Multi-agent cooperative construction safety collaborative scheduling and risk disposal method and system

Abstract

The present application relates to the technical field of intelligent construction management, and more particularly to a construction safety collaborative scheduling and risk disposal method based on multi-agent cooperation. The method identifies atomic work units by constructing a multi-dimensional construction state space and creates a corresponding agent set. Agents share information and conduct multi-round negotiations through a hierarchical message passing architecture to generate a global collaborative scheduling plan. During execution, a risk prediction model triggers an emergency response mechanism to quickly suppress risks by freezing high-risk operations, adjusting work sequences, and reallocating resources, and feeds back the results to update safety parameters. The present application realizes dynamic collaboration in construction scheduling and proactive disposal of risks, improving construction safety and efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction management technology, and in particular to a method and system for collaborative scheduling and risk management of construction safety involving multiple agents. Background Technology

[0002] In large and complex construction sites, multi-agent technology has been introduced to assist in construction scheduling and safety management. Current conventional practices typically rely on centralized or distributed scheduling systems, combined with field data collected by sensor networks, to manage construction tasks, resource allocation, and potential risks. These systems often decompose the construction process into a series of tasks and assign an agent to each task or resource unit, making it responsible for local decision-making and status monitoring. Agents exchange information through pre-defined communication protocols and coordinate their actions according to certain rules or optimization algorithms, such as contract network protocols, market-based auction mechanisms, or simple negotiation strategies, aiming to achieve goals such as construction efficiency and resource utilization, while also responding to identified safety risks.

[0003] However, these conventional multi-agent scheduling and risk management methods reveal several inherent flaws in practical applications. A major problem lies in the fact that the coordination and conflict resolution mechanisms among agents are often not deep or flexible enough. Existing methods mostly employ single-round or finite-round negotiation, with agents primarily making decisions based on their own local goals and states, lacking a full understanding and sharing of the global security situation and the deeper intentions of other agents. This easily leads to negotiation deadlock or the generation of suboptimal global scheduling schemes, making it difficult to effectively balance the complex relationship between individual operational efficiency and overall construction safety. When multiple agents' tasks involve resource competition or spatiotemporal conflicts, simple priority rules or compromises are insufficient to achieve a consensus that is both fair and ensures global optimal safety.

[0004] Another significant drawback lies in the insufficient real-time and collaborative nature of risk response. Conventional systems typically employ an independent response model based on threshold triggers. This means that when a sensor or agent detects an excessive risk level, it often only takes pre-defined countermeasures within its localized area of ​​responsibility or alerts the central controller. This model lacks the ability to proactively predict future risk evolution trends, and emergency response actions are usually sequential and isolated, failing to fully mobilize and coordinate multiple related agents for parallel and joint intervention. For example, when dealing with a high-risk work site, the impact on surrounding work unit sequences and the dynamic reallocation of emergency resources may not be considered simultaneously, resulting in slow risk suppression, low efficiency, and potentially even secondary risks or scheduling chaos. Furthermore, the decision-making and response modules of existing methods are often fragmented, making it difficult to provide timely and effective feedback on risk management results and influence the generation of subsequent scheduling strategies. Summary of the Invention

[0005] The embodiments of the present invention provide a method and system for collaborative scheduling and risk management of construction safety through multi-agent cooperation, which can solve the problems in the prior art.

[0006] A first aspect of this invention provides a method for collaborative scheduling and risk management of construction safety through multi-agent cooperation, comprising: Spatiotemporal alignment and semantic annotation are performed on real-time data collected by the sensor network deployed at the construction site to form a multi-dimensional construction state space that includes work progress, resource occupancy, and risk exposure dimensions. Atomic work units with clear work boundaries and resource requirements are identified in the multi-dimensional construction state space. A corresponding set of intelligent agents is created according to the number of atomic work units. A hierarchical message passing architecture that supports bidirectional information exchange between intelligent agents is established. The hierarchical message passing architecture enables state synchronization and intent sharing among intelligent agents. Based on the shared information, a multi-round negotiation mechanism is used to resolve scheduling conflicts. The multi-round negotiation mechanism balances the local interests of each intelligent agent with the global security goal through concession and compensation strategies, ultimately forming a conflict-free global collaborative scheduling plan. During the execution of the global collaborative scheduling plan, a rolling time-domain risk prediction model is used to predict the risk evolution trend of each work unit within the future time window. When the predicted risk level exceeds the warning line, an emergency response mechanism jointly participated in by multiple agents is triggered. The emergency response mechanism achieves rapid risk suppression through three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution result of the emergency response mechanism is fed back to the inference engine of each agent to update the safety parameters for subsequent scheduling decisions.

[0007] The real-time data collected by the sensor network deployed at the construction site is spatiotemporally aligned and semantically labeled to form a multi-dimensional construction state space containing dimensions of work progress, resource occupancy, and risk exposure. Atomic work units with clear work boundaries and resource requirements are identified within this multi-dimensional construction state space, including: The system acquires real-time data collected by a sensor network deployed at the construction site. The real-time data includes timestamps, spatial coordinates, and measurement values. The system performs spatiotemporal alignment processing on the real-time data to unify the data collected by different sensors to a preset time reference and spatial coordinate system. Real-time data that has completed spatiotemporal alignment is mapped to at least one semantic category among job type, equipment status, personnel behavior, and material attributes. A multi-dimensional construction state space is constructed based on the data that has completed the semantic mapping. The multi-dimensional construction state space includes job progress dimension, resource occupancy dimension, and risk exposure dimension. The job progress dimension is quantified by the deviation between the planned completion ratio and the actual completion ratio. The resource occupancy dimension is quantified by equipment utilization rate and personnel presence rate. The risk exposure dimension is quantified by the frequency of safety monitoring parameters exceeding preset thresholds. In the multidimensional construction state space, atomic work units are identified. By detecting data clusters in the multidimensional construction state space that simultaneously meet the following conditions within a preset time window: continuous change in the work progress dimension, stable resource occupation dimension, and no interruption condition triggered in the risk exposure dimension, the spatiotemporal region corresponding to the data cluster is determined as an atomic work unit with clear work boundaries and resource requirements. The work boundaries include the start time, end time, and work space range, and the resource requirements include equipment type combinations and personnel skill combinations.

[0008] Create a corresponding set of agents according to the number of atomic work units, and establish a hierarchical message passing architecture that supports bidirectional information exchange between agents, including: Each agent is equipped with a perception engine, an inference engine, and a control engine. The perception engine continuously monitors the execution status of the corresponding atomic work unit and the dynamic changes of surrounding work units. The inference engine generates local scheduling decisions based on the monitoring results and preset safety rules. The control engine converts the local scheduling decisions into control signals for the work equipment. The design employs a layered messaging architecture. The bottom layer is the state synchronization layer, where each agent broadcasts state messages containing execution progress, resource usage snapshots, and current risk values ​​to neighboring agents at fixed intervals. The middle layer is the intent sharing layer, where agents send intent messages containing planned operations, required resource types and quantities, and expected execution time windows to potential conflicting parties after generating local scheduling decisions. The top layer is the negotiation layer, where conflicting agents exchange concession schemes through negotiation messages.

[0009] A layered messaging architecture is used to synchronize the states and share intentions among agents. Based on shared information, a multi-round negotiation mechanism is employed to resolve scheduling conflicts, including: A distributed consistency maintenance mechanism for the global state view is established. Each agent locally maintains a copy of the state of its neighboring agents. Vector clocks are used to mark the causal order of messages. Expiration and duplication of messages are detected by comparing vector clocks to ensure the causal consistency of state updates. Automatic detection rules for scheduling conflicts are defined, including resource conflict detection rules, timing conflict detection rules, and spatial conflict detection rules. The resource conflict detection rules determine whether two intentions request the same resource and whether their time windows overlap. The timing conflict detection rules determine whether the execution order of two intentions violates process constraints. The spatial conflict detection rules determine whether the operating spaces of two work units overlap and do not meet the safety distance requirement. Each agent constructs a local utility function based on the scheduling objective and constraints. During the negotiation process, the agent calculates the impact of different concession schemes on local utility, selects the concession scheme with the minimum utility loss to generate a negotiation message, and adds compensation conditions to the negotiation message to balance the utility loss. A negotiation termination condition is introduced: when the difference in utility between the concession proposals put forward by both parties in two consecutive rounds of negotiation is less than a preset threshold, or when the number of negotiation rounds reaches the upper limit, the negotiation is terminated and the final scheduling scheme is decided by the global optimization module.

[0010] Each agent constructs a local utility function based on the scheduling objective and constraints. During the negotiation process, the agent calculates the impact of different concession schemes on local utility, selects the concession scheme with the minimum utility loss, and generates a negotiation message, including: A multi-objective local utility function is constructed, which includes three sub-objectives: task completion time, resource utilization rate, and safety margin. These sub-objectives are aggregated into a single utility value through a weighted summation method, and the weights of each sub-objective are dynamically adjusted according to the priority of the current construction stage. In response to the detected conflict, the agent enumerates a set of feasible concession schemes, which include three basic concession actions and their combinations: delaying execution time, reducing resource requests, and shortening operation time. For each option in the set of concession options, the agent calculates the local scheduling plan after executing the option and evaluates the value of the local utility function under the new scheduling plan. By comparing the utility values ​​before and after the concession, the utility loss of each option is obtained. From the set of concession options, candidate options with utility loss less than the tolerable loss threshold are selected. The candidate options are then sorted in ascending order of utility loss, and the option with the smallest utility loss is selected as the proposal for this round of negotiation. Based on the proposed solution, compensation conditions are generated. These conditions include requesting the other party to prioritize resource allocation in subsequent time periods, requesting the other party to assume partial safety monitoring responsibility, and requesting the other party to adjust the order of subsequent operations to reduce waiting time. The specific content of the compensation conditions is determined according to the magnitude and type of utility loss. The proposed solution and its compensation conditions are encapsulated into a negotiation message and sent to the opposing intelligent agent. At the same time, the proposed solution and utility loss value of this round of negotiation are recorded to determine the convergence status in subsequent negotiation rounds.

