Engineering resource allocation optimization method and system based on artificial intelligence

Abstract

The invention discloses an artificial intelligence-based engineering resource allocation optimization method and system, and relates to the technical field of artificial intelligence and resource management crossing. The problems of resource configuration dynamic change and resource conflict coordination are solved. And processing the multi-modal heterogeneous data in the target engineering scene through space-time alignment and semantic coding to generate a dynamic data stream. A task, resource and environment node relationship is constructed based on a dynamic heterogeneous graph network, and a node dependency weight is updated through an event-driven mechanism. Two-stage collaborative optimization is executed, in the first stage, a baseline scheme is generated through Monte Carlo sampling, and in the second stage, resource conflicts are eliminated through back propagation negotiation. The resource dynamic entropy is monitored in real time, a rebalance algorithm is triggered during local overload, and a distribution scheme is adjusted in combination with security constraints. And finally, outputting the optimization scheme to an engineering management system for execution, dynamically returning updated graph network parameters, forming a data acquisition, optimization decision and feedback closed loop, and improving the intelligence and robustness of resource allocation.

Description

Technical Field

[0001] The present invention relates to the technical field of intersection of artificial intelligence and resource management, and specifically to an engineering resource allocation optimization method and system based on artificial intelligence. Background Art

[0002] As global engineering projects continue to expand in scale and complexity, the efficient allocation and dynamic optimization of engineering resources have become core issues for ensuring project progress, cost control, and resource utilization. Especially in areas such as large-scale infrastructure construction, intelligent manufacturing, and smart city development, engineering projects face challenges such as multi-trade collaboration, fluctuating resource supply chains, and high uncertainty in the construction environment. Traditional static resource allocation models struggle to adapt to dynamically changing engineering scenarios. Artificial intelligence technologies, particularly the development of multimodal data processing, graph neural networks, and intelligent optimization algorithms, are providing new technical means for intelligent perception, dynamic modeling, and optimized decision-making of engineering resources, driving the transformation of engineering management models from experience-driven to data-driven and intelligent decision-making.

[0003] Existing engineering resource allocation methods are often based on static rule-making, heuristic algorithms, or simplified assumptions for optimization, which presents numerous limitations. On the one hand, traditional methods often struggle to fully integrate multimodal, heterogeneous data, resulting in decision-making models with poor adaptability to complex engineering scenarios. On the other hand, existing optimization algorithms often employ single-stage static solutions and lack the ability to model resource demand uncertainty, dynamic conflict mediation, and local anomaly response mechanisms. This can easily lead to deviations between resource allocation results and the actual engineering execution process, resulting in wasted resources, project delays, and increased risk. Furthermore, existing systems generally lack closed-loop optimization mechanisms based on real-time feedback, making it impossible to dynamically adjust optimization strategies based on execution data, limiting the system's adaptability and optimization accuracy. Summary of the Invention

[0004] In response to the deficiencies of the prior art, the present invention provides an engineering resource allocation optimization method and system based on artificial intelligence, which solves the problems of the above-mentioned background technology.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: an engineering resource allocation optimization method based on artificial intelligence, comprising the following steps: S1: real-time collection of multimodal heterogeneous data in the target engineering scenario, including task attribute data, resource status data and environmental interference data, spatiotemporal alignment and semantic encoding of the multimodal heterogeneous data, and generation of a multimodal dynamic data stream; S2: constructing a dynamic heterogeneous graph network based on the multimodal dynamic data stream, defining the heterogeneous interaction relationship between task nodes, resource nodes and environmental nodes, calculating the dynamic dependency weights between nodes based on the multimodal graph attention mechanism, and updating the edge weights through an event-driven mechanism; S3: performing two-stage collaborative optimization to map the dynamic dependency relationship output by the dynamic heterogeneous graph network into quantifiable constraints; S3.1: the first stage: based on Monte Carlo sampling Simulate external interference scenarios, generate a probability distribution of resource demands, and construct a stochastic optimization model to solve the baseline allocation plan that minimizes the risk-weighted cost; S3.2: The second stage: decompose the baseline plan into a local decision-making problem of multiple seed agents, coordinate resource conflicts between agents through the backpropagation negotiation protocol, and output the allocation plan after the conflict is resolved; S4: Monitor the dynamic resource entropy indicator of resource distribution in real time. When a local overload is detected, trigger the gradient-driven resource rebalancing algorithm, adjust the allocation weight in the direction of utility maximization, and embed a safety constraint verification layer to filter illegal solutions, and output the optimized allocation plan; S5: Output the optimized allocation plan to the engineering management system for execution, and at the same time collect resource utilization and construction period deviation indicators in the actual execution data, and dynamically transmit them back to the dynamic heterogeneous graph network to update the edge weight parameters.

[0006] Furthermore, S1 includes the following contents: S1.1: Collect multimodal heterogeneous data through the interface between the Internet of Things terminal and the engineering management system: Task attribute data includes the duration, priority, task dependency, and real-time progress deviation rate in the task decomposition structure; Resource status data includes the human resource skill matrix, the availability timing status of mechanical equipment, and the dynamic consumption rate of material inventory; Environmental interference data includes real-time meteorological monitoring indicators, supply chain delay warning signals, and equipment failure event logs; S1.2: Perform spatiotemporal alignment on the collected heterogeneous data, and build a unified spatiotemporal index based on timestamp synchronization and spatial position mapping rules; S1.3: Apply semantic coding strategies to the spatiotemporally aligned data to extract task description features, resource attribute features, and environmental status features to generate a multimodal dynamic data stream in a unified format.

[0007] Furthermore, a dynamic heterogeneous graph network is constructed based on multimodal dynamic data flows, and the specific process of defining the heterogeneous interaction relationships among task nodes, resource nodes, and environment nodes is as follows: S2.1: Node classification and attribute definition: Task nodes are generated based on atomic tasks in the task decomposition structure, and node attributes include duration, priority, and dependency strength; resource nodes are divided into three sub-nodes: manpower, machinery, and materials, and attributes include skill level, available period, and safety stock threshold; environment nodes dynamically generate external interference event entities, and attributes include event type, geographical impact radius, and duration; S2.2: Heterogeneous relationship modeling and edge attribute binding: The task-resource edge defines the demand intensity coefficient and binds spatiotemporal constraints to limit resources to take effect in specific geographical areas or time windows; the resource-environment edge establishes a vulnerability score for resource nodes to environmental interference, and quantifies the impact of environmental events on resource availability; the task-environment edge associates the direct weight of external events on task progress, and dynamically adjusts the priority of task nodes.

[0008] Furthermore, the dynamic dependency weights between nodes are calculated according to the multimodal graph attention mechanism, and the logical process of the event-driven edge weight update mechanism is as follows: S2.3: Multimodal attention calculation: A scalar attention mechanism is used for numerical resource data to calculate the linear dependency strength between task nodes and resource nodes; a temporal attention mechanism is used for sequential task data to capture the forward and backward driving dynamic relationship between tasks; a spatial attention mechanism is used for image-type environmental data to extract the resource distribution hotspot features in the construction site monitoring screen; S2.4: Event-driven dynamic update of edge weights: When equipment failure or sudden weather change events are detected, the affected subgraph area in the dynamic heterogeneous graph network is located; according to the predefined event impact rule library, the event type and node attributes are matched, and the weight parameters of the associated edges are locally updated; the original weights of the unaffected areas are retained, and only the event-related subgraphs are incrementally updated to avoid global model retraining.

[0009] Furthermore, S3.1 includes the following contents: S3.1.1: Define a set of external interference scenarios: Based on the historical engineering event library and real-time environmental monitoring data, extract three typical interference events of equipment failure, supply chain interruption, and extreme weather, and construct the probability of event occurrence and impact intensity distribution; S3.1.2: Perform Monte Carlo sampling: Within the preset time window, generate multiple sets of interference scenario sequences according to the event probability distribution to simulate the fluctuation pattern of resource demand; S3.1.3: Construct a risk-weighted cost model: Integrate the resource demand probability distribution output by the dynamic heterogeneous graph network with the interference scenario simulation results, define the cost function as the weighted sum of the expected value of the resource overrun cost and the conditional risk value, and generate a baseline allocation plan.

[0010] Furthermore, S3.2 includes the following contents: S3.2.1: Define a dynamic utility function for each worker seed agent, which includes three configurable priority dimensions: task progress gain, resource conflict penalty, and safety constraint compliance, and dynamically calibrate weight parameters based on real-time engineering events; S3.2.2: Construct a directional communication link network, and each agent only exchanges the current resource occupancy weight, utility gradient direction, and conflict intensity index with adjacent agents with resource conflicts; S3.2.3: Iteratively correct the local allocation weight according to the gradient negotiation message, and judge convergence when the conflict intensity change rate of multiple consecutive iterations is lower than the preset threshold, and output the optimization plan after conflict resolution.

