A multi-agent construction simulation optimization method, system, device and storage medium

Abstract

The application discloses a kind of multi-agent construction simulation optimization method, system, equipment and storage medium, it is related to building construction informatization and artificial intelligence cross technical field, including the construction environment basic data based on target engineering, corresponding virtual construction environment is generated;Construction task is parsed based on management agent, and standardization task description unit and semantic feature are output;Standardized task description unit is received in virtual construction environment by executing agent, and collaborative physical simulation is executed, and operating state feature is output;Operating state feature and semantic feature are input into executing agent to carry out optimization, and target construction scheme is output.The method described in the application extracts semantic features through management agents to ensure compliance with instructions;Collaborative simulation is used by executing agents, and dynamic interference game is truly restored;Finally, joint operating state and semantic features execute multi-objective optimization, improve convergence efficiency, and ensure that the output scheme has global optimality and engineering feasibility.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of building construction informatization and artificial intelligence, specifically to a multi-agent construction simulation optimization method, system, equipment, and storage medium. Background Technology

[0002] With the digital transformation of the construction industry, the formulation and optimization of construction plans are gradually shifting from traditional manual experience-driven approaches to information-based and intelligent methods. Currently, Building Information Modeling (BIM) technology and computer-aided design software are widely used in the field to construct three-dimensional, visualized static models of projects, and these are combined with conventional simulation tools to pre-simulate construction progress, resource allocation, and spatial layout. This digital construction technology, based on pre-set processes and idealized conditions, has improved the visualization level of construction planning to a certain extent and has become an important technical means for the pre-construction preparation and scheme design of modern large-scale construction projects.

[0003] However, existing construction simulation technologies still have significant shortcomings when dealing with complex and ever-changing real-world engineering projects. Traditional simulation environments lack the fusion and mapping of dynamic on-site data, resulting in rigid simulation processes that fail to reflect real physical interference and dynamic constraints. In actual engineering management, macro-level construction task instructions are often issued in unstructured form, with complex construction method constraints and highly coupled implicit logical conflicts hidden between process nodes. Existing virtual construction entities typically only follow fixed, hard numerical rules, lacking the professional semantic parsing capability for unstructured instructions. They cannot effectively identify and assess the aforementioned implicit risks before physical execution. When facing cross-operation of multiple trades and equipment, they cannot autonomously understand intent and resolve collaborative conflicts, easily leading to local deadlocks or simulation distortions. When optimizing solutions for multiple objectives, existing technologies often rely solely on single numerical quantitative indicators for blind trial and error, failing to deeply explore and utilize the implicit semantic risk characteristics behind the construction state. This makes it difficult for optimization algorithms such as reinforcement learning to converge when faced with a large and complex construction state space, making it difficult to truly and efficiently output a globally optimal construction solution that balances schedule, cost, and safety. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is: existing construction scheme simulation methods lack dynamic physical mapping in the simulation environment, the execution entities lack autonomous collaborative game ability leading to easy deadlock, the scheme optimization relies on a single numerical index leading to blind search and difficulty in convergence, and how to combine large language model semantic reasoning and multi-agent reinforcement learning mechanism to achieve global optimal solution in complex construction state space.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a multi-agent construction simulation optimization method, comprising generating a corresponding virtual construction environment based on the basic construction environment data of the target project; parsing the construction tasks based on the management agent, outputting standardized task description units and semantic features; the execution agent receiving the standardized task description units in the virtual construction environment and performing collaborative physical simulation, outputting operational status features; and inputting the operational status features and semantic features into the execution agent for optimization, outputting the target construction scheme.

[0007] As a preferred embodiment of the multi-agent construction simulation optimization method described in this invention, the generation of the corresponding virtual construction environment includes: parsing and extracting the basic data of the construction environment and outputting spatiotemporal mapping parameters; using a spatial coordinate system to perform entity space reconstruction and process logic connection on the spatiotemporal mapping parameters; and outputting a virtual construction environment with spatial physical interference boundaries and dynamic operation constraints.

[0008] As a preferred embodiment of the multi-agent construction simulation optimization method described in this invention, the step of parsing the construction task based on the management agent includes: the management agent receiving construction task instructions; using the built-in construction big language model to call hierarchical embedded knowledge to execute instruction parsing and implicit risk assessment; and outputting standardized task description units that match the underlying execution protocol and semantic features that characterize the engineering feasibility boundary.

[0009] As a preferred embodiment of the multi-agent construction simulation optimization method of the present invention, the execution of collaborative physical simulation includes: the executing agents receiving standardized task description units and perceiving the operation interference state in the virtual construction environment during operation; when operation interference occurs, the executing agents perform distributed consensus negotiation on the interactive task attribute parameters, output avoidance strategies to eliminate operation interference, complete physical simulation based on avoidance strategies, and output operating status characteristics.