[0011] During the execution of the global collaborative scheduling plan, a rolling time-domain risk prediction model is used to predict the risk evolution trend of each work unit within the future time window, including: A rolling time-domain risk prediction model is constructed. This model takes the current time as the starting point and sets a fixed-length prediction time-domain window. Within the prediction time-domain window, risk simulation is performed in discrete time steps. The input of the model for each time step includes the planned execution actions of the work unit, the resource allocation status, the predicted values ​​of environmental parameters, and the dynamic impact of adjacent work units. A risk propagation graph is established, with work units as nodes and directed edges representing the physical proximity, resource sharing, and process dependence relationships between work units. The weight of each edge represents the intensity coefficient of risk propagation between nodes. Dynamic risk simulation is performed on the risk propagation graph. The risk level of each node at the initial moment is taken from the current real-time monitoring value. The risk level at subsequent time steps is calculated based on the node's own risk growth function and the risk increment passed from adjacent nodes through directed edges. The risk growth function takes into account the inherent risk rate of the job type, the cumulative effect of the job duration, and the modulating effect of environmental factors.

[0012] When a risk level is predicted to exceed the warning threshold, an emergency response mechanism involving multiple agents is triggered, including: The risk level time series of each work unit obtained from the simulation are compared with the warning line to identify the work unit that first exceeds the warning line and the time of the exceedance. The work unit that exceeds the warning line is marked as a high-risk work unit. The associated work units that contribute to the risk of the high-risk work unit are identified by reverse tracing through the risk propagation map. An emergency trigger message is sent to multiple agents responsible for high-risk work units and their associated work units. The emergency trigger message includes the predicted breakthrough time, risk type, risk level, and correlation diagram. The emergency response mechanism is activated, and multiple agents participating in the emergency response simultaneously execute three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution results of each action are then summarized to form an emergency response report, which is fed back to the inference engine of each agent.

[0013] A second aspect of this invention provides a multi-agent collaborative construction safety scheduling and risk management system, comprising: The state space unit is used to perform spatiotemporal alignment and semantic annotation on real-time data collected by the sensor network deployed at the construction site, forming a multi-dimensional construction state space that includes the dimensions of work progress, resource occupancy, and risk exposure. In the multi-dimensional construction state space, atomic work units with clear work boundaries and resource requirements are identified. The collaborative scheduling unit is used to create a corresponding set of intelligent agents according to the number of atomic job units, establish a hierarchical message passing architecture that supports bidirectional information exchange between intelligent agents, realize state synchronization and intent sharing of each intelligent agent through the hierarchical message passing architecture, and resolve scheduling conflicts by adopting a multi-round negotiation mechanism based on shared information. The multi-round negotiation mechanism balances the local interests of each intelligent agent with the global security goal through concession strategy and compensation strategy, and finally forms a conflict-free global collaborative scheduling plan. The risk response unit is used to predict the risk evolution trend of each work unit within a future time window during the execution of the global collaborative scheduling plan using a rolling time-domain risk prediction model. When the predicted risk level exceeds the warning line, it triggers an emergency response mechanism jointly participated in by multiple agents. The emergency response mechanism achieves rapid risk suppression through three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution result of the emergency response mechanism is fed back to the inference engine of each agent to update the safety parameters for subsequent scheduling decisions.

[0014] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0015] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0016] This method constructs a multi-dimensional construction state space, enabling precise characterization of the complex and dynamic environment of the construction site. By identifying atomic work units, large-scale construction tasks are decomposed into independently manageable modules, laying the foundation for refined scheduling. The multi-dimensional state description comprehensively reflects the work progress, resource consumption, and risk exposure, providing accurate data support for subsequent intelligent agent collaboration.

[0017] The layered messaging architecture supports bidirectional information exchange between agents, ensuring the real-time performance and reliability of state synchronization and intent sharing. A multi-round negotiation mechanism, combined with concession and compensation strategies, effectively balances local operational efficiency with global security goals, systematically resolving resource contention and spatiotemporal conflicts. The resulting global collaborative scheduling plan possesses inherent consistency, avoiding decision-making biases caused by information lag in traditional centralized scheduling.

[0018] The rolling time-domain risk prediction model can proactively capture the risk evolution trend of each operational unit, realizing a shift from passive response to proactive early warning. When the risk exceeds the warning threshold, the triggered multi-agent joint emergency response mechanism achieves rapid intervention and suppression of the risk through three parallel actions: freezing, adjustment, and reallocation. This parallel approach significantly shortens the emergency response time and curbs the spread and chain reactions of the risk.

[0019] Emergency response results are fed back to the inference engines of each agent in real time for dynamic updates to safety parameters, enabling the scheduling system to continuously learn and adaptively optimize. The entire method forms a closed-loop management system of "state awareness - collaborative scheduling - risk prediction - emergency response - feedback optimization," which systematically improves the overall safety level and risk resistance capability of complex construction sites while ensuring construction progress. Attached Figure Description

[0020] Figure 1 A flowchart illustrating a multi-agent collaborative scheduling and risk management method for construction safety. Figure 2 This is a flowchart illustrating the process of resolving scheduling conflicts using a multi-round negotiation mechanism. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0023] Figure 1 This is a flowchart illustrating the multi-agent collaborative scheduling and risk management method for construction safety according to an embodiment of the present invention. Figure 1 As shown, the multi-agent collaborative method for construction safety scheduling and risk management includes: Spatiotemporal alignment and semantic annotation are performed on real-time data collected by the sensor network deployed at the construction site to form a multi-dimensional construction state space that includes work progress, resource occupancy, and risk exposure dimensions. Atomic work units with clear work boundaries and resource requirements are identified in the multi-dimensional construction state space. A corresponding set of intelligent agents is created according to the number of atomic work units. A hierarchical message passing architecture that supports bidirectional information exchange between intelligent agents is established. The hierarchical message passing architecture enables state synchronization and intent sharing among intelligent agents. Based on the shared information, a multi-round negotiation mechanism is used to resolve scheduling conflicts. The multi-round negotiation mechanism balances the local interests of each intelligent agent with the global security goal through concession and compensation strategies, ultimately forming a conflict-free global collaborative scheduling plan. During the execution of the global collaborative scheduling plan, a rolling time-domain risk prediction model is used to predict the risk evolution trend of each work unit within the future time window. When the predicted risk level exceeds the warning line, an emergency response mechanism jointly participated in by multiple agents is triggered. The emergency response mechanism achieves rapid risk suppression through three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution result of the emergency response mechanism is fed back to the inference engine of each agent to update the safety parameters for subsequent scheduling decisions.

[0024] The real-time data collected by the sensor network deployed at the construction site is spatiotemporally aligned and semantically labeled to form a multi-dimensional construction state space containing dimensions of work progress, resource occupancy, and risk exposure. Atomic work units with clear work boundaries and resource requirements are identified within this multi-dimensional construction state space, including: The system acquires real-time data collected by a sensor network deployed at the construction site. The real-time data includes timestamps, spatial coordinates, and measurement values. The system performs spatiotemporal alignment processing on the real-time data to unify the data collected by different sensors to a preset time reference and spatial coordinate system. Real-time data that has completed spatiotemporal alignment is mapped to at least one semantic category among job type, equipment status, personnel behavior, and material attributes. A multi-dimensional construction state space is constructed based on the data that has completed the semantic mapping. The multi-dimensional construction state space includes job progress dimension, resource occupancy dimension, and risk exposure dimension. The job progress dimension is quantified by the deviation between the planned completion ratio and the actual completion ratio. The resource occupancy dimension is quantified by equipment utilization rate and personnel presence rate. The risk exposure dimension is quantified by the frequency of safety monitoring parameters exceeding preset thresholds. In the multidimensional construction state space, atomic work units are identified. By detecting data clusters in the multidimensional construction state space that simultaneously meet the following conditions within a preset time window: continuous change in the work progress dimension, stable resource occupation dimension, and no interruption condition triggered in the risk exposure dimension, the spatiotemporal region corresponding to the data cluster is determined as an atomic work unit with clear work boundaries and resource requirements. The work boundaries include the start time, end time, and work space range, and the resource requirements include equipment type combinations and personnel skill combinations.

[0025] Real-time data is acquired from a sensor network deployed at the construction site. This sensor network includes various types of sensing nodes, such as vibration sensors, temperature sensors, displacement sensors, video acquisition devices, laser scanning devices, and RFID readers. Each sensing node generates a data record containing a timestamp, spatial coordinates, and measured values ​​according to a preset sampling frequency. The timestamp uses the UTC standard time format to record the moment the data was generated. The spatial coordinates are obtained using RTK positioning technology to obtain the sensor's three-dimensional position in the construction site coordinate system. The measured values, depending on the sensor type, include specific physical quantities such as acceleration amplitude, temperature value, displacement, image frame, or RF signal strength. In practical applications, different sensors experience inconsistencies in time and spatial reference due to clock drift, data transmission delays, and installation location deviations.

