Modular construction dynamic hoist scheduling decision system

Abstract

The application provides a kind of modular construction dynamic hoisting scheduling decision system, it is related to industrial automation technical field.The system includes environment perception layer, digital twin modeling layer, intelligent agent decision layer and scheduling output layer.Through real-time collection of three-dimensional point cloud, image and equipment state data by on-site unmanned aerial vehicle and camera, high-fidelity construction scene digital twin model is driven to be built;Further, hoisting, transportation and weather intelligent agent based on multi-agent reinforcement learning algorithm, collaborative decision and counter-training are carried out in digital twin environment, hoisting sequence, tower crane path and transportation plan are dynamically optimized;Finally, the optimal scheduling instruction is issued through message queue and the execution feedback is received.The application solves the problem of response lag and multi-dimensional decision fragmentation of traditional scheduling method in complex mountainous environment, realizes adaptive dynamic optimization of hoisting operation, significantly improves construction efficiency, resource utilization and operation safety.

Description

Technical Field

[0001] This invention relates to the field of industrial automation technology, specifically to a modular construction dynamic hoisting scheduling decision system. Background Technology

[0002] In modular construction in mountainous areas, dynamic hoisting scheduling is a core element affecting the overall efficiency and safety of the project. Currently, the field relies on pre-planning equipment paths and work sequences based on design drawings, with limited adjustments made during execution using local environmental perception data. This approach is still applicable in conventional environments with stable construction conditions and few changing factors.

[0003] However, when facing construction scenarios in mountainous areas with complex terrain and variable weather, there is a serious disconnect between the environmental model on which scheduling decisions are based and the real-time dynamics of the construction site. Due to the lack of an effective mechanism for continuously integrating and perceiving multi-dimensional factors such as terrain, equipment status, component location, and instantaneous weather to drive synchronous updates to the decision-making model, the system often makes judgments based on lagging information. The resulting scheduling plans are poorly adaptable to sudden changes, often leading to equipment idleness, path conflicts, or project delays. Secondly, because of the lack of real-time panoramic status, the three key decision dimensions of hoisting operations, material transportation, and weather impact are treated in isolation during actual scheduling. The inability to characterize and optimize the complex interrelationships among these three elements within a unified framework that reflects real dynamic interactions prevents the system from achieving integrated scheduling that achieves optimal resource synergy and proactive risk control in dynamic and complex mountainous environments. Summary of the Invention

[0004] To achieve the above objectives, this invention proposes a modular construction dynamic hoisting scheduling decision-making system, comprising: The environmental perception layer continuously collects 3D point cloud data, image data, equipment status data, and weather data through data acquisition equipment equipped with lidar and cameras deployed at the construction site. The digital twin modeling layer constructs and dynamically updates a digital twin model that includes a static geometric model imported from BIM files and a real-time data-driven solid model of the construction site. The intelligent agent decision-making layer comprises a hoisting agent, a transportation agent, and a weather agent. The weather agent predicts future weather conditions based on historical and real-time weather data and outputs its results as dynamic environmental constraints. The hoisting and transportation agents make decisions in a joint state space composed of a digital twin model and dynamic environmental constraints. Specifically, the hoisting agent's decision output changes the tower crane's task status and space occupancy in the digital twin model in real time, constituting dynamic obstacles and time windows that the transportation agent must adhere to when planning routes and times. The transportation agent's decision output is used to plan the transport vehicle's entry schedule and update the expected arrival sequence and positioning status of each module to be hoisted in the digital twin model, thus providing the necessary prerequisites and resource constraints for the hoisting agent to generate hoisting tasks. The weather agent's prediction results dynamically modify the safe operation boundary and equipment performance parameters in the joint state space, coupling them in real time to the action selection and reward calculation of the hoisting and transportation agents. Based on the interaction between agents and combined with dynamic environmental constraints, the intelligent decision-making layer generates a collaborative scheduling strategy for future time periods in the digital twin model. The scheduling output layer generates a scheduling plan for future time periods based on the collaborative scheduling strategy of the intelligent agent decision layer, and sends it to the field execution system through the message queue telemetry transmission protocol.

