Cast-in-place box girder construction progress management method based on BIM and edge network

Abstract

The invention discloses a cast-in-place box girder construction progress management method based on a BIM and an edge network, and relates to the technical field of construction progress management, and the method comprises the steps: 1, extracting a construction process, resource dependence and space-time logic based on the BIM, collecting sensor data and environment data in real time through the edge network, and storing the sensor data and the environment data in a database; based on the sensor data and the environment data, constructing a dynamic space-time knowledge graph with a construction process as a node and space-time logic and resource dependence as an edge; 2, deploying a lightweight graph neural network model at each node, performing embedded learning on the dynamic space-time knowledge graph, capturing progress state characteristics of each node, aggregating the progress state characteristics of each node through a federated learning framework, forming a global progress state recognition model, and outputting a real-time progress state based on the global progress state recognition model; and step 3, mapping the real-time progress state to the BIM digital twinborn model, and outputting an approximately optimal progress regulation and control strategy through a multi-agent deep reinforcement learning algorithm.

Description

Technical Field

[0001] This invention relates to the field of construction progress management technology, specifically a method for managing the construction progress of cast-in-place box girders based on BIM and edge networks. Background Technology

[0002] Cast-in-place box girders are reinforced concrete beams formed by directly binding steel bars, pouring concrete, and curing them on supports or formwork at the bridge construction site. They serve as the main load-bearing structure of the bridge, supporting traffic loads such as vehicles and pedestrians, and evenly distributing the loads to the piers and abutments to ensure bridge traffic safety. Compared to precast beams, they are more flexible and adaptable to long-span, curved, and irregularly shaped bridges, such as urban interchanges and approach bridges across rivers. The structural dimensions can be adjusted according to the site topography and construction conditions. The box section has excellent torsional and bending resistance, which can improve the overall rigidity of the bridge and reduce deformation and vibration during operation. In construction sites with narrow spaces or difficulties in transporting or hoisting precast beams, they can be directly cast on-site, reducing construction limitations.

[0003] Traditional cast-in-place box girder progress management relies on manual statistics and on-site inspections, which suffers from problems such as delayed progress data collection and large errors, making it difficult to grasp the real-time status of each process in a timely manner. Existing BIM technology focuses more on static model construction and lacks dynamic connection with real-time data on the construction site, making it difficult to reflect the dynamic changes in progress in sync. The output format of progress information is singular and cannot meet the needs of front-line personnel for simplicity and the multi-dimensional decision-making needs of management personnel. Moreover, it lacks a data feedback and iteration mechanism, and the progress management model cannot be continuously optimized based on on-site data. This results in insufficient accuracy of resource scheduling strategies, low efficiency in progress correction, and is prone to problems such as project delays and resource waste. Summary of the Invention

[0004] To achieve the above objectives, the present invention provides the following technical solution: Step 1: Extract construction procedures, resource dependencies, and spatiotemporal logic based on BIM, collect sensor data and environmental data in real time using edge networks, and construct a dynamic spatiotemporal knowledge graph with construction procedures as nodes and spatiotemporal logic and resource dependencies as edges based on sensor data and environmental data. Step 2: Deploy lightweight graph neural network models on each node, perform embedding learning on the dynamic spatiotemporal knowledge graph, capture the progress status features of each node, aggregate the progress status features of each node through a federated learning framework to form a global progress status recognition model, and output the real-time progress status based on the global progress status recognition model. Step 3: Map the real-time progress status to the BIM digital twin model, and output an approximately optimal progress control strategy through a multi-agent deep reinforcement learning algorithm.

[0005] Furthermore, the construction process is decomposed into a WBS structure, and the sub-tasks in each WBS structure are associated with the corresponding model components in the BIM. Resource dependency is extracted from the component attributes of the BIM model, including the resources required for construction procedures, such as manpower, equipment and material parameters. At the same time, BIM is used to identify the resource constraints between construction procedures and form a dependency mapping table between construction procedures and resources. Resource dependency includes temporal logic and spatial logic. Temporal logic in BIM binds the start or end time and duration of construction procedures, and calculates the earliest or latest time and critical path through the logical overlap between construction procedures. Spatial logic, based on the three-dimensional coordinates of the model in BIM, identifies the spatial operation range of construction procedures and the spatial conflict constraints of different construction procedures, and parameterizes the spatiotemporal constraints into the BIM model.

[0006] Furthermore, the process of collecting sensor data and environmental data is as follows: Edge nodes actively acquire raw data from various sensors at a preset sampling frequency. The edge nodes perform preliminary cleaning and format conversion on the acquired raw data. The pre-processed raw data is then transmitted in real time to the edge computing nodes at the construction site for temporary storage via the local communication link of the edge network.

[0007] Furthermore, the construction of a dynamic spatiotemporal knowledge graph: By tagging and binding sensor data and environmental data collected by the edge network with construction procedures and resource dependencies extracted by BIM, and using construction procedures as nodes of the knowledge graph, dynamic attributes are filled into each node based on the associated data. Based on the spatiotemporal constraints of BIM and the real-time progress data of sensors, temporal or spatial relationship edges between procedures are constructed. Based on the binding relationship between resource data and procedures, a dynamic spatiotemporal knowledge graph is constructed.

