, and then triggers the generation of candidate strategies, construction of counterfactual scenarios, prediction of future material fields, strategy evaluation, and deterministic control solution for the next control cycle.

Images