[0005] As a further technical solution, the digital twin modeling layer, based on the standardized real-time data stream provided by the environment perception layer, and combined with static geometric information such as building structure and terrain imported from the project BIM model, dynamically constructs and continuously updates a high-precision three-dimensional virtual construction scene. The digital twin model includes static and dynamic parts; the static part is directly derived from the BIM file, forming an unchanging construction environment background; the dynamic part represents all movable or variable entities on the construction site, mainly including tower cranes, modular transport vehicles, and building modules to be hoisted and already in place. The system maintains an independent state structure for each dynamic entity instance in the model. This state structure records the entity type, unique identifier, current three-dimensional position in the global coordinate system, velocity vector, orientation angle, and other basic attributes, and also includes a label representing its task status. The system identifies entities such as idle, in transit, awaiting hoisting, in hoisting, and completed, along with an environmental feature vector closely associated with the entity's current state. The content of the environmental feature vector is dynamically filled based on real-time sensing data. For example, for a tower crane, its feature vector includes the current position of its hook, the current load weight, the boom angle, and the shortest distance calculated based on point cloud data to other dynamic entities within a certain radius centered on its current position, such as another tower crane, a transport vehicle, or static obstacles like installed modules or mountains. The system uses a dynamic update engine to map and update the latest state sequence, which includes timestamps, device IDs, coordinates, and status identifiers, obtained from the environmental sensing layer, to the state structure of the corresponding entity in real time. This drives the position, posture, and animation state of the corresponding 3D model to change synchronously, thereby achieving virtual-real synchronization between the virtual model and the physical entity.

[0006] As a further technical solution, the intelligent agent decision layer comprises three closely cooperating agents: a hoisting agent, a transportation agent, and a weather agent. These three agents are placed in a dynamic joint state space defined by a digital twin model, interacting and learning collaboratively through a multi-agent reinforcement learning framework to jointly generate a globally optimized scheduling strategy. The weather agent continuously receives real-time data such as wind speed, wind direction, temperature, and humidity uploaded by the environmental perception layer, and aligns it with the historical weather database by timestamp to form a continuous time series. This time series is input into a time series prediction model built on a Long Short-Term Memory (LSTM) network. The LSTM model learns the periodic and trend features in historical weather data and outputs a predicted sequence of various weather parameters for a future period. Simultaneously, the system calculates the mean absolute error between the predicted sequence and subsequent actual monitoring values. And through the reward function To evaluate prediction accuracy, MAE tThe mean absolute error at deadline t is used; this reward signal is backpropagated to optimize the parameters of the LSTM model, making its predictions increasingly accurate; the output of the weather agent is transformed into dynamic environmental constraints that can directly affect the decision-making of other agents; specifically, the system pre-sets a comparison table of wind load and safe operating radius based on engineering standards and equipment manuals, as well as an empirical model of temperature and humidity and equipment performance degradation; based on the predicted wind speed, the safe swing arm operating boundary radius of the tower crane is dynamically calculated and adjusted at the current and future times; based on the predicted temperature and humidity, the degradation coefficients of performance parameters such as the theoretical maximum lifting weight and motor efficiency of the hoisting equipment are dynamically calculated; these dynamically changing safe radii and performance degradation coefficients, as key state variables, are injected into the joint state space of the digital twin model in real time, directly affecting the hoisting and transportation agents' perception of the environment and decision-making costs.

[0007] As a further technical solution, the hoisting agent makes decisions in a joint state space infused with dynamic weather constraints; its decision objective is to efficiently and conflict-free complete the installation of all modules to be hoisted while satisfying safety constraints; the hoisting agent is trained using a deep deterministic policy gradient algorithm, and its decision strategy is adjusted by continuously optimizing a hoisting agent reward function; the hoisting agent reward function is: Encourage successful completion of the hoisting task, N t,collision Severe penalties will be imposed for any form of collision or spatial conflict, including collisions with static obstacles, other tower cranes or transport vehicles, as well as risky conflicts determined due to wind speeds exceeding the dynamic safety radius; The system penalizes unnecessary long-distance movements of the tower cranes and encourages path optimization, with α being the distance penalty coefficient. At each decision time step, the lifting agent outputs an action based on the current joint state obtained from the digital twin model, including the positional state of each tower crane, the positioning state of each module to be lifted, and the dynamic safety boundary provided by the weather agent. This action involves two key decisions: selecting the next module number to be lifted, and planning a swing arm spatial path for the tower crane performing the lifting task from its current position to the module grab point and then to the target installation point. This path must undergo rapid collision detection in the digital twin model to ensure that it does not intrude into the dynamic safety boundaries of itself and other equipment, nor intersect with any static or dynamic obstacles. Once the action is adopted and executed, the system immediately updates the task state of the selected module in the digital twin model to "lifting" and temporarily locks the module and the spatial area involved in its lifting path, forming a new dynamic obstacle. These updates change the joint state space in real time, creating time windows and spatial no-go zones that the transportation agent must adhere to in its decision-making.