[0008] Furthermore, the process of deploying the lightweight graph neural network model at each node is as follows: Based on the features of nodes and edges in a dynamic spatiotemporal knowledge graph, redundant fully connected layers are removed. The model size is compressed through model quantization and pruning. Adaptation to the hardware architecture of edge nodes is performed on each edge node at the construction site. Each lightweight graph neural network model package is pushed to the corresponding edge node through the local file transfer protocol of the edge network. Based on the unique ID of the construction process, a mapping relationship is established between the model and the process node corresponding to the edge node. The model is started at the edge node, pre-trained parameters are loaded, and fine-tuning is performed using 1 to 2 sets of real-time data of the current process to adapt the model to the feature distribution of the current construction scenario.

[0009] Furthermore, the process of capturing the progress status characteristics of each node is as follows: Centered on the target construction process node, based on spatiotemporal logic and resource dependency edges, select its predecessor construction process, parallel construction process, resource dependency, and 1-2 hop neighbor nodes to form a local subgraph; encode the dynamic attributes of the nodes and the relational attributes of the edges in the local subgraph into numerical feature vectors, which are used as inputs to the lightweight graph neural network. For the local subgraph, the graph convolutional layer aggregates the feature vectors of neighboring nodes to the target process node based on the relational weights of the edges, updating the feature representation of the target node. Through graph pooling, the updated features of the target node are fused with the global features of the local subgraph to generate an intermediate feature vector. From the embedded feature vector, key dimensions that can represent the progress status of the process are extracted. The extracted progress status features of each edge node are temporarily stored in the local database and labeled with the corresponding timestamp and process ID. Some key features are fed back to the dynamic spatiotemporal knowledge graph to update the progress status label of the corresponding node.

[0010] Furthermore, the formation process of the global progress status identification model is as follows: A federated learning coordinator is deployed at the cloud management node of the construction site. Each edge node acts as a client of federated learning and establishes an encrypted communication link with the coordinator. The coordinator sends unified training configuration parameters to all edge nodes, so that the models of all clients start from the same point. Each edge node independently completes model training based on local data, without transmitting raw construction data externally, and generates model parameter update quantities. Each edge node encrypts and uploads the locally trained model parameters to the federated coordinator. The coordinator receives the parameters and verifies their validity. The federated coordinator uses a preset aggregation algorithm to weight and fuse the valid parameters of each node to generate global model parameters. The converged global model parameters are then distributed to each edge node to complete the update and synchronization of the node models, forming the final global progress status recognition model.

[0011] Furthermore, the process of outputting the real-time progress status is as follows: Each edge node continuously collects real-time data of the construction process and adapts it to the model input format. Each edge node calls the locally deployed global progress status recognition model to complete the inference of real-time data and progress status recognition. The edge node compares the progress status output by the model with the locally stored construction logs and manual inspection records to verify the consistency of the results. The progress status results after verification and correction are output and fed back to the construction management system through multiple channels.

[0012] Furthermore, the process of outputting a near-optimal schedule control strategy is as follows: Based on a BIM digital twin model with mapped real-time progress, a BIM twin simulation training environment is constructed to reproduce the construction process. A two-layer architecture of global coordination agent and process execution agent is adopted. The overall control objectives of the global agent and the local control tasks of the process agents are defined. The action space and state space of the agents are clarified. A local feedback and global decision-making collaboration mechanism is established. A constraint perception module is introduced to filter illegal actions. A dual network structure of strategy network and value network is constructed. A multi-dimensional reward and punishment function with progress and resource dimensions is designed. The agents are trained offline in the simulation environment. The training is terminated when the preset conditions are met, and the global and single-process control strategies are output.

[0013] The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks provided by this invention has the following beneficial effects: (1) This invention uses edge networks to complete the local collection and processing of sensor data at the construction site. At the same time, it combines the dynamic mapping capability of BIM digital twin model to realize the synchronous connection between progress status and physical construction process, so as to grasp the real-time progress of each process in a timely and accurate manner and avoid the problems of progress disconnection and process connection confusion caused by information lag.

[0014] (2) This invention uses the BIM digital twin model as a unified data hub to share progress information in real time. At the same time, it constructs a closed loop of data collection, strategy training and execution feedback to improve the efficiency of multi-party collaborative decision-making, reduce engineering changes and rework caused by information asymmetry, and promote the upgrading of construction management to lean and dynamic.

[0015] (3) The multi-agent deep reinforcement learning of this invention completes the adaptation training of scheduling strategy in the BIM simulation environment. Combined with the global collaborative optimization mechanism of federated learning, it realizes the precise planning of cross-process resource allocation and process overlap, making resource allocation more in line with the actual construction needs, reducing the idle waste of manpower and equipment, and ensuring the smooth progress of each process. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1This application provides a method for managing the construction progress of cast-in-place box girders based on BIM and edge networks. The method includes: Step 1: Extract construction procedures, resource dependencies, and spatiotemporal logic based on BIM, collect sensor data and environmental data in real time using edge networks, and construct a dynamic spatiotemporal knowledge graph with construction procedures as nodes and spatiotemporal logic and resource dependencies as edges based on sensor data and environmental data. Construction process breakdown: The cast-in-place box girder project is decomposed into a WBS structure according to the construction process, such as formwork installation, rebar tying, and concrete pouring. In BIM software, each WBS structure sub-task is associated with the corresponding model components, such as box girder formwork and rebar cage, to clarify the level of construction procedures and the scope of corresponding components. Specifically, by importing the BIM 3D model of the cast-in-place box girder, including the main body of the box girder, formwork, reinforcement, and embedded parts, the construction task decomposition function is enabled in the BIM software. According to the actual construction process of the cast-in-place box girder, the overall project is divided into 8 core procedures: site leveling of the box girder construction area; procedure 2: scaffolding erection and foundation reinforcement; bottom formwork installation and elevation calibration; reinforcement cage processing and on-site binding; side formwork and end formwork installation and sealing; concrete mixing and pumping; concrete curing and strength monitoring; and formwork and scaffolding removal. Each process is bound to the corresponding component in the BIM model, such as binding the steel reinforcement skeleton to the steel reinforcement component group in the model, and marking a unique construction area for each process in the BIM model, such as the left and right sections of the box girder.