[0011] Furthermore, S3.2.3 includes the following contents: S3.2.3.1: Structured design of gradient negotiation message: The gradient negotiation message is defined to contain three core fields: resource occupancy weight, utility gradient direction and conflict intensity score. The utility gradient direction is generated by calculating the partial derivative of the dynamic utility function on the resource weight; the message transmission cycle is dynamically bound to the engineering scenario, high-frequency events trigger instant transmission, and low-frequency scenarios are transmitted at fixed intervals; S3.2.3.2: Iterative weight correction rule: Each agent updates the resource weight according to the comprehensive value of the local utility gradient direction and the gradient reverse compensation of the conflicting party based on the received gradient message, and introduces a momentum factor to smooth the iterative process to avoid weight oscillation; S3.2.3.3: Convergence judgment and output: When the change rate of the conflict intensity score for three consecutive iterations is lower than the preset threshold, it is judged to be converged, and the final weight is output to the dynamic heterogeneous graph network to generate an optimization plan after conflict resolution.

[0012] Furthermore, S4 includes the following contents: quantifying the degree of chaos caused by local resource overload based on the spatial density difference and time decay effect of resource distribution, where the spatial density difference is calculated through the variance of regional resource occupancy rate, and the time decay effect is generated by weighting the idle time of resources; adjusting the allocation weight along the gradient ascent direction of the resource utility function, and using the projected gradient method to constrain the weight correction within the safety and compliance range to avoid the generation of illegal solutions; for high entropy areas, prioritizing the allocation of idle resources and synchronously updating the resource occupancy status of adjacent areas; defining hard constraints through the engineering safety specification library, including the minimum manpower configuration threshold, equipment safety redundancy and the lower limit of material inventory; calling the linear programming solver to verify the allocation plan, filtering out illegal solutions that violate the constraints, and outputting the final compliant allocation plan.

[0013] Furthermore, the resource utilization and duration deviation indicators in the actual execution data are collected and dynamically transmitted back to the dynamic heterogeneous graph network to update the edge weight parameters. The specific process is as follows: resource utilization indicators, including the proportion of mechanical equipment working hours, manpower load rate and material consumption rate, are collected in real time through the Internet of Things terminal of the engineering management system, the duration deviation characteristics are extracted, and the deviation rate between the actual progress of the critical path task and the planned progress is calculated; the demand intensity coefficient of the task-resource edge is reversely adjusted according to the resource utilization deviation, and the demand weight of the task node with lower-than-expected resource utilization is reduced, and vice versa, the priority is increased; the dependency weight of the task node is dynamically corrected in combination with the duration deviation rate, and the dependency intensity of the task node with a deviation rate exceeding the threshold is weakened; the updated edge weight parameters are transmitted back to the dynamic heterogeneous graph network, and only the edges of the affected subgraph are incrementally adjusted to maintain the stability of the global network structure, and historical execution data is filtered based on the sliding window mechanism to prevent noise data from interfering with the model parameters.

[0014] The engineering resource allocation optimization system based on artificial intelligence includes the following modules: heterogeneous data acquisition and encoding module, dynamic heterogeneous graph construction module, collaborative optimization decision module, resource dynamic monitoring module, and scheme execution module; the heterogeneous data acquisition and encoding module is used to collect multimodal heterogeneous data in the target engineering scenario in real time, including task attribute data, resource status data and environmental interference data, perform spatiotemporal alignment and semantic encoding on the multimodal heterogeneous data, and generate a multimodal dynamic data stream; the dynamic heterogeneous graph construction module is used to construct a dynamic heterogeneous graph network according to the multimodal dynamic data stream, define the heterogeneous interaction relationship between task nodes, resource nodes, and environment nodes, calculate the dynamic dependency weights between nodes according to the multimodal graph attention mechanism, and update the edge weights through an event-driven mechanism: the collaborative optimization decision module is used to perform two-stage collaborative optimization, and the dynamic dependency relationship output by the dynamic heterogeneous graph network is updated. The system is mapped into quantifiable constraints, and external interference scenarios are simulated according to Monte Carlo sampling to generate a probability distribution of resource requirements. A stochastic optimization model is constructed to solve a baseline allocation plan that minimizes risk-weighted costs. The baseline plan is decomposed into a local decision-making problem of multiple seed agents. Resource conflicts between agents are coordinated through a backpropagation negotiation protocol, and an allocation plan after conflict resolution is output; the resource dynamic monitoring module is used to monitor the dynamic resource entropy index of resource distribution in real time. When a local overload is detected, a gradient-driven resource rebalancing algorithm is triggered to adjust the allocation weight in the direction of utility maximization, and a safety constraint verification layer is embedded to filter illegal solutions, and the optimized allocation plan is output; the plan execution module is used to output the optimized allocation plan to the engineering management system for execution, while collecting resource utilization and construction period deviation indicators in the actual execution data, and dynamically transmitting them back to the dynamic heterogeneous graph network to update the edge weight parameters.

[0015] The present invention has the following beneficial effects:

[0016] (1) The engineering resource allocation optimization method based on artificial intelligence collects multimodal heterogeneous data in the target engineering scenario in real time, aligns the multimodal heterogeneous data in time and space, and performs semantic encoding on the multimodal heterogeneous data to generate a multimodal dynamic data stream, thereby achieving unified fusion and dynamic modeling of multi-source data on tasks, resources, and environments, improving the continuity and expression accuracy of the data foundation, and providing high-quality input for subsequent optimization. By constructing a dynamic heterogeneous graph network based on the multimodal dynamic data stream, defining the heterogeneous interaction relationship between task nodes, resource nodes, and environment nodes, and calculating the dynamic dependency weights between nodes based on the multimodal graph attention mechanism, and adopting an event-driven edge weight update mechanism, the complex dynamic relationship between tasks, resources, and environmental factors can be accurately characterized and updated, which greatly improves the responsiveness and adaptability of resource allocation decisions to environmental changes.

[0017] (2) The AI-based engineering resource allocation optimization system monitors the dynamic resource entropy indicators of resource distribution in real time, triggers a gradient-driven resource rebalancing algorithm when a local overload is detected, adjusts the allocation weights in the direction of utility maximization, and embeds a safety constraint verification layer to filter illegal solutions, thus achieving dynamic adaptive adjustment of resource allocation status and prevention of illegal allocations, effectively reducing the risk of local resource overload and the abnormal rate of engineering execution. By outputting the optimized allocation plan to the engineering management system for execution, and collecting resource utilization and construction period deviation indicators from the actual execution data, and dynamically transmitting them back to the dynamic heterogeneous graph network to update the edge weight parameters, a closed-loop feedback mechanism between allocation optimization and actual execution data is achieved, promoting system self-learning and continuous decision-making optimization, and further improving the accuracy of resource allocation and the intelligence level of engineering management.

[0018] Of course, any product implementing the present invention does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 This is a flow chart of the engineering resource allocation optimization method based on artificial intelligence of the present invention.

[0020] Figure 2 This is a flow chart of the engineering resource allocation optimization system based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0021] The embodiment of the present application solves the problems of low resource utilization and high project duration risk caused by insufficient heterogeneous data fusion, inefficient resource conflict coordination, and delayed response to external disturbances in the dynamic configuration of engineering resources through an artificial intelligence-based engineering resource allocation optimization method and system. By collecting multimodal heterogeneous data and performing spatiotemporal alignment and semantic encoding, a dynamic heterogeneous graph network is constructed to model the dynamic relationship between tasks, resources, and the environment; based on the multimodal graph attention mechanism and event-driven edge weight update, rapid modeling of resource demand changes is achieved; a two-stage collaborative optimization strategy is adopted to take into account external disturbance simulation and local conflict coordination to generate a globally optimized and highly robust allocation plan; by real-time monitoring of resource dynamic entropy and triggering gradient-driven resource rebalancing, it is ensured that the system can still stably optimize resource allocation under local overload conditions; at the same time, resource utilization and project duration deviation data are collected and transmitted back to the dynamic heterogeneous graph network to form a closed-loop optimization mechanism, which significantly improves the intelligent level of resource allocation and the timeliness and reliability of project execution.

[0022] The overall idea of the solution in the embodiments of this application is as follows:

[0023] First, based on the target engineering scenario, multimodal heterogeneous data, including task attributes, resource status, and environmental interference, is collected in real time. Through spatiotemporal alignment and semantic encoding, a continuously updated multimodal dynamic data stream is formed, providing high-dimensional input for subsequent modeling. Subsequently, a dynamic heterogeneous graph network is constructed, defining the heterogeneous interactions between task nodes, resource nodes, and environment nodes. A multimodal graph attention mechanism and an event-driven edge weight update mechanism are applied to dynamically characterize changes in node dependencies, enhancing the model's sensitivity and adaptability to environmental fluctuations.

[0024] During the optimization phase, a two-stage collaborative optimization strategy is employed. First, Monte Carlo sampling is used to simulate external interference scenarios, generating a probability distribution of resource demands. Based on this distribution, a stochastic optimization model is constructed to determine a baseline resource allocation solution that minimizes risk-weighted costs. Sub-agent local decision-making and backpropagation negotiation mechanisms are then used to resolve resource conflicts and generate a coordinated allocation solution. Subsequently, a dynamic resource entropy indicator, reflecting the resource distribution status, is monitored in real time. When local resource overload is detected, a gradient-driven resource rebalancing algorithm is triggered. A built-in security constraint validation layer is used to eliminate illegal solutions, ensuring the rationality and legality of resource adjustments.