[0008] As a further technical solution, the transport agent and the hoisting agent operate synchronously in the same joint state space to plan the arrival schedule of the module transport vehicles, ensuring that the modules to be hoisted arrive accurately at the designated placement point at the time required by the hoisting agent, avoiding crane waiting or transport vehicle congestion; the transport agent also adopts a deep deterministic policy gradient algorithm, and its decision-making strategy is achieved by optimizing the reward function of the other transport agent. The function is adjusted to reward timely arrival of transport vehicles, penalize delays, and penalize excessively long transport routes, where β is the path penalty coefficient. The transport agent's decisions depend on the dynamic environment generated by the actions of the hoisting agent. The joint state it observes includes the working area and time occupancy of each tower crane within a future period, defined by the latest decision of the hoisting agent, i.e., dynamic obstacles and time windows. Based on this, the transport agent outputs an action for a module to be transported, which includes: the departure time of the transport vehicle and the detailed travel route from the yard or processing area to the construction site's placement point. When planning the route, it is necessary to avoid the digital twin module. The model includes all static obstacles, other moving transport vehicles, and restricted areas locked by the actions of the hoisting agent. When planning the departure time, the travel time and preparation time at the landing site must be considered, and the starting point of the hoisting time window reserved for the module by the hoisting agent must be targeted to ensure punctuality rather than being too early or too late. After the transport agent executes its actions, the expected arrival time and landing status of the corresponding module in the digital twin model are updated. This update, in turn, creates new decision-making conditions for the hoisting agent: a module can only be selected as an executable task by the hoisting agent when its landing status is marked as ready, thus forming a resource constraint closed loop.

[0009] As a further technical solution, during the training phase, the hoisting agent, transportation agent, and weather agent share the global joint state information provided by the digital twin model; the system periodically starts a training cycle; within each training cycle, the system simulates a decision-making process: each agent outputs actions based on the current joint state, and these actions are simulated in the digital twin model for a fixed duration T; during the simulation, the system records the single-step instantaneous reward calculated by each agent according to its reward function at each step; after the simulation ends, the reward is calculated according to the formula... Calculate the expected cumulative return J(π) over the next T steps under the current multi-agent joint policy π, where γ is the discount factor and R0 is the cumulative return over the next T steps. tThe sum of the instantaneous rewards for each agent at time step t; the training objective is to find a joint policy that maximizes J(π); the system employs gradient backpropagation to simultaneously adjust the parameters of the decision-making networks of the three agents, enabling them to learn in a virtual twin environment how to achieve the global goal of maximizing overall construction efficiency, minimizing conflicts, and ensuring the best safety through competition for hoisting resources and cooperation in transportation to meet hoisting needs; after training, the network parameters of each agent are fixed; in the actual execution phase, each agent makes rapid decisions based solely on its observed local state, generating action combinations to form the final collaborative scheduling strategy; this strategy ensures that each agent's decision fully considers the potential behaviors of other agents and the dynamic changes in the weather environment.

[0010] As a further technical solution, the scheduling output layer is responsible for transforming the collaborative scheduling strategy generated by the intelligent agent decision layer based on future time period predictions into precise instructions that can be recognized and executed by field equipment, and for providing feedback on the execution status. The received collaborative scheduling strategy is a structured data packet, which contains a list of planned hoisting module sequences for the next few hours, a sequence of key coordinate points for each tower crane's boom path in each hoisting task, a planned departure time for each transport vehicle, key points of its travel path, and the corresponding module identifier. The scheduling output layer first parses this data packet, breaking down each scheduling element, such as "Tower crane A will hoist module M3 at 10:15, with a path point sequence of P1->P2->P3," into discrete, atomic instruction elements. Then, according to the instruction template agreed upon with the field execution system, the instruction elements are combined with the current system timestamp and the unique number of this scheduling batch, and encapsulated into a structured XML document. This XML document serves as the message payload. The system publishes instructions via the lightweight MQTT message queue telemetry transport protocol. The target topic name is formatted as device type / device number, such as Crane / A001 or Truck / T005, ensuring accurate delivery of instructions to the designated field device terminal. Simultaneously, the scheduling output layer immediately subscribes to a specific topic named after the current scheduling batch number to receive instruction confirmation messages from the field execution system. Upon receiving the instruction, the field device publishes a feedback message to this confirmation topic, typically including the device ID, instruction received time, planned start time, or an error code indicating execution failure. After listening to these confirmation messages, the scheduling output layer updates the task status of the corresponding device entity in the digital twin model, for example, updating the status of tower crane A from idle to planned, with the task starting at 10:15. This ensures the digital twin model maintains consistency with the execution progress in the physical world, providing an accurate initial state for the next round of real-time perception and dynamic decision-making.

