An industrial intelligent agent scheduling method and system based on end-side cloud task graph arrangement

Abstract

The application discloses an industrial intelligent agent scheduling method and system based on end-edge-cloud task graph arrangement, and belongs to the technical field of industrial intelligent scheduling.The application constructs a global state matrix through time-space alignment and feature fusion by collecting industrial multi-source heterogeneous data, and accurately represents field conditions.The complex production task is disassembled into an atomic task unit, a task hypergraph topology with multiple constraint couplings is constructed, and fine modeling of the complex production task is realized;an end-edge-cloud three-level intelligent agent collaborative architecture is built, a scheduling strategy for maximizing global benefits and minimizing local resource conflicts is solved based on a Nash equilibrium game algorithm, and a scheduling instruction is generated.A closed-loop dynamic iteration mechanism is introduced, a hypergraph edge weight gradient correction and a game strategy dynamic update are completed through task execution feedback error, the problems of low modeling accuracy, poor collaboration, static strategy and weak dynamic adaptation capability of traditional industrial scheduling are solved, and industrial resource utilization and task scheduling stability are effectively improved.

Description

Technical Field

[0001] This invention belongs to the field of industrial intelligent scheduling technology, specifically relating to an industrial intelligent agent scheduling method and system based on edge-cloud task graph orchestration. Background Technology

[0002] With the deep integration of intelligent manufacturing and industrial internet technologies, industrial production scenarios are gradually exhibiting characteristics of heterogeneous equipment, complex tasks, distributed resources, and dynamic networks. Edge-cloud collaborative architecture has become the core supporting architecture for computing power scheduling, task collaboration, and intelligent management in industrial settings. Industrial intelligent agents, as the core carriers for achieving autonomous scheduling of production tasks, autonomous control of equipment, and autonomous allocation of resources, directly determine the production efficiency, energy consumption level, and operational stability of intelligent manufacturing production lines through their scheduling efficiency, collaboration accuracy, and dynamic adaptability. High-precision, adaptive, and low-conflict intelligent agent scheduling technology has become a research hotspot and core requirement in the current field of industrial intelligence.

[0003] Currently, most existing industrial intelligent agent scheduling solutions adopt traditional fixed-time scheduling, static resource allocation, or single cloud-based centralized scheduling modes, which have many technical shortcomings in practical industrial applications. First, traditional scheduling methods are mostly based on static task lists and fixed resource constraints for task allocation, which can only adapt to standardized production scenarios with simple processes and stable operating conditions. They cannot fully perceive and quantify the multi-dimensional dynamic states of industrial sites, such as multi-source heterogeneous data, dynamic equipment status, real-time network fluctuations, and differentiated energy consumption thresholds. It is difficult to build a state assessment system covering all dimensions of production, resulting in insufficient data support for scheduling decisions and poor adaptability.

[0004] Secondly, existing task modeling methods mostly use ordinary directed acyclic graphs to model task relationships, which can only represent simple task temporal dependencies. They cannot adapt to the complex correlation characteristics of multi-process coupling, multi-resource sharing, and multi-constraint intersection in industrial composite production tasks. They are difficult to quantify multi-dimensional constraint relationships such as process dependence, time delay constraints, energy consumption thresholds, and business priorities. They suffer from low task modeling accuracy, incomplete constraint coverage, and weak topological representation capabilities, which can easily lead to production problems such as task temporal disorder, resource contention, and process violation.

[0005] Meanwhile, current mainstream scheduling strategies mostly employ static scheduling algorithms with a single optimization objective, focusing solely on maximizing production efficiency or minimizing energy consumption. They lack a hierarchical game-theoretic collaborative mechanism among edge, cloud, and end-user agents, failing to balance the dual optimization objectives of global production gains and local resource conflicts. Centralized cloud-based scheduling suffers from high latency, poor real-time performance, and significant computing power bottlenecks, while edge-local scheduling suffers from insufficient global coordination capabilities, fragmented resource scheduling, and difficulties in cross-regional collaboration. The balance and optimality of multi-level agent collaborative scheduling are difficult to guarantee.

