Civil construction task automatic scheduling method and system

Abstract

The application discloses a kind of civil engineering construction task automatic scheduling method and system, including the construction of representing construction task element and its associated relationship Heterogeneous graph model;Utilize graph neural network to process heterogeneous graph model, to excavate the interdependence between process and extract the risk associated features related to uncertainty, form construction state atlas;Based on construction state atlas, construct bayesian network;Through bayesian network, probability reasoning is carried out, the influence of uncertain factor is quantified, and the key information representing risk transmission is fed back to graph neural network, to dynamically correct the identification of process interdependence;Fusion graph neural network output dependency constraint and the risk quantification result output by bayesian network, generate construction task scheduling scheme;According to field construction feedback data, heterogeneous graph model, graph neural network and bayesian network are updated synchronously, to realize the closed-loop iterative optimization of scheduling scheme.

Description

Technical Field

[0001] This invention belongs to the field of intelligent scheduling technology for civil engineering construction, and in particular relates to an automatic scheduling method and system for civil engineering construction tasks. Background Technology

[0002] Civil engineering construction task scheduling is a core aspect of project management. Its goal is to rationally determine the execution sequence, timing, and resource allocation of work processes under various constraints to optimize key performance indicators such as schedule and cost. Traditional scheduling methods are primarily based on network planning techniques, such as Critical Path Method (CPM) and Program Evaluation and Review Technique (PERT). These methods calculate the critical path and schedule by constructing a logical relationship network between work processes, supplemented by Gantt charts for visualization and manual adjustment. With the development of information technology, Building Information Modeling (BIM) technology provides richer 3D spatial and component information for scheduling, supporting 4D construction simulation to identify potential conflicts in advance. In recent years, machine learning, especially deep learning, has been introduced into this field. For example, recurrent neural networks (RNNs) are used to learn from historical progress data to predict project duration, or optimization algorithms are used to solve resource-constrained project scheduling problems. Furthermore, Bayesian networks are also applied independently to construction risk assessment, constructing probabilistic relationships between risk factors and consequences for qualitative risk reasoning and early warning.

[0003] However, the aforementioned existing technologies and methods still have significant limitations and disconnects when dealing with the inherent high complexity, dynamism, and uncertainty of civil engineering construction. First, traditional network planning and BIM models rely on predefined, explicit process logic, failing to automatically identify and express implicit dependencies arising from competition for shared resources, overlapping workspaces, or collaborative workflows between work teams, resulting in an incomplete scheduling foundation model. Second, existing risk assessment methods (such as standalone Bayesian networks) are often separated from the schedule generation process; risk information is mostly qualitative or static probability, making it difficult to dynamically quantify the specific impact and transmission path of particular risk events on complex process chains, leading to delayed and untargeted risk responses. Furthermore, even with advanced algorithms, existing systems often operate independently, lacking a collaborative closed-loop mechanism that can continuously learn from data and feed risk perceptions back to the scheduling logic in real time, resulting in insufficient adaptive capabilities. In addition, existing methods often focus on single or limited objectives such as schedule and cost, making it difficult to collaboratively optimize multiple objectives such as safety, resource balance, and green emission reduction in decision-making. Finally, at the implementation level, existing intelligent solutions are often highly complex and slow to respond, making it difficult to meet the real-time and agile scheduling needs of construction sites. Furthermore, their user-unfriendly human-computer interaction and opaque decision-making processes limit their widespread adoption and application in actual engineering projects. These shortcomings collectively constitute the key technological bottleneck currently facing the improvement of the intelligence and precision of construction scheduling. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, this invention provides an automatic scheduling method and system for civil engineering construction tasks, which solves the problems of inaccurate scheduling and delayed response caused by the difficulty in identifying implicit dependencies, dynamically quantifying risks, and achieving multi-objective collaborative optimization in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] An automatic scheduling method for civil engineering construction tasks includes the following steps:

[0007] S1. Construct a heterogeneous graph model of construction tasks, in which construction process entities, construction resource entities, and construction site condition entities serve as nodes of the heterogeneous graph, and constraints and relationships between processes serve as connecting edges.

[0008] S2, using graph neural networks to perform representation learning on the heterogeneous graph model, and using a feedforward neural network to aggregate and propagate node features in multiple rounds to mine the temporal and resource competition dependencies between processes, and extract risk-related features from the high-dimensional feature vectors of nodes;

[0009] S3, based on the output of step S2, collaboratively constructs a dynamic Bayesian network; specifically including:

[0010] S3.1, Each directed edge in the process dependency relationship identified in step S2 is defined as a parent-child node connection representing conditional probability dependency in the dynamic Bayesian network.

[0011] S3.2, through a parameter generation network, the risk association features extracted in step S2 are mapped and quantified into conditional probability table parameters of the corresponding nodes in the dynamic Bayesian network; the conditional probability table defines the probability distribution of a node in different states given its parent node state.