[0011] This invention provides a modular construction dynamic hoisting scheduling decision system, which has the following beneficial effects: 1. This invention solves the core problem of slow decision-making response caused by the reliance on static or lagging environmental data in scheduling methods by establishing a digital twin model driven by real-time sensing data. This transforms the scheduling basis from historical information to real-time field situation, significantly improving the system's dynamic adaptability and operational robustness in the face of complex and ever-changing terrain conditions.

[0012] 2. This invention solves the fundamental problem of failing to perform global collaborative optimization by constructing a multi-agent collaborative decision-making framework composed of hoisting, transportation, and weather agents. This framework enables agents to interact and learn in a unified twin environment, achieving global dynamic optimization across processes and resources, and significantly improving the efficiency and scheduling flexibility of multi-equipment collaborative operations.

[0013] 3. This invention enables the system to proactively predict and avoid safety risks caused by severe weather and spatial conflicts, transforming passive response into proactive prevention, and greatly improving the automation level of the scheduling process and the ability to proactively control safety risks. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the intelligent decision-making process. Detailed Implementation

[0015] The technical solutions in the embodiments of the present invention will be clearly and completely described below. 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.

[0016] like Figure 1 and Figure 2As shown, at a modular building project site in a mountainous area, to implement dynamic hoisting scheduling, a site perception network consisting of multiple multi-rotor drones and several fixed high-definition cameras was first deployed. The drones adopted a hexacopter configuration, each equipped with a high-precision 32-line lidar with a detection range between 0.5 meters and 100 meters and an RGB-D depth camera. These drones patrolled along a pre-defined flight path at a cruising speed of 30 meters per minute, performing a full-coverage 3D scan of the core construction area every 10 minutes to collect raw 3D point cloud data, with an initial point cloud density of approximately 100 points per square meter. The fixed cameras used 4K resolution wide-angle lenses with infrared night vision capabilities and a horizontal field of view of 120 degrees, continuously capturing 2D image information of the construction site at a rate of 1 frame per second. All collected raw data, including point cloud files, image sequences, and real-time status data such as equipment number, latitude and longitude coordinates, hook height, and travel speed read from the tower crane's programmable logic controller and the transport vehicle's onboard GPS module, were transmitted in real time to the cloud server through a 5G wireless communication network established at the construction site.

[0017] In the data preprocessing stage on the cloud server, the system first performs a pass-through filter on the raw point cloud data collected by the LiDAR, removing discrete noise points that are more than 100 meters away from the sensor or less than 0.5 meters away. Then, a voxel grid filtering algorithm is applied to downsample the point cloud, setting the side length of the voxel cube to 0.1 meters, thereby reducing the average density of the point cloud data from approximately 100 points per square meter to approximately 10 points per square meter. For synchronously acquired image frames, the system calls a target detection model that has undergone transfer learning based on specific construction site data. This target detection model can accurately identify and label tower cranes, transport vehicles of different tonnages, and various types of building modules to be hoisted in the image, and output the pixel coordinates of their bounding boxes. Next, using the camera intrinsic and extrinsic parameter matrix obtained through prior calibration, the two-dimensional pixel coordinates of each target identified in the image are back-projected onto the image through a pass-through filter. After time synchronization and spatial alignment, the corresponding 3D point cloud data clusters are segmented to identify 3D point cloud subsets belonging to each independent physical entity. Taking a tower crane numbered Crane01 as an example, the system calculates the 3D coordinates of the crane's rotation center in the local coordinate system of the lidar from its associated 3D point cloud clusters. Then, using a pre-determined coordinate transformation matrix, the local coordinates of all target objects are uniformly transformed to the global UTM coordinate system with the project-defined reference point as the origin. For a small number of missing data records in the equipment status data stream due to network fluctuations, the system uses a timestamp-based linear interpolation algorithm to complete them. Finally, Z-score standardization is performed on all numerical feature data, including spatial coordinate values, velocity values, wind speed values, etc., so that the mean of the processed data is 0 and the standard deviation is 1, thus completing the data preprocessing process.