[0010] As a preferred embodiment of the multi-agent construction simulation optimization method described in this invention, the execution of the collaborative physical simulation further includes: the executing agent receiving individual action feedback results and global optimization constraints; calculating a global guidance reward signal using a policy gradient algorithm and mapping the global guidance reward signal to the behavior space of the executing agent; and outputting a collaborative action sequence restricted by the global optimization constraints.

[0011] As a preferred embodiment of the multi-agent construction simulation optimization method of the present invention, the output running state features include: inputting the position parameters and attribute parameters of the executing agent in the virtual construction environment to the physics engine; using the physics engine to perform simulation calculations of kinematic and dynamic features; and outputting a set of quantified numerical values ​​representing the kinematic and dynamic features of the entity as running state features.

[0012] As a preferred embodiment of the multi-agent construction simulation optimization method of the present invention, the output target construction scheme includes: receiving running state features and semantic features, injecting the semantic features into the reward function of the reinforcement learning model in the executing agent to generate a knowledge-enhanced reward function; using the reinforcement learning model in the executing agent to perform state space optimization based on the knowledge-enhanced reward function and running state features; and outputting the target construction scheme.

[0013] Another objective of this invention is to provide a multi-agent construction simulation optimization system, which can drive the execution agents to carry out dynamic inference and combine semantic features to perform multi-objective balance iteration. This solves the problems of current construction scheme simulation technology, such as the lack of dynamic constraints in the simulation environment, the lack of autonomous collaborative game ability of the execution entities, and the blind search and difficulty in convergence caused by the reliance on a single numerical index.

[0014] As a preferred embodiment of the multi-agent construction simulation optimization system described in this invention, it includes: a virtual environment generation module, a task parsing module, a collaborative simulation module, and a scheme optimization module; the virtual environment generation module is used to receive physical and spatiotemporal related basic data, reconstruct the space, and output a virtual construction environment with physical interference boundaries and dynamic operation constraints; the task parsing module is used to receive macro-level construction task instructions, use the management agent to perform professional logic parsing, output low-level executable standardized task description units, and simultaneously output semantic features representing engineering risks; the collaborative simulation module is used to drive the execution agent to conduct dynamic deduction and interactive game based on the received task units in the virtual construction environment, and output operating state features; the scheme optimization module is used to receive operating state features and semantic features, drive the execution agent to perform multi-objective balance iteration in the state space, and output the target construction scheme.

[0015] Another object of the present invention is to provide a multi-agent construction simulation optimization device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the multi-agent construction simulation optimization method.

[0016] Another object of the present invention is to provide a multi-agent construction simulation optimization storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the multi-agent construction simulation optimization method.

[0017] The beneficial effects of this invention are as follows: The multi-agent construction simulation optimization method provided by this invention generates a corresponding virtual construction environment based on the basic data of the target project's construction environment, achieving a precise mapping between the real construction physical space and constraints in the virtual environment; by analyzing construction tasks based on the management agent and outputting standardized task description units and semantic features, it overcomes the problem of traditional instruction issuance lacking professional domain logic, ensuring the compliance of engineering specifications for the underlying execution tasks; by receiving standardized task description units and performing collaborative physical simulation in the virtual construction environment, the execution agent is endowed with the ability of autonomous interaction and collaborative game, realistically restoring the dynamic interference process of complex cross-operations; by jointly inputting the running state features and semantic features into the execution agent for optimization, it changes the limitation of traditional iteration blindly relying on a single numerical quantitative indicator, effectively guiding the multi-objective search direction in a large state space using semantic features, significantly improving the model convergence efficiency and ensuring that the final output target construction scheme has global optimality and engineering feasibility. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The above is an overall flowchart of a multi-agent construction simulation optimization method provided in Embodiment 1 of the present invention.

[0020] Figure 2 This is an interaction timing diagram of a multi-agent construction simulation optimization method provided in Embodiment 1 of the present invention.

[0021] Figure 3 This is a timing diagram of the execution agent's collaborative avoidance of interference in the operation of a multi-agent construction simulation optimization method provided in Embodiment 1 of the present invention.

[0022] Figure 4 This is a comparison chart of the convergence effects of multi-agent multi-objective balance iterative optimization for a multi-agent construction simulation optimization method provided in Embodiment 1 of the present invention.

[0023] Figure 5 This is a data flow diagram between modules of a multi-agent construction simulation optimization system provided in Embodiment 2 of the present invention.

[0024] In the diagram: 101, Virtual construction environment; 201, Management agent; 202, Standardized task description unit; 203, Semantic features; 301, Execution agent; 301a, First execution agent; 301b, Second execution agent; 302, Operating status features; 401, Target construction plan; A, Virtual environment generation module; B, Task parsing module; C, Collaborative simulation module; D, Plan optimization module. Detailed Implementation

[0025] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0026] Example 1, referring to Figures 1-4 As an embodiment of the present invention, a multi-agent construction simulation optimization method is provided, comprising: S1: Based on the basic construction environment data of the target project, generate the corresponding virtual construction environment 101.