[0012] S4. Probabilistic reasoning is performed through the dynamic Bayesian network. Variable elimination or sampling algorithms are used to calculate the posterior probability and influence gradient of the process state change caused by uncertain factors. At the same time, the path formed by the node sequence with the largest change in posterior probability in the probability graph is used as a feedback signal to input into the graph neural network to adjust its attention weights in the next training cycle.

[0013] S5. By integrating the set of dependency constraints output by the graph neural network with the risk posterior probability distribution output by the dynamic Bayesian network, a mixed integer programming model is established with the goal of minimizing the total project duration, resource usage variance, and risk expectation value, and the scheduling scheme is generated by solving the model.

[0014] S6. Based on the on-site construction feedback data, the node features and edge weights of the heterogeneous graph model are incrementally updated, and the conditional probability table parameters of the dynamic Bayesian network are corrected using the expectation-maximization algorithm, thereby triggering a new round of joint training and inference of the graph neural network and the dynamic Bayesian network to achieve closed-loop iterative optimization.

[0015] Preferably, in step S1, the node has a multi-dimensional feature vector to characterize the state and attributes of the entity; the edge has a weight scalar to characterize the strength or type of the relationship.

[0016] Preferably, in step S2, the representation learning of the graph neural network is specifically implemented using a graph attention network or a graph convolutional network architecture.

[0017] Preferably, in step S3.2, the parameter generation network is a multilayer perceptron.

[0018] Preferably, in step S5, the optimization objective of the mixed integer programming model also includes minimizing the estimated total carbon emissions during the construction phase.

[0019] Preferably, in step S4, the probabilistic reasoning process also incorporates causal effect analysis based on do-calculus to quantify the net impact of specific risk factors on key processes.

[0020] Preferably, the method further includes step S0: receiving and parsing the dispatch intention instruction from the on-site management personnel, converting it into structured parameters, and synchronously inputting them into the graph neural network in step S2 and the dynamic Bayesian network in step S4 to temporarily adjust their internal calculation benchmarks.

[0021] Preferably, the method further includes step S7: after generating the scheduling scheme in step S5, automatic compliance verification is performed based on preset security specifications and resource quota rules, and the conflict entries found in the verification are input as negative feedback information into the iterative optimization process of step S6.

[0022] Preferably, the graph neural network and dynamic Bayesian network models in steps S1 to S6, after knowledge distillation and quantization compression, are deployed on edge computing devices at the construction site for real-time scheduling and response; the global training and updating of the models are completed on cloud servers.

[0023] Preferably, an automatic scheduling system for civil engineering construction tasks includes:

[0024] The heterogeneous graph construction and management module is used to perform step S1, constructing a heterogeneous graph model of the construction task and describing it with feature vectors and weight scalars.

[0025] The graph neural network processing module is used to execute step S2, which uses a graph attention network or graph convolutional network architecture to perform representation learning on heterogeneous graphs in order to mine dependencies and extract risk features.

[0026] The dynamic Bayesian network collaborative construction module is used to execute steps S3.1 and S3.2, mapping the dependencies to a directed acyclic graph structure of the Bayesian network, and generating network initialization conditional probability table parameters through the parameters constructed by the multilayer perceptron.

[0027] The dynamic Bayesian network reasoning and feedback module is used to execute step S4, using variable elimination or sampling algorithms to perform probabilistic reasoning and causal effect analysis, and outputting information on the critical path of risk transmission as a feedback signal.

[0028] The scheduling scheme generation and optimization module is used to execute step S5, establish and solve a multi-objective mixed integer programming model that integrates dependency constraints, risk probabilities and carbon emission targets, and generate a scheduling scheme.

[0029] The closed-loop iterative optimization module is used to execute step S6, drive the incremental update of the heterogeneous graph model and the parameter correction of the dynamic Bayesian network based on the on-site feedback data, and trigger a new round of joint training.

[0030] Optionally, it also includes at least one of the following auxiliary modules:

[0031] The human-computer interaction instruction parsing module is used in step S0 to parse the scheduling intent and generate structured parameters;

[0032] The automatic compliance verification module is used in step S7 to verify the scheme based on preset rules and generate feedback.

[0033] The edge-cloud collaborative deployment management module is used to manage the deployment of lightweight models at the edge and the collaborative updating of global models in the cloud.

[0034] The technical effects and advantages of the automatic scheduling method and system for civil construction tasks of this invention are as follows:

[0035] 1. This invention significantly improves the accuracy and interpretability of construction scheduling models through a unique collaborative construction mechanism between graph neural networks and Bayesian networks. The complex dependencies automatically mined by graph neural networks are directly converted into interpretable probabilistic graph structures of Bayesian networks. At the same time, the risk features extracted are quantified into explicit probability parameters, making scheduling decisions not only data-driven but also supported by clear probabilistic causal logic.