[0006] Furthermore, existing industrial intelligent agent scheduling systems generally lack a closed-loop dynamic iterative optimization mechanism. The scheduling strategies and task topology models are mostly static and fixed, and cannot be adaptively updated and corrected according to the actual execution error of the task, changes in equipment status, and fluctuations in the network environment. The scheduling scheme lags behind changes in on-site working conditions, and under long-term operation, problems such as decreased scheduling accuracy, reduced resource utilization, and increased task conflicts are likely to occur, making it difficult to meet the actual production needs of modern industrial flexible production, dynamic scheduling, and high-precision collaboration.

[0007] In summary, existing industrial intelligent agent scheduling technologies have shortcomings such as insufficient global state perception capabilities, low accuracy in complex task modeling, poor balance in edge-cloud collaborative scheduling, static and fixed scheduling strategies, and lack of closed-loop dynamic optimization capabilities, making them unsuitable for complex, dynamic, and flexible industrial production scenarios. Summary of the Invention

[0008] To address the aforementioned technical deficiencies, this application proposes an industrial intelligent agent scheduling method and system based on edge-cloud task graph orchestration, which enables full-domain state quantification perception, accurate modeling of complex task hypergraphs, optimal scheduling through multi-agent game theory, and dynamic optimization of the entire process closed loop.

[0009] The technical solution adopted in this invention is as follows: In one implementation of the first aspect, this application proposes an industrial intelligent agent scheduling method based on edge-cloud task graph orchestration, comprising the following steps: S1: Collect multi-source heterogeneous data from industrial production scenarios, construct a global state matrix covering production operations, equipment status, resources, network environment, and historical iteration dimensions, and generate a global state feature vector through spatiotemporal alignment algorithm and feature fusion processing. S2: Based on the global state matrix, the composite production task is decomposed into atomic task units, and a task hypergraph topology is constructed that includes process dependence, resource constraints, latency determinism, energy consumption threshold and business priority association; the task hypergraph topology is a directed acyclic network structure with atomic tasks as nodes and various constraints as hyperedges. S3: Build a cloud-based overall planning intelligent agent, an edge-based collaborative intelligent agent, and a terminal execution intelligent agent. Based on Nash equilibrium, solve for the optimal scheduling strategy of each atomic task in the task hypergraph at the terminal, edge, and cloud levels and generate scheduling instructions. The optimal scheduling strategy simultaneously satisfies the dual objectives of maximizing global benefits and minimizing local resource conflicts. S4: Based on the scheduling instructions, the industrial intelligent agent is flexibly reorganized into atomic units and tasks are executed. The task execution status is collected and fed back to the global state matrix to drive the dynamic update of the task hypergraph topology and game strategy. The dynamic update includes the correction of the hypergraph edge weight gradient based on the feedback error.

[0010] In one implementation of the second aspect, this application proposes an industrial intelligent agent scheduling system based on edge-cloud task graph orchestration, which is constructed based on the industrial intelligent agent scheduling method described above, including: a global state perception module, a task hypergraph orchestration module, a multi-agent game scheduling module, and an iterative execution module; The global state perception module is used to collect multi-source heterogeneous industrial data, complete data cleaning, spatiotemporal alignment and feature fusion, construct and update the global state matrix in real time, and output the global state feature vector to provide data support for task orchestration and scheduling decisions. The task hypergraph orchestration module is communicatively connected to the global state perception module. It is used to decompose composite tasks based on global state feature vectors, extract multi-dimensional constraints, construct a directed acyclic task hypergraph topology with multi-constraint hyperedges, and realize quantitative modeling of the relationship between tasks and constraints. The multi-agent game scheduling module is connected to the task hypergraph orchestration module and the global state perception module respectively. It is used to build a three-level intelligent agent architecture of edge cloud, construct a game scheduling model and solve the Nash equilibrium optimal scheduling strategy, and generate and issue standardized hierarchical scheduling instructions. The iterative execution module, together with the multi-agent game scheduling module and the task hypergraph orchestration module, forms a closed-loop linkage to perform flexible reorganization of agents and atomic task scheduling, provide real-time feedback on task status to update the global matrix, complete hypergraph weight correction and iterative optimization of game strategy, and realize system adaptive dynamic updates.

[0011] In one implementation of the third aspect, this application proposes an electronic device including a processor, a memory, and a communication bus coupled to each other, wherein the processor and the memory communicate with each other through the communication bus; the memory stores a computer program, which, when executed by the processor, implements the industrial intelligent agent scheduling method as described above.