[0027] Furthermore, generating the corresponding virtual construction environment 101 includes parsing and extracting the basic data of the construction environment, outputting spatiotemporal mapping parameters; using a spatial coordinate system to reconstruct the entity space and connect the process logic to the spatiotemporal mapping parameters; and outputting a virtual construction environment 101 with spatial physical interference boundaries and dynamic operation constraints, such as... Figure 1 As shown.

[0028] It should be noted that the geometric and structural information of the target project is extracted from the BIM database, and dynamic data on site is collected through IoT devices and integrated into the construction organization schedule. A spatiotemporal alignment operation is performed: using a unified Cartesian coordinate system, the entity position parameters are mapped to spatial nodes in the model, and the process progress (i.e., logical relationship) is attached as a time dimension label to the corresponding spatial nodes, generating a virtual construction environment that possesses both "spatial physical interference boundaries" and "dynamic operational constraints."

[0029] Taking the construction of the core tube of a super high-rise building as an example, when implementing the method of this embodiment, the BIM 3D model data of the core tube wall is extracted, the UWB positioning data of the tower crane on site is accessed in real time, and the "core tube construction progress Gantt chart" is retrieved. The data is parsed and extracted into spatiotemporal mapping parameters containing location, time, and attributes. The climbing formwork entity being lifted is reconstructed in the virtual space coordinate system (defining the physical boundary), and a dynamic process logic of "lifting can only be carried out after the concrete strength of the lower wall reaches 15MPa" is attached to the virtual node corresponding to the climbing formwork (setting the operation constraint). This outputs a high-fidelity virtual environment in which the intelligent agent can legally operate.

[0030] It should also be noted that by extracting and constructing unified spatiotemporal mapping parameters, the heterogeneous data barriers between traditional BIM static spatial data, fragmented IoT dynamic data and macro-level schedule plans have been successfully broken down, achieving standardized fusion of multi-source data in the same coordinate system.

[0031] S2: Based on the management intelligent agent 201, the construction task is analyzed, and standardized task description unit 202 and semantic features 203 are output.

[0032] Furthermore, the parsing of construction tasks based on the management agent 201 includes: the management agent 201 receiving construction task instructions; using the built-in construction big language model (LLM) to call hierarchical embedded knowledge to execute instruction parsing and implicit risk assessment; and outputting standardized task description units 202 that match the underlying execution protocol and semantic features 203 that characterize the engineering feasibility boundary.

[0033] It should be noted that the management agent 201 receives macro-level construction task instructions as input, triggering the built-in construction big language model. Hierarchical embedded knowledge is pre-constructed as a construction-domain-specific knowledge graph and vector database, including a foundation layer primarily based on hard constraints, a decision layer primarily based on the spatiotemporal topological relationships of processes (such as the logic network of pre- and post-processes), and a method layer primarily based on the underlying control protocols of equipment. In the downward decomposition direction, the construction big language model extracts semantic entities from the input unstructured macro-level task instructions and performs vector mapping and graph alignment with the hierarchical embedded knowledge. By extracting the process topological relationships in the decision layer and the control parameters in the method layer, the natural language instructions are transformed into fine-grained operational steps constrained by construction method logic, forming standardized task description units 202 that the bottom-level execution agent 301 can directly respond to, possessing strict timestamps and spatial coordinates. In the upward reasoning direction, combining the contextual environment state and the safety specification boundaries of the foundation layer, a latent risk assessment is performed on the potential hazards of the current process combination within the limited spatiotemporal constraints, and semantic features 203 representing the engineering feasibility boundary are extracted. Figure 2As shown, after completing instruction parsing, the management agent 201 outputs the standardized task description unit 202 and then flows it down to the collaborative physical simulation stage. At the same time, it outputs the semantic features 203, such as the engineering feasibility boundary, and directly flows them into the subsequent reinforcement learning model.

[0034] Taking the task issuance process during the construction of the core tube of a super high-rise building as an example. When executing the method of this embodiment, the management agent 201 first receives the construction task instruction of "completing the concrete pouring of the 10th floor core tube bottom slab". The management agent 201 calls the construction process logic knowledge graph in the hierarchical embedded knowledge, and the construction big language model identifies the implicit pre-dependent node (i.e., bottom slab reinforcement binding and acceptance) and automatically converts it into machine instructions containing pre-triggered conditions. Then, it breaks it down into a series of standardized task description units 202 such as "pump truck enters the site and is positioned", "deploy the concrete placing machine to the designated coordinates", and "pour concrete at the set speed", and issues them to the corresponding execution agent 301. At the same time, combined with the standard analysis of the current input on-site meteorological environment data in the hierarchical embedded knowledge, it is found that the wind speed value is near the critical threshold of the standard. Then, an implicit risk assessment is performed, and the semantic feature 203 of "the current high-altitude wind speed is relatively large. If the surrounding tower cranes are operating simultaneously, there is a high risk of swaying interference of the suspended object" is output. Thus, the feasibility boundary of the current working environment is defined by the logic of business language.