[0036] 2. This invention establishes a closed-loop optimization capability through bidirectional feedback and continuous learning. The key risk paths inferred by the Bayesian network can dynamically adjust the learning focus of the graph neural network, while on-site data can simultaneously update both networks, enabling the entire scheduling system to possess strong self-evolution and adaptability, and continuously respond to changes in the construction environment.

[0037] 3. This invention achieves automated collaborative optimization of multi-dimensional engineering objectives. The system can automatically quantify and solve multiple key indicators such as project duration, resource utilization, risk cost, and carbon emissions in the scheduling optimization model, thereby generating a comprehensive optimal scheduling scheme, effectively overcoming the limitations of manual scheduling in taking multiple objectives into account. Attached Figure Description

[0038] Figure 1 This is a system block diagram of an automatic scheduling system for civil engineering construction tasks proposed in this invention;

[0039] Figure 2 This is a flowchart of an automatic scheduling method for civil engineering construction tasks proposed in this invention. Detailed Implementation

[0040] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0041] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0042] refer to Figures 1-2This invention provides an automatic scheduling method and system for civil engineering construction tasks, aiming to solve the problem that traditional scheduling methods cannot handle implicit dependencies and dynamic risks. The method first constructs a heterogeneous graph model of construction tasks that integrates processes, resources, site conditions, and risk factors. Then, a graph neural network is used to learn the model's representation, uncovering explicit and implicit dependencies between processes and extracting risk-related features to form a construction state map. Next, a Bayesian network is constructed based on this map, mapping dependencies to its topological structure and quantifying risk features into probability parameters. Real-time probabilistic inference is performed through the Bayesian network to quantify the impact of uncertain risks and analyze their transmission paths, while key information is fed back to the graph neural network to dynamically adjust its learning weights. Furthermore, the dependency constraints of the graph neural network and the risk quantification results of the Bayesian network are integrated to perform multi-objective optimization, generating the final scheduling scheme. Finally, the heterogeneous graph model, graph neural network, and Bayesian network are updated synchronously based on field feedback data, achieving closed-loop iteration and dynamic optimization of the scheduling scheme. Through a deep collaboration and feedback mechanism between two networks, the system achieves accurate characterization of construction complexity, dynamic quantitative assessment of risks, and intelligent balancing of conflicts among multiple objectives.

[0043] Example 1: Core Collaborative Scheduling Process (Construction of Standard Floor in High-Rise Residential Buildings):

[0044] Purpose of implementation:

[0045] This embodiment aims to demonstrate the core technology closed loop of the present invention from data to decision, and to clarify how graph neural networks and Bayesian networks work together to achieve the leap from static planning to dynamic intelligent scheduling.

[0046] Implementation System:

[0047] The system software platform comprises five core modules: digital twin modeling module, intelligent mining module, probabilistic reasoning module, optimization decision-making module, and closed-loop learning module, which are deployed on the project center server.

[0048] Implementation method:

[0049] Taking the standard floor cyclic construction of a 30-story residential building as an example, each floor includes processes such as steel reinforcement, formwork, and concrete.

[0050] S1: Create a construction digital twin model (heterogeneous graph). The system automatically imports BIM and the schedule, generating a graph model with "wall and column reinforcement binding," "wall and column formwork erection," and "beam and slab concrete pouring" as process nodes, and "tower crane A" and "carpentry team 3" as resource nodes. Node attributes include planned duration and required resource quantity; edges represent process logical relationships or resource allocation relationships. Simultaneously, the "probability of strong winds the next day (30%)" provided by the meteorological API is used as a risk attribute and associated with the "external wall formwork erection" node, which is highly susceptible to wind impact.

[0051] S2: Graph Neural Networks Uncover Latent Patterns. The system uses a graph attention network to process the model. After training, the network not only identifies explicit sequences in construction specifications (such as "formwork can only be erected after rebar is completed"), but more importantly, it automatically discovers two implicit dependencies by analyzing historical data: ① Due to the limited capacity of tower cranes, there is an implicit time competition between "core tube shear wall pouring" and "external frame beam and slab pouring"; ② "Precast staircase installation" temporarily occupies "floor passageways," affecting the passage of other trades. These are things that traditional planning cannot reflect.

[0052] S3: Construct a dynamic risk probability model. Transform all dependencies (including implicit dependencies) discovered in the previous step into a Bayesian network structure. Each process step is a variable node, and dependencies form directed edges. Utilize features learned from historical wind events using a graph neural network to initialize the conditional probability parameter for "winds affecting formwork construction efficiency" in the network.