[0018] The digital twin modeling layer constructs and updates a 3D virtual scene based on the preprocessed data stream. For example, this digital twin modeling layer uses the Unity 3D engine as its core visualization and basic physical simulation platform, while integrating a point cloud library for real-time rendering and 3D spatial relationship queries of point cloud data. The static environment model is directly exported from the project's building information model in a common 3D format and imported, including high-precision terrain meshes, completed building foundations, and the outline of the planned building structure. The dynamic entity model is generated and updated entirely by real-time perception data. The system creates a corresponding virtual entity object and a control script attached to it in the digital twin model for each real-world physical entity. The script defines a data structure to store all the dynamic attributes of the entity. Taking a tower crane as an example, its status data structure includes the following fields: entity type identifier is tower crane, unique equipment number, a 3D vector representing its position in the global coordinate system, a 3D vector representing its speed, the swing angle of the boom, and its current working status. The system generates a task status code, the current weight of the crane, and a feature vector describing the characteristics of its surrounding environment. This environmental feature vector is a floating-point array of length 50, which is automatically calculated and filled by the system periodically. The calculation mainly includes querying point cloud data to calculate the distance between the crane hook point and the nearest static obstacle within a certain radius, calculating the distance between the crane hook point and other nearest dynamic entities on the construction site, and integrating the radius scaling factor that currently affects safe operation output by the weather agent. The system runs a continuously working data update process, listening to the data stream from the preprocessing stage in real time. When it receives the latest status data packet for Crane01, it immediately updates the position, velocity, and other fields in its status data structure and calls the interface of the 3D engine to drive the corresponding Crane01 3D model in the digital twin scene to move to the new coordinates and adjust its posture, thereby achieving synchronization between the virtual model and the physical entity state.

[0019] The intelligent agent decision layer comprises three core intelligent agent units. The weather agent starts working first, continuously receiving chronologically ordered weather data sequences uploaded by the environmental perception layer. Each data sequence generates a record every 5 minutes, containing measurements of four dimensions: wind speed, wind direction, ambient temperature, and relative humidity. Internally, the weather agent maintains a time-series prediction model based on a long short-term memory network. This model takes historical weather data sequences from the past 12 hours as input and outputs predicted values ​​for the four weather parameters at 36 time points within the next 3 hours. Simultaneously, the system calculates the mean absolute error (MAE) between the predicted and actual monitored values ​​every 5 minutes. t The calculation formula is: Where N is the number of samples within the calculation window, yi This is the actual monitored value. i The model predicts the value; the decision reward value of the weather agent is based on the formula. The calculation is performed, where the reward coefficient γ is set to 0.95 in this embodiment. This reward value is used to periodically fine-tune the parameters of the long short-term memory network model through the time backpropagation algorithm. The prediction results are further transformed into specific dynamic environmental constraints. The system has a preset table of wind load and safe operating radius relationship based on engineering standards. For example, when the wind speed is predicted to reach 15 meters per second at a certain time in the future, the safe operating radius should be adjusted to 60% of the standard radius by looking up the table. The system uses this scaling factor as a dynamic variable and injects it into the environmental feature vector of the corresponding tower crane entity in the digital twin model.

[0020] The hoisting agent is trained and makes decisions using a deep deterministic policy gradient algorithm framework; its decision-making strategy is continuously optimized through the hoisting agent's reward function. The value is adjusted accordingly; the reward function is specifically... In this embodiment, the distance penalty coefficient α is set to 0.1; at each decision time step t, the system first obtains the cumulative reward value up to step t-1. The lifting agent, based on the current joint state obtained from the digital twin model, outputs decision actions, including selecting the module to be lifted and planning the tower crane boom path; these actions are simulated and executed in the digital twin model; if the lifting task is successfully completed, then N... t,success Increase by 1; if a spatial collision or conflict occurs during the simulation, then N t,collision Increase by 1; simultaneously, the system calculates the Euclidean distance by which the tower crane moved due to this decision. Subsequently, the system calculates the cumulative reward for the current time step, and then obtains the incremental reward for each step. The reward increment is fed into its policy network, which updates the network parameters by maximizing the expected cumulative return of future discounts, thereby optimizing subsequent decisions.