[0025] Finally, the optimized resource allocation plan is output to the engineering management system for execution. At the same time, resource utilization and construction period deviation indicators are collected during the actual execution process, and the edge weight parameters in the heterogeneous graph network are dynamically updated to form a closed-loop system of decision-making-execution-feedback, continuously improving the accuracy of resource allocation and the intelligence level of engineering execution.

[0026] See also Figure 1, an embodiment of the present invention provides a technical solution: an engineering resource allocation optimization method based on artificial intelligence, comprising the following steps: S1: real-time collection of multimodal heterogeneous data in the target engineering scenario, including task attribute data, resource status data and environmental interference data, spatiotemporal alignment and semantic encoding of the multimodal heterogeneous data, and generation of a multimodal dynamic data stream; S2: constructing a dynamic heterogeneous graph network based on the multimodal dynamic data stream, defining the heterogeneous interaction relationship between task nodes, resource nodes and environment nodes, calculating the dynamic dependency weights between nodes based on the multimodal graph attention mechanism, and updating the edge weights through an event-driven mechanism; S3: performing two-stage collaborative optimization, mapping the dynamic dependency relationship output by the dynamic heterogeneous graph network into quantifiable constraints; S3.1: the first stage: simulating the external Interference scenario, generate the probability distribution of resource demand, build a stochastic optimization model to solve the baseline allocation plan that minimizes the risk-weighted cost; S3.2: The second stage: decompose the baseline plan into a local decision-making problem of multiple seed agents, coordinate resource conflicts between agents through the backpropagation negotiation protocol, and output the allocation plan after the conflict is resolved; S4: Real-time monitoring of the dynamic resource entropy indicator of resource distribution, triggering the gradient-driven resource rebalancing algorithm when local overload is detected, adjusting the allocation weight in the direction of utility maximization, and embedding a safety constraint verification layer to filter illegal solutions, and outputting the optimized allocation plan; S5: Output the optimized allocation plan to the engineering management system for execution, and at the same time collect resource utilization and construction period deviation indicators in the actual execution data, and dynamically transmit them back to the dynamic heterogeneous graph network to update the edge weight parameters.

[0027] In this embodiment, S1: In this step, multimodal heterogeneous data refers to data types with different sources and forms, specifically including task attribute data (such as task type, priority, deadline, etc.), resource status data (such as equipment availability, personnel load, material inventory, etc.), and environmental interference data (such as weather changes, traffic conditions, on-site accidents, etc.). By performing spatiotemporal alignment (i.e., unifying the time scale and spatial scale of different data sources) and semantic encoding (i.e., converting the original data into a vector representation with unified expression meaning) on the above multimodal heterogeneous data, a multimodal dynamic data stream with dynamic temporal characteristics is generated, laying the data foundation for subsequent graph network modeling. S2: In this step, a dynamic heterogeneous graph network refers to a graph structure that can describe different types of nodes and their interactive relationships and is updated over time. Among them, tasks are defined as task nodes, resources are defined as resource nodes, and environmental interference factors are defined as environment nodes. Heterogeneous interaction edges are established between different types of nodes. Through the multimodal graph attention mechanism, the dynamic dependency weights between nodes are calculated based on node features and their contextual information, realizing adaptive modeling of the influence relationship between nodes. At the same time, an event-driven edge weight update mechanism is introduced. This mechanism dynamically adjusts edge weights based on new events (such as resource failures and task changes) to ensure that the graph network can reflect changes in the project site status in real time. S3: Specifically, it includes: S3.1: Phase 1: Using Monte Carlo sampling, a random simulation of possible external interference scenarios is performed to generate a probabilistic distribution model of resource demand. Based on this, a stochastic optimization model is established to minimize the risk-weighted cost and solve a baseline allocation plan that is highly robust under uncertain environments. S3.2: Phase 2: The baseline plan is refined into a local decision-making problem for multiple sub-agents. A backpropagation negotiation protocol is designed to enable each sub-agent to make resource allocation decisions based on local information. Through multiple rounds of backpropagation negotiation, conflicting areas are adjusted, achieving coordination and resolution of resource conflicts and outputting a globally consistent allocation plan. S4: A dynamic resource entropy indicator is used to monitor resource distribution in real time. In this step, dynamic resource entropy refers to a quantitative indicator that measures the balance of resource distribution and local load pressure. When an abnormal increase in resource entropy in a certain area is detected (i.e., local resource overload), the gradient-driven resource rebalancing algorithm is triggered to adjust the resource allocation weights in the direction of utility maximization (i.e., to maximize the efficiency of overall resource allocation). During the adjustment process, a safety constraint verification layer is embedded to verify the legitimacy of the adjustment results in real time to prevent the generation of illegal plans that violate construction site safety regulations or resource unavailability constraints, thereby outputting an optimized allocation plan that meets safety requirements. S5: Output the optimized allocation plan to the engineering management system for execution. During the execution of the plan, the resource utilization rate and construction period deviation indicators in the actual execution data are continuously collected.By dynamically transmitting data back to the dynamic heterogeneous graph network, edge weight parameters are updated in real time based on feedback data, thereby achieving closed-loop optimization and improving the adaptability and execution efficiency of engineering resource management.

[0028] Specifically, S1 includes the following contents: S1.1: Collect multimodal heterogeneous data through the interface between the Internet of Things terminal and the engineering management system: Task attribute data includes the duration, priority, task dependency, and real-time progress deviation rate in the task decomposition structure; Resource status data includes the human resource skill matrix, the availability timing status of mechanical equipment, and the dynamic consumption rate of material inventory; Environmental interference data includes real-time meteorological monitoring indicators, supply chain delay warning signals, and equipment failure event logs; S1.2: Perform spatiotemporal alignment on the collected heterogeneous data, and build a unified spatiotemporal index based on timestamp synchronization and spatial position mapping rules; S1.3: Apply semantic coding strategies to the data after spatiotemporal alignment to extract task description features, resource attribute features, and environmental status features to generate a multimodal dynamic data stream in a unified format.

[0029] In this implementation scheme, S1.1: In this step, the Internet of Things terminal (IoT terminal) refers to a sensing node installed at the project site that can sense and collect environment, equipment, and personnel status information in real time. The engineering management system interface refers to the data interaction channel with the engineering project management platform. The collected multimodal, heterogeneous data specifically includes: Task attribute data, covering the duration (i.e., the planned duration of each subtask), priority (the urgency of each task), task dependencies (the preceding and subsequent logic between different tasks), and the real-time progress deviation rate (i.e., the ratio of actual progress to planned progress) in the task breakdown structure, which is used to dynamically assess task execution status; Resource status data, including the human resource skill matrix (i.e., the mapping between personnel and skill categories), the time series status of mechanical equipment availability (i.e., the time-varying record of equipment availability / unavailability), and the dynamic consumption rate of material inventory (i.e., the change in material consumption per unit time), which is used to dynamically characterize resource supply capacity; Environmental interference data, including real-time meteorological monitoring indicators (such as temperature, humidity, and wind speed), supply chain delay warning signals (e.g., supplier delivery anomaly warnings, logistics disruption alerts), and equipment failure event logs (recording the time, type, and impact range of equipment failures), which are used to characterize the impact of external uncertainties on project execution. Comprehensive, multi-source data collection provides comprehensive and dynamic data support for subsequent allocation optimization. S1.2: In this step, spatiotemporal alignment refers to the standardization of different data sources in terms of time and spatial location. Specifically, it includes: based on the timestamp synchronization rules, mapping data with different collection frequencies and different sampling times to a unified time scale to ensure the consistency of the temporal correspondence of the data; based on the spatial location mapping rules, standardizing the mapping of data involving physical locations (such as equipment, work blocks, and personnel locations) through a unified coordinate system to ensure that data in the spatial dimension can be connected. Spatiotemporal alignment processing can eliminate the inconsistency of heterogeneous data sources in sampling time and location, improve the accuracy of data fusion and the reliability of subsequent analysis and modeling. S1.3: In this step, the semantic encoding strategy refers to the unified feature extraction and vectorization processing of original numerical, textual or categorical data through machine learning feature engineering or embedded representation methods. Specifically, it includes: extracting task description features (such as task urgency coding, task complexity index); extracting resource attribute features (such as personnel skill matching, equipment health score, material supply stability score); extracting environmental status features (such as weather condition severity index, supply chain reliability level). Through semantic encoding, the originally heterogeneous and messy raw data is converted into a multimodal dynamic data stream with a unified vector representation and standard format. Multimodal dynamic data streams have temporal traceability, spatial localization, and feature consistency, providing high-quality, structured data input for the subsequent construction of dynamic heterogeneous graph networks.