[0053] S4: Risk Inference and Model Enhancement. On the day of construction, the high wind warning was upgraded to a reality. The system input this evidence into the Bayesian network for real-time inference. Network calculations showed that the high probability of delays in "exterior wall formwork erection" would propagate to all subsequent processes, increasing the estimated total project delay risk from 15% to 65%. Simultaneously, the system fed back the conclusion that "exterior wall formwork erection" was identified as a key risk propagation node to the graph neural network. In the next round of training, the network will pay more attention to such key nodes and their associated edges, thus making its future detection more accurate.

[0054] S5: Generate a risk-resistant scheduling plan. Integrating new dependency constraints and high-risk node information, the optimization module recalculates with the goal of minimizing the total project delay due to strong winds. The system-generated plan suggests: temporarily suspending the "outer frame portion of the work" on the non-critical path, prioritizing the concentration of all tower crane and carpentry resources to ensure the construction progress of the "core tube" area, thereby mitigating the impact of strong winds on peripheral operations, and reducing the expected total project delay to 1.5 days.

[0055] S6: Closed-Loop Learning and Adaptation. After the strong winds subside, the system collects actual delay data (e.g., the actual time for formwork erection was 4 hours longer than planned), updates the efficiency parameters of the "Exterior Wall Formwork Erection" node, and uses this data to fine-tune the Bayesian network and graph neural network. In this way, the system increasingly "understands" the characteristics of the project, and its scheduling capabilities continuously evolve.

[0056] Implementation results:

[0057] This embodiment achieves a fundamental shift in scheduling decisions from "based on fixed rules and experience" to "based on data-driven and real-time simulation." The system can automatically identify potential conflicts, quantify risk impacts, and generate proactive response strategies, transforming project management from reactive response to proactive defense.

[0058] Example 2: Fine-grained modeling integrating multi-source data (construction of the main tower of a cross-river bridge):

[0059] Purpose of implementation:

[0060] This embodiment details how to integrate multi-source data such as IoT, BIM, and external databases to construct a highly detailed construction digital twin model that deeply depicts reality and is computable and analyzable.

[0061] Implementation System:

[0062] Based on the system in Example 1, the data access and fusion capabilities have been enhanced, integrating multiple data interfaces such as GPS, sensors, and image recognition.

[0063] Implementation method:

[0064] Take the hydraulic climbing formwork construction of a cross-river bridge tower as an example.

[0065] Extreme refinement of nodes and attributes:

[0066] For the process node "concrete pouring of segment N", its feature vector includes not only volume and duration, but also "real-time data of concrete temperature upon placement" from sensors.

[0067] For the resource node "concrete pump truck", its feature vector dynamically integrates its "real-time GPS location", "fuel balance" and "recent fault code history".

[0068] For the site node "floating concrete mixing plant", its feature vector is linked to "the next day's waterway control plan issued by the maritime department of a certain region" and "the supplier's cementitious material inventory early warning information".

[0069] Edge relationships with dynamic weights:

[0070] The weight of the "Concrete Transportation" edge (connecting the mixing plant and the pump truck) is no longer a fixed value, but a "dynamic transportation time" calculated based on the real-time traffic API.

[0071] The weight of the "spatial conflict" edge (connecting two adjacent work surfaces) is dynamically generated by analyzing the spacing between the working machinery through image recognition of orthophotos taken daily by drones.

[0072] Quantitative embedding of uncertainties: The major risk of "main vessel failure" is linked to the resource node of "large floating crane". The risk value is not estimated manually, but is predicted by a machine learning model based on the floating crane's "equipment life cycle management database" of operating hours, maintenance records and average failure interval of similar equipment, which is the "probability of failure in the next 72 hours".

[0073] Implementation results:

[0074] Through the construction of this embodiment, the status of all elements of the construction site—personnel, machinery, materials, methods, and environment—can be mapped in the digital space in real time and with precision. This ensures that subsequent intelligent analysis is not "water without a source," but is built on a foundation of fresh, high-fidelity data, greatly improving the authenticity and reliability of scheduling decisions.

[0075] Example 3: Green Intelligent Scheduling Embedded with Carbon Emission Constraints (Green Construction of Data Centers):

[0076] Purpose of implementation:

[0077] This embodiment demonstrates how to deeply integrate full life-cycle carbon emission analysis into an intelligent scheduling system, enabling construction organizations to proactively fulfill their social responsibility for green emission reduction while pursuing efficiency.

[0078] Implementation System:

[0079] The system adds a "carbon footprint calculation engine" and a "green optimization target library," which are deeply integrated with the core optimization module.

[0080] Implementation method:

[0081] Take, for example, the prefabricated construction of steel structure and electromechanical components for a data center that is pursuing a high level of green certification.