[0035] It should also be noted that by utilizing the built-in construction language model and invoking hierarchical embedded knowledge, the application limitations of traditional large language models, which are limited to natural language text generation, are overcome. This embodiment uses the construction language model as the compilation hub connecting unstructured engineering intentions with the underlying rigorous mechanism model. Through graph alignment and vector retrieval, it automatically extracts implicit process logic and attaches spatiotemporal constraint nodes, constructing a reliable conversion mechanism connecting macro-level schedule plans and underlying entity actions. This ensures that the instructions received by the underlying execution agent 301 possess absolute engineering specification compliance and precise granularity, eliminating execution deviations or physical deadlocks caused by ambiguity in general instructions. In the upward reasoning direction, implicit risk assessment is performed and semantic features 203 are output, overcoming the technical deficiency of traditional pure numerical simulations lacking professional business qualitative analysis.

[0036] S3: The execution agent 301 receives the standardized task description unit 202 in the virtual construction environment 101 and performs collaborative physical simulation, outputting the running status characteristics 302.

[0037] Furthermore, the collaborative physical simulation includes the following steps: the execution agent 301 receives the standardized task description unit 202 and senses the operation interference state in the virtual construction environment 101 during operation; when operation interference occurs, the execution agents 301 perform distributed consensus negotiation by exchanging task attribute parameters, output avoidance strategies to eliminate operation interference, complete physical simulation based on avoidance strategies, and output operating state features 302.

[0038] It should be noted that after parsing the standardized task description unit 202, the executing agent 301 drives itself to perform deductions along a predetermined trajectory or time sequence in the virtual construction environment 101. During this period, it continuously calculates the spatial intersection of its own three-dimensional motion envelope with other surrounding executing entities or static physical boundaries to perceive the operation interference status in real time. If a trajectory intersection warning or resource monopoly conflict is detected in the short-term spatiotemporal range, the underlying collaborative negotiation mechanism is triggered. The parties involved in the interference conflict establish peer-to-peer communication and exchange task attribute parameters. The task attribute parameters are a multi-dimensional vector matrix, and the core parameters include the baseline task priority, the urgency index of a single node process, and the delay penalty weight associated with subsequent critical paths. Based on the above interactive data, the executing agent 301 runs a distributed consensus protocol (such as the improved Raft algorithm) to perform distributed consensus negotiation. Each conflicting party comprehensively and weights the received task attribute parameters according to a preset value assessment function. The party with the higher value score (i.e., the one that causes the greatest global delay cost) gains the dominant position in the communication vote and is confirmed as the priority passage party. The party with the lower value score is automatically determined as the yielding party. The yielding party autonomously calculates and outputs an avoidance strategy based on the expected time and space range occupied by the priority passage party. The avoidance strategy is either a time-dimensional calculation of peak waiting time or a spatial dimension of obstacle avoidance trajectory replanning. After eliminating the operation interference, each executing agent 301 continues to drive the built-in physics engine to complete the subsequent simulation based on the updated avoidance strategy, and finally quantifies and outputs the operating status characteristics 302, which include entity displacement deviation, actual operation time, load changes, and overall energy consumption indicators.

[0039] Taking the multi-equipment cross-operation during the construction of the core tube of a super high-rise building as an example. The execution agent 301, representing the tower crane, receives the standardized task description unit 202 for hoisting steel bars into the core tube and begins simulation; simultaneously, the execution agent 301, representing the climbing formwork, reaches the predetermined jacking time node. At this point, both parties perceive a spatial overlap of operations within the confined space of the core tube, causing operational interference. The execution agents 301 corresponding to the tower crane and climbing formwork interact with each other, calculating and comparing task attribute parameters: obstructing the climbing formwork jacking will cause all subsequent wall construction processes to stagnate (a highly correlated penalty weight). The tower crane execution agent 301 outputs an avoidance strategy of "adjusting the crane boom to perform secondary tasks in non-core tube areas" to relinquish space. After eliminating the operational interference, the execution agent 301 completes physical simulation based on this avoidance strategy and finally outputs operating status characteristics 302 representing the actual waiting time, rotation angle, and motor energy consumption load of the tower crane.