[0019] Resource dependency extraction involves extracting the manpower, equipment, and material parameters required for each construction process from the component attributes and resource database of the BIM model. For example, formwork installation requires 5 workers and 1 crane. At the same time, through the association function of BIM, the resource constraint relationship between construction processes is identified. For example, concrete pouring depends on the idle state of resources after the steel reinforcement is tied, forming a dependency mapping table between processes and resources. Specifically, the resource configuration data for each process is retrieved through the resource library module of the BIM model. For example, scaffolding erection requires 20 scaffolders, 5 steel pipe cutting machines, and 80 tons of steel pipe support materials; rebar tying requires 15 rebar workers and 2 rebar bending machines, and there is a waiting period until the bottom formwork is installed, after which the working space and support resources in the bottom formwork area are available; concrete pouring requires continuous material supply from 3 concrete pump trucks and 1 mixing plant, and the sealing and acceptance resources for side formwork installation must be completed simultaneously by the quality inspector. In BIM software, resource dependency trigger conditions are set, such as the arrival status of steel reinforcement materials. These conditions need to be associated with the entry attribute label of the steel reinforcement components in the BIM model. The process is initiated when the label status is updated to "arrived on site".

[0020] Spatiotemporal logical connections are established. In BIM, start or end times and durations are bound to each process, and logical overlap relationships between processes are set, such as completion to start and start to start. The earliest or latest time and critical path are calculated using 4D BIM tools, such as Navisworks and Synchro. Based on the three-dimensional coordinates of the BIM model, the spatial operation range of the process is identified, such as the area boundary of box girder segmented casting, as well as the spatial conflict constraints of different processes, such as the inability of formwork installation and pipeline laying to be parallel in the same area. The spatiotemporal constraints are parameterized and embedded into the BIM model. Specifically, in the 4D progress module of BIM, the duration of each process is set, such as 2 days for bottom formwork installation and 1 day for concrete pouring; the time connection between processes is defined, with a 4-hour interval after completion before the bottom formwork stability check is restarted, and curing must be carried out continuously; the work space boundaries of each process are selected in the BIM model, such as the rebar binding area, which is limited to the three-dimensional space 0.5-2.0 meters above the bottom formwork to avoid overlap with the high-altitude work space of the scaffolding; a space conflict warning is set, and when the work area of ​​side formwork installation overlaps with the unfinished rebar area, the BIM model automatically marks the space occupation conflict and locks the process that starts afterward.

[0021] Deployment of edge nodes and sensors: In the construction area of ​​cast-in-place box girder, such as the formwork, rebar work area and concrete pouring area, various sensors are deployed, such as progress RFID or positioning sensors, environmental temperature, humidity or wind speed sensors, and equipment vibration or energy consumption sensors. Edge computing nodes, such as edge gateways and small edge servers, are also deployed nearby on the construction site. The sensors are connected to the edge nodes via wired or wireless methods, such as LoRa and 5G.

[0022] Edge nodes actively acquire raw data from various sensors at preset sampling frequencies, such as second-level or minute-level, including personnel or equipment location, ambient temperature and humidity, and equipment operating status. The edge nodes perform preliminary cleaning of the collected raw data, such as removing outliers, filling in missing values, and format conversion, such as standardizing the data to reduce the amount of invalid data transmitted. This data is then transmitted in real-time to edge computing nodes at the construction site for temporary storage via local communication links in the edge network, such as LANs or edge cloud dedicated lines. Simultaneously, edge nodes push some critical data, such as abnormal environmental parameters and equipment fault signals, to the remote management platform based on data priority.

[0023] Constructing a dynamic spatiotemporal knowledge graph: Sensor data collected by the edge network, such as personnel location, equipment status, process progress, and environmental data, such as temperature, humidity, and wind speed, are tagged and bound to construction processes and resource information extracted from BIM. Each construction process, such as box girder rebar tying, is assigned a unique ID, and sensor data of the corresponding work area, such as personnel attendance RFID data, rebar material weighing sensor data, and environmental data, such as work area temperature and humidity, are associated with that process ID. For resource dependencies, such as rebar tying relying on cranes to lift rebar, the crane's energy consumption or positioning sensor data is bound to the rebar tying process ID, and the crane's resource occupation time period is also associated, establishing a mapping relationship between sensor or environmental data and construction elements. Using construction procedures as the core nodes of the knowledge graph, dynamic attributes are populated for each node based on associated data, including basic attributes, real-time status attributes, environmental attributes, and spatiotemporal attributes; among which, basic attributes include procedure ID, procedure name, and work area, which are obtained from BIM spatial logic; Real-time status attributes are extracted from sensor data, such as the current progress, including the completion rate of rebar tying and resource occupancy status, such as the number of workers and equipment operating status. Environmental attributes are extracted from environmental data, including parameters such as temperature, humidity, and wind speed in the work area. Spatiotemporal attributes are obtained from the combination of BIM spatiotemporal logic and sensor data, including the actual start or duration of operation and the current work space coordinates. For example, the attributes of a box girder formwork installation node may include: progress: 60%, resources: 2 cranes, 8 workers, ambient temperature: 25℃, wind speed: 3m / s, spatiotemporal: 4 hours of operation completed, and work area: box girder 0 to 10m section.