[0026] The real-time data undergoes spatiotemporal alignment processing. First, a global time reference is established at the construction site, and clock calibration is achieved by broadcasting time synchronization signals to each sensor node. Specifically, a time server deployed at the construction site command center is selected as the master clock. Each sensor node, acting as a slave clock, receives the synchronization signal and calculates the time deviation between its local clock and the master clock. This deviation is used as a correction factor to retrospectively adjust the timestamps of the locally collected data. For time deviations caused by network latency, the round-trip delay measurement method is used to estimate the transmission delay, and half of the one-way delay is used as the compensation value for the timestamp. Spatial alignment processing transforms the spatial coordinates in the data collected by each sensor to a three-dimensional Cartesian coordinate system established at the construction site. The origin of this coordinate system is set at the reference control point in the construction area, with the X-axis pointing due east, the Y-axis pointing due north, and the Z-axis vertically upward along the direction of gravity. For sensor data recorded using relative coordinates, the known coordinates of the sensor's installation location in the global coordinate system are obtained as reference points, and a coordinate transformation matrix is ​​used to convert the relative coordinates to global coordinates. After completing the spatiotemporal alignment, the sensor data that were originally scattered under different time bases and spatial reference systems are unified into the same spatiotemporal framework, providing a consistent data foundation for subsequent semantic annotation.

[0027] Real-time data with spatiotemporal alignment is mapped to at least one semantic category among job type, equipment status, personnel behavior, and material attributes. Job type mapping is achieved by analyzing the characteristic patterns of sensor data. For example, periodic high-frequency vibration signals collected by vibration sensors correspond to piling operations, continuous high-temperature areas detected by temperature sensors correspond to welding operations, and vertical movement trajectories in 3D point cloud data captured by laser scanning equipment correspond to hoisting operations. Equipment status mapping is based on the numerical range of equipment operating parameters. Parameters such as equipment power consumption, speed, and load rate are compared with preset status thresholds. When power consumption exceeds 80% of the rated power, it is marked as a high-load operating state; when the speed is lower than the minimum operating speed, it is marked as an idling state. Personnel behavior mapping is achieved by analyzing the activity trajectories and posture characteristics of personnel recorded by video acquisition equipment, identifying typical behavioral patterns such as walking, climbing, operating equipment, and staying still, and combining this with information from RFID tags worn by personnel to confirm their identity and work team. Material attribute mapping classifies and labels materials entering and leaving the construction site. It obtains material number, specifications, storage location and quantity information by scanning material tags with an RFID reader, and completes attribute labeling by matching this information with the construction material list.

[0028] A multi-dimensional construction state space is constructed based on the semantically mapped data. This space organizes the data using a three-dimensional tensor structure. The first dimension corresponds to the time series, the second dimension corresponds to the spatial grid of the construction site, and the third dimension corresponds to three state components: work progress, resource occupancy, and risk exposure. The work progress dimension is quantified by the deviation between the planned completion rate and the actual completion rate. Specifically, it is calculated by counting the difference between the number of completed work units and the planned number at each time point. For example, if the plan is to complete 20 work units and 18 are actually completed at a certain time, the work progress deviation is: Negative values ​​indicate delays. Resource utilization is quantified through equipment utilization and personnel presence. Equipment utilization is calculated as the ratio of actual working hours to planned working hours, and personnel presence is calculated as the ratio of the number of personnel on duty to the planned number of personnel. Risk exposure is quantified by the frequency of safety monitoring parameters exceeding preset thresholds. It involves statistically analyzing the number of times temperature exceeds limits, vibration exceeds limits, and personnel cross boundaries within a unit time window for each spatial grid, using the cumulative frequency of these events as the risk exposure indicator.

[0029] Atomic work units are identified in the multidimensional construction state space by detecting data clusters that simultaneously meet three criteria within a preset time window. The first criterion is continuous change in the work progress dimension. Specifically, this is determined by calculating the difference in work progress values ​​between adjacent time steps. When the differences across multiple consecutive time steps remain non-zero and have the same sign, the work progress dimension is considered to exhibit continuous change. The second criterion is stable resource usage. This is determined by calculating the coefficient of variation of the resource usage index within the time window. The coefficient of variation is defined as the ratio of the standard deviation to the mean. When the coefficient of variation is less than a preset threshold, such as 0.15, the resource usage dimension is considered to be in a stable state. The third criterion is that the risk exposure dimension has not triggered an interruption condition. An interruption threshold is set for the risk exposure index. When the maximum value of this index within the time window does not exceed the interruption threshold, this condition is considered met. The multidimensional construction state space is scanned grid-by-grid to identify continuous spatiotemporal regions that simultaneously meet the above three conditions, and these regions are marked as candidate data clusters.

[0030] The candidate data clusters are boundary-determined and resource requirements extracted to establish a complete description of the atomic work unit. The start time of the work boundary is determined as the first time node in the data cluster that meets the condition of continuous change in work progress, and the end time is determined as the last time node in the data cluster that simultaneously meets all three judgment conditions. The work space range is determined by calculating the circumscribed rectangle or circumscribed polygon of the spatial grid covered by the data cluster. For a spatially continuous set of grids, the coordinates of the boundary grids are extracted to form a closed region. The equipment type combination in the resource requirements is determined by analyzing all equipment types appearing in the spatiotemporal range of the data cluster and filtering out the equipment types that are continuously present during the work. The personnel skill combination is determined by analyzing the work group information of the personnel active in the spatiotemporal range of the data cluster, extracting the skill tags corresponding to each work group, and merging all appearing skill tags into the personnel skill combination of the atomic work unit. After the above processing, each atomic work unit is represented as a structured data object containing time boundaries, spatial boundaries, and resource requirements. These objects constitute the basic unit for subsequent agent creation and scheduling decisions. In a real-world construction scenario, a steel structure hoisting operation is identified as an atomic work unit, with a start time of 9:15 AM and an end time of 11:30 AM. The work space is the grid set on the east side of area 3 of the construction site. The equipment type combination includes tower cranes and auxiliary slings, and the personnel skill combination includes crane operators, signalmen, and welders. This structured description clearly defines the spatiotemporal boundaries and resource dependencies of the operation, supporting subsequent collaborative scheduling decisions.

[0031] Create a corresponding set of agents according to the number of atomic work units, and establish a hierarchical message passing architecture that supports bidirectional information exchange between agents, including: Each agent is equipped with a perception engine, an inference engine, and a control engine. The perception engine continuously monitors the execution status of the corresponding atomic work unit and the dynamic changes of surrounding work units. The inference engine generates local scheduling decisions based on the monitoring results and preset safety rules. The control engine converts the local scheduling decisions into control signals for the work equipment. The design employs a layered messaging architecture. The bottom layer is the state synchronization layer, where each agent broadcasts state messages containing execution progress, resource usage snapshots, and current risk values ​​to neighboring agents at fixed intervals. The middle layer is the intent sharing layer, where agents send intent messages containing planned operations, required resource types and quantities, and expected execution time windows to potential conflicting parties after generating local scheduling decisions. The top layer is the negotiation layer, where conflicting agents exchange concession schemes through negotiation messages.

[0032] When creating the agent set, the total number of atomic task units identified in the preceding steps is first counted. Assuming there are M identified atomic task units, M agent objects are instantiated, with each agent forming a one-to-one mapping with an atomic task unit. To ensure the stability of this mapping, a globally unique identifier is assigned to each agent. (i = 1, 2, ..., M), and record the spatial coordinates, job type code, and estimated duration of the atomic work unit corresponding to the agent. These static attributes are loaded into the local knowledge base during agent initialization as the basis for subsequent reasoning.

[0033] Each agent integrates three core functional engines, each responsible for processing tasks at different levels. The perception engine monitors the status of its corresponding atomic work unit by subscribing to sensor data streams. This includes changes in equipment coordinates reported by position sensors deployed within the work area, construction load data fed back by force sensors, and personnel activity information captured by vision sensors. After acquiring raw data using high-frequency sampling, the perception engine first filters to remove measurement noise, then converts physical quantities into standardized state descriptors. For example, it converts the three-dimensional coordinates of the equipment into offsets relative to the work reference point, and force sensor readings into the percentage of work progress completed. In addition to monitoring its own work unit, the perception engine continuously tracks the dynamic changes of other work units in its surrounding neighborhood. The neighborhood is defined as a spatial area with a radius of R_n meters centered on the current work unit; work units within this area are considered potential resource competitors or sources of risk transmission.

[0034] The inference engine receives state descriptors from the perception engine and combines them with a pre-loaded safety rule base to generate local scheduling decisions. The safety rule base is stored as a set of rules, each consisting of a trigger condition and a response action. The trigger condition is typically represented as a logical expression, involving parameters such as job progress thresholds, resource conflict determination conditions, and risk level boundaries. Internally, the inference engine implements a forward-reasoning-based rule matching algorithm, traversing the rule set to find all rules whose trigger conditions are met in the current state. When multiple rules are active simultaneously, the rule with the highest priority is selected and its corresponding response action is executed. Response actions may include specific operational instructions such as maintaining the current job rhythm, prematurely terminating a subtask, or requesting additional resource support. The inference engine outputs these abstract instructions as structured decision objects, containing fields such as operation type identifier, execution parameter list, and expected effective time.

[0035] The control engine acts as the interface layer between local scheduling decisions and physical equipment, responsible for translating the abstract decisions generated by the inference engine into control signals that the operating equipment can recognize. For different types of construction equipment, the control engine maintains corresponding drive adapters. For example, control signals for tower cranes include hook speed vector, braking torque setpoint, and safety limit switch status; control signals for concrete pump trucks include pumping pressure, delivery flow rate, and boom attitude angle. During the conversion process, the control engine needs to perform parameter boundary checks to ensure that the generated control signals do not exceed the physical performance limits of the equipment. Simultaneously, it considers the dynamic response characteristics of the equipment and introduces necessary buffer time to avoid mechanical shocks caused by abrupt parameter jumps.