[0030] Specifically, a dynamic heterogeneous graph network is constructed based on multimodal dynamic data flows, and the specific process of defining the heterogeneous interaction relationships among task nodes, resource nodes, and environment nodes is as follows: S2.1: Node classification and attribute definition: Task nodes are generated based on atomic tasks in the task decomposition structure, and node attributes include duration, priority, and dependency strength; resource nodes are divided into three sub-nodes: manpower, machinery, and materials, and attributes include skill level, available time period, and safety stock threshold; environment nodes dynamically generate external interference event entities, and attributes include event type, geographical impact radius, and duration; S2.2: Heterogeneous relationship modeling and edge attribute binding: The task-resource edge defines the demand intensity coefficient and binds spatiotemporal constraints to limit resources to take effect in specific geographical areas or time windows; the resource-environment edge establishes a vulnerability score for resource nodes to environmental interference, and quantifies the impact of environmental events on resource availability; the task-environment edge associates the direct weight of external events on task progress, and dynamically adjusts the priority of task nodes.

[0031] In this embodiment, specifically, S2.1: In this step, the dynamic heterogeneous graph network consists of three types of nodes: task nodes, resource nodes, and environment nodes. The specific definitions are as follows: Task nodes are generated based on atomic tasks (i.e., the smallest executable unit that cannot be further subdivided) in the task breakdown structure (WBS). Each task node is accompanied by the following attribute information: Duration: refers to the time required for the planned completion of the task; Priority: reflects the urgency and importance of the task; Dependency Strength: quantifies the importance of the logical relationship between the task and its predecessor tasks. Resource nodes are subdivided into three types of subnodes based on resource type: Human resource nodes: attributes include personnel skill level and available working hours; Mechanical equipment nodes: attributes include equipment performance level, maintenance status, and available hours; Material resource nodes: attributes include material type, dynamic inventory level, and safety stock threshold (i.e., inventory level below which timely replenishment is required). Environment nodes: Dynamically generated nodes represent external disturbance factors, such as meteorological disasters, supply chain anomalies, and mechanical failure events. Each environment node has: event type (e.g., typhoon, rainstorm, logistics delay); geographic impact radius (i.e., impact range); and duration. The purpose of node classification and attribute definition is to lay a clear foundation for the subsequent modeling of complex heterogeneous interaction relationships between nodes, and at the same time, to depict the actual heterogeneity and dynamics in the engineering scene through rich node attributes. S2.1 Node Classification and Attribute Definition. Node Set Definition: Task Node Set: Resource node collection: Environment node collection: Node attribute definition: Task node v τi Attribute: Duration: d τi ; Priority: p τi ; Dependence strength: r τi ; Resource node vρj Attributes: Skill Level: S ρj ; Available period: t ρj ; Safety stock threshold: q ρj ; Environment node v ηk Attributes: Event Type: e ηk ; Geographical influence radius: g ηk ; Duration: h ηk . S2.2: Heterogeneous relationship modeling and edge attribute binding. In this step, based on the defined nodes, heterogeneous interaction relationships between nodes of various types are established (i.e., connecting edges between nodes of different categories), and specific attributes are bound to each edge to enhance the expressive power of the graph network. Specifically including: Task-resource edge: defining the demand intensity coefficient to indicate the degree of demand for a certain resource for a certain task; binding spatiotemporal constraints, that is, limiting resources to specific geographical locations and specific time windows to provide support for tasks, reflecting the spatial limitations and schedule arrangements of actual engineering operations. Resource-environment edge: establishing a vulnerability score, that is, evaluating the degree of impact of specific resource nodes under specific environmental interference events (for example: the availability of open-air mechanical equipment decreases under heavy rain conditions), and providing a basis for considering environmental risks in subsequent resource allocation decisions. Task-environment edge: associating the direct impact weight of external events on task progress, dynamically adjusting the priority of task nodes to reflect the dynamic impact of environmental changes on construction progress management. The role of heterogeneous relationship modeling and edge attribute binding is to make the complex dependencies between task execution, resource allocation, and environmental disturbances explicit in the graph structure and quantify them into computable features, providing a fine-grained and dynamically adaptive foundation for subsequent graph network reasoning and allocation optimization. Task-resource edge modeling comprehensively defines the matching degree of task demand for resources Φ τρ (v τi ,v ρj ): Where: Φ τρ (v τi ,v ρj ):Task node v τi With resource node v ρj Demand matching degree; τi :Task v τi The intensity of demand for resource categories (preset value, quantifying the intensity of manpower / machinery / materials required); ρj :Resources ρj Ability index (by skill level s ρj , inventory threshold q ρj etc.);∈ τi,ρj : The degree of incompatibility between tasks and resources in space and time (for example, if the time periods do not overlap or the geographical areas do not match, the penalty value will be increased); Ω τi,ρj: Resource spatiotemporal availability coefficient (1 means fully available, 0 means unavailable, and partially available is between 0 and 1); α1, α2: Control parameters that balance the weights of tasks on resource capabilities and resource accessibility. Resource-environment edge modeling, defining the comprehensive vulnerability score Ψ of resource nodes to environment nodes ρη (v ρj ,v ηk ): Among them: ρη (v ρj ,v ηk ): resource node v ρj For environment node v ηk Vulnerability score; v ρj : Environmental anti-interference capability of resource nodes (the larger the value, the more resistant to environmental impact); κ ηk : Destruction intensity coefficient of environmental events (based on event type e ηk Duration h ηk quantification); g ηk : geographical impact radius of environmental events; ξ ρj : The geographical movement radius of resources (such as the range of mechanical movement, the radius of material transportation); β1, β2: adjustment parameters, weighing direct vulnerability and geographical exposure risk. Define the dynamic correction amount of the environment to the task priority Θ τη (v τi ,v ηk ): Where: Θ τη (v τi ,v ηk ):Environment node v ηk For task node v τi The priority adjustment amount; Mission v τi basic sensitivity to external disturbances; Environment Node v ηk The impact complexity (indicators of event type, geographical superposition, and multiple interference sources); ρ τi,ηk : Geographic intersection coefficient (the degree of overlap between the task construction site and the environmental event impact area, between 0 and 1); γ1, γ2: Control parameters to control the weight of the task's internal vulnerability and external exposure. The final priority dynamic update formula is: This formula reflects the multiple dynamic adaptation mechanisms of tasks to environmental changes.

[0032] Specifically, the dynamic dependency weights between nodes are calculated according to the multimodal graph attention mechanism, and the logical process of the event-driven edge weight update mechanism is as follows: S2.3: Multimodal attention calculation: A scalar attention mechanism is used for numerical resource data to calculate the linear dependency strength between task nodes and resource nodes; a temporal attention mechanism is used for sequential task data to capture the forward and backward driving dynamic relationship between tasks; a spatial attention mechanism is used for image-type environmental data to extract the resource distribution hotspot features in the construction site monitoring screen; S2.4: Event-driven dynamic update of edge weights: When equipment failure or sudden weather change events are detected, the affected subgraph area in the dynamic heterogeneous graph network is located; according to the predefined event impact rule library, the event type and node attributes are matched, and the weight parameters of the associated edges are locally updated; the original weights of the unaffected areas are retained, and only the event-related subgraphs are incrementally updated to avoid global model retraining.

[0033] In this implementation scheme, S2.3 Multimodal Attention Calculation This step constructs corresponding multimodal attention mechanisms for different types of input data (numerical, sequence, image) to finely capture the dynamic dependency between task nodes and resource nodes. Explanation of the scalar attention mechanism for numerical resource data: For static numerical resources (such as equipment load rate, remaining working hours, etc.), a scalar attention mechanism is used to calculate the linear dependency strength between task nodes and resource nodes. This mechanism emphasizes the direct impact of numerical size on resource scheduling priority. Formula: ω tr =σ(λ1·η t +λ2·ρ r +b1); where: ωtr: scalar attention weight of task node vt and resource node vr; σ(): Sigmoid activation function, ensuring the output range is (0,1); ηt: numerical feature representation of task node vt (such as task urgency indicator); ρr: numerical attribute of resource node vr (such as remaining capacity); λ1, λ2: trainable scalar parameters, controlling feature importance; b1: bias term. Explanation of the temporal attention mechanism for sequential task data: For task sequences with temporal characteristics (such as process schedules), a temporal attention mechanism is introduced to model the causal and temporal dependencies between tasks. Considering the impact of time intervals on dependency strength, the importance of closely connected tasks is emphasized. Formula representation: Where: tt' :Task node v t With v t' The temporal attention weight of t ,τ t': The planned start timestamp of the corresponding task; γ: The time decay coefficient (adjustable time sensitivity between tasks); The denominator is a normalization term to ensure that the sum of all attention weights is 1. Explanation of the spatial attention mechanism for image-based environmental data: For environmental monitoring images of construction sites, the spatial attention mechanism is used to extract the hotspot area features of resource distribution, focusing on locations with high physical density and high resource concentration in the image to assist in resource scheduling decisions. S2.4 Event-driven dynamic update of edge weights. After detecting an emergency (such as equipment failure or sudden weather change), this step locally updates the edge weights in the dynamic heterogeneous graph to reflect the impact of environmental changes on resource scheduling logic in real time. Subgraph positioning: Based on the event source node, the k-hop neighbor expansion algorithm is used to quickly determine the affected local subgraph area and limit the update range. When the event node v e When changes occur, the k-order neighbor search method is used to screen all distances v e A set of nodes not exceeding k is used as the updated subgraph Explanation of local update of edge weights based on event impact rule base: According to the preset event-node attribute rule table, the event impact strength is judged and the edge weights in the affected subgraph are dynamically adjusted. Formula: Let the original edge weight be θ mn , then the updated edge weight θ mn 'Expressed as: θ mn '=θ mn ×(1-ξ e ·χ mn );where: θ mn :node v m With v n Original edge weight; θ mn ': Update edge weight after event driving; ξ e : the intensity coefficient of the impact of event type e on edge relations (set according to the rule base); mn :Node pair (v m ,v n ) matches the event attribute (1 if matched, 0 if not matched). The unaffected area is retained. In order to avoid global model retraining, the subgraph is not updated. The nodes and edges in the graph retain their original weights. All edges (v i ,v j ), keep the original weight θ ij constant.