[0024] Connecting relationships based on spatiotemporal logic and resource dependencies: Based on BIM's spatiotemporal constraints and real-time progress data from sensors, we construct temporal or spatial relationship edges between processes, such as the time sequence edge from the completion of box girder formwork installation to the start of rebar binding, and the time sequence edge from the installation of the 0 to 10m section of box girder formwork to the installation of the 10 to 20m section of box girder formwork. At the same time, we mark the edges with attributes such as preceding, parallel, and subsequent spatial conflicts or compatibility to construct spatiotemporal logical edges. Based on the binding relationship between resource data and processes, resource association edges are constructed between processes and resources, and between processes, such as crane, rebar tying, rebar tying, and concrete pouring. Resource type dependency strength is marked on the edges, such as mandatory dependency and optional dependency, and resource dependency edges are constructed. The properties of the edges will be dynamically updated based on sensor or environmental data. For example, when the rebar binding progress reaches 100%, the timing edges of formwork installation and rebar binding will be marked as completed, and at the same time, the edge status from rebar binding to concrete pouring will be updated to pending activation.

[0025] Real-time data graph iteration: When sensors detect changes in process progress, such as a completion rate increasing from 60% to 80%, or changes in resource status, such as crane malfunctions or abnormal environmental parameters, such as wind speed exceeding a safety threshold, edge nodes will automatically push data to the graph engine to update the attributes of the corresponding nodes or the relationships between edges. For example, resource edges for cranes and rebar tying will be marked as interrupted. The graph engine will combine the new data to verify spatiotemporal constraints. For example, if the actual start time of a process is later than the planned time, it will automatically adjust the temporal edge attributes of subsequent processes. If resource conflicts occur after the update, such as the same crane being assigned to two parallel processes or spatiotemporal conflicts, such as two processes occupying the same space at the same time, the graph will mark the conflicting nodes and edges and push early warning information.

[0026] Step 2: Deploy lightweight graph neural network models on each node, perform embedding learning on the dynamic spatiotemporal knowledge graph, capture the progress status features of each node, aggregate the progress status features of each node through a federated learning framework to form a global progress status recognition model, and output the real-time progress status based on the global progress status recognition model. The process of deploying a lightweight graph neural network model is as follows: Based on the features of nodes and edges in a dynamic spatiotemporal knowledge graph, including process node attributes and spatiotemporal or resource edge relationships, the hidden layer dimensions of the basic graph neural network are simplified. For example, the original 1024-dimensional hidden layer is compressed to 256 dimensions. Redundant fully connected layers are removed, and the core graph convolutional layers are retained to extract node neighbor relationships. Pooling layers are also used to aggregate local features. For the hardware architecture of edge nodes, such as ARM chips, operator fusion is used to merge multiple graph computation operators into a single operation and hardware acceleration instructions are adapted. At each edge node of the construction site, such as the edge gateway and small server in the work process area, the deployment environment is adapted; the CPU computing power, memory capacity, and storage space of the edge nodes are tested to ensure that they meet the running requirements of the lightweight model, such as at least 1 CPU core, 2GB memory, and 200MB of free storage; lightweight inference frameworks, such as TensorFlowLite and ONNXRuntime, are installed on the edge nodes, and the dependency libraries required for model operation, such as the Python lightweight runtime and graph computing toolkit, are configured, and the environment is quickly deployed and isolated through containerization, such as Docker lightweight images; a data receiving service is deployed on the edge nodes, and an interface is established with the local data storage module of the dynamic spatiotemporal knowledge graph to ensure that the model can read the work process attributes and edge relationship data of the node in real time; The modified lightweight graph neural network model is distributed and bound to the corresponding construction process node. Each lightweight graph neural network model package is pushed to the corresponding edge node through the local file transfer protocol of the edge network, such as FTP Lightweight Edition. Based on the unique ID of the construction process, a mapping relationship is established between the model and the process node corresponding to the edge node. For example, the box girder reinforcement binding process node is bound to the lightweight graph neural network model on its work area edge node. The model is started at the edge node, pre-trained parameters are loaded, and pre-trained based on historical progress data of similar box girder construction. It is then fine-tuned using 1 to 2 sets of real-time data of the current process, such as progress completion rate and resource occupancy status, to adapt the model to the feature distribution of the current construction scenario. After model deployment, lightweight monitoring scripts are deployed on edge nodes to collect inference latency and resource utilization of the model in real time. When an anomaly occurs, such as latency exceeding 200ms, the model is automatically restarted or resources are rescheduled. The inference results of each node model, such as progress status features, are collected periodically. The model accuracy is evaluated in conjunction with actual progress data. When the accuracy drops by more than 10%, the lightweight model is retrained based on the latest construction data and pushed to the corresponding node for update via the edge network. Backup edge nodes are configured for key process nodes. When the main node model fails, the model is automatically switched to the backup node model to ensure uninterrupted extraction of progress status features.