[0012] In one implementation of the fourth aspect, this application provides a computer storage medium storing computer program instructions that can be loaded and executed by a processor. When the program instructions are executed by the processor, they implement the industrial intelligent agent scheduling method.

[0013] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: This invention employs an integrated technical solution combining an edge-cloud collaborative architecture with task hypergraph orchestration and multi-agent game scheduling. By collecting multi-source heterogeneous industrial data and constructing a multi-dimensional global state matrix through spatiotemporal alignment and feature fusion, it achieves accurate quantitative perception of the dynamic state of the entire industrial production scenario, solving the problems of one-sided state perception, spatiotemporal data misalignment, and lagging condition representation in traditional scheduling technologies. Simultaneously, it abandons the traditional single-constraint modeling method of directed acyclic graphs, adopting a task hypergraph topology with atomic tasks as nodes and multi-dimensional production constraints as hyperedges to complete the refined decomposition of composite tasks and multi-constraint coupled modeling, effectively improving the modeling accuracy and logical representation capability of complex industrial coupled tasks. Furthermore, by constructing a three-level hierarchical intelligent agent architecture of cloud, edge, and terminal, based on Nash... The equilibrium game theory algorithm solves a dual-layer optimal scheduling strategy that balances maximizing global production benefits with minimizing local resource conflicts. This overcomes the technical drawbacks of traditional single-level scheduling, such as insufficient global coordination, frequent local conflicts, and suboptimal scheduling schemes. Ultimately, relying on a closed-loop adaptive mechanism of real-time feedback on task execution status, gradient correction of hypergraph edge weights, and dynamic iterative updates of the game theory strategy, the algorithm achieves dynamic adaptive optimization of the task topology model and scheduling strategy according to the on-site working conditions. This completely solves the defects of traditional industrial intelligent agent scheduling methods, such as static rigidity, poor adaptability, and continuous decay of scheduling accuracy under dynamic working conditions. It significantly improves the utilization rate of industrial production resources, the rationality of task scheduling, and the stability of edge-cloud collaborative operations, adapting to the high-precision intelligent scheduling needs of complex, dynamic, and flexible industrial production scenarios. Attached Figure Description

[0014] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the method in this invention; Figure 2 This is a block diagram of the system architecture in this invention.

[0015] Figure 3 This is a block diagram of the hardware composition of the electronic device of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0017] Example 1:

[0018] This embodiment provides an industrial intelligent agent scheduling method based on edge-cloud task graph orchestration. It can be widely applied to complex dynamic industrial production scenarios such as intelligent manufacturing production lines, flexible industrial production, collaborative equipment processing, and distributed industrial computing power scheduling. It effectively solves the technical problems of traditional industrial scheduling methods, such as static rigidity, significant resource conflicts, weak cross-level collaboration capabilities, and poor dynamic adaptability. (See reference...) Figure 1 Specifically, it includes the following steps: S1: Real-time acquisition of multi-source heterogeneous operational data in industrial production scenarios, systematic integration of key information from the entire production process, and construction of a global state matrix covering five core dimensions: production business operation status, real-time status of industrial equipment, hardware and software resource utilization, edge-cloud global network environment, and historical scheduling iteration optimization. For the multi-source heterogeneous data with chaotic time series, large dimensional differences, and inconsistent formats, a precise spatiotemporal alignment algorithm is used to complete the unified calibration of data dimensions, and a deep feature fusion algorithm is used to complete the removal of redundant information and effective feature coupling processing, ultimately generating a global state feature vector with unified dimensions, synchronized time series, accurate features, and direct support for subsequent task modeling and intelligent scheduling decisions.

[0019] S2: Based on the constructed global state matrix and corresponding global state feature vectors, and combined with the inherent process flow and production specifications of industrial production, complex production tasks with high complexity, strong process coupling, and multiple resource cross-occupancy are finely decomposed to obtain the smallest atomic task unit with uniform granularity, independent function, and independent scheduling and execution. At the same time, combined with the actual operation rules of industrial production, a task hypergraph topology is constructed that includes process dependencies, hardware and software resource occupancy constraints, task transmission and execution latency deterministic constraints, equipment operating energy consumption threshold constraints, and production business urgency priority correlation. Unlike the single association representation method of traditional ordinary directed acyclic graphs, this task hypergraph topology uses independent atomic tasks as topology nodes and various quantitative constraints as multi-dimensional association hyperedges, effectively representing the complex association relationships of multiple tasks sharing resources and multiple processes cross-coupling, and finally forming a directed acyclic mesh topology structure with clear structure, full constraint coverage, and rigorous logic.