[0036] The layered messaging architecture adopts a three-layer structure, with each layer handling different types of information exchange tasks. The bottom layer, state synchronization, operates in a time-driven manner, with a fixed broadcast period. Each agent sends a state message packet to other agents in its neighborhood at the end of the cycle. The state message packet is encapsulated in a compact binary format and contains the agent identifier. timestamp Execution progress percentage Resource usage snapshot and current risk value The execution progress percentage is obtained by calculating the ratio of completed work to total work. The resource usage snapshot records the various resources locked by the agent at the current moment and their usage quantities, represented in key-value pair form: Current risk value The risk assessment function, which comprehensively considers factors such as the degree of hazard exposure of the work unit, the interference intensity of surrounding work units, and the resource constraint tension, is calculated. Upon receiving status messages broadcast by other agents in the neighborhood, the local agent stores these messages in the neighborhood status cache table and updates the latest status record of the corresponding agent. The cache table is equipped with a timeliness check mechanism; if the time limit exceeds [a certain threshold], [the message will be deleted]. Records that have not received an update will be marked as invalid.

[0037] The middle-layer intent-sharing layer operates in an event-driven manner. When the inference engine generates a new local scheduling decision, the agent needs to determine whether the decision involves resource contention or spatial conflict. This determination is based on comparing the resource requirement list in the decision with the current resource occupancy of neighboring agents, and comparing the job space planned in the decision with the job boundaries of neighboring agents. If a potential conflict is detected, the agent sends an intent message to all conflict stakeholders. The intent message contains the agent's identifier. Decision sequence number Planned Operation Description List of required resource types and quantities Expected execution time window The planned operation description uses standardized operation codes to represent specific job types. The required resource list details the type code and requested quantity of each resource. The expected execution time window defines the start and end times when the decision is expected to take effect, leaving room for coordination in subsequent negotiations. The agent receiving the intent message checks for conflicts between the message content and its own plan. The conflict detection algorithm calculates the intersection of resource requirements and the overlap of job time windows. If the intersection is not empty and the time windows overlap, a conflict is confirmed, and preparations are made to enter the negotiation process.

[0038] The top-level negotiation layer is specifically designed to handle conflict resolution. When an agent confirms a conflict with another agent, the parties exchange negotiation messages to reach a compromise. The structure of a negotiation message includes a negotiation round identifier, a set of proposed solutions, and evaluation metrics for each solution. The set of proposed solutions contains one or more candidate adjustments, each describing a specific modification to the original plan; for example, a solution might be to postpone the expected execution window. The time unit can be used to reduce the number of requests for a certain type of resource or to adjust the scope of workspace usage. The evaluation metric quantifies the impact of the proposed solution on the agent's own goals, represented by a scalar value; the smaller the value, the lower the cost of the concession. The negotiation process employs a multi-round iterative mechanism. The first round is initiated by the conflict detection party, which sends an initial proposed solution to the other party. The receiving party evaluates the feasibility of each solution. If a solution can resolve the conflict without seriously harming its own interests, it replies with an acceptance message along with a selection flag for that solution. If all solutions are unacceptable, it generates a reverse proposed solution and sends it to the other party, entering the next round of negotiation. To prevent the negotiation from falling into an infinite loop, a maximum negotiation round limit is set. When the negotiation rounds exceed the limit and no agreement is reached, the superior arbitration mechanism is triggered, and the global coordination module intervenes to forcibly allocate resources or adjust the work order.

[0039] Information exchange between agents via a message-passing architecture ensures that each agent has a consistent understanding of the overall situation at the construction site. The state synchronization layer provides real-time state snapshots, enabling agents to perceive neighborhood dynamics. The intent-sharing layer proactively exposes potential conflict points, and the negotiation layer avoids local optima traps through a benefit-balancing mechanism. This three-layer architecture works collaboratively to form a coordinated scheduling plan that satisfies the execution needs of each work unit while adhering to global security constraints. In actual deployment, message passing is achieved through a wireless communication network built on-site. Agent nodes run on edge computing servers or mobile computing devices. Messages are encapsulated using lightweight protocols to ensure transmission efficiency, while message encryption and integrity verification mechanisms are configured to prevent malicious tampering.

[0040] A layered messaging architecture is used to achieve state synchronization and intent sharing among agents. Based on shared information, a multi-round negotiation mechanism is employed to resolve scheduling conflicts, including: A distributed consistency maintenance mechanism for the global state view is established. Each agent locally maintains a copy of the state of its neighboring agents. Vector clocks are used to mark the causal order of messages. Expiration and duplication of messages are detected by comparing vector clocks to ensure the causal consistency of state updates. Automatic detection rules for scheduling conflicts are defined, including resource conflict detection rules, timing conflict detection rules, and spatial conflict detection rules. The resource conflict detection rules determine whether two intentions request the same resource and whether their time windows overlap. The timing conflict detection rules determine whether the execution order of two intentions violates process constraints. The spatial conflict detection rules determine whether the operating spaces of two work units overlap and do not meet the safety distance requirement. Each agent constructs a local utility function based on the scheduling objective and constraints. During the negotiation process, the agent calculates the impact of different concession schemes on local utility, selects the concession scheme with the minimum utility loss to generate a negotiation message, and adds compensation conditions to the negotiation message to balance the utility loss. A negotiation termination condition is introduced: when the difference in utility between the concession proposals put forward by both parties in two consecutive rounds of negotiation is less than a preset threshold, or when the number of negotiation rounds reaches the upper limit, the negotiation is terminated and the final scheduling scheme is decided by the global optimization module.

[0041] For complex scenarios involving multiple work units operating simultaneously on a construction site, a global state view is needed to ensure information synchronization among all agents. Each agent allocates a dedicated cache in its local storage space, storing the latest state data of neighboring agents with potential interaction capabilities. The neighborhood is determined by analyzing resource dependencies and spatial proximity relationships between work units. Specific criteria include: overlapping planned execution time windows of two work units, minimum spatial distance between two work units less than 15 meters, or two work units sharing the same type of construction machinery resources. An agent corresponding to a work unit meeting any of these conditions is included in the neighborhood.

[0042] To address the inconsistency in state caused by the uncertain arrival order of messages in a distributed environment, a vector clock mechanism is used to mark the causal relationship of each message. Each agent maintains a... dimensional vector ,in The total number of agents, Indicates the first A vector clock for each agent. When an agent experiences a local event, it increments its corresponding vector component by 1; when an agent sends a message, it appends the current vector clock to the message; when an agent receives a message, it takes the maximum of the local and received values ​​for each component of the vector clock, and then increments its corresponding component by 1. The causal order of messages can be determined by comparing the two vector clocks: if the vector clocks... All components are less than or equal to If the corresponding components are such that at least one component is strictly less than a certain value, then the vector is considered to be... The corresponding event occurred Before.

[0043] Before processing a message, the receiving agent checks the vector clock carried in the message. If it finds that some components of the message's vector clock are less than the corresponding components of locally processed messages, it determines that the message is expired and discards it. If the vector clocks are exactly the same, it determines that the message is a duplicate and discards it as well. This mechanism ensures that the agent updates its state copy according to the true causal order of events, avoiding state rollback or erroneous overwriting caused by out-of-order messages.

[0044] Based on state synchronization, an automatic scheduling conflict detection mechanism is constructed. Resource conflict detection focuses on the mutual exclusion of physical resources; when an agent... Planned within the time frame Internal use of tower crane resources, intelligent agents Planned within the time frame If the same tower crane is used within the same time interval, the detection algorithm calculates the intersection of two time intervals; if the intersection is not empty, then... If a conflict is detected, a resource conflict is determined. This rule is used to detect resources with exclusive attributes, such as construction machinery, temporary power interfaces, and unloading platforms.

[0045] Temporal conflict detection focuses on the technological dependencies between work units. In the construction field, some operations have strict sequential requirements; for example, concrete pouring must be carried out after formwork erection, and exterior wall painting must be carried out after exterior wall plastering is completed and has reached a certain curing time. The system pre-establishes a process constraint diagram, where nodes represent work units, directed edges represent predecessor-successor relationships, and minimum time intervals are marked on the edges. During detection, the relationship between two work units in the constraint diagram is extracted; if the work units... It is a work unit The successor node, and the minimum interval of the constraint icon annotation is In the scheduling intent submitted by the agent start time and End time The interval between If so, it is determined that the timing constraint is violated.

[0046] Spatial conflict detection addresses the physical space occupancy of work units. Each work unit registers its three-dimensional envelope of operating space during initialization, using its center coordinates. Length, width and height and orientation angle Description. The detection algorithm first calculates the Euclidean distance between the center points of the two envelopes: If the distance is less than the sum of the maximum diagonal lengths of the two envelopes, a more precise intersection determination is performed. The separation axis theorem is used, projecting onto the three coordinate axes and the edge direction of the envelope. If the projections on all separation axes overlap, the two envelopes are determined to intersect. Furthermore, a safety distance parameter is introduced. Different values ​​are set according to the type of work. The safe distance between high-altitude operations and ground transportation routes is set at 3 meters, and the safe distance between hot work operations and flammable material storage areas is set at 10 meters. When the envelopes do not directly intersect but the center distance is less than the specified safe distance, it is also judged as a spatial conflict.

[0047] Upon detecting a conflict, a multi-round negotiation mechanism is initiated. Each agent constructs a local utility function based on the characteristics of the work units it manages. The utility function comprehensively considers three dimensions: work schedule deviation, resource utilization efficiency, and risk exposure level. Work schedule deviation measures the difference between the actual start time and the planned start time; the greater the deviation, the lower the utility. Resource utilization efficiency is reflected by the proportion of idle time; the longer the idle time, the lower the utility. Risk exposure level is calculated based on the hazard density and personnel exposure duration in the work unit's environment; the higher the risk, the lower the utility. The local utility function is expressed as a weighted sum of the three terms, with the weighting coefficients dynamically adjusted according to the overall project objectives. The weight of schedule deviation is increased during periods of tight deadlines, and the weight of risk exposure level is increased during safety inspections.