[0027] Capture the progress status characteristics of each node: The process of extracting process progress status is achieved by transforming the node and edge relationships of a graph into quantifiable feature vectors through a lightweight graph neural network model.

[0028] Local subgraph sampling and input construction of knowledge graphs: Since the dynamic spatiotemporal knowledge graph is a network of connections across the entire process, the lightweight graph neural network of each edge node only needs to process the local subgraph, namely the current process node and its directly related neighbor nodes and edges. Centered on the target process node, based on spatiotemporal logic and resource dependency edges, select its predecessor processes, parallel processes, resource supply nodes, etc., with 1 to 2 hop neighbor nodes to form a local subgraph, controlling the number of nodes to within 20 to adapt to the computing power of the lightweight model. The dynamic attributes of the nodes in the subgraph, such as progress completion rate, resource utilization rate, and edge relationship attributes, such as the completion of spatiotemporal logic corresponding to the start and strong dependency of resource dependence, are encoded into numerical feature vectors. For example, node attributes are normalized to vectors in the range of 0 to 1, and edge relationships are converted into one-hot encodings, which are used as inputs to the lightweight graph neural network.

[0029] The embedding learning process of lightweight graph neural networks is as follows: For local subgraphs, the graph convolutional layer aggregates the feature vectors of neighboring nodes to the target process node based on edge relation weights, such as the weight of edges with strong resource dependencies being higher than those with weak dependencies. This updates the feature representation of the target node. For example, the rebar tying node will aggregate the completion status of its preceding process, formwork installation, and the crane operation status features of the resource node. Through graph pooling, the updated features of the target node are fused with the global features of the local subgraph, such as the average progress and average resource usage within the subgraph, to generate an intermediate feature vector containing the node's own state and the states of its neighbors. Since construction data changes in real time, the lightweight graph neural network repeats the above process at a frequency of minutes, continuously updating the embedded features of the target process node, achieving a synchronous mapping from dynamic knowledge graph to dynamic feature vectors.

[0030] Extraction and quantification of progress status features: The embedding vectors are used to extract indicators that directly reflect progress, such as the deviation rate between actual completion rate and planned schedule and the critical path matching degree. Resource adequacy ratio, such as the proportion of workers or equipment on site, and environmental adaptability, such as the influence coefficient of temperature and humidity on construction efficiency, are extracted to indirectly affect progress. Abnormal fluctuations in the embedding vectors, such as a sudden increase in the progress deviation rate and a sharp drop in the resource adequacy ratio, are identified and transformed into features such as the risk level of schedule delay and resource shortage warning signals. These features are then quantified into numerical indicators that can be recognized by subsequent models, such as a score from 0 to 00 and a risk level from 1 to 5. Each edge node temporarily stores the extracted progress status features in the local database and marks them with the corresponding timestamp and process ID, which facilitates parameter aggregation in subsequent federated learning. Some key features, such as high-risk warnings, are fed back to the dynamic spatiotemporal knowledge graph to update the progress status labels of the corresponding nodes. For example, the status of the rebar binding node is marked from normal to risk, realizing bidirectional iteration between the graph and features.

[0031] The process of building the global progress status recognition model is as follows: A federated learning coordinator, such as FedAvg or FedProx, is deployed on the cloud management nodes at the construction site. Each edge node, corresponding to a different construction process, acts as a client for federated learning, establishing an encrypted communication link with the coordinator, such as based on TLS / SSL protocol, to ensure secure parameter transmission. The coordinator distributes unified training configuration parameters to all edge nodes, including: model training epochs, such as 30 epochs; local training batch size, such as 16 batches; learning rate, such as 0.001; parameter aggregation weights, allocated according to the importance of each node's process or the proportion of data volume, such as setting the weight of key process nodes to 0.2 and ordinary process nodes to 0.1; and parameter upload cycle, such as uploading parameters once every 2 epochs of local training. The coordinator also distributes unified lightweight graph neural network initial model parameters to each edge node, i.e., the pre-trained parameters when each node is deployed. Each node accesses locally stored progress status feature data, learned through embedding from a dynamic spatiotemporal knowledge graph. The training and validation sets are divided in an 8:2 ratio, with data labels representing actual construction status such as progress level and progress deviation type. Based on the initial model, a lightweight graph neural network model is iteratively trained using the local training set. Model parameters are optimized through backpropagation, such as convolutional layer weights and pooling layer parameters, minimizing the error between the model's predicted progress status and the actual status labels, for example, using a cross-entropy loss function. After each preset number of local training rounds (e.g., 2 rounds), the updated model parameters of the current node are recorded, representing the difference between the current parameters and the initial or previous round's global parameters. The performance of the trained model is evaluated using the local validation set, such as accuracy and F1 score. If the performance does not reach a preset threshold (e.g., accuracy ≥ 85%), local iteration continues. If the threshold is reached, the current model parameters are marked as parameters to be uploaded, along with metadata such as node ID, data volume, and performance metrics.