[0082] Construct a carbon inventory for each process: For each process such as "precast column hoisting", "modular computer room assembly", and "photovoltaic panel installation", calculate an accurate "carbon emission equivalent" label based on its material usage, machinery model and energy consumption, and transportation distance.

[0083] Extended multi-objective optimization model: In the scheduling optimization model, "total carbon emissions during the construction phase" is formally introduced as an optimization objective alongside schedule and cost. The system needs to find the optimal schedule under carbon emission budget constraints.

[0084] Graph Neural Networks Learn Low-Carbon Correlation: When training a graph neural network, in addition to learning the spatiotemporal and resource dependencies between processes, it is also guided to identify "low-carbon patterns." For example, the network may automatically conclude that "large-scale hoisting operations (dependent on grid power) during peak daytime electricity load periods will lead to a surge in indirect carbon emissions," thus implicitly linking "operation time" with "carbon emissions."

[0085] Bayesian network carbon risk assessment: A "daily carbon emission peak exceeding the limit" risk node is set in the Bayesian network. The state probability of this node is affected by multiple parent nodes, such as "concentration of high-carbon processes", "whether it is during peak electricity consumption period", and "real-time power generation of renewable energy (on-site photovoltaic)". The system can simulate the probability of triggering carbon risk under different scheduling schemes.

[0086] Generating and Selecting Green Solutions: After optimization, the system provides a set of Pareto optimal solutions in three-dimensional coordinates (project duration, cost, and carbon emissions). The project team can choose one of the "green preference" solutions: This solution intelligently adjusts the sequence of work processes, scheduling most high-energy-consuming operations at night (utilizing off-peak electricity prices and a higher proportion of clean electricity), and matches it with the on-site photovoltaic power generation cycle. Ultimately, it is expected to reduce phased carbon emissions by 12% while keeping the total project duration unchanged.

[0087] Implementation results:

[0088] This embodiment transforms green construction from a concept and slogan and post-event statistics into a quantifiable, optimizable, and executable proactive intelligent decision-making process. It enables project teams to proactively design and select low-carbon construction paths during the planning phase, providing core technological tools for achieving higher green building standards.

[0089] Example 4: An agile response system with edge-cloud collaboration (urban subway tunnel excavation):

[0090] Purpose of implementation:

[0091] This embodiment addresses the engineering challenge of achieving "real-time response" for complex AI models on construction sites with limited resources and variable network conditions. Through an edge-cloud collaborative architecture, it ensures that intelligent scheduling can not only quickly respond to changes on-site but also continuously accumulate and evolve knowledge.

[0092] Implementation System:

[0093] A distributed hybrid cloud architecture is adopted. Edge computing stations equipped with lightweight inference engines are deployed at the tunnel entrances; the group headquarters has a cloud-based AI training platform with full data and heavy-duty algorithms.

[0094] Implementation method:

[0095] It was applied in the construction of a subway shield tunnel section in a certain city.

[0096] Edge: Millisecond-level real-time response:

[0097] Deploying lightweight graph neural networks and Bayesian networks after model compression and pruning results in a size that is only 1 / 10th the size of the cloud version.

[0098] It receives real-time data from the tunnel boring machine's PLC (thrust, torque, and excavated soil) and data from tunnel convergence monitoring sensors.

[0099] A rapid inference is run every 30 seconds: when the monitored data is abnormal, the risk is immediately assessed based on the latest local heterogeneity map (such as "the strata ahead may soften"), and an instant command is generated to adjust the tunnel boring machine parameters (such as "it is recommended to reduce the advance speed by 10% and increase the synchronous grouting pressure by 5%) to ensure construction safety.

[0100] Cloud-based: Global Optimization and Continuous Learning

[0101] We collect edge data from all lines under construction and retrain and optimize the entire model every night.

[0102] Deep-seated patterns were discovered across projects, such as: "When traversing a specific silty sand stratum, if the soil excavation rate exceeds 103% for three consecutive rings, the probability of excessive surface subsidence within the subsequent three rings increases to 40%." This new knowledge was solidified into the knowledge base of the Bayesian network.

[0103] Develop macro-level plans for the next 24 hours, such as the supply schedule of tunnel sections and the scheduling plan for construction waste transportation vehicles, and distribute them to the edge as boundary constraints.

[0104] Highly efficient collaboration: The edge acts like "nerve endings," processing rapid, reflexive decisions; the cloud acts like the "brain," conducting deep thinking and strategic planning. Edge model parameters are synchronized and updated weekly from the cloud, ensuring that the "ends" also possess the latest "collective intelligence."

[0105] Implementation results:

[0106] This architecture perfectly balances the conflict between real-time performance and intelligence. Edge computing ensures that on-site operations still receive basic intelligent safety guarantees in the event of network interruptions or delays; cloud learning ensures that the entire system becomes smarter with use, achieving a leap from "project intelligence" to "enterprise intelligence".