[0048] The negotiation process employs a proposal-response interactive model. The party detecting the conflict acts as the initiator, calculating several concession options. Each option corresponds to a different scheduling adjustment strategy, including delaying the start time of the task, shortening the task duration, changing the resources used, and adjusting the task's spatial location. The initiator calculates the adjusted local utility value for each option and selects the option with the minimum utility loss as the initial proposal. To balance the utility loss from concessions, compensation conditions are attached to the proposal. Compensation includes priority access to resources for a subsequent time period, priority improvement in the next scheduling round, and additional emergency resource quotas. Upon receiving the proposal, the receiving party assesses its impact on its own utility. If accepting the proposal results in an acceptable utility loss, it sends an acceptance message to complete the negotiation; otherwise, it generates a counter-proposal, outlining its willingness to make concessions and its expected further concessions from the other party.

[0049] The negotiation process may involve multiple iterations. To avoid infinite loops, negotiation termination conditions are set. The first termination condition is utility convergence, meaning that in two consecutive rounds of negotiation, the change in utility corresponding to the concessions proposed by both parties is less than a preset threshold, typically set at 2% of the initial utility value. The second termination condition is a round limit to prevent prolonged negotiation from affecting real-time performance. The round limit is set based on the urgency of the task, generally not exceeding 5 rounds. When a termination condition is triggered but no consensus is reached, the conflict is submitted to the global optimization module. The global optimization module maintains a complete state model of the entire construction site and uses a constraint satisfaction algorithm to solve for a scheduling scheme that maximizes global utility. This module constructs an optimization problem encompassing all work units, resources, and constraints. The objective function is the weighted sum of the local utilities of all agents, and the constraints include resource mutual exclusion constraints, temporal dependency constraints, and spatial safety constraints. A branch-and-bound algorithm or a genetic algorithm is used to solve this optimization problem, and the solution is distributed as a mandatory decision to the relevant agents for execution. The adjudication scheme records the time of the conflict, the agents involved, the type of conflict, and the final solution. These records are used to optimize subsequent negotiation strategies. By analyzing historical conflict patterns through machine learning methods, the parameters of the concession strategy and the rules for setting compensation conditions are adjusted to gradually improve negotiation efficiency and success rate.

[0050] Each agent constructs a local utility function based on the scheduling objective and constraints. During the negotiation process, the agent calculates the impact of different concession schemes on local utility, selects the concession scheme with the minimum utility loss, and generates a negotiation message, including: A multi-objective local utility function is constructed, which includes three sub-objectives: task completion time, resource utilization rate, and safety margin. These sub-objectives are aggregated into a single utility value through a weighted summation method, and the weights of each sub-objective are dynamically adjusted according to the priority of the current construction stage. In response to the detected conflict, the agent enumerates a set of feasible concession schemes, which include three basic concession actions and their combinations: delaying execution time, reducing resource requests, and shortening operation time. For each option in the set of concession options, the agent calculates the local scheduling plan after executing the option and evaluates the value of the local utility function under the new scheduling plan. By comparing the utility values ​​before and after the concession, the utility loss of each option is obtained. From the set of concession options, candidate options with utility loss less than the tolerable loss threshold are selected. The candidate options are then sorted in ascending order of utility loss, and the option with the smallest utility loss is selected as the proposal for this round of negotiation. Based on the proposed solution, compensation conditions are generated. These conditions include requesting the other party to prioritize resource allocation in subsequent time periods, requesting the other party to assume partial safety monitoring responsibility, and requesting the other party to adjust the order of subsequent operations to reduce waiting time. The specific content of the compensation conditions is determined according to the magnitude and type of utility loss. The proposed solution and its compensation conditions are encapsulated into a negotiation message and sent to the opposing intelligent agent. At the same time, the proposed solution and utility loss value of this round of negotiation are recorded to determine the convergence status in subsequent negotiation rounds.

[0051] In actual operation at the construction site, when resource competition or time conflicts are detected between different work units, each agent needs to quantitatively assess its own concession costs and choose the negotiation strategy that has the least impact on the overall scheduling. To achieve this goal, a utility evaluation system that reflects the core demands of each agent corresponding to a work unit needs to be constructed. This utility evaluation system adopts the form of a multi-objective local utility function, integrating the three key dimensions of work completion time, resource utilization rate, and safety margin into a single, quantifiable and comparable value.

[0052] The calculation of the task completion time is based on the planned start and expected end times of the task units managed by the agent. By comparing the actual execution progress with the contractually agreed milestones, the contribution of the task unit to the overall project duration is determined. When a task unit is on the critical path, its completion time value has a direct impact on the overall project duration, and in this case, the weight of this item needs to be set at a higher level. The resource utilization rate focuses on the efficiency of the agent's use of resources such as manpower, machinery, and materials during task execution. By calculating the ratio of the actual amount of resources invested per unit time to the theoretically optimal resource allocation, it is determined whether the current scheduling scheme has problems with resource idleness or over-utilization. The safety margin measures the distance from the risk threshold during task execution. The larger the distance, the more sufficient the safety reserve, and the higher the corresponding utility value.

[0053] When aggregating the three sub-objectives using a weighted summation method, the determination of the weighting coefficients needs to consider the characteristics of the construction phase. In the initial stage of the project, since the on-site infrastructure has not yet been fully established, the weight of the safety margin item is usually set to above 0.5 to ensure that each work unit does not blindly pursue progress before safety protection measures are in place. After entering the main construction phase, if the overall project schedule lags behind the plan, the weight of the work completion time item can be dynamically increased to the range of 0.4 to 0.5. At this time, the weights of the resource utilization rate item and the safety margin item are reduced accordingly, but still need to be kept within an acceptable range. When there is a severe weather warning or a major safety hazard investigation at the construction site, the weight of the safety margin item is immediately adjusted to the highest priority, and the weights of the other two items are reduced to the level of maintaining basic operational continuity.

[0054] Upon detecting a scheduling conflict, the agent first needs to identify the specific type of conflict. Resource contention conflicts occur when multiple work units simultaneously request the same type of resource, and the resource supply is insufficient. In this case, the focus of the concession plan design is to adjust the time window for resource requests or reduce the number of requests per request. Time-related conflicts occur when the scheduled execution time of a work unit overlaps with that of other work units, and the two cannot work concurrently. Concession plans for this type of conflict focus on modifying the start and end times of the work execution. Spatial conflicts occur when the work areas of different work units physically overlap, and safety regulations prohibit simultaneous construction. The corresponding concession plan requires re-dividing the work areas or adjusting the work sequence.

[0055] For a clearly defined conflict type, the agent enumerates a set of feasible concession schemes. The concession action of delaying execution time is achieved by postponing the planned start time of the work unit by a specified duration. Candidate values ​​for the delay duration are typically set as discrete levels such as 15 minutes, 30 minutes, 1 hour, and 2 hours. The agent selects the appropriate delay level based on the severity of the conflict. The concession action of reducing resource requests is suitable for resource-competitive conflicts. The agent calculates the minimum acceptable resource input while ensuring work quality. For example, the original plan to request three tower cranes to work simultaneously can be adjusted to two tower cranes working in batches. Although the output per unit time decreases, resource conflicts are avoided. The concession action of shortening operation time requires the agent to compress the execution time of non-critical processes. This is achieved by increasing manpower or optimizing work processes to accelerate the completion of unit tasks, freeing up time resources for other conflicting parties.

[0056] In actual negotiation, a single type of concession may not be sufficient to resolve the conflict. In such cases, it is necessary to combine multiple concessions to form a composite solution. For example, combining delayed execution time with reduced resource requests can both reduce the intensity of resource competition through time-shifting and reserve resource space for other work units by reducing the total resource demand. The generation of the composite solution follows the principle of prioritizing safety; the superposition of any concessions must not cause the safety margin to fall below the warning threshold.

[0057] For each concession option in the set, the agent needs to simulate the scheduling state after executing that option. During the simulation, the agent modifies its own parameters such as job start time, resource allocation list, and operation duration according to the concession option, and recalculates the multi-objective local utility function value under the modified scheduling plan. The difference between the utility value after the concession and the baseline utility value before the concession is the utility loss of that option. A positive utility loss indicates that the concession behavior reduces the agent's local gains, and a larger loss value means a higher concession cost.

[0058] After calculating the utility loss for all candidate solutions, the agent sets a tolerable loss threshold as a selection criterion. This threshold is determined by comprehensively considering the urgency of the task unit and the agent's historical negotiation record. For agents with ample timelines and who have received compensation in previous negotiations, a higher tolerable loss threshold is set, indicating that they are willing to accept larger concessions in the current negotiation round. Conversely, if the task unit is on the critical path and has already made multiple concessions in previous negotiations, the tolerable loss threshold is lowered to protect the agent's core interests. The selected candidate solutions are ranked in ascending order of utility loss, and the solution with the lowest utility loss is selected as the proposed solution for this round of negotiation.

[0059] After selecting a proposed solution, the agent generates compensation conditions based on the magnitude and source of the utility loss. When the utility loss primarily stems from decreased resource utilization, the compensation condition requests the other agent to prioritize resource allocation in subsequent time periods. Specifically, this means that the agent's resource requests will have priority approval rights for several consecutive scheduling cycles after the conflict is resolved. When the utility loss originates from reduced safety margin, the compensation condition requires the other agent to assume partial safety monitoring responsibility. For example, during the execution of the agent's work unit, the other agent must deploy additional safety observers or share its safety sensor data to assist in monitoring potential risks. When the utility loss manifests as a delay in work completion time, the compensation condition requests the other agent to adjust the sequence of subsequent work. After the agent's work resumes normal progress, the other agent proactively reduces the duration of its non-critical processes or postpones the start time of secondary tasks, creating conditions for the agent to catch up on the schedule.