[0040] Furthermore, such as Figure 3 As shown, the management agent 201 distributes the parsed and generated standardized task description units 202 to the first execution agent 301a and the second execution agent 301b, which have spatial overlap. The first execution agent 301a and the second execution agent 301b execute work actions according to their respective tasks in the virtual construction environment 101, and obtain dynamic work constraints and work status from environmental feedback in real time. When the virtual construction environment 101 detects that the first execution agent 301a and the second execution agent 301b have work interference in the future spatiotemporal range, a collaborative negotiation mechanism is triggered. The first execution agent 301a and the second execution agent 301b establish peer-to-peer communication, bidirectionally exchange task attribute parameters (such as process urgency indicators, delay penalty weights, etc.), and perform distributed consensus negotiation based on a preset consistency protocol. After negotiation calculation and comparison, the first execution agent 301a is confirmed as the priority proceeder, and the second execution agent 301b is the yielding party. The second execution agent 301b autonomously generates and outputs an avoidance strategy, including waiting or trajectory replanning, based on the expected spatiotemporal range occupied by the first execution agent 301a. Both agents eliminate interference based on this avoidance strategy and input updated cooperative actions into the virtual construction environment 101 for subsequent simulation. The first and second execution agents 301a and 301b input the position and attribute parameters of the current simulation frame into the underlying physics engine, which performs kinematic and dynamic simulation calculations and ultimately provides closed-loop feedback of quantified operational state features 302 to each execution agent 301.

[0041] It should also be noted that, through the distributed consensus negotiation mechanism, multiple conflicting entities can complete dynamic game theory in a decentralized peer-to-peer network, recreating the interaction process of multi-machine collaborative operation on site, and effectively resolving the local physical deadlock problem that is very easy to cause in complex cross-construction. During the negotiation process, the task attribute parameters (such as the delay penalty weight associated with subsequent critical paths) are compared, so that the avoidance strategy calculated and output by the yielding party strictly obeys the maximization of the global process value of the system. Based on the real avoidance action after the interference is eliminated, the physics engine continues to drive and output the running state feature 302, ensuring that the final collected kinematic and dynamic data fully conforms to objective physical laws, and greatly eliminating the objective deviation between theoretical scheme deduction and real physical conditions.

[0042] Furthermore, the execution of the cooperative physical simulation also includes the following steps: the executing agent 301 receives individual action feedback results and global optimization constraints; calculates a global guidance reward signal using a policy gradient algorithm and maps the global guidance reward signal to the behavior space of the executing agent 301; and outputs a cooperative action sequence constrained by the global optimization constraints.

[0043] It should be noted that during the simulation process, each time the execution agent 301 performs an action, it simultaneously receives individual action feedback results from its own sensors or physics engine (e.g., data on the local efficiency improvement brought about by a crane's self-acceleration under full load). At the same time, the execution agent 301 receives global optimization constraints issued by the project's top-level authority (e.g., peak power load limits for the entire work area and safety red line restrictions to avoid multiple equipment clusters). If relying solely on individual action feedback results, the execution agent 301 is highly susceptible to resource contention in order to quickly clear its own tasks. Therefore, a policy gradient algorithm is introduced for performance evaluation. When an individual action improves local efficiency but approaches or even violates global optimization constraints, the policy gradient algorithm calculates a global guiding reward signal with a negative penalty value. This global guiding reward signal is directly mapped to the action space of the execution agent 301 to update its action probability distribution. Actions that violate global constraints are given a very low selection probability, while actions that take into account global benefits are given a higher selection probability. Through high-frequency simulation iterations, the executing agent 301 continuously adjusts its action choices, abandoning selfish behaviors that yield short-term high returns but harm overall efficiency. Ultimately, it converges and outputs a set of collaborative action sequences that, while ensuring the progress of individual tasks, do not exceed the overall macro-level red line and are subject to global optimization constraints.

[0044] Taking the simulation of multiple tower cranes and construction elevators operating simultaneously at a construction site as an example, in the simulation, the agent 301 corresponding to tower crane No. 1 attempts to perform a lifting action with a large-angle, full-speed rotation. The individual action feedback it receives indicates that the lifting time for this single action is extremely short. However, the global optimization constraints at this time require controlling the total instantaneous power load on site to prevent tripping and strictly ensuring the physical collision avoidance redundancy of adjacent equipment. Using the policy gradient algorithm, it was found that the full-speed action of tower crane No. 1 caused a surge in instantaneous power grid load on site, exceeding the limit and forcing the nearby construction elevator to trigger emergency braking, seriously damaging the overall efficiency of the project. Based on this, the algorithm calculates a global guidance reward signal with a high penalty and maps it to the behavior space of tower crane No. 1, significantly reducing the probability of it subsequently performing a full-speed action. After multiple policy iterations, the agent of tower crane No. 1 finally outputs "reducing speed to 80% of rated power and actively hovering and waiting for 2 minutes when the construction elevator passes through a critical elevation," achieving an optimal balance between overall energy consumption and safety of cross-operations while steadily advancing its own task.