[0021] The transportation agent is also based on the deep deterministic policy gradient algorithm, and its decision-making policy is achieved by continuously optimizing the transportation agent's reward function. The value is adjusted accordingly; the reward function is specifically... In this embodiment, the path penalty coefficient β is set to 0.05; at decision time step t, the system records the cumulative reward up to step t-1. The transport agent, based on the current joint state and specifically considering the dynamic obstacles and time windows defined by the actions of the hoisting agent, outputs the departure time and route plan for the transport vehicle. After the action simulation is executed, if the transport vehicle arrives within ±5 minutes of the hoisting time window, then N... t,ontimeIncrease by 1; if there is a delay, then N t,delay Increase by 1; simultaneously, accumulate the length of the planned path. Calculate the current cumulative reward and obtain the instant reward for each step. This is used to optimize its policy network.

[0022] The system employs a centralized training framework to collaboratively optimize the training of multiple agents. The training objective is to maximize the global expected cumulative reward J(π); the objective function is... The discount factor γ is set to 0.95, and R t The sum of the instantaneous rewards of all agents at each time step, i.e. The training cycle is set to start once per hour. During each training session, the system samples data from the experience replay pool, and each agent performs centralized training under the condition of sharing global information. The gradient of the expected cumulative reward J(π) is calculated, and the parameters of the hoisting, transportation, and weather agent decision networks are updated simultaneously using the backpropagation algorithm. The training continues to iterate until the expected cumulative reward J(π) converges to a stable state.

[0023] Once the agent training reaches satisfactory performance, the system enters the real-time online decision-making and scheduling instruction generation phase. Suppose that at 2 PM on a certain weekday afternoon, the weather agent predicts that the wind speed will remain at 15 meters per second for the next two hours. The system updates the safety radius scaling factor of the relevant tower cranes to 0.6. Based on this updated joint state, the hoisting agent decides to prioritize hoisting modules on lower floors that are less affected by wind. The transportation agent simultaneously adjusts its plan to delay the transportation of modules that are more affected by wind.

[0024] The scheduling output layer generates a scheduling scheme for future time periods based on the collaborative scheduling strategy generated by the agent decision layer. The strategy data packets received by this layer contain a list of hoisting sequences for the next three hours, such as module M2, module M3, and module M1. For each hoisting task, the data packet contains the specified tower crane number, such as Crane01, and the sequence of critical path coordinates planned for that tower crane from its current location to the module grabbing point and then to the installation point, such as P1, P2, and P3. Simultaneously, the data packet also contains the planned departure time of each transport vehicle and the corresponding module identifier; for example, transport vehicle Truck05 should depart at 14:20 to transport module M2. The instruction generation module of the scheduling output layer first parses this structured strategy data into discrete instruction elements; then, according to the preset instructions... The template combines these elements with the current system timestamp of 14:00 and the current scheduling batch number Sched-20231027-1400, encapsulating them into a structured Extensible Markup Language (EXPLAIN) document. This document serves as the message payload and is published to topics named using a combination of device type and device number via the Message Queuing Telemetry Transport Protocol (MQTP). For example, a path command for tower crane Crane01 is published to the topic Crane / Crane01 / Command, and a scheduling command for truck Truck05 is published to the topic Truck / Truck05 / Command. The on-board control terminals of the corresponding tower cranes and trucks subscribe to their respective command topics, parse the commands immediately upon receipt, and begin automatic execution or prompt the operators to execute them.