[0034] Specifically, S3.1 includes the following contents: S3.1.1: Define a set of external interference scenarios: Based on the historical engineering event library and real-time environmental monitoring data, extract three typical interference events: equipment failure, supply chain interruption, and extreme weather, and construct the probability of event occurrence and impact intensity distribution; S3.1.2: Perform Monte Carlo sampling: Within the preset time window, generate multiple sets of interference scenario sequences according to the event probability distribution to simulate the fluctuation pattern of resource demand; S3.1.3: Construct a risk-weighted cost model: Integrate the resource demand probability distribution output by the dynamic heterogeneous graph network with the interference scenario simulation results, define the cost function as the weighted sum of the expected value of the resource overrun cost and the conditional risk value, and generate a baseline allocation plan.

[0035] In this implementation plan, S3.1 generates a resource baseline allocation plan. This step aims to generate a preliminary resource allocation baseline by constructing external interference scenarios, simulating resource demand changes through Monte Carlo sampling, and optimizing risk-weighted costs. S3.1.1 Define a set of external interference scenarios: Based on the historical engineering event database and real-time environmental monitoring data, three major types of external interference events are screened out: equipment failure (such as sudden shutdown of key equipment); supply chain disruptions (such as delayed delivery of materials); extreme weather (such as heavy rain and strong winds causing construction interruptions). For each type of event, statistics are separately made: probability distribution of occurrence (for example, classification by time, region, and equipment type); impact intensity distribution (for example, the magnitude of the impact on construction period and resource consumption). Let the set of external interference events be Corresponding to equipment failure, supply chain disruption and extreme weather respectively; for each event e b , establish its probability distribution P(e b ) and influence intensity distribution S(e b S3.1.2 Perform Monte Carlo sampling: Within a given time window ΔT, perform multiple rounds of random sampling based on the probability distribution of each event to generate a possible interference scenario sequence, simulating the random fluctuation process of resource demand. Formula: The interference scenario sequence obtained by the ath sampling is expressed as: in: The interference scene obtained by sampling for the ath time; e b : bth event type; The sampling time point at which the event occurs within the time window; Event impact intensity (from S(e b ) distribution). Further, the resource demand disturbance simulation formula is: in: The demand for resource r at time t under the ath sampling; Basic resource requirements without interference; Event e b The disturbance function for resource r at time t (the decay function depends on the duration of the disturbance); Interference intensity. S3.1.3 Constructing a risk-weighted cost model Explanation: By integrating the probability distribution of resource demand obtained by dynamic heterogeneous graph prediction with the disturbance results of the interference scenario from Monte Carlo simulation, a risk-weighted cost function for resource overrun is defined as the optimization basis for generating a baseline resource allocation plan. Formula: The resource allocation cost objective function is set as: Where: C: comprehensive cost index; α∈(0,1): weight coefficient of expected value and conditional risk value (adjustable risk preference); Mathematical expectation of resource overrun cost (OC); CVaR β (OC): Conditional risk value at confidence level β, measuring the tail loss of overspending risk; OC: Resource overspending cost calculated under each sampling scenario. Specific definition of resource overspending cost: For each sampling a, the resource overspending cost is calculated as: Among them: r : penalty cost coefficient for overspending per unit resource r; The actual peak demand of resource r under the a-th sampling; a r : The baseline supply of resource r pre-allocated. Finally, by optimizing C, the initial resource allocation {a r}.

[0036] Specifically, S3.2 includes the following contents: S3.2.1: Define a dynamic utility function for each worker seed agent, which includes three configurable priority dimensions: task progress gain, resource conflict penalty, and safety constraint compliance, and dynamically calibrate weight parameters based on real-time engineering events; S3.2.2: Construct a directional communication link network, and each agent only exchanges the current resource occupancy weight, utility gradient direction, and conflict intensity index with adjacent agents with resource conflicts; S3.2.3: Iteratively correct the local allocation weight based on the gradient negotiation message, and judge convergence when the conflict intensity change rate of multiple consecutive iterations is lower than the preset threshold, and output the optimization plan after conflict resolution.

[0037] In this implementation, S3.2.1 defines the sub-agent dynamic utility function: for each worker sub-agent A j Establish a dynamic utility function U j , comprehensively considering the following three aspects: task progress gain (encouraging faster work); resource conflict penalty (suppressing resource competition); and safety constraint compliance (ensuring safe and compliant operations). After a real-time engineering event (such as equipment failure or extreme weather) occurs, the weights of these three indicators in the utility function are dynamically calibrated to ensure the real-time adaptability of the agent decision-making strategy. The formula is: Among them: U j :Subagent A j The current utility value of G j: Task progress gain indicators (such as task completion rate, output efficiency); C j : Resource conflict penalty index (such as resource occupancy overlap); V j : Security constraint violation indicators (e.g., security risk scores); The dynamic weight coefficients corresponding to the three indicators satisfy Weights are dynamically adjusted based on real-time events. S3.2.2 Constructing a Directed Communication Link Network Explanation: To improve collaboration efficiency, subagents only establish directed communication links with adjacent subagents that have resource conflicts. Negotiation information transmitted on the link includes: current resource occupancy weight; utility gradient direction; conflict intensity index. Communication link establishment condition text description: Only the current agent A j With A k There is a shared resource collection When the occupation weight conflict exceeds the set threshold γ0, a unidirectional link L is established. j→k Conflict intensity definition formula: Where: Φ jk :Subagent A j With A k The intensity of resource conflict between jr :Agent A j The occupation weight of resource r; ω kr :Agent A k The occupation weight of resource r. jk When >γ0, the link is established and local information exchange is allowed. S3.2.3 Gradient negotiation and local allocation weight correction. In this step, each sub-agent interacts by receiving and sending gradient negotiation messages (including resource occupancy weights, utility gradient directions and conflict intensity scores) based on the established directional communication link network. Each sub-agent iteratively corrects the local resource allocation weights based on its own utility optimization direction and the compensation information of the conflicting party. In order to ensure the stability of the correction process, the momentum factor is introduced to smooth the weight update and avoid drastic fluctuations. The system monitors the rate of change of the conflict intensity score in real time. When the rate of change is lower than the preset threshold in multiple consecutive iterations, it is determined that the gradient negotiation process has converged, and finally outputs the optimized resource allocation plan after the conflict is resolved.

[0038] Specifically, S3.2.3 includes the following contents: S3.2.3.1: Structured design of gradient negotiation message: The gradient negotiation message is defined to contain three core fields: resource occupancy weight, utility gradient direction and conflict intensity score. The utility gradient direction is generated by calculating the partial derivative of the dynamic utility function on the resource weight; the message transmission cycle is dynamically bound to the engineering scenario, high-frequency events trigger instant transmission, and low-frequency scenarios are transmitted at fixed intervals; S3.2.3.2: Iterative weight correction rule: Each agent updates the resource weight according to the comprehensive value of the local utility gradient direction and the gradient reverse compensation of the conflicting party based on the received gradient message, and introduces a momentum factor to smooth the iterative process to avoid weight oscillation; S3.2.3.3: Convergence judgment and output: When the change rate of the conflict intensity score for three consecutive iterations is lower than the preset threshold, it is judged to be converged, and the final weight is output to the dynamic heterogeneous graph network to generate an optimization plan after conflict resolution.