[0045] It should also be noted that enabling the executing agent 301 to autonomously perceive interference and complete dynamic game through distributed consensus negotiation reduces local deadlocks in multi-machine cross-operations. During negotiation, task attribute parameters are compared and global optimization constraints are introduced simultaneously. The policy gradient algorithm is used to transform these constraints into global guiding reward signals and map them to the individual's behavior space. This guides entities to avoid high-risk actions through algorithmic rewards and punishments, reducing the tendency to sacrifice the system's global optimum by pursuing individual local optima. This mechanism outputs a sequence of collaborative actions limited to macroscopic conditions without the need for pre-setting massive amounts of artificial rules, and drives the physics engine to generate operating state characteristics 302 that conform to objective laws, achieving an effective balance between the underlying physical simulation and the top-level engineering global benefits.

[0046] Furthermore, the output running state feature 302 includes inputting the position parameters and attribute parameters of the executing agent 301 in the virtual construction environment 101 into the physics engine; using the physics engine to perform simulation calculations of kinematic and dynamic features; and outputting a set of quantified numerical values ​​representing the kinematic and dynamic features of the entity as the running state feature 302.

[0047] It should be noted that the data transmission interface extracts the position and attribute parameters of the executing agent 301 in the current simulation frame and formats them into rigid body definition parameters and constraints that the physics engine can recognize. Simultaneously, it accesses global environmental parameters preset by the environment generation module (such as gravitational acceleration, ambient wind speed, and contact surface friction coefficient). At the kinematic simulation level, the physics engine, based on the action sequence instructions of the executing agent 301 and the set time step, applies numerical integration methods to solve the motion equations, calculating the instantaneous displacement, linear velocity, and angular velocity of each joint in three-dimensional space. At the dynamic simulation level, the physics engine performs dynamic force analysis based on the Newton-Euler equations or the Lagrange mechanical system. The calculation process comprehensively considers the equipment's self-weight, the weight of the suspended / effective load, inertial forces, and external environmental resistance, calculating the underlying driving force necessary to maintain the current kinematic state, the output torque borne by each hinge point, and the accumulated work and energy consumption of the equipment over time. The physics engine discretizes and samples the continuous simulated physical quantities. The metrics, including runtime, peak instantaneous speed, maximum load, and total energy consumption, are aligned and normalized, and then encapsulated to generate a structured set of quantified values. This set of quantified values ​​serves as the final runtime state feature 302.

[0048] It should also be noted that by using the physics engine to perform simulation calculations of kinematic and dynamic characteristics and outputting a set of quantified numerical values, the simulation process is expanded from a single spatial geometric boundary detection to an objective physical calculation dimension that includes force analysis and energy consumption calculation. By coupling the position parameters, attribute parameters, and global environment parameters of the executing agent 301, the underlying driving force, torque, and energy consumption are solved based on physical equations, which improves the degree to which the simulation model reflects the real engineering physical laws. The continuous simulated physical quantities are processed to generate a structured set of quantified numerical values, which are output as the operating state feature 302, thus reducing the theoretical deviation between the simulation scheme and the actual engineering conditions.

[0049] S4: Input the running state features 302 and semantic features 203 into the execution agent 301 for optimization, and output the target construction plan 401.

[0050] Furthermore, the output target construction scheme 401 includes receiving the running state features 302 and semantic features 203, injecting the semantic features 203 into the reward function of the reinforcement learning model in the executing agent 301 to generate a knowledge-enhanced reward function; using the reinforcement learning model in the executing agent 301, performing state space optimization based on the knowledge-enhanced reward function and the running state features 302; and outputting the target construction scheme 401.

[0051] It should be noted that the semantic features 203 (such as the overload risk threshold indicated in the assessment report) output by the large language model are extracted and converted into mathematical constraint terms (such as setting a precipitous negative penalty constant) using reward shaping techniques. Simultaneously, the quantified operational state features 302 (such as the actual construction period increment) calculated from physical simulation are received. These are then converted into scalar numerical semantic penalty terms and directly superimposed onto the basic reward function of the MDP, thereby reconstructing and generating a knowledge-enhanced reward function containing prior knowledge of engineering safety and compliance. The reinforcement learning model specifically employs the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. During the optimization iteration process, the operational state features 302 are input as environmental states into the Actor network (policy network) and Critic network (value network) of the MADDPG algorithm. The reinforcement learning model calculates the reward value of the current action selection under the knowledge-enhanced reward function, thereby deriving the temporal difference (TD) error, and performing gradient backpropagation through a deep neural network. The high penalty imposed by semantic feature 203 allows the model to truncate high-risk decision branches that violate norms early in the exploration phase, enabling the Actor network to update the action probability distribution along the value gradient direction in the environment state space. Combined with... Figure 2 In the iterative optimization process at the bottom of the reinforcement learning model, after the MADDPG algorithm parameters are updated multiple times, the policy gradient of the reinforcement learning model gradually approaches zero. The algorithm converges in the limited legal state space and reaches the Pareto optimal state among the mutually constraining evaluation indicators such as construction period and cost. Finally, it outputs the target construction scheme 401 composed of the optimal cooperative action sequence of each agent.