[0032] Each node encrypts the parameters to be uploaded using a pre-defined encryption algorithm, such as homomorphic encryption or differential privacy encryption, to prevent the parameters from being stolen or tampered with during transmission. The encrypted parameters and metadata are then uploaded to the federated coordinator via an encrypted communication link. After receiving all uploaded data from all nodes, the coordinator performs a validity check, including: whether the node ID is in the authorized list, whether the parameter format conforms to the agreement, whether the metadata is complete, and whether the model performance meets the standards. Illegal parameters, such as those with incorrect formats or substandard performance, are removed. Based on the metadata of each valid node, such as the percentage of data volume and the importance of the process, the aggregation weight is dynamically adjusted. For example, if a node accounts for 30% of the data volume, its weight is increased to 0.3; if the performance of a key process node meets the standards, its weight is increased by an additional 0.05. The coordinator decrypts the encrypted parameters of each node and, based on the calibrated aggregation weights, calculates the weighted average of the parameters using a federated aggregation algorithm, such as the commonly used FedAvg algorithm, to obtain the updated global model parameters, which is the sum of the updated parameters of each node multiplied by their corresponding weights. The updated global parameters are then superimposed with the previous round's global model parameters or the initial parameters to generate new global progress status recognition model parameters. If it is the first round of aggregation, the weighted fused parameters are directly used as the first round's global model parameters. The coordinator calls preset global evaluation metrics, such as cross-node average accuracy and global progress deviation recognition accuracy, and combines this with the local performance metrics uploaded by each node to comprehensively evaluate the newly generated global model. If the evaluation result does not meet the global convergence condition, such as global accuracy ≥ 90% and fluctuation ≤ 1% for three consecutive rounds, the current global parameters are used as the initial parameters for the next round of local training, and the process returns to the local model training stage for further iteration. If the convergence condition is met, the process proceeds to the next step of model solidification. The coordinator encrypts the converged global model parameters and sends them along with update instructions to all authorized edge nodes via an encrypted communication link. Each edge node receives the global parameters, decrypts them, and replaces its local model parameters, completing the local model upgrade. Simultaneously, it fine-tunes the global model based on a small amount of local real-time data, such as the latest 10 progress feature data points, to ensure the model adapts to subtle differences in the local construction scenario—a secondary optimization from global to local. The fine-tuned node model maintains consistency with the core parameters of the global model and is formally solidified as the distributed node version of the global progress status recognition model. Each node uses this model to achieve real-time recognition of its local progress status, while the federated coordinator continuously monitors the running status of each node's model, such as triggering a new round of parameter aggregation weekly.

[0033] The process of outputting real-time progress status is as follows: Relying on the sensor acquisition link of the edge network, dynamic data of the target process is continuously acquired, including: process progress feature data, extracted from the latest embedding learning results of the dynamic spatiotemporal knowledge graph, such as current completion rate, resource occupancy status, raw sensor data, i.e., personnel attendance status, equipment operating parameters, and environmental data, including temperature, humidity, wind speed, etc. The acquisition frequency is synchronized with the model inference frequency, such as once per minute. The edge nodes quickly clean the acquired real-time data, remove abnormal sensor data, fill in missing values, and complete the standardization processing according to the input format agreed upon by the global model. Unstructured data, such as equipment operating status descriptions, are transformed into structured values. All feature data are normalized to the range of 0 to 1, and finally a fixed-dimensional feature vector is generated that is consistent with the input dimension of the model training stage, ensuring that the data can be directly recognized by the model.

[0034] The system employs both timed and event-triggered modes. Under normal circumstances, model inference is triggered at a preset frequency, such as once every minute. When sensors detect abnormal data, such as sudden resource shortages or excessive environmental parameters, emergency inference is immediately triggered to ensure rapid response to emergencies. Preprocessed real-time feature vectors are input into the global model. The model extracts dynamic relationship features between the current process and related nodes through graph convolutional layers. Combined with the global feature aggregation capability of the pooling layer, it completes multi-dimensional identification of the progress status and outputs core identification results, including: basic progress status (e.g., normal progress, minor delays, and severe delays), progress completion rate, deviation from planned progress (e.g., 2-hour delay), and key factors affecting progress (e.g., crane malfunction leading to resource shortages). Insufficient resources and high temperatures reduce construction efficiency. Edge nodes compare the progress status output by the model with the locally stored construction logs and manual inspection records to verify consistency. For example, if the model identifies a rebar binding completion rate of 80% and it matches the manual inspection record, the verification passes. If the deviation exceeds 5%, a second inference is triggered. For progress status identification results across processes, if subsequent processes are affected by delays in preceding processes, edge nodes upload the local inference results to the federated coordinator. The coordinator performs a global verification based on the progress status of other related nodes. If a logical conflict is found, such as a process being identified as progressing normally, but its dependent resource node being identified as having interrupted supply, a correction command is issued, and the edge nodes adjust the output results based on global correlation information.