[0039] In this implementation scheme, S3.2.3.1 Gradient Negotiation Message Structure Design Explanation: For resource conflict negotiation between sub-agents, a unified gradient negotiation message structure is designed. The core fields of the message include: resource occupancy weight ω; utility gradient direction The conflict intensity score Φ. The utility gradient direction is determined by the dynamic utility function U within the agent. j The partial derivative of the resource occupancy weight is calculated as follows: Message transmission mechanism: If a high-frequency dynamic event is detected (such as sudden equipment failure, sudden weather changes, etc.), the message will be immediately triggered for instant transmission; for low-frequency or steady-state scenarios, the message will be transmitted periodically at fixed time intervals to ensure the adaptability of the communication load to the system dynamics. Explanation of iterative correction based on gradient negotiation: Through gradient negotiation, the sub-agent locally corrects its own resource allocation decision based on the utility gradient direction and conflict intensity transmitted by the adjacent agent, gradually reducing the conflict. In each round of iteration, the sub-agent updates its own allocation weight: Resource allocation weight correction formula: in: At the nth iteration, subagent A j The occupation weight of resource r; η j : Step size parameter, controlling the amplitude of a single iteration; Subagent A j The local utility gradient at the nth iteration; Neighbor Agent A k The gradient direction is used as the reverse compensation term; μ j : Momentum factor (0≤μ j <1), used to smooth the iterative process and suppress weight oscillations caused by local optimality or high-frequency oscillations; The resource weight value at the last iteration. The significance of the momentum term: Introducing the weight change trend of the previous round, maintaining the continuity of resource adjustment, and improving convergence stability. In the process of resource conflict resolution, in order to ensure the stability and timeliness of optimization iterations, a convergence threshold ε of the conflict intensity change rate is set. The specific convergence judgment criterion is: in the process of three consecutive iterations, for any subagent A on the communication link j With A k The conflict intensity score Φ jk , if the following conditions are met: The resource weight optimization process is then determined to have converged. The parameters are explained as follows: At the nth iteration, subagent A j With A k The conflict intensity score between the two sub-agents indicates the current conflict degree between the two sub-agents in resource usage. The higher the value, the more serious the conflict. At the n+1th iteration, the conflict intensity score between the same pair of sub-agents represents the new conflict level after a round of resource weight modification. The relative rate of change of the conflict intensity score. This measures the rate of change in conflict intensity between two consecutive iterations and is used to determine whether conflict resolution has stabilized. ε: The preset convergence threshold for the conflict intensity change rate (ε > 0). This threshold is used to define the convergence criterion and is typically a small positive number between 0.01 and 0.05. If the rate of change falls below this value, the conflict is considered sufficiently small and the system is stable. A set of subagent communication links. This includes all subagent pairs that have resource conflicts and established a directional communication relationship. j, k: The indexes of the two subagents participating in the conflict negotiation in the link. They represent the specific resource contention relationship between the two subagents. n: The iteration round index. This indicates that the current iteration is the nth round of resource conflict optimization.

[0040] Specifically, S4 includes the following contents: quantifying the degree of chaos caused by local resource overload based on the spatial density difference and time decay effect of resource distribution, where the spatial density difference is calculated through the variance of regional resource occupancy rate, and the time decay effect is generated by weighting the idle time of resources; adjusting the allocation weight along the gradient ascent direction of the resource utility function, and using the projected gradient method to constrain the weight correction within the safety and compliance range to avoid the generation of illegal solutions; for high entropy areas, prioritizing the allocation of idle resources and synchronously updating the resource occupancy status of adjacent areas; defining hard constraints through the engineering safety specification library, including the minimum manpower configuration threshold, equipment safety redundancy and the lower limit of material inventory; calling the linear programming solver to verify the allocation plan, filtering out illegal solutions that violate the constraints, and outputting the final compliant allocation plan.

[0041] In this implementation, the degree of local resource overload and disorder is quantified, and the spatial density difference is quantified by calculating the variance of resource occupancy rates in each area to quantify the degree of spatial imbalance in resource distribution. The formula is expressed as: The variance of resource occupancy rate reflects the difference in spatial resource density; M: number of regions; r m : Resource occupancy rate of the mth region; The mean of resource utilization rates across all regions. Explanation: The larger the variance, the more uneven the resource distribution and the higher the risk of local chaos. 2) Time decay effect weighting: Based on the length of time resources are idle, a decay weighting factor is generated to quantify the impact of excess resources on the degree of system chaos. The decay factor is defined as: d n : the time decay weight of the nth resource; λ: time decay coefficient (positive value, reflecting the weighted speed of idle growth); t n : The cumulative idle time of the nth resource. The longer the idle time, the larger the weighting factor, indicating that the resource's contribution to the disorder is more significant. Adjust the weight along the gradient of the resource utility function. The allocation weight of each resource is adjusted along the gradient of the utility function to improve resource allocation efficiency. Using the projected gradient method, the formula is expressed as: ω (k) : resource allocation weight vector for the kth iteration; α: step size factor; The gradient of the resource utility function U(·) with respect to the weight ω; The projection operator projects the updated weights back into the safe and compliant feasible region Ω. After each weight adjustment, projection is used to ensure that the results meet engineering specifications for safety, manpower, and equipment redundancy, and that no illegal solutions are generated. Prioritized resource replenishment in high-entropy regions is synchronized with neighboring regions. High-entropy region identification: Based on a comprehensive score of local resource density and time decay, regions with high resource dislocation are identified. Idle resource priority scheduling: Idle resources are allocated to high-entropy regions. Neighborhood synchronization updates: Resource occupancy status in adjacent regions is simultaneously corrected to avoid new conflicts caused by local resource influx. Engineering safety specification constraints are defined using the engineering safety specification library, setting the following hard constraints: Minimum manpower allocation threshold: Each area or task must meet a minimum number of workers; Equipment safety redundancy: Critical equipment must have backup units to meet failover requirements; Material inventory minimum: Dynamically ensures that construction material inventory levels do not fall below safety warning levels. These conditions limit the range of resource weights to ensure the engineering feasibility of the allocation plan. S4.5 Linear Programming Verification and Compliance Solution Output: Finally, the preliminary allocation plan is input into the linear programming solver to verify that all constraints are met. The solution model is briefly described as follows: Objective function (maximize resource utilization or minimize idle rate): max∑ n β n ω n ; Constraints: ∑ n ω n ≤Total resource capacity;ω n: the allocation weight of the nth resource unit; β n : Resource utilization or utility coefficient; The upper and lower bounds of resource allocation are determined by engineering specifications; N is the set of resource units. After solving, the optimal allocation solution that meets all hard constraints is retained, and invalid solutions are directly filtered out.

[0042] Specifically, the specific process of collecting resource utilization and duration deviation indicators from actual execution data and dynamically transmitting them back to the dynamic heterogeneous graph network to update edge weight parameters is as follows: resource utilization indicators, including the proportion of mechanical equipment working hours, manpower load rate and material consumption rate, are collected in real time through the Internet of Things terminal of the engineering management system, the duration deviation characteristics are extracted, and the deviation rate between the actual progress of critical path tasks and the planned progress is calculated; the demand intensity coefficient of the task-resource edge is reversely adjusted according to the resource utilization deviation, and the demand weight of the task node with lower-than-expected resource utilization is reduced, and vice versa, the priority is increased; the dependency weight of the task node is dynamically corrected in combination with the duration deviation rate, and the dependency strength of the task node with a deviation rate exceeding the threshold is weakened; the updated edge weight parameters are transmitted back to the dynamic heterogeneous graph network, and only the edges of the affected subgraph are incrementally adjusted to retain the stability of the global network structure, and historical execution data is filtered based on the sliding window mechanism to prevent noise data from interfering with the model parameters.

[0043] In this implementation scheme, the dynamic heterogeneous graph network is dynamically updated to collect actual execution process data, and is dynamically updated in the following ways: through the Internet of Things terminals deployed in the engineering management system, various resource utilization indicators are collected in real time, including: the working hour ratio of mechanical equipment (that is, the ratio of the actual operating time of the equipment to the planned operating time), the manpower load rate (that is, the ratio of the actual workload of the personnel to the standard load), and the material consumption rate (that is, the material consumption per unit time). The construction period deviation characteristics are extracted simultaneously, that is, for each task node on the critical path, the deviation rate between its actual progress and the planned progress is calculated to quantify the progress deviation of the task execution. Subsequently, dynamic parameter updates are performed based on the above-mentioned collected data: based on the resource utilization deviation, the demand intensity coefficient of the task-resource edge is adjusted in reverse: when the actual utilization rate of the resources associated with a task node is lower than expected, the demand weight of the task node for the resource is reduced; when the actual utilization rate is higher than expected, its resource demand priority is increased accordingly. Based on the duration deviation rate, the dependency weights between task nodes are dynamically corrected: if the progress deviation rate of a task node exceeds the set threshold, the dependency strength between it and the subsequent task node is weakened, reflecting its negative impact on the subsequent progress in the overall task chain. During the parameter update process, a subgraph incremental adjustment mechanism is adopted, that is, only the edge weights in the local subgraph affected by the current execution data are updated, avoiding large-scale disturbances to the global heterogeneous graph network, thereby maintaining the stability of the overall network structure. In addition, a sliding window mechanism is introduced to filter historical execution data, limiting the data within the most recent time windows to participate in edge weight adjustment, in order to filter out the adverse effects of abnormal fluctuations or noise data on model parameters, ensuring that the update process of the dynamic heterogeneous graph network has good robustness and adaptability.