[0052] Taking the optimization of the construction process rhythm of the core tube of a super high-rise building as an example, in the simulation, the MADDPG algorithm outputs the decision to "increase the concrete pouring speed of the wall to the theoretical extreme value to compress the construction period" in a certain iteration. At this time, the input operation status feature 302 shows that the construction period of a single floor is shortened; however, the semantic feature 203 received at the same time indicates that "the pouring speed is too fast, resulting in insufficient initial setting strength of the lower layer, which poses a risk of structural failure". The knowledge-enhanced reward function reconstructed by the reward shaping technique returns an extremely high negative TD error penalty to this action branch. The Critic network of the MADDPG algorithm evaluates that the value of this strategy is extremely low, and then guides the Actor network to update the parameters and abandon this high-risk exploration direction. After updating the network parameters along the safety value gradient, the algorithm converges and outputs the target construction scheme 401, which is to "increase the wall pouring speed by 10% and allow the tower crane to prioritize the hoisting of non-core tube areas during the climbing formwork lifting".

[0053] Furthermore, such as Figure 4As shown in the figure, the convergence trajectory of the reinforcement learning model built into the executive agent 301 during multi-objective balancing iteration is illustrated. The horizontal axis represents the number of iterations (Timesteps) in the simulation, and the vertical axis represents the cumulative reward calculated by the model. In the early stages of model iteration, the semantic features 203 parsed by the management agent are injected into the reward function of the reinforcement learning model to generate a knowledge-enhanced reward function. When the model attempts to explore high-risk decision branches that violate engineering specifications or safety red lines, the semantic features 203 are transformed into high negative penalties, manifested as significant negative fluctuations in the 0 to 1M step range in the figure. This effectively avoids blind trial and error in the vast state space of the algorithm, accurately truncates high-risk action exploration in the early stages. After clarifying the feasibility boundary, the reinforcement learning model continuously receives the quantized running state features 302 output by the physics engine and uses this as the environmental state input. The model performs high-frequency temporal difference error calculation and network gradient backpropagation based on the knowledge-enhanced reward function and running state features 302. The Actor network rapidly updates the action probability distribution along the value gradient direction, as shown by the steep rise of the cumulative reward curve in the 1M to 3M step interval of the figure. This indicates that the system is rapidly approaching the global optimum. After sufficient iterative optimization, the policy gradient of the reinforcement learning model gradually approaches zero, and the algorithm converges to the Pareto optimal state within the constrained state space. This is represented by the curve stabilizing at a high level after 3M steps in the figure. At this point, the optimization process ends, and the system finally outputs the target construction plan 401, which balances schedule, cost, and safety.

[0054] It should also be noted that by transforming semantic features 203 into quantified conditional constraints under the MDP framework through reward shaping techniques, the limitations of traditional reinforcement learning, which relies solely on simple operational state features 302 for trial and error, are overcome. During the optimization iteration process using the MADDPG algorithm, the Critic network can promptly evaluate the low-reward value of high-risk strategies by utilizing the negative TD error feedback from the knowledge-enhanced reward function, thereby guiding the Actor network to adjust the action probability distribution. This reduces invalid computations in the vast environmental state space and improves the convergence efficiency of algorithm parameter updates. Furthermore, by guiding the algorithm to approach and converge to the Pareto optimal state within a limited legal state space, the final output target construction scheme 401 achieves an objective balance among mutually constraining evaluation indicators such as construction period and cost, ensuring the compliance and feasibility of the scheme in actual engineering projects.

[0055] Example 2, refer to Figure 5 As one embodiment of the present invention, a multi-agent construction simulation optimization system is provided, including a virtual environment generation module A, a task parsing module B, a collaborative simulation module C, and a scheme optimization module D. Figure 5 As shown.

[0056] Among them, the virtual environment generation module A is used to receive physical and spatiotemporal related basic data for spatial reconstruction and output a virtual construction environment 101 with physical interference boundaries and dynamic operation constraints.

[0057] Task parsing module B is used to receive macro-level construction task instructions, use management intelligent agent 201 to perform professional logic parsing, output the underlying executable standardized task description unit 202, and simultaneously output semantic features 203 representing engineering risks.