[0107] Example 5: Natural Interaction and Visualized Decision-Making Cockpit (General Contracting Management of Large Stadiums):

[0108] Purpose of implementation:

[0109] This embodiment focuses on improving the system's ease of use and decision-making intuitiveness. Through natural human-computer interaction and powerful visualization technology, it makes the complex AI decision-making process transparent and credible, achieving a deep integration of human wisdom and machine intelligence.

[0110] Implementation System:

[0111] The system is equipped with a BIM-based 3D visualization decision-making cockpit and integrates natural interaction interfaces such as voice recognition and gesture recognition.

[0112] Implementation method:

[0113] During the peak construction period of the steel structure and membrane structure of a large stadium, the chief commander used this system.

[0114] Natural language command input: The commander-in-chief observed the delay in the hoisting of the east grandstand in front of the 3D model and said directly: "The hoisting of the east roof is the key this month, give it priority, and send a 400-ton crane from the west to provide support." The voice recognition module converted the command into a structured command: {"Strengthen the target": "East roof hoisting", "Resource adjustment": "Deploy crane CX-400 from the west to the east"}.

[0115] Intelligent Fusion and Solution Regeneration: The command was simultaneously sent to the system. The graph neural network temporarily increased the attention given to nodes and edges related to "East Zone Roof Erection"; the Bayesian network temporarily increased the probability of derivative risks that might arise from "Reduction of Crane Resources in the West Zone". Based on this new "preference", the system quickly recalculated and generated multiple alternative solutions. One of the solutions showed that moving the crane would delay the installation of some substructures in the West Zone by 2 days, but would have no impact on the overall project schedule.

[0116] Augmented Reality (AR) Sand Table Simulation and Compliance Verification: The chief commander put on AR glasses, and the plan was overlaid as a virtual image on the real construction site sand table. He could see the virtual crane movement path and the new work sequence. At the same time, the system automatically performed compliance verification and popped up a prompt: "The plan is feasible. Note: The relocation of the CX-400 crane requires a temporary road occupancy permit, which is expected to take 4 hours; under the new hoisting sequence, the hoisting operation of the 5th section in the East Zone will overlap with the membrane structure tensioning operation for 24 hours, with a safety risk level of 'medium'. It is recommended to add temporary protective netting."

[0117] Visualized Decision-Making and Order Issuance: Ultimately, all information converges on the large screen in the cockpit: on the left is a 4D dynamic simulation, vividly demonstrating the new plan; in the middle is a risk heat map, showing the changes in risk values ​​in each area after resource adjustments; on the right is a resource load curve. After comprehensive perception, the commander-in-chief confirms the plan and issues it with one click. Dispatch instructions are automatically broken down into task cards and pushed to the mobile terminals of relevant foremen and crane operators.

[0118] Implementation results:

[0119] This embodiment changes the way people use intelligent systems. Managers do not need to understand complex algorithms; they only need to give instructions, and the system will transform them into precise solutions and present them in the most intuitive way.

[0120] Comparative Example 1

[0121] Comparison with traditional critical path method and independent early warning system:

[0122] Purpose of comparison:

[0123] By comparing the present invention with the most commonly used traditional methods in current engineering practice under the same scenarios and conditions, the present invention objectively and quantitatively demonstrates the actual technical problems solved by the present invention and the significant benefits it brings.

[0124] Comparison Systems and Methods:

[0125] The comparison adopts a classic combination commonly used in the industry: using mainstream project management software to develop detailed schedules based on the critical path method, and using the company's unified SMS mass-messaging early warning platform to receive external early warning information such as weather forecasts. The management team relies entirely on regular meetings, manual analysis of reports, and telephone communication for scheduling and adjustments.

[0126] Comparison of scenarios and processes:

[0127] Imagine a commercial complex project identical to that in Example 1, encountering the same continuous heavy rainfall. Compare the entire process of the two methods from the onset of the risk to its resolution.

[0128] During the risk warning phase, traditional management teams receive text messages such as "Orange Rainstorm Warning Issued." This information is qualitative; the team only knows "the rain is heavy," but cannot quantify the impact on specific aspects like "concrete curing," "exterior wall plastering," and "tower crane efficiency," let alone predict the cascading effects on the overall project timeline. Decisions are based on vague, empirical judgments. In contrast, this invention's system outputs quantitative probability reports, clearly identifying key vulnerabilities.

[0129] During the scheduling and adjustment phase, traditional methods require emergency meetings and manual task rearrangement in software, a process that takes several hours and offers limited options, often only allowing for simple delays. This invention's system can automatically generate multiple data-driven alternative optimization solutions within minutes, providing the expected results for each solution (delay days, cost changes, and risk level).