[0044] See also Figure 2, the engineering resource allocation optimization system based on artificial intelligence includes the following modules: heterogeneous data acquisition and encoding module, dynamic heterogeneous graph construction module, collaborative optimization decision module, resource dynamic monitoring module, and scheme execution module; the heterogeneous data acquisition and encoding module is used to collect multimodal heterogeneous data in the target engineering scenario in real time, including task attribute data, resource status data and environmental interference data, perform spatiotemporal alignment and semantic encoding on multimodal heterogeneous data, and generate multimodal dynamic data streams; the dynamic heterogeneous graph construction module is used to construct a dynamic heterogeneous graph network according to the multimodal dynamic data stream, define the heterogeneous interaction relationship between task nodes, resource nodes, and environment nodes, calculate the dynamic dependency weights between nodes according to the multimodal graph attention mechanism, and update the edge weights through an event-driven mechanism: the collaborative optimization decision module is used to perform two-stage collaborative optimization, and the dynamic dependency relationship output by the dynamic heterogeneous graph network is updated. The system maps resources into quantifiable constraints, simulates external interference scenarios based on Monte Carlo sampling, generates a probability distribution of resource requirements, constructs a stochastic optimization model to solve the baseline allocation plan that minimizes the risk-weighted cost, decomposes the baseline plan into a local decision-making problem of multiple seed agents, coordinates resource conflicts between agents through the backpropagation negotiation protocol, and outputs the allocation plan after the conflict is resolved; the resource dynamic monitoring module is used to monitor the dynamic resource entropy index of resource distribution in real time. When a local overload is detected, the gradient-driven resource rebalancing algorithm is triggered, and the allocation weight is adjusted in the direction of utility maximization. A safety constraint verification layer is embedded to filter illegal solutions, and the optimized allocation plan is output; the plan execution module is used to output the optimized allocation plan to the engineering management system for execution, while collecting resource utilization and construction period deviation indicators in the actual execution data, and dynamically transmitting them back to the dynamic heterogeneous graph network to update the edge weight parameters.

[0045] In this implementation, the heterogeneous data acquisition and encoding module introduces a spatiotemporal alignment and semantic encoding mechanism to address the difficulty of data fusion caused by heterogeneous data sources and inconsistent granularity in engineering scenarios, providing a high-quality, uniformly expressed data input foundation for subsequent dynamic heterogeneous graph modeling. Compared with traditional static data acquisition, it has higher timeliness and dynamic adaptability. The dynamic heterogeneous graph construction module constructs a dynamic heterogeneous graph network based on the collected multimodal dynamic data stream, defining three types of heterogeneous nodes (task nodes, resource nodes, and environment nodes) and their interaction relationships (such as resource allocation relationships, task dependencies, and environmental impact relationships). A multimodal graph attention mechanism is introduced to adaptively calculate the dynamic dependency weights between nodes, emphasizing important nodes and important connections. Through an event-driven edge weight update mechanism, the heterogeneous graph structure and parameters are locally adjusted in real time in response to key events in the scenario (such as task delays and equipment failures). Mapping multimodal dynamic engineering information into a structured, computable dynamic heterogeneous graph breaks the limitations of the single static graph model in previous engineering management. The attention mechanism and event-driven updates significantly enhance the model's adaptability to sudden changes (such as temporary work stoppages and resource failures). The collaborative optimization decision module performs a two-stage collaborative optimization. In the first stage, the dynamic node dependencies output by the dynamic heterogeneous graph are mapped into quantifiable resource constraints. Monte Carlo sampling is used to simulate external disruption scenarios (such as supply delays and changes in the construction environment) to generate a probability distribution of resource demands. A stochastic optimization model is then established to solve a baseline allocation plan that minimizes risk-weighted costs. In the second stage, the baseline allocation plan is subdivided into specialized subagents (such as civil engineering and electrical trades). Resource conflicts are resolved among the subagents through a backpropagation negotiation protocol, and a final allocation plan is output after conflicts are resolved. The introduction of Monte Carlo scenario simulation and stochastic optimization significantly enhances the system's robustness to future uncertainties (such as environmental changes and resource fluctuations). By decomposing the resource allocation problem into a multi-agent local negotiation problem, the computational bottlenecks caused by centralized optimization are avoided, improving real-time decision-making efficiency in large-scale engineering scenarios. The dynamic resource monitoring module monitors resource distribution in real time and dynamically calculates a resource entropy index to quantify the degree of resource distribution chaos. When an increase in resource entropy is detected in a local area (i.e., resource overload or imbalanced distribution), the gradient-driven resource rebalancing algorithm is triggered to adjust the allocation weight along the gradient direction of the resource utility function. A safety constraint verification layer is embedded to perform compliance verification on the adjusted allocation results (such as minimum manpower configuration, equipment redundancy, and safety inventory requirements), and automatically filter out illegal solutions. Introducing resource entropy as a monitoring indicator enables automatic perception of resource imbalance status, which is different from the traditional approach of relying on manual monitoring or fixed threshold alarms. Combining gradient optimization with safety verification takes into account both the flexibility of resource rebalancing and the safety compliance of project execution. Plan execution module: Push the optimized allocation plan to the project management system to guide the actual construction resource configuration.At the same time, actual execution data is continuously collected through IoT terminals, including resource utilization indicators (such as equipment utilization, labor load rate, and material consumption rate) and construction period deviation indicators (the deviation rate between the critical path progress and the planned progress). The above execution data is dynamically transmitted back to the dynamic heterogeneous graph network, and the model parameters are continuously optimized through the local edge weight incremental update mechanism to maintain the high responsiveness and accuracy of the model to real-world changes. The closed-loop feedback mechanism between resource allocation plans and project execution has been opened up, which not only realizes the dynamic optimization of resource allocation, but also can adapt to the evolution of scenarios, avoiding the problem of disconnection between allocation and execution in traditional systems. The incremental update mechanism of edge weights effectively balances the local agility and overall stability of the network, improving the operating efficiency of the system.

[0046] In summary, this application has at least the following effects:

[0047] This AI-based engineering resource allocation optimization method and system utilizes a heterogeneous data acquisition and encoding module to spatially and temporally align and semantically encode multi-source heterogeneous data, including task attributes, resource status, and environmental interference. This improves the adaptability and accuracy of data input to changes in real-world engineering scenarios. A dynamic heterogeneous graph construction module defines heterogeneous interactions among task, resource, and environmental nodes based on multimodal dynamic data flows. Incorporating a graph attention mechanism, it dynamically adjusts inter-node weights, enabling real-time reflection of changes in resource distribution, task dependencies, and environmental interference, improving the resource scheduling model's ability to perceive and respond to complex working conditions. The collaborative optimization decision-making module employs a two-stage optimization strategy: first, a baseline allocation plan is generated based on a stochastic optimization model. Then, resource conflicts are resolved through a backpropagation negotiation protocol between subagents. This effectively balances global optimality with local feasibility, improving the coordination and convergence efficiency of resource allocation. The dynamic resource monitoring module monitors the dynamic entropy of resource distribution in real time. When local resources are overloaded, a gradient-driven rebalancing algorithm is triggered, which is then corrected based on safety and compliance constraints. This significantly enhances the system's adaptive control capabilities in the face of resource imbalances. The solution execution module collects actual execution data (including resource utilization and schedule deviation indicators). Based on incremental subgraph adjustments and a sliding window data filtering strategy, the heterogeneous graph edge weights are dynamically updated, enabling the optimization model to continuously self-calibrate, effectively suppressing noise interference and maintaining long-term optimization performance. The engineering safety specification library is introduced as a hard constraint standard at each stage, and a linear programming validator is used to filter out illegal solutions. This ensures that the final resource allocation solution meets key safety requirements such as minimum manpower allocation, equipment safety redundancy, and material inventory minimums, thus ensuring the safety and standardization of the project implementation process.

[0048] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0049] The present invention is described with reference to flowcharts and / or block diagrams of systems, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0050] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0051] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0052] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0053] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

Claims

1. An engineering resource allocation optimization method based on artificial intelligence, characterized in that: The following steps are involved: S1: Real-time collection of multimodal heterogeneous data in the target engineering scenario, including task attribute data, resource status data, and environmental interference data, performs spatiotemporal alignment and semantic encoding on the multimodal heterogeneous data, and generates a multimodal dynamic data stream; S2: Construct a dynamic heterogeneous graph network based on multimodal dynamic data streams, define the heterogeneous interaction relationships between task nodes, resource nodes, and environment nodes, calculate the dynamic dependency weights between nodes based on the multimodal graph attention mechanism, and implement an event-driven edge weight update mechanism; S3: Perform a two-stage collaborative optimization to map the dynamic dependencies output by the dynamic heterogeneous graph network into quantifiable constraints; S3.1: Phase 1: Simulate external interference scenarios based on Monte Carlo sampling, generate a probability distribution of resource requirements, and construct a stochastic optimization model to solve the baseline allocation plan that minimizes the risk-weighted cost; S3.2: The second stage: decompose the baseline solution into a local decision-making problem of multiple seed agents, coordinate resource conflicts between agents through the back-propagation negotiation protocol, and output the allocation plan after the conflict is resolved; S4: A dynamic resource entropy indicator that monitors resource distribution in real time. When a local overload is detected, a gradient-driven resource rebalancing algorithm is triggered to adjust allocation weights in the direction of utility maximization. A safety constraint verification layer is embedded to filter out illegal solutions, and the optimized allocation plan is output. S5: The optimized allocation plan is output to the engineering management system for execution. At the same time, the resource utilization and construction period deviation indicators in the actual execution data are collected and dynamically transmitted back to the dynamic heterogeneous graph network to update the edge weight parameters.

2. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 1, characterized in that: S1 includes the following: S1.1: Collect multimodal heterogeneous data through the IoT terminal interface with the engineering management system: Task attribute data includes duration, priority, task dependencies in the task breakdown structure, and real-time progress deviation rate; Resource status data includes human resource skill matrix, time series status of mechanical equipment availability, and dynamic consumption rate of material inventory; Environmental interference data includes real-time weather monitoring indicators, supply chain delay warning signals, and equipment failure event logs; S1.2: Perform spatiotemporal alignment on the collected heterogeneous data and build a unified spatiotemporal index based on timestamp synchronization and spatial location mapping rules; S1.3: Apply semantic encoding strategies to the spatiotemporally aligned data to extract task description features, resource attribute features, and environmental state features to generate a multimodal dynamic data stream in a unified format.

3. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 2, characterized in that: The specific process of constructing a dynamic heterogeneous graph network based on multimodal dynamic data streams and defining the heterogeneous interaction relationships between task nodes, resource nodes, and environment nodes is as follows: S2.1: Node classification and attribute definition: Task nodes are generated based on atomic tasks in the task breakdown structure. Node attributes include duration, priority, and dependency strength. Resource nodes are divided into three sub-nodes: manpower, machinery, and materials. Attributes include skill level, available time period, and safety stock threshold. The environment node dynamically generates external interference event entities, whose attributes include event type, geographical impact radius and duration; S2.2: Heterogeneous relationship modeling and edge attribute binding: Task-resource edges define demand intensity coefficients and bind spatiotemporal constraints to restrict resources to specific geographic areas or time windows. The resource-environment edge establishes a vulnerability score for resource nodes to environmental disturbances, quantifying the degree to which environmental events affect resource availability; The task-environment edge associates the direct weight of external events on task progress and dynamically adjusts the priority of task nodes.

4. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 3, characterized in that: The logical process of calculating the dynamic dependency weights between nodes based on the multimodal graph attention mechanism and updating the edge weights through the event-driven mechanism is as follows: S2.3: Multimodal attention calculation: Scalar attention mechanism is used for numerical resource data to calculate the linear dependency strength between task nodes and resource nodes; A temporal attention mechanism is used for sequential task data to capture the forward and backward driving relationship between tasks; A spatial attention mechanism is used for image-based environmental data to extract resource distribution hotspot features in construction site monitoring images. S2.4: Event-driven dynamic update of edge weights: When a device failure or sudden weather change event is detected, the affected subgraph area in the dynamic heterogeneous graph network is located; According to the predefined event impact rule library, the event type and node attributes are matched, and the weight parameters of the associated edges are locally updated; The original weights of the unaffected regions are retained, and only the event-related subgraphs are incrementally updated to avoid global model retraining.

5. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 4, characterized in that: S3.1 includes the following: S3.1.1: Define a set of external interference scenarios: Based on a historical engineering event database and real-time environmental monitoring data, extract three typical interference events: equipment failure, supply chain disruption, and extreme weather, and construct a distribution of event probability and impact intensity; S3.1.2: Perform Monte Carlo sampling: Generate multiple interference scenario sequences based on event probability distribution within a preset time window to simulate the fluctuation pattern of resource demand; S3.1.3: Construct a risk-weighted cost model: Integrate the resource demand probability distribution output by the dynamic heterogeneous graph network with the interference scenario simulation results, define the cost function as the weighted sum of the expected value of resource overrun cost and the conditional risk value, and generate a baseline allocation plan.

6. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 5, characterized in that: S3.2 includes the following: S3.2.1: Define a dynamic utility function for each worker seed agent, including three configurable priority dimensions: task progress gain, resource conflict penalty, and safety constraint compliance, and dynamically calibrate the weight parameters based on real-time engineering events; S3.2.2: Construct a directional communication link network, where each agent exchanges the current resource occupancy weight, utility gradient direction, and conflict intensity index only with neighboring agents with resource conflicts; S3.2.3: Iteratively modify the local allocation weights based on the gradient negotiation message. When the rate of change of the conflict intensity for multiple consecutive iterations is lower than the preset threshold, convergence is determined and the optimized solution after conflict resolution is output.

7. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 6, characterized in that: S3.2.3 includes the following: S3.2.3.1: Structural design of gradient negotiation message: Gradient negotiation message is defined to contain three core fields: resource occupancy weight, utility gradient direction, and conflict intensity score. The utility gradient direction is calculated by taking the partial derivative of the dynamic utility function with respect to the resource weight. The message transmission cycle is dynamically bound to the engineering scenario. High-frequency events trigger immediate transmission, while low-frequency scenarios are transmitted at fixed intervals. S3.2.3.2: Iterative weight correction rule: Each agent updates the resource weight according to the received gradient message, based on the comprehensive value of the local utility gradient direction and the reverse compensation of the conflicting party's gradient. The momentum factor is introduced to smooth the iterative process and avoid weight oscillation. S3.2.3.3: Convergence judgment and output: When the rate of change of the conflict intensity score for three consecutive iterations is lower than the preset threshold, it is judged to be converged, and the final weight is output to the dynamic heterogeneous graph network to generate an optimized solution after conflict resolution.

8. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 7, characterized in that: S4 includes the following: The degree of local resource overload chaos is quantified based on the spatial density difference and time decay effect of resource distribution. The spatial density difference is calculated by the variance of regional resource occupancy rate, and the time decay effect is generated by weighting the idle time of resources. Adjust the allocation weights along the gradient ascent direction of the resource utility function, and use the projected gradient method to constrain the weight correction within the safety and compliance range to avoid the generation of illegal solutions; For high entropy areas, idle resources are allocated first and the resource occupancy status of adjacent areas is updated synchronously; Define hard constraints through the engineering safety specification library, including minimum manpower configuration thresholds, equipment safety redundancy, and material inventory minimums; Call the linear programming solver to verify the allocation plan, filter out illegal solutions that violate constraints, and output the final compliant allocation plan.

9. The method for optimizing engineering resource allocation based on artificial intelligence according to claim 8, characterized in that: The specific process of collecting resource utilization and duration deviation indicators from actual execution data and dynamically transmitting them back to the dynamic heterogeneous graph network to update edge weight parameters is as follows: The IoT terminal of the engineering management system collects resource utilization indicators in real time, including the proportion of mechanical equipment working hours, labor load rate, and material consumption rate, extracts the characteristics of construction period deviation, and calculates the deviation rate between the actual progress of critical path tasks and the planned progress; The demand intensity coefficient of the task-resource edge is adjusted inversely based on the resource utilization deviation. Task nodes with lower-than-expected resource utilization have their demand weight reduced, while those with lower-than-expected resource utilization have their priority increased. The dependency weights of task nodes are dynamically modified based on the duration deviation rate. Task nodes with deviation rates exceeding the threshold are weakened in terms of the dependency strength of their subsequent tasks. The updated edge weight parameters are propagated back to the dynamic heterogeneous graph network, and only the edges of the affected subgraphs are incrementally adjusted to maintain the stability of the global network structure. Historical execution data is filtered based on the sliding window mechanism to prevent noise data from interfering with model parameters.

10. An artificial intelligence-based engineering resource allocation optimization system, applying the artificial intelligence-based engineering resource allocation optimization method according to any one of claims 1 to 9, characterized in that: It includes the following modules: heterogeneous data acquisition and encoding module, dynamic heterogeneous graph construction module, collaborative optimization decision module, resource dynamic monitoring module, and solution execution module; The heterogeneous data acquisition and encoding module is used to collect multimodal heterogeneous data in the target engineering scenario in real time, including task attribute data, resource status data and environmental interference data, perform spatiotemporal alignment and semantic encoding on the multimodal heterogeneous data, and generate a multimodal dynamic data stream; The dynamic heterogeneous graph construction module is used to construct a dynamic heterogeneous graph network based on multimodal dynamic data streams, define the heterogeneous interaction relationships between task nodes, resource nodes, and environment nodes, calculate the dynamic dependency weights between nodes based on the multimodal graph attention mechanism, and update the edge weights through an event-driven mechanism: The collaborative optimization decision module is used to perform two-stage collaborative optimization, mapping the dynamic dependencies output by the dynamic heterogeneous graph network into quantifiable constraints, simulating external interference scenarios based on Monte Carlo sampling, generating a probability distribution of resource requirements, and constructing a stochastic optimization model to solve a baseline allocation plan that minimizes risk-weighted costs. The baseline plan is decomposed into local decision-making problems for multiple seed agents, and resource conflicts between agents are coordinated through a backpropagation negotiation protocol to output a conflict-resolved allocation plan. The resource dynamic monitoring module is used to monitor the dynamic resource entropy index of resource distribution in real time. When a local overload is detected, the gradient-driven resource rebalancing algorithm is triggered to adjust the allocation weights in the direction of utility maximization. A safety constraint verification layer is embedded to filter illegal solutions and output an optimized allocation plan. The scheme execution module is used to output the optimized allocation scheme to the engineering management system for execution, and at the same time collect resource utilization and construction period deviation indicators in the actual execution data, and dynamically transmit them back to the dynamic heterogeneous graph network to update the edge weight parameters.

Images