[0060] The quantification level of the compensation condition is positively correlated with the value of the utility loss. When the utility loss is within 50% of the tolerable threshold, the compensation condition is set to a single type of mild compensation, such as requesting resource priority for only one scheduling cycle. When the utility loss exceeds 80% of the tolerable threshold, the compensation condition is upgraded to a multi-type combined compensation, covering multiple dimensions such as resource priority, shared safety responsibility, and job order adjustment, to ensure that the loss of the yielding agent can be fully compensated in subsequent execution.

[0061] The generated proposal and compensation conditions are encapsulated as a structured negotiation message. The message explicitly specifies the parameters of the concession actions, the expected utility loss value, the type of compensation condition, and its duration. The negotiation message is sent to the opposing agent via a layered messaging architecture, while simultaneously recording complete information for this round of negotiation locally, including the proposal number, the utility loss value, and the sending timestamp. These records are used to determine whether the negotiation has converged. When the change in utility loss for each agent is below a set threshold across multiple consecutive rounds of negotiation, the negotiation is considered to have reached a stable state, and all parties accept the current proposal and execute the corresponding concession actions and compensation commitments.

[0062] During the execution of the global collaborative scheduling plan, a rolling time-domain risk prediction model is used to predict the risk evolution trend of each work unit within the future time window, including: A rolling time-domain risk prediction model is constructed. This model takes the current time as the starting point and sets a fixed-length prediction time-domain window. Within the prediction time-domain window, risk simulation is performed in discrete time steps. The input of the model for each time step includes the planned execution actions of the work unit, the resource allocation status, the predicted values ​​of environmental parameters, and the dynamic impact of adjacent work units. A risk propagation graph is established, with work units as nodes and directed edges representing the physical proximity, resource sharing, and process dependence relationships between work units. The weight of each edge represents the intensity coefficient of risk propagation between nodes. Dynamic risk simulation is performed on the risk propagation graph. The risk level of each node at the initial moment is taken from the current real-time monitoring value. The risk level at subsequent time steps is calculated based on the node's own risk growth function and the risk increment passed from adjacent nodes through directed edges. The risk growth function takes into account the inherent risk rate of the job type, the cumulative effect of the job duration, and the modulating effect of environmental factors.

[0063] The rolling time-domain risk prediction model uses a dynamic window mechanism to continuously predict future construction risks. The length of the prediction time-domain window is determined according to the characteristic cycle of different types of operations. For continuous operations such as concrete pouring, the prediction time-domain is set to 4 to 6 hours, while for instantaneous operations such as equipment hoisting, the prediction time-domain is compressed to 30 to 90 minutes.

[0064] The window is set to the current time. Starting from this point, extending into the future to every moment. ,in This represents the total length of the prediction time domain. Within this time domain, predictions are made at fixed time intervals. Discretization is performed, dividing the continuous time axis into several independent derivation steps. Time interval The value of needs to balance computational efficiency and prediction accuracy, and is usually set to 5 to 15 minutes, so that there are 16 to 72 discrete time steps in a single prediction time domain.

[0065] Risk projection for each discrete time step relies on the fusion of multi-source data. The planned execution actions of each work unit are extracted from the global collaborative scheduling plan, including whether a new process is initiated within that time step, the current process's progress percentage, the planned number of personnel, and the type of equipment. Resource allocation status describes the specific configuration of temporary facilities, energy supply, and safety equipment used by the work unit at that time step. For example, a tower crane work unit might use distribution box #3, two sets of safety nets, and one emergency lighting device at a certain time step. Environmental parameter predictions are obtained through access to a meteorological service interface, including temperature, humidity, wind speed, rainfall probability, and visibility data for future time steps. These parameters significantly modulate the risk level of high-altitude operations and outdoor welding operations. The dynamic impact of adjacent work units is reflected by real-time monitoring of their execution deviations. When the actual progress of a work unit lags behind the planned progress by more than 20%, the waiting risk it generates for downstream dependent work units is presented in the model input as a delay coefficient.

[0066] The risk propagation map is constructed based on the spatial topology and technological logic of the construction site. Each atomic work unit is abstracted as a node in the map, with node attributes including work type label, spatial coordinate range, and baseline risk level. Three types of relationships exist between work units: physical proximity, determined by calculating the shortest distance between work area boundaries; when the safety protection zones of two work units overlap or the distance is less than the specified safety isolation distance, a directed edge is established between the corresponding nodes, pointing towards the more vulnerable work unit; resource sharing, establishing bidirectional edges for work units simultaneously using the same large equipment or sharing the same temporary passage, representing the bidirectional risk transmission that may result from resource conflicts; and process dependency, established based on the process sequence constraints in the construction organization design, with directed edges pointing from preceding work units to subsequent work units, reflecting the risk path of process delays or quality defects propagating downstream.

[0067] The weights of directed edges quantify the strength of risk propagation between nodes. For physical proximity relationships, the weights... Distance between the two work units The propagation weight of the risk of falling objects from heights is calculated by considering both the inverse relationship and the type of risk source. ,in Indicates elevation difference, This represents the propagation coefficient of the risk of falling objects. For resource-sharing relationships, the weight depends on the scarcity of resources. When the load rate of a tower crane exceeds 75%, the edge weight between two work units sharing that tower crane increases to 1.8 times the baseline value. For process dependencies, the weight reflects the sensitivity of subsequent processes to the output quality of preceding processes. The edge weight between rebar tying and concrete pouring is set to 0.9, while the weight between formwork installation and concrete pouring is increased to 1.2 because formwork defects have a more direct impact on the risk of pouring operations.

[0068] Dynamic risk simulation is executed progressively step-by-step on the risk propagation map. Initial moment. The risk level of each node is directly assigned to the real-time risk indicator monitored by the sensor network. For example, the initial risk value of a welding operation unit is set as the weighted sum of the current monitored concentration of harmful gases exceeding the standard by a factor of 1 and the number of workers not wearing protective masks. Starting from the second time step, the nodes... At time step risk level It consists of two parts: the node's own endogenous risk growth and the external risk growth from neighboring nodes. The endogenous risk growth is calculated using a risk growth function, which has the following form: ,in Assignment type The inherent risk rate of high-altitude demolition operations The value is 0.12, while the value for ordinary plastering operations is... The value is only 0.03; The cumulative effect function representing the duration of the task is in increasing form: ,in This represents the number of hours the work unit has been continuously operating, reflecting the increased risk due to accumulated fatigue. For the modulation function of environmental factors, when time step When the wind speed exceeds level 5 in the corresponding environmental parameters, The value is 1.6, 1.4 under rainy conditions, and 1.0 under normal weather conditions.

[0069] The increase in external risk comes from all the nodes that point to it. The risk contribution propagated through directed edges. For each node... Pointing to node The edge, its time step The amount of risk transmitted is: ,in For edge weights, For nodes At the risk level of the previous time step, This is the risk propagation attenuation factor, with a default value of 0.7, representing the natural attenuation of risk during propagation. The sum of the risk propagated by all incoming edges yields the node... Total exogenous risk increment: ,in Represents all pointer nodes The set of neighboring nodes. Nodes At time step The final risk level is calculated as follows This recursive relationship starts from the initial time. Initially, the process proceeds step by step forward until the end of the prediction time domain. .

[0070] Throughout the prediction time domain, the model continuously tracks the risk evolution trajectory of each node. When a node reaches a certain time step in the future... Predicted risk value The first time the preset warning threshold was exceeded At that time, the system records the node's identifier, the moment the threshold is exceeded, and the predicted risk peak. The warning threshold is set according to the job type and on-site safety level: the threshold for Level 1 risk jobs is set at 0.65, the threshold for Level 2 risk jobs is increased to 0.80, and the threshold for Level 3 routine jobs is relaxed to 0.90. The prediction results are output as a visual curve, with the horizontal axis representing the future time step sequence and the vertical axis representing the risk level of each job unit. Different job units are distinguished by different colored curves. When a curve touches the warning threshold level, the corresponding agent for that job unit immediately receives a risk warning signal and initiates the subsequent emergency response mechanism.

[0071] The core feature of the rolling temporal mechanism lies in the continuous sliding update of the prediction window. After each execution cycle, the system changes the prediction starting point from... Move forward Based on the new real-time monitoring data, the risk levels of each node are reinitialized, and the complete dynamic simulation process is executed again. This rolling update strategy enables the model to dynamically absorb the latest changes in the field conditions and promptly correct deviations in early predictions. When the actual monitored risk level of a work unit deviates from the previous prediction value by more than 15%, the model automatically adjusts the inherent risk rate parameter corresponding to that work unit. This enables adaptive calibration of the prediction model, ensuring that the accuracy of subsequent predictions remains within an acceptable range.

[0072] When a risk level is predicted to exceed the warning threshold, an emergency response mechanism involving multiple agents is triggered, including: The risk level time series of each work unit obtained from the simulation are compared with the warning line to identify the work unit that first exceeds the warning line and the time of the exceedance. The work unit that exceeds the warning line is marked as a high-risk work unit. The associated work units that contribute to the risk of the high-risk work unit are identified by reverse tracing through the risk propagation map. An emergency trigger message is sent to multiple agents responsible for high-risk work units and their associated work units. The emergency trigger message includes the predicted breakthrough time, risk type, risk level, and correlation diagram. The emergency response mechanism is activated, and multiple agents participating in the emergency response simultaneously execute three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution results of each action are then summarized to form an emergency response report, which is fed back to the inference engine of each agent.