[0035] Step 3: Map the real-time progress status to the BIM digital twin model, and output an approximately optimal progress control strategy through a multi-agent deep reinforcement learning algorithm; Mapping real-time progress status to the BIM digital twin model: By extracting the core progress status data from the model output, including unique process IDs, progress completion rates, progress deviation durations, progress status tags (normal progress, minor delays, and severe delays), influencing factors, and data collection timestamps, redundant information is removed to form a standardized data field set. The standardized progress data is then converted from the numerical or text format output by the model into a data format supported by the BIM digital twin platform, such as IFC or JSON, and encapsulated through communication protocols agreed upon by the BIM platform, such as HTTP, REST, or WebSocket. Using the unique ID of each process as the primary key, a one-to-one mapping is established between the real-time progress status data of each process and the corresponding component in the BIM model. For example, the template installation process of the 0 to 10m section of the box girder corresponds to the template component of that section in the BIM model. At the same time, a mapping is established between the progress influencing factors and the corresponding resource components. The above mapping relationships are organized into a rule base, which includes rule items such as process ID and BIM component ID, progress field and model display attributes, and stored in the backend database of the BIM digital twin platform. The BIM digital twin platform continuously receives standardized progress status data uploaded by various edge nodes through pre-defined interfaces. It filters the latest data based on timestamps and discards outdated and redundant data. According to the mapping rule library, it matches real-time progress data to the corresponding BIM components and updates them synchronously. This includes updating component attributes by adding fields such as real-time completion rate and progress deviation to the attribute list of BIM components to achieve data traceability; and updating the visual status through color rendering, such as green for normal progress, yellow for minor delays, and red for severe delays. Progress bars are overlaid on the components to display the completion rate progress bar, and influencing factors are labeled next to the components to make the progress status intuitively identifiable. Finally, it summarizes the real-time progress data of all processes and generates a progress heatmap and critical path progress curve in the global view of the BIM model, intuitively presenting the progress distribution of the entire construction process. Real-time progress status data is pushed to the dynamic spatiotemporal knowledge graph engine to update the progress status labels of corresponding process nodes. For example, the label of the rebar tying node is updated from normal to slightly delayed. At the same time, the attributes of the associated edges are adjusted. For example, the temporal edge status of rebar tying and concrete pouring is updated to pending activation. This ensures that the knowledge graph is synchronized with the actual progress. The real-time progress status data, after format conversion and verification, is stored in the local database of the edge nodes and synchronized to the cloud database at the agreed period. This data serves as a feedback sample for model optimization. The BIM platform regularly compares the progress data in the model with the original progress data uploaded by edge nodes and manual inspection records. If the deviation exceeds a preset threshold, such as a completion rate deviation of >5%, data remapping is triggered to correct the erroneous state in the model. When the model identifies abnormal progress states such as serious delays, the corresponding components are highlighted in the BIM model, and an early warning mechanism is triggered simultaneously. The early warning information, including the location of the abnormal process, influencing factors, and deviation duration, is pushed to the cloud management platform, on-site display screens, and mobile terminals of management personnel, providing accurate basis for progress control.

[0036] The process of outputting a near-optimal schedule control strategy is as follows: Based on the BIM digital twin model with mapped real-time progress, simulation engines such as construction process deduction, resource scheduling, and environmental disturbance are embedded to reproduce the operation process such as formwork installation and rebar binding. It supports the simulation of working conditions such as resource increase and decrease, process overlap adjustment, and equipment failure. At the same time, parameters such as time and space constraints, resource constraints, and schedule targets are embedded, and real-time progress data is imported to ensure that the simulation environment fits the actual site conditions. A two-tier architecture of global coordination agent and process execution agent is adopted. The global coordination agent provides overall control of the entire process, with the goal of optimizing the total project duration and maximizing resource utilization. The process execution agents correspond to individual processes such as formwork and rebar, and are responsible for local progress correction and resource allocation. They are scheduled by the global agent and provide status feedback. The action space of the agents is defined. The global coordination agent adjusts cross-process resource allocation and process overlap. The local coordination agents adjust resource input for their own process, optimize scheduling, and manage the state space, real-time progress, resource usage, and environmental parameters to establish a collaborative mechanism between local feedback and global decision-making. A constraint perception module is introduced to filter actions that violate spatiotemporal or resource constraints; a dual-network structure is constructed, consisting of a policy network that outputs scheduling actions and a value network that evaluates the benefits of actions; a multi-dimensional reward and punishment function is designed: positive rewards are given for reducing schedule deviations and improving resource utilization; high penalties are given for resource waste and process conflicts; global agent rewards and punishments are linked to the performance of all processes to promote collaborative optimization. The agent learns iteratively in the BIM simulation environment, mastering basic scheduling rules through a cycle of perceiving states, outputting actions, receiving BIM feedback results, and updating parameters with rewards and penalties. It also introduces disruptive conditions such as equipment failure and severe weather. Training terminates when the agent's strategy parameters are stable, the post-scheduling schedule deviation rate is ≤5%, and resource utilization is ≥85%. Multiple extreme conditions, such as resource shortages and equipment downtime, are imported into the simulation environment to verify the strategy's effectiveness. After fine-tuning based on engineering experience, a global control strategy is output, including resource allocation plans, schedule correction paths, and local resource adjustments and scheduling schemes for single-process strategies. The strategy's deduction effect is then visualized in the BIM model.

[0037] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0038] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for managing the construction progress of cast-in-place box girders based on BIM and edge networks, characterized in that, The method includes: Step 1: Extracting construction procedures, resource dependencies and spatiotemporal logic based on BIM, collecting sensor data and environmental data in real time using edge networks, and constructing a dynamic spatiotemporal knowledge graph with construction procedures as nodes and spatiotemporal logic and resource dependencies as edges based on sensor data and environmental data; Step 2: Deploy lightweight graph neural network models on each node, perform embedding learning on the dynamic spatiotemporal knowledge graph, capture the progress status features of each node, aggregate the progress status features of each node through a federated learning framework to form a global progress status recognition model, and output the real-time progress status based on the global progress status recognition model. Step 3: Map the real-time progress status to the BIM digital twin model, and output an approximately optimal progress control strategy through a multi-agent deep reinforcement learning algorithm.

2. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 1, characterized in that, The construction process is decomposed into a WBS structure, and the sub-tasks in each WBS structure are associated with the corresponding model components in BIM. Resource dependency is extracted from the component attributes of the BIM model, including the resources required for construction procedures, such as manpower, equipment and material parameters. At the same time, BIM is used to identify the resource constraints between construction procedures and form a dependency mapping table between construction procedures and resources. Resource dependency includes temporal logic and spatial logic. In BIM, temporal logic binds the start or end time and duration of construction procedures, and calculates the earliest or latest time and critical path through the logical overlap between construction procedures. Spatial logic, based on the three-dimensional coordinates of the model in BIM, identifies the spatial operation range of construction procedures and the spatial conflict constraints of different construction procedures, and parameterizes the spatiotemporal constraints into the BIM model.

3. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 2, characterized in that, The process of collecting sensor data and environmental data is as follows: Edge nodes actively acquire raw data from various sensors at a preset sampling frequency. The edge nodes perform preliminary cleaning and format conversion on the acquired raw data. The pre-processed raw data is then transmitted in real time to the edge computing nodes at the construction site for temporary storage via the local communication link of the edge network.

4. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 3, characterized in that, Construction of dynamic spatiotemporal knowledge graphs: By tagging and binding sensor data and environmental data collected by the edge network with construction procedures and resource dependencies extracted by BIM, and using construction procedures as nodes of the knowledge graph, dynamic attributes are filled into each node based on the associated data. Based on the spatiotemporal constraints of BIM and the real-time progress data of sensors, temporal or spatial relationship edges between procedures are constructed. Based on the binding relationship between resource data and procedures, a dynamic spatiotemporal knowledge graph is constructed.

5. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 1, characterized in that, The process of deploying a lightweight graph neural network model at each node is as follows: Based on the features of nodes and edges in a dynamic spatiotemporal knowledge graph, redundant fully connected layers are removed. The model size is compressed through model quantization and pruning. Adaptation to the hardware architecture of edge nodes is performed on each edge node at the construction site. Each lightweight graph neural network model package is pushed to the corresponding edge node through the local file transfer protocol of the edge network. Based on the unique ID of the construction process, a mapping relationship is established between the model and the process node corresponding to the edge node. The model is started at the edge node, pre-trained parameters are loaded, and fine-tuning is performed using 1 to 2 sets of real-time data of the current process to adapt the model to the feature distribution of the current construction scenario.

6. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 1, characterized in that, The process of capturing the progress status characteristics of each node is as follows: Centered on the target construction process node, based on spatiotemporal logic and resource dependency edges, its preceding construction process, parallel construction process, resource dependency, and 1 to 2 hop neighbor nodes are selected to form a local subgraph; the dynamic attributes of the nodes and the relational attributes of the edges in the local subgraph are encoded into numerical feature vectors, which are used as inputs to the lightweight graph neural network. For the local subgraph, the graph convolutional layer aggregates the feature vectors of neighboring nodes to the target process node based on the relational weights of the edges, updating the feature representation of the target node. Through graph pooling, the updated features of the target node are fused with the global features of the local subgraph to generate an intermediate feature vector. From the embedded feature vector, key dimensions that can represent the progress status of the process are extracted. The extracted progress status features of each edge node are temporarily stored in the local database and labeled with the corresponding timestamp and process ID. Some key features are fed back to the dynamic spatiotemporal knowledge graph to update the progress status label of the corresponding node.

7. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 1, characterized in that, The formation process of the global progress status identification model is as follows: A federated learning coordinator is deployed at the cloud management node of the construction site. Each edge node acts as a client of federated learning and establishes an encrypted communication link with the coordinator. The coordinator sends unified training configuration parameters to all edge nodes, so that the models of all clients start from the same point. Each edge node independently completes model training based on local data, without transmitting raw construction data externally, and generates model parameter update quantities. Each edge node encrypts and uploads the locally trained model parameters to the federated coordinator. The coordinator receives the parameters and verifies their validity. The federated coordinator uses a preset aggregation algorithm to weight and fuse the valid parameters of each node to generate global model parameters. The converged global model parameters are then distributed to each edge node to complete the update and synchronization of the node models, forming the final global progress status recognition model.

8. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 1, characterized in that, The process of outputting real-time progress status is as follows: Each edge node continuously collects real-time data of the construction process and adapts it to the model input format. Each edge node calls the locally deployed global progress status recognition model to complete the inference of real-time data and progress status recognition. The edge node compares the progress status output by the model with the locally stored construction logs and inspection records to verify the consistency of the results. The progress status results after verification and correction are output and fed back to the construction management system through multiple channels.

9. The method for managing the construction progress of cast-in-place box girders based on BIM and edge networks according to claim 1, characterized in that, The process of outputting a near-optimal schedule control strategy is as follows: Based on a BIM digital twin model with mapped real-time progress, a BIM twin simulation training environment is constructed to reproduce the construction process. A two-layer architecture of global coordination agent and process execution agent is adopted. The overall control objectives of the global agent and the local control tasks of the process agents are defined. The action space and state space of the agents are clarified. A local feedback and global decision-making collaboration mechanism is established. A constraint perception module is introduced to filter illegal actions. A dual network structure of strategy network and value network is constructed. A multi-dimensional reward and punishment function with progress and resource dimensions is designed. The agents are trained offline in the simulation environment. The training is terminated when the preset conditions are met, and the global and single-process control strategies are output.

Images