[0025] Simultaneously, at the dispatch output layer, the system automatically initiates a subscription to the acknowledgment topic named with the current dispatch batch number Sched-20231027-1400. When a field device begins executing an instruction or completes a critical step, its control terminal will publish a feedback message to this acknowledgment topic. The feedback message uses JavaScript object notation and includes at least the device ID, the received dispatch batch number, the instruction reception status, and the timestamp of the start of execution. For example, if tower crane Crane01 starts executing a movement instruction at 14:02, its terminal will publish the following feedback message: Device ID:C rane01, instruction batch: Sched-20231027-1400, status: executing, start time: 14:02:15; the digital twin modeling layer continuously listens for this confirmation topic; upon receiving this feedback message, the system immediately searches for the tower crane entity with the number Crane01 in the digital twin model and updates its task status from idle to executing; more importantly, the system will start a corresponding simulation execution process in the digital twin model based on the start execution time in the feedback message and the pre-stored instruction content, in order to predict and display the future movement trajectory of the equipment in the virtual environment.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A modular construction dynamic hoisting scheduling decision-making system, characterized in that, include: The environmental perception layer continuously collects 3D point cloud data, image data, equipment status data, and weather data through data acquisition equipment equipped with lidar and cameras deployed at the construction site. The digital twin modeling layer constructs and dynamically updates a digital twin model that includes a static geometric model imported from BIM files and a real-time data-driven entity model of the construction site. The intelligent agent decision-making layer comprises a hoisting agent, a transportation agent, and a weather agent. The weather agent predicts future weather conditions based on historical and real-time weather data and outputs its results as dynamic environmental constraints. The hoisting and transportation agents make decisions in a joint state space composed of a digital twin model and dynamic environmental constraints. Specifically, the hoisting agent's decision output changes the tower crane's task status and space occupancy in the digital twin model in real time, constituting dynamic obstacles and time windows that the transportation agent must adhere to when planning routes and times. The transportation agent's decision output is used to plan the transport vehicle's entry schedule and update the expected arrival sequence and positioning status of each module to be hoisted in the digital twin model, thus providing the necessary prerequisites and resource constraints for the hoisting agent to generate hoisting tasks. The weather agent's prediction results dynamically modify the safe operation boundary and equipment performance parameters in the joint state space, coupling them in real time to the action selection and reward calculation of the hoisting and transportation agents. Based on the interaction between agents and combined with dynamic environmental constraints, the intelligent decision-making layer generates a collaborative scheduling strategy for future time periods in the digital twin model. The scheduling output layer generates a scheduling plan for future time periods based on the collaborative scheduling strategy of the intelligent agent decision layer, and sends it to the field execution system through the message queue telemetry transmission protocol.

2. The modular construction dynamic hoisting scheduling decision system according to claim 1, characterized in that, In the environmental perception layer, the point cloud data and image data synchronously acquired by the lidar and camera are fused and processed, including: downsampling the acquired point cloud data; spatiotemporally aligning the downsampling point cloud with the acquired synchronous image frames, and identifying targets in the construction site entity model in the image frames to obtain their pixel position information; segmenting point cloud clusters associated with each target in the corresponding point cloud data based on the pixel position information, and calculating the three-dimensional coordinates of the targets based on the spatial points of each point cloud cluster; uniformly transforming the three-dimensional coordinates to the global coordinate system of the construction site; interpolating and completing the missing parts of the target coordinates in the continuous acquisition sequence, and normalizing the obtained equipment status sequence data.

3. The modular construction dynamic hoisting scheduling decision-making system according to claim 1, characterized in that: In the digital twin modeling layer, the construction and dynamic updating of the digital twin model includes: extracting state sequences of the tower crane, transport vehicle, and hoisting module based on the data processed by the environmental perception layer. Each state sequence includes a timestamp, equipment identification number, three-dimensional coordinates, and state identifier. A corresponding state structure is maintained for each dynamic entity in the digital twin model. This state structure includes at least the entity type, current position, speed, task status, and an environmental feature vector associated with the entity. The latest data in the state sequence is updated to the corresponding state structure, and the three-dimensional posture and position transformation of the corresponding dynamic entity is driven according to this state structure. The environmental feature vector is filled according to the relative spatial relationship between the dynamic entity and the static geometric model, as well as the state structures of other dynamic entities at the same timestamp, to form the joint state space required by each agent in the agent decision-making layer when making decisions.

4. The modular construction dynamic hoisting scheduling decision-making system according to claim 3, characterized in that: When the hoisting agent makes decisions in the joint state space, its decision strategy is adjusted by continuously optimizing the value of a hoisting agent reward function; the hoisting agent reward function is: , where N t,success N represents the number of hoisting operations completed up to time t. t,collision This represents the number of collisions or conflicts that have occurred up to time t. This represents the distance the tower crane travels in the k-th time step. Let be the distance penalty coefficient, and >0; at each decision time step t, calculate the reward increment for the current step. The incremental reward is the single-step instant reward obtained by the hoisting agent at this time step. The hoisting agent selects the module to be hoisted and plans the tower crane boom path as an action output based on the current joint state obtained from the digital twin model. After executing this action, the task status of the selected module is updated to occupied in real time in the digital twin model, and the relevant space area is locked. At the same time, based on the single-step instant reward, the subsequent decision-making strategy is optimized by maximizing the cumulative expected return, so that it adjusts the hoisting and tower crane boom path.