[0130] In terms of conflict foresight, traditional manual adjustments are prone to causing secondary problems, such as overcrowding and decreased efficiency due to the relocation of large numbers of workers to indoor operations. The system of this invention, leveraging its graph neural network's understanding of implicit relationships, can automatically avoid such conflicts during the solution generation phase.

[0131] In terms of balancing objectives, traditional methods often sacrifice cost and safety to meet deadlines under pressure. The system of this invention, however, consistently adheres to multi-objective optimization, outputting a solution that achieves the optimal balance among multiple dimensions such as schedule, cost, risk, and resources.

[0132] Comparison results:

[0133] Simulation results show that, in response to the same heavy rainfall event: projects using traditional methods experienced a 4-day delay in the overall project duration, with costs exceeding the budget by approximately 15% due to rushing and inefficiency, and a minor safety incident occurred. In contrast, projects using the system of this invention saw the overall project duration delay controlled within 2 days, with a cost overrun of only 7%. The entire process was smooth and orderly, with no secondary problems or safety incidents.

[0134] Compared to Examples 1-5 and Comparative Example 1, this invention, through five progressive examples and one comparative example, clearly outlines the fundamental generational difference between traditional construction scheduling methods and the intelligent collaborative scheduling system of this invention. The traditional method represented by the comparative example is static, linear, and reliant on human experience. It formulates plans based on fixed critical paths and passively responds through isolated early warning information. Managers, like those in a fog, rely on experience and intuition to direct operations, possessing only a qualitative perception of risk and unable to quantify its specific impact. Adjustments are generated through manual meetings and dragging Gantt charts, which is time-consuming, labor-intensive, and makes it difficult to foresee secondary conflicts. The decision-making process often falls into a single-objective mindset of "ensuring the project deadline," easily sacrificing cost and safety. The entire process is reactive, localized, and knowledge cannot be effectively accumulated. In contrast, the system of this invention, as shown in Examples 1-5, is dynamic, network-based, and data-driven at its core. It transforms the construction site into a deeply interconnected heterogeneous graph digital twin model, automatically mining implicit dependencies, including resource competition and spatial conflicts, through graph neural networks, providing managers with an "X-ray machine" to understand the internal connections of complex systems. More importantly, the system uses Bayesian networks to transform uncertain risks (such as weather and equipment failure) into calculable and predictable probabilistic events, achieving a quantitative leap from "it might rain" to "rain has a 70% probability of causing a 2-hour delay in process A, thus increasing the total project duration risk by 40%." This transforms risk response from post-event remediation to pre-event planning.

[0135] This fundamental difference directly leads to a stark contrast in practical results. Traditional methods, when dealing with sudden risks (such as the heavy rainfall in the implementation case), resemble a chaotic skirmish, characterized by slow response, crude decision-making, often at the cost of high cost overruns and project delays, and prone to secondary safety issues. In contrast, the intelligent system of this invention acts like a "scheduling brain" with supercomputing power and a global perspective, capable of multi-objective collaborative optimization, generating multiple alternative solutions within minutes that balance project time, cost, resource load, and carbon emissions. It not only responds quickly but also anticipates and avoids secondary conflicts arising from adjustments. Examples 4 and 5 further demonstrate that this system achieves a combination of "local real-time response" and "global continuous evolution" through an edge-cloud collaborative architecture, and integrates "human strategic intent" with "precise machine calculation" through a natural interactive interface. The final results are strikingly different: under the same risk impact, traditional methods result in a 4-day project delay, a 15% cost overrun, and accompanying safety incidents; while the system of this invention controls the delay to within 2 days, with a cost overrun of only 7%, and the process is smooth and orderly. This proves that the present invention is not only an efficiency tool, but also a paradigm shift that can improve the overall resilience and scientific decision-making of construction projects.

[0136] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations 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. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0137] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic scheduling method for civil engineering construction tasks, characterized in that, Includes the following steps: S1, construct a heterogeneous graph model of construction tasks, where... Construction process entities, construction resource entities, and construction site condition entities are used as nodes in the heterogeneous graph, and the constraints and relationships between processes are used as connecting edges. S2, using graph neural networks to perform representation learning on the heterogeneous graph model, and using a feedforward neural network to aggregate and propagate node features in multiple rounds to mine the temporal and resource competition dependencies between processes, and extract risk-related features from the high-dimensional feature vectors of nodes; S3, based on the output of step S2, collaboratively constructs a dynamic Bayesian network; specifically including: S3.1, Each directed edge in the process dependency relationship identified in step S2 is defined as a parent-child node connection representing conditional probability dependency in the dynamic Bayesian network. S3.2, through a parameter generation network, the risk association features extracted in step S2 are mapped and quantified into conditional probability table parameters of the corresponding nodes in the dynamic Bayesian network; the conditional probability table defines the probability distribution of a node in different states given its parent node state. S4. Probabilistic reasoning is performed through the dynamic Bayesian network. Variable elimination or sampling algorithms are used to calculate the posterior probability and influence gradient of the process state change caused by uncertain factors. At the same time, the path formed by the node sequence with the largest change in posterior probability in the probability graph is used as a feedback signal to input into the graph neural network to adjust its attention weights in the next training cycle. S5. By integrating the set of dependency constraints output by the graph neural network with the risk posterior probability distribution output by the dynamic Bayesian network, a mixed integer programming model is established with the goal of minimizing the total project duration, resource usage variance, and risk expectation value, and the scheduling scheme is generated by solving the model. S6. Based on the on-site construction feedback data, the node features and edge weights of the heterogeneous graph model are incrementally updated, and the conditional probability table parameters of the dynamic Bayesian network are corrected using the expectation-maximization algorithm, thereby triggering a new round of joint training and inference of the graph neural network and the dynamic Bayesian network to achieve closed-loop iterative optimization.