[0073] When comparing the risk level time series of each work unit obtained from the simulation with the early warning line, the predicted risk level value of each work unit within the future time window is first extracted from the output of the rolling time-domain risk prediction model. The length of the time window is dynamically set according to the work rhythm of the construction site. For units with short work cycles, such as earthwork excavation and steel structure hoisting, the time window is set to 15 to 30 minutes, while for units with longer work cycles, such as concrete pouring and underground pipeline laying, the time window is extended to 1 to 2 hours. The predicted risk level values ​​are organized in the form of a time series, with each time point corresponding to a specific risk value, which comprehensively reflects the degree of safety threat faced by the work unit at that moment. The early warning line is set based on historical accident data and on-site safety regulations. Different risk types correspond to different early warning thresholds. For example, the early warning line for the risk of falling from height is set with a risk index of 0.65, the early warning line for the risk of electric shock is set with a risk index of 0.72, and the early warning line for the risk of being struck by an object is set with a risk index of 0.58. Each predicted value in the time series is compared with the early warning line of the corresponding risk type one by one, and the existence of breach behavior is identified by judging the magnitude of the value.

[0074] A time-series scanning algorithm is used to identify the work unit that first exceeds the warning line and the time of the exceedance. Starting from the beginning of the time window, the risk level prediction value of each work unit is checked in chronological order. When a work unit exceeds the warning line at time t... k Risk prediction value R k First satisfaction of R k > R threshold At that time, record the identifier of the work unit and the breakthrough time t. k And the specific risk value R kTo avoid false triggers caused by frequent fluctuations in risk values ​​near the warning line, a time persistence judgment mechanism is introduced, requiring the risk value to exceed the warning line for three consecutive predicted times before a true breach is confirmed. Work units meeting the breach conditions are marked as high-risk work units, their safety status labels are updated in the system database, and a warning record containing the breach time, risk type, and current risk value is generated.

[0075] When identifying associated work units that contribute to high-risk work units through reverse tracing of the risk propagation graph, a pre-constructed directed risk propagation graph is used for path search. The risk propagation graph uses work units as nodes and risk propagation relationships as directed edges, with edge weights representing the intensity of risk propagation. Starting from the high-risk work unit node, a breadth-first search is performed along the reverse direction of the directed edges to identify all source nodes that can reach the high-risk node. The search depth is limited to 3 layers to avoid a surge in computational complexity due to an excessively large tracing range. For each source node found, its risk contribution to the high-risk work unit is calculated. The contribution is calculated by accumulating the path weights; specifically, the contribution value of a single path is obtained by multiplying the weights of all edges on the path. If multiple paths connect the same source node and the high-risk node, the maximum value of all path contribution values ​​is taken as the final contribution of the source node. A contribution threshold of 0.15 is set; only source nodes with a contribution exceeding this threshold are marked as associated work units to ensure the validity of the association relationships. The identification results of associated work units are stored in the form of a list. Each entry in the list contains the identifier of the associated work unit, the risk contribution value, and the connection path information.

[0076] When sending emergency trigger messages to multiple agents responsible for high-risk work units and their associated work units, the first step is to determine the set of agents that need to receive the message based on the mapping relationship between work units and agents. This mapping relationship is recorded in a distributed database, with each work unit corresponding to a unique agent identifier. Emergency trigger messages are encapsulated in a structured data format, with the message header containing a message type identifier, a timestamp, and a priority field. The priority is set to the highest level to ensure immediate message delivery.

[0077] The message body includes the predicted breakout moment. It uses Unix timestamp format to record data accurate to the second; the risk type field is represented by enumerated values, covering 13 common construction risk types such as falls from heights, electric shocks, falling objects, and mechanical injuries; the risk level field is divided into four levels, based on risk values. The size mapping is divided into four levels: low, medium, high, and very high. The mapping rule is as follows: Corresponding to higher levels, Corresponds to the highest level.

[0078] The relationship graph is embedded in the message in the form of an adjacency list, recording the directed connections between high-risk task units and all related task units, as well as their corresponding risk contributions. Messages are sent through an emergency channel in the hierarchical messaging architecture. The emergency channel uses an independent network connection and a priority queue to ensure that messages are delivered to all target agents within 100 milliseconds.

[0079] When the emergency response mechanism is activated, multiple intelligent agents participating in the emergency response immediately enter emergency handling mode upon receiving the emergency trigger message. Freezing the current operation of a high-risk work unit is achieved by sending a stop command to the associated machinery and personnel. The stop command includes three key parameters: equipment identifier, stop type, and execution time. Stop types are divided into immediate stop and gradual stop. For operations with high inertia, such as hoisting and excavation, the gradual stop mode is selected, allowing the equipment to smoothly decelerate to a stop within 5 to 10 seconds. For operations that can be interrupted immediately, such as welding and cutting, the immediate stop mode is selected, requiring the energy supply to be cut off within 1 second. The intelligent agents send the command to the equipment controller via the IoT control interface, and simultaneously push audio and vibration alerts to the personnel's mobile terminals to ensure timely awareness of the stop command. After the freeze operation is completed, the intelligent agents record the current status parameters of the equipment, including equipment location, operating parameters, and load status, providing a status baseline for subsequent resumption of operations.

[0080] When dynamically adjusting the execution order of related work units, the agent first obtains the current scheduling plan of the related work units. The scheduling plan is organized in the form of a task queue, and each task in the queue contains four elements: start time, duration, prerequisites, and resource requirements. Based on the contribution information in the risk propagation graph, the related work units are prioritized and reordered. Work units with higher contributions are postponed or temporarily removed from the scheduling queue, while work units with lower contributions but less impact on the overall progress are executed earlier to fill time gaps. The adjustment process must satisfy the temporal dependency constraints between work units. Dependency checks ensure that work units executed earlier do not depend on postponed work units, avoiding logical conflicts. The adjusted execution order is published in the form of an updated task queue. The new start time of each task in the queue is calculated using a time offset. The offset is dynamically determined based on the expected freeze duration of high-risk work units and the resource release status of related work units, with a typical offset range between 10 minutes and 1 hour.

[0081] When reallocating emergency resources, the agent retrieves a list of available emergency resources from the global resource pool. Emergency resources include physical resources such as backup safety equipment, emergency lighting facilities, medical rescue equipment, and temporary support structures, as well as human resources such as professional safety officers and emergency response personnel. Resource allocation prioritizes the safety reinforcement needs of high-risk work units, such as adding safety nets and fall arrestors to high-altitude work units, and adding leakage current devices and insulation protective equipment to electrical work units. The allocation algorithm matches resources based on their reachability and the urgency of the work unit. Reachability is calculated by dividing the distance between the current location of the resource and the location of the work unit by the transportation speed. Urgency is determined by both the risk level and the proximity of the breach point. The matching result is generated in the form of a resource scheduling instruction, which specifies the resource type, quantity, target work unit, and required arrival time. The agent triggers the actual transfer of resources through the logistics scheduling system and tracks the transportation status of resources in real time to ensure that resources arrive within the predetermined time.

[0082] When summarizing the execution results of each action to form an emergency response report, each agent encapsulates the execution status of its responsible action into a structured data record. The execution result of a freeze operation includes the freeze completion time, equipment stop status confirmation flag, and anomaly description; the result of execution order adjustment includes a comparison of task queues before and after adjustment, a list of affected work units, and estimated delay time; the result of resource allocation includes a list of allocated resources, resource arrival time, and statistics of remaining available resources. Each agent uploads its execution result records to the central coordination node through a hierarchical messaging architecture. The central coordination node aggregates the records according to the agent identifier and timestamp to generate a complete emergency response report. The report is organized in a hierarchical structure. The top layer is an overall overview of the emergency event, including the trigger time, the number of work units involved, and the emergency response duration; the second layer is an execution summary of each parallel action, statistically analyzing the success rate and average execution time of each action; the third layer is a detailed execution log, recording... A second aspect of this invention provides a multi-agent collaborative construction safety scheduling and risk management system, comprising: The state space unit is used to perform spatiotemporal alignment and semantic annotation on real-time data collected by the sensor network deployed at the construction site, forming a multi-dimensional construction state space that includes the dimensions of work progress, resource occupancy, and risk exposure. In the multi-dimensional construction state space, atomic work units with clear work boundaries and resource requirements are identified. The collaborative scheduling unit is used to create a corresponding set of intelligent agents according to the number of atomic job units, establish a hierarchical message passing architecture that supports bidirectional information exchange between intelligent agents, realize state synchronization and intent sharing of each intelligent agent through the hierarchical message passing architecture, and resolve scheduling conflicts by adopting a multi-round negotiation mechanism based on shared information. The multi-round negotiation mechanism balances the local interests of each intelligent agent with the global security goal through concession strategy and compensation strategy, and finally forms a conflict-free global collaborative scheduling plan. The risk response unit is used to predict the risk evolution trend of each work unit within a future time window during the execution of the global collaborative scheduling plan using a rolling time-domain risk prediction model. When the predicted risk level exceeds the warning line, it triggers an emergency response mechanism jointly participated in by multiple agents. The emergency response mechanism achieves rapid risk suppression through three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution result of the emergency response mechanism is fed back to the inference engine of each agent to update the safety parameters for subsequent scheduling decisions.

[0083] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0084] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0085] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0086] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for collaborative scheduling and risk management of construction safety involving multiple agents, characterized in that... ,include: Spatiotemporal alignment and semantic annotation are performed on real-time data collected by the sensor network deployed at the construction site to form a multi-dimensional construction state space that includes work progress, resource occupancy, and risk exposure dimensions. Atomic work units with clear work boundaries and resource requirements are identified in the multi-dimensional construction state space. A corresponding set of intelligent agents is created according to the number of atomic work units. A hierarchical message passing architecture that supports bidirectional information exchange between intelligent agents is established. The hierarchical message passing architecture enables state synchronization and intent sharing among intelligent agents. Based on the shared information, a multi-round negotiation mechanism is used to resolve scheduling conflicts. The multi-round negotiation mechanism balances the local interests of each intelligent agent with the global security goal through concession and compensation strategies, ultimately forming a conflict-free global collaborative scheduling plan. During the execution of the global collaborative scheduling plan, a rolling time-domain risk prediction model is used to predict the risk evolution trend of each work unit within the future time window. When the predicted risk level exceeds the warning line, an emergency response mechanism jointly participated in by multiple agents is triggered. The emergency response mechanism achieves rapid risk suppression through three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution result of the emergency response mechanism is fed back to the inference engine of each agent to update the safety parameters for subsequent scheduling decisions.