5. The modular construction dynamic hoisting scheduling decision-making system according to claim 3, characterized in that: When a transportation agent makes decisions in a joint state space, its decision-making strategy is adjusted by continuously optimizing the value of a transportation agent's reward function; the transportation agent's reward function is: , where N t,ontime N represents the number of hoisting operations completed up to time t. t,delay This represents the number of collisions or conflicts that have occurred up to time t. This represents the distance the tower crane traveled in the k-th time period. Let be the distance penalty coefficient, and >0; At each decision time step t, the system calculates the reward increment for the current step. The incremental reward is the instant reward obtained by the transport agent at this time step. Based on the current joint state obtained from the digital twin model, and the dynamic obstacles and time windows generated by the action output of the hoisting agent, the transport agent plans the departure time and travel route of the transport vehicle as an action output. After executing the action, the expected arrival time and positioning status of the corresponding building module to be hoisted are updated in the digital twin model. At the same time, the subsequent decision-making strategy is optimized by maximizing the cumulative expected return based on the instant reward of the transport agent to generate a transport vehicle entry plan that dynamically matches the hoisting sequence.

6. The modular construction dynamic hoisting scheduling decision-making system according to claim 1, characterized in that, The process of generating dynamic environmental constraints for a weather intelligence agent includes: continuously receiving real-time weather data collected by the environmental perception layer and aligning it with historical weather data by timestamp to form a continuous time series; inputting the processed time series into a time series prediction model to output a predicted value series of wind speed, wind direction, temperature, and humidity for a preset future time period; and simultaneously, based on the average absolute error between the predicted value series and the actual monitored values... Through the weather agent reward function Calculate the decision reward value, where MAE t It is the mean absolute error of weather forecasts up to time t, where N is the number of samples within the calculation window, and y is the mean absolute error of weather forecasts up to time t. i This represents the actual monitored value at the i-th time point. For the model prediction value at time point i, Let be the distance penalty coefficient, and >0; Based on the predicted value sequence and decision reward value, the dynamic shrinkage radius of the tower crane's safe operating boundary is calculated through the wind load and safety radius relationship table, and the dynamic attenuation coefficient of the hoisting equipment performance parameters is calculated through temperature, humidity and equipment efficiency; The dynamic shrinkage radius and dynamic attenuation coefficient are used as dynamic environmental constraints and updated in real time to the corresponding state variables in the joint state space, directly participating in the action selection and reward calculation process of the hoisting agent and the transportation agent, while the decision reward value is used to optimize the parameters of the time series prediction model.

7. The modular construction dynamic hoisting scheduling decision system according to claim 1, characterized in that, The intelligent decision-making layer generates a collaborative scheduling strategy for future time periods, including: in each decision cycle, acquiring the actions output by the hoisting agent, transportation agent, and weather agent based on the current joint state space; performing a simulation of the actions for a fixed duration T in a digital twin model; recording the single-step instantaneous reward obtained by each agent at each time step t during the simulation, and applying the formula... Calculate the expected cumulative reward J(π) under the action combination policy π, where γ is the discount factor, π represents the agent's policy, i.e., the mapping relationship from state to action, and R t The sum of the instantaneous rewards of each agent at time step t; under the centralized training framework, with the goal of maximizing the expected cumulative reward J(π), the parameters of the decision network of each agent are adjusted through gradient backpropagation; the above simulation, evaluation and parameter adjustment process is iteratively executed until the expected cumulative reward J(π) converges or reaches the preset number of iterations, and the action combination output by each agent at this time is determined as the final cooperative scheduling strategy.

8. The modular construction dynamic hoisting scheduling decision-making system according to claim 1, characterized in that, The scheduling output layer generates and distributes future time-slot scheduling plans, including: receiving collaborative scheduling strategies from the intelligent decision-making layer and parsing the hoisting sequence, tower crane path coordinate sequence, transport vehicle arrival time, and corresponding module identifiers into discrete instruction elements; combining the instruction elements with the current system timestamp and scheduling batch number according to a preset instruction template, and encapsulating them into a structured extensible markup language document; publishing the extensible markup language document as a message payload to a topic named after a combination of equipment type and equipment number via a message queue telemetry transmission protocol; simultaneously, initiating a subscription to a confirmation topic named after the scheduling batch number at the scheduling output layer to receive instruction reception status and start execution time returned from the field execution system; and updating the task status of the corresponding equipment entity in the digital twin model based on the received confirmation information.

Images