2. The automatic scheduling method for civil construction tasks as described in claim 1, characterized in that, In step S1, the node has a multi-dimensional feature vector to characterize the state and attributes of the entity; the edge has a weight scalar to characterize the strength or type of the relationship.

3. The automatic scheduling method for civil construction tasks as described in claim 2, characterized in that, In step S2, the representation learning of the graph neural network is specifically implemented using a graph attention network or a graph convolutional network architecture.

4. The automatic scheduling method for civil construction tasks as described in claim 1, characterized in that, In step S3.2, the parameter generation network is a multilayer perceptron.

5. The automatic scheduling method for civil construction tasks as described in claim 4, characterized in that, In step S5, the optimization objective of the mixed integer programming model also includes minimizing the estimated total carbon emissions during the construction phase.

6. The automatic scheduling method for civil construction tasks as described in claim 1, characterized in that, In step S4, the probabilistic reasoning process also incorporates causal effect analysis based on do-calculus to quantify the net impact of specific risk factors on key processes.

7. The automatic scheduling method for civil construction tasks as described in claim 1, characterized in that, The method further includes step S0: receiving and parsing the dispatch intention instruction from the on-site management personnel, converting it into structured parameters, and synchronously inputting it into the graph neural network in step S2 and the dynamic Bayesian network in step S4 to temporarily adjust their internal calculation benchmarks.

8. The automatic scheduling method for civil construction tasks as described in claim 1, characterized in that, The method further includes step S7: after generating the scheduling scheme in step S5, automatic compliance verification is performed based on the preset security specifications and resource quota rules, and the conflict entries found in the verification are input as negative feedback information into the iterative optimization process of step S6.

9. The automatic scheduling method for civil construction tasks as described in claim 1, characterized in that, The graph neural network and dynamic Bayesian network models in steps S1 to S6, after knowledge distillation and quantization compression, are deployed on edge computing devices at the construction site for real-time scheduling and response; the global training and updating of the models are completed on cloud servers.

10. An automatic scheduling system for civil engineering construction tasks, characterized in that, For implementing the method of any one of claims 1 to 9, comprising: The heterogeneous graph construction and management module is used to perform step S1, constructing a heterogeneous graph model of the construction task and describing it with feature vectors and weight scalars. The graph neural network processing module is used to execute step S2, which uses a graph attention network or graph convolutional network architecture to perform representation learning on heterogeneous graphs in order to mine dependencies and extract risk features. The dynamic Bayesian network collaborative construction module is used to execute steps S3.1 and S3.2, mapping the dependencies to a directed acyclic graph structure of the Bayesian network, and generating network initialization conditional probability table parameters through the parameters constructed by the multilayer perceptron. The dynamic Bayesian network reasoning and feedback module is used to execute step S4, using variable elimination or sampling algorithms to perform probabilistic reasoning and causal effect analysis, and outputting information on the critical path of risk transmission as a feedback signal. The scheduling scheme generation and optimization module is used to execute step S5, establish and solve a multi-objective mixed integer programming model that integrates dependency constraints, risk probabilities and carbon emission targets, and generate a scheduling scheme. The closed-loop iterative optimization module is used to execute step S6, drive the incremental update of the heterogeneous graph model and the parameter correction of the dynamic Bayesian network based on the on-site feedback data, and trigger a new round of joint training. When the system is used to implement the method of claim 7, the system further includes: The human-computer interaction instruction parsing module is used to execute step S0, parse the scheduling intent, and generate structured parameters; When the system is used to implement the method of claim 8, the system further includes: The automatic compliance verification module is used to execute step S7, verify the scheme based on preset rules and generate feedback; The system also includes: The edge-cloud collaborative deployment management module is used to manage the deployment of lightweight models at the edge and the collaborative updating of global models in the cloud.

Images