2. The method according to claim 1, characterized in that, The real-time data collected by the sensor network deployed at the construction site is spatiotemporally aligned and semantically labeled to form a multi-dimensional construction state space containing dimensions of work progress, resource occupancy, and risk exposure. Atomic work units with clear work boundaries and resource requirements are identified within this multi-dimensional construction state space, including: The system acquires real-time data collected by a sensor network deployed at the construction site. The real-time data includes timestamps, spatial coordinates, and measurement values. The system performs spatiotemporal alignment processing on the real-time data to unify the data collected by different sensors to a preset time reference and spatial coordinate system. Real-time data that has completed spatiotemporal alignment is mapped to at least one semantic category among job type, equipment status, personnel behavior, and material attributes. A multi-dimensional construction state space is constructed based on the data that has completed the semantic mapping. The multi-dimensional construction state space includes job progress dimension, resource occupancy dimension, and risk exposure dimension. The job progress dimension is quantified by the deviation between the planned completion ratio and the actual completion ratio. The resource occupancy dimension is quantified by equipment utilization rate and personnel presence rate. The risk exposure dimension is quantified by the frequency of safety monitoring parameters exceeding preset thresholds. In the multidimensional construction state space, atomic work units are identified. By detecting data clusters in the multidimensional construction state space that simultaneously meet the following conditions within a preset time window: continuous change in the work progress dimension, stable resource occupation dimension, and no interruption condition triggered in the risk exposure dimension, the spatiotemporal region corresponding to the data cluster is determined as an atomic work unit with clear work boundaries and resource requirements. The work boundaries include the start time, end time, and work space range, and the resource requirements include equipment type combinations and personnel skill combinations.

3. The method according to claim 1, characterized in that, Create a corresponding set of agents according to the number of atomic work units, and establish a hierarchical message passing architecture that supports bidirectional information exchange between agents, including: Each agent is equipped with a perception engine, an inference engine, and a control engine. The perception engine continuously monitors the execution status of the corresponding atomic work unit and the dynamic changes of surrounding work units. The inference engine generates local scheduling decisions based on the monitoring results and preset safety rules. The control engine converts the local scheduling decisions into control signals for the work equipment. The design employs a layered messaging architecture. The bottom layer is the state synchronization layer, where each agent broadcasts state messages containing execution progress, resource usage snapshots, and current risk values ​​to neighboring agents at fixed intervals. The middle layer is the intent sharing layer, where agents send intent messages containing planned operations, required resource types and quantities, and expected execution time windows to potential conflicting parties after generating local scheduling decisions. The top layer is the negotiation layer, where conflicting agents exchange concession schemes through negotiation messages.

4. The method according to claim 1, characterized in that... A layered messaging architecture is used to achieve state synchronization and intent sharing among agents. Based on shared information, a multi-round negotiation mechanism is employed to resolve scheduling conflicts, including: A distributed consistency maintenance mechanism for the global state view is established. Each agent locally maintains a copy of the state of its neighboring agents. Vector clocks are used to mark the causal order of messages. Expiration and duplication of messages are detected by comparing vector clocks to ensure the causal consistency of state updates. Automatic detection rules for scheduling conflicts are defined, including resource conflict detection rules, timing conflict detection rules, and spatial conflict detection rules. The resource conflict detection rules determine whether two intentions request the same resource and whether their time windows overlap. The timing conflict detection rules determine whether the execution order of two intentions violates process constraints. The spatial conflict detection rules determine whether the operating spaces of two work units overlap and do not meet the safety distance requirement. Each agent constructs a local utility function based on the scheduling objective and constraints. During the negotiation process, the agent calculates the impact of different concession schemes on local utility, selects the concession scheme with the minimum utility loss to generate a negotiation message, and adds compensation conditions to the negotiation message to balance the utility loss. A negotiation termination condition is introduced: when the difference in utility between the concession proposals put forward by both parties in two consecutive rounds of negotiation is less than a preset threshold, or when the number of negotiation rounds reaches the upper limit, the negotiation is terminated and the final scheduling scheme is decided by the global optimization module.

5. The method according to claim 4, characterized in that... Each agent constructs a local utility function based on the scheduling objective and constraints. During the negotiation process, the agent calculates the impact of different concession schemes on local utility, selects the concession scheme with the minimum utility loss, and generates a negotiation message, including: A multi-objective local utility function is constructed, which includes three sub-objectives: task completion time, resource utilization rate, and safety margin. These sub-objectives are aggregated into a single utility value through a weighted summation method, and the weights of each sub-objective are dynamically adjusted according to the priority of the current construction stage. In response to the detected conflict, the agent enumerates a set of feasible concession schemes, which include three basic concession actions and their combinations: delaying execution time, reducing resource requests, and shortening operation time. For each option in the set of concession options, the agent calculates the local scheduling plan after executing the option and evaluates the value of the local utility function under the new scheduling plan. By comparing the utility values ​​before and after the concession, the utility loss of each option is obtained. From the set of concession options, candidate options with utility loss less than the tolerable loss threshold are selected. The candidate options are then sorted in ascending order of utility loss, and the option with the smallest utility loss is selected as the proposal for this round of negotiation. Based on the proposed solution, compensation conditions are generated. These conditions include requesting the other party to prioritize resource allocation in subsequent time periods, requesting the other party to assume partial safety monitoring responsibility, and requesting the other party to adjust the order of subsequent operations to reduce waiting time. The specific content of the compensation conditions is determined according to the magnitude and type of utility loss. The proposed solution and its compensation conditions are encapsulated into a negotiation message and sent to the opposing intelligent agent. At the same time, the proposed solution and utility loss value of this round of negotiation are recorded to determine the convergence status in subsequent negotiation rounds.

6. The method according to claim 1, characterized in that... During the execution of the global collaborative scheduling plan, a rolling time-domain risk prediction model is used to predict the risk evolution trend of each work unit within the future time window, including: A rolling time-domain risk prediction model is constructed. This model takes the current time as the starting point and sets a fixed-length prediction time-domain window. Within the prediction time-domain window, risk simulation is performed in discrete time steps. The input of the model for each time step includes the planned execution actions of the work unit, the resource allocation status, the predicted values ​​of environmental parameters, and the dynamic impact of adjacent work units. A risk propagation graph is established, with work units as nodes and directed edges representing the physical proximity, resource sharing, and process dependence relationships between work units. The weight of each edge represents the intensity coefficient of risk propagation between nodes. Dynamic risk simulation is performed on the risk propagation graph. The risk level of each node at the initial moment is taken from the current real-time monitoring value. The risk level at subsequent time steps is calculated based on the node's own risk growth function and the risk increment passed from adjacent nodes through directed edges. The risk growth function takes into account the inherent risk rate of the job type, the cumulative effect of the job duration, and the modulating effect of environmental factors.

7. The method according to claim 6, characterized in that, When a risk level is predicted to exceed the warning threshold, an emergency response mechanism involving multiple agents is triggered, including: The risk level time series of each work unit obtained from the simulation are compared with the warning line to identify the work unit that first exceeds the warning line and the time of the exceedance. The work unit that exceeds the warning line is marked as a high-risk work unit. The associated work units that contribute to the risk of the high-risk work unit are identified by reverse tracing through the risk propagation map. An emergency trigger message is sent to multiple agents responsible for high-risk work units and their associated work units. The emergency trigger message includes the predicted breakthrough time, risk type, risk level, and correlation diagram. The emergency response mechanism is activated, and multiple agents participating in the emergency response simultaneously execute three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution results of each action are then summarized to form an emergency response report, which is fed back to the inference engine of each agent.

8. A multi-agent collaborative construction safety scheduling and risk management system, used to implement the method as described in any one of claims 1-7, characterized in that, include: The state space unit is used to perform spatiotemporal alignment and semantic annotation on real-time data collected by the sensor network deployed at the construction site, forming a multi-dimensional construction state space that includes the dimensions of work progress, resource occupancy, and risk exposure. In the multi-dimensional construction state space, atomic work units with clear work boundaries and resource requirements are identified. The collaborative scheduling unit is used to create a corresponding set of intelligent agents according to the number of atomic job units, establish a hierarchical message passing architecture that supports bidirectional information exchange between intelligent agents, realize state synchronization and intent sharing of each intelligent agent through the hierarchical message passing architecture, and resolve scheduling conflicts by adopting a multi-round negotiation mechanism based on shared information. The multi-round negotiation mechanism balances the local interests of each intelligent agent with the global security goal through concession strategy and compensation strategy, and finally forms a conflict-free global collaborative scheduling plan. The risk response unit is used to predict the risk evolution trend of each work unit within a future time window during the execution of the global collaborative scheduling plan using a rolling time-domain risk prediction model. When the predicted risk level exceeds the warning line, it triggers an emergency response mechanism jointly participated in by multiple agents. The emergency response mechanism achieves rapid risk suppression through three parallel actions: freezing the current operation of high-risk work units, dynamically adjusting the execution order of related work units, and reallocating emergency resources. The execution result of the emergency response mechanism is fed back to the inference engine of each agent to update the safety parameters for subsequent scheduling decisions.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 7.

Images