[0020] S3: Build a three-tiered intelligent agent system adapted to the industrial edge-cloud layered architecture, deploying a cloud-based intelligent agent with global coordination capabilities, an edge-based intelligent agent with regional collaboration capabilities, and a terminal-based intelligent agent with on-site execution capabilities. Based on the completed task hypergraph topology model, multi-agent game theory is introduced, and the optimal allocation scheme and scheduling execution strategy for each atomic task in the terminal, edge, and cloud three-tier architecture are accurately solved based on the Nash equilibrium iterative solution mechanism, outputting standardized executable scheduling instructions. The optimal scheduling strategy solved in this step strictly follows the dual optimization objectives: under the premise of maximizing the overall global production benefits and optimizing production efficiency, it minimizes cross-level and cross-task resource contention conflicts, timing conflicts, and energy consumption conflicts, achieving local resource conflicts and achieving a global and local scheduling balance.

[0021] S4: Strictly based on the optimal scheduling instructions obtained from the solution, the functional units, resource permissions, and collaborative relationships of the three-level industrial intelligent agents at the edge, cloud, and end levels are dynamically and flexibly reorganized into atomic units to adapt to the execution requirements of the current batch of atomic tasks and initiate collaborative task execution. Throughout the task execution process, real-time feedback data such as the task running status, equipment condition, and resource consumption of each level of intelligent agents are continuously collected, and the feedback data is transmitted back in real time and iteratively updated to the global state matrix. This drives the dynamic adaptive update of the front-end task hypergraph topology and the multi-agent game scheduling strategy. The core mechanism of dynamic update includes deviation analysis based on the actual task execution error and the theoretical preset index, and adaptive optimization of the weights of each hyperedge of the hypergraph is completed through the gradient correction algorithm, realizing the closed-loop iterative optimization of the entire scheduling model.

[0022] In a further embodiment, step S1 specifically includes: S1.1: Multi-source heterogeneous data full-domain acquisition. Relying on various terminal sensors, data acquisition modules, edge gateway devices, and cloud-based production management platforms deployed in industrial sites, a multi-level synchronous data acquisition system is constructed to collect all-dimensional operational data in industrial production scenarios. The collected data specifically covers five core categories: production business process data, real-time equipment operating status data, hardware and software computing power and storage resource usage data, cross-level network bandwidth and latency environment data between the terminal, edge, and cloud, and historical task scheduling iteration optimization record data. The raw collected data includes structured data stored in industrial databases, semi-structured data such as equipment logs, and unstructured data such as image monitoring and operating waveforms. This completes the accurate collection of all elements and dimensions of raw data in industrial production, providing a complete data foundation for subsequent state modeling.

[0023] S1.2: Construction of the global state matrix involves standardizing and preprocessing the massive, multi-source, heterogeneous raw data that has been collected. This includes unifying data formats, cleaning dirty data, intelligently repairing missing data, and removing abnormal fluctuations, ensuring the validity and accuracy of the raw data. Using continuous time as the matrix row dimension and spatial device nodes of various terminal devices, edge nodes, and cloud computing nodes in the industrial field as the matrix column dimension, the matrix accurately associates the core attribute parameters of five dimensions: production business, equipment status, resource occupancy, network environment, and historical iteration. This constructs a global state matrix that is dimensionally comprehensive, updates data in real time, and is dynamically iterative, enabling the quantitative, structured, and visual representation of all scenarios, elements, and states of industrial production.

[0024] S1.3: Multi-source data spatiotemporal alignment processing addresses the data timing misalignment caused by different sampling frequencies of different acquisition devices and different transmission delays of different transmission links. A high-precision timestamp synchronization algorithm is adopted to unify the time base of the entire domain data, completely eliminating the difference in acquisition delay and data asynchrony deviation of multi-source data. At the same time, combined with the actual topology installation location of industrial production equipment and the actual execution space area corresponding to each task, the spatial dimension of multi-source data is accurately registered, effectively solving the spatiotemporal misalignment and dimension mismatch problems of cross-terminal, cross-edge, and cross-cloud data. Ultimately, it achieves a high degree of unification of the temporal and spatial dimensions of multi-source data across the entire domain, laying the foundation for subsequent feature fusion.