[0058] The collaborative simulation module C is used to drive the execution agent 301 to carry out dynamic inference and interactive game based on the received task units in the virtual construction environment 101, and output the running status characteristics 302.

[0059] The scheme optimization module D is used to receive the running state features 302 and semantic features 203, drive the execution agent 301 to perform multi-objective balance iteration in the state space, and output the target construction scheme 401.

[0060] This embodiment also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements a multi-agent construction simulation optimization system as proposed in the above embodiment.

[0061] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements a multi-agent construction simulation optimization system as proposed in the above embodiment.

[0062] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0063] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0064] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0065] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0066] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A multi-agent construction simulation optimization method, characterized in that, include: Based on the basic construction environment data of the target project, a corresponding virtual construction environment (101) is generated. The construction task is analyzed based on the management agent (201), and standardized task description units (202) and semantic features (203) are output. The execution agent (301) receives a standardized task description unit (202) in the virtual construction environment (101) and performs a collaborative physical simulation, outputting the running status characteristics (302). The running status features (302) and semantic features (203) are input into the execution agent (301) for optimization, and the target construction plan (401) is output.

2. The multi-agent construction simulation optimization method as described in claim 1, characterized in that: The generation of the corresponding virtual construction environment (101) includes, Perform parsing and extraction on the basic data of the construction environment, and output spatiotemporal mapping parameters; Utilize spatial coordinate systems to perform entity space reconstruction and process logic connection on spatiotemporal mapping parameters; Output a virtual construction environment with spatial physical interference boundaries and dynamic operation constraints (101).

3. The multi-agent construction simulation optimization method as described in claim 2, characterized in that: The process of parsing construction tasks based on the management agent (201) includes, The management agent (201) receives the construction task instructions; The built-in construction big language model is used to call hierarchical embedded knowledge to execute instructions for parsing and implicit risk assessment; The output is a standardized task description unit (202) that matches the underlying execution protocol and a semantic feature (203) that characterizes the engineering feasibility boundary.

4. The multi-agent construction simulation optimization method as described in claim 3, characterized in that: The execution of collaborative physics simulation includes, The execution agent (301) receives the standardized task description unit (202) and senses the operation interference state in the virtual construction environment (101) during operation; When a job interference occurs, the interaction of task attribute parameters between the execution agents (301) is carried out in a distributed consensus negotiation, and an avoidance strategy is output to eliminate the job interference. Based on the avoidance strategy, physical simulation is completed and the running state characteristics (302) are output.

5. The multi-agent construction simulation optimization method as described in claim 1 or 4, characterized in that: The execution of collaborative physics simulation also includes, The executing agent (301) receives individual action feedback results and global optimization constraints; The global guidance reward signal is calculated using the policy gradient algorithm and then mapped to the behavior space of the executing agent (301). Output a sequence of cooperative actions subject to global optimization constraints.

6. The multi-agent construction simulation optimization method as described in claim 5, characterized in that: The output operating status feature (302) includes, The position and attribute parameters of the executing agent (301) in the virtual construction environment (101) are input into the physics engine; The physics engine is used to perform simulations of kinematic and dynamic characteristics. Output a set of quantified numerical values ​​representing the kinematic and dynamic characteristics of the entity as the running state characteristics (302).

7. The multi-agent construction simulation optimization method as described in any one of claims 1, 4, or 6, characterized in that: The output target construction plan (401) includes, Receive running state features (302) and semantic features (203), and inject the semantic features (203) into the reward function of the reinforcement learning model in the executing agent (301) to generate a knowledge-enhanced reward function; Using the reinforcement learning model in the executive agent (301), the state space is optimized based on the knowledge-enhanced reward function and the running state characteristics (302); the target construction plan is output (401).

8. A multi-agent construction simulation optimization system, employing the multi-agent construction simulation optimization method as described in any one of claims 1 to 7, characterized in that: It includes a virtual environment generation module (A), a task parsing module (B), a co-simulation module (C), and a solution optimization module (D); The virtual environment generation module (A) is used to receive physical and spatiotemporal related basic data for spatial reconstruction and output a virtual construction environment (101) with physical interference boundaries and dynamic operation constraints. The task parsing module (B) is used to receive macro-level construction task instructions, use the management agent (201) to perform professional logic parsing, output the underlying executable standardized task description unit (202), and simultaneously output the semantic features (203) representing engineering risks. The collaborative simulation module (C) is used to drive the execution agent (301) to carry out dynamic deduction and interactive game based on the received task unit in the virtual construction environment (101) and output the running status characteristics (302). The scheme optimization module (D) is used to receive the running state features (302) and semantic features (203), drive the execution agent (301) to perform multi-objective balance iteration in the state space, and output the target construction scheme (401).

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the multi-agent construction simulation optimization method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-agent construction simulation optimization method according to any one of claims 1 to 7.

Images