[0025] S1.4: Multi-dimensional feature fusion and feature vector generation. A lightweight deep feature fusion algorithm is used to perform hierarchical feature mining on standardized data after spatiotemporal alignment. First, shallow basic data features are extracted, and then deep correlation and coupling features are mined. Production business correlation features, equipment status time-series evolution features, resource dynamic occupancy fluctuation features, network environment real-time fluctuation features, and historical scheduling iteration optimization experience features are accurately extracted. Invalid, duplicate, and redundant features are removed through a feature screening mechanism, and core effective features with high correlation and high contribution rate are retained. The normalization and fusion of multi-dimensional core features are completed, and finally, a global state feature vector with unified dimensions, accurate features, comprehensive representation, and can be directly used for subsequent task decomposition, topology construction, and intelligent scheduling decisions is generated.

[0026] In a further embodiment, step S2 specifically includes: S2.1: Composite Production Task Analysis and Atomic Task Decomposition. Based on the production business rules, industrial process standards, and equipment execution capability parameters carried by the global state matrix, composite production tasks with high overall complexity are decomposed hierarchically and in detail. During the decomposition process, redundant and repetitive processes in the production process are eliminated, and complex task modules with deep coupling of multiple processes are broken down. The complex composite production tasks covering multiple processes, constraints, and procedures are decomposed into the smallest atomic task units with uniform granularity, independent functions, clear responsibilities, and can be individually scheduled and executed by intelligent agents at various levels of edge, cloud, and endpoint. At the same time, each atomic task after decomposition is attribute-labeled to accurately record the task type, resource execution requirements, process timing requirements, basic operating parameters, and initial constraint attributes of each atomic task, providing a unit foundation for subsequent constraint extraction and topology modeling.

[0027] S2.2: Multi-dimensional task constraint parameter extraction. Combining real-time updated equipment operation parameters, resource occupancy status, network transmission indicators, energy consumption monitoring data, and production business priority data in the global state matrix, for each decomposed atomic task, the corresponding multi-dimensional constraints are precisely extracted one by one. Specifically, these include process dependency constraints of process sequence, resource occupancy constraints of hardware computing power and storage, time delay determinism constraints of cross-level data transmission and task execution, maximum threshold constraints of equipment operating energy consumption, and urgency priority constraints of production business. Various constraints are quantified, interval-calibrated, and standardized to construct a complete set of atomic task constraint parameters, providing comprehensive and quantitative constraint data support for the accurate construction of subsequent task hypergraph topology.

[0028] S2.3: Construction of Task Hypergraph Nodes and Hyperedges. Using all decomposed independent atomic tasks as the topological foundation nodes, and five types of constraints corresponding to each atomic task—quantified process dependence, resource consumption, latency limits, energy consumption thresholds, and business priorities—as multi-dimensional associative hyperedges, this approach addresses the complex associative scenarios commonly found in industrial production, such as multiple tasks sharing the same equipment resources, multiple processes interleaving, and multiple constraints overlapping. It constructs a multi-branch, multi-associative hyperedge structure, distinguishing itself from the limitations of traditional topological structures with their single-edge associations. This structure accurately represents the multi-dimensional constraint relationships between complex tasks in industrial production, achieving comprehensive modeling of task associations, resource associations, and constraint associations.

[0029] S2.4: The directed acyclic task hypergraph topology is formed, strictly following the inherent sequential execution logic of industrial production processes and the constraint association order between atomic tasks. The hypergraph nodes and multi-dimensional hyperedges are precisely labeled in a directed manner, clearly defining the legal execution flow and scheduling order of each atomic task. Through the topology verification mechanism, invalid topologies with cyclic associations, reverse constraints, and logical contradictions are automatically identified and eliminated. Finally, a directed acyclic task hypergraph mesh topology structure with acyclic paths, full constraint coverage, logical traversability, and adaptability to flexible production scenarios is constructed, providing an accurate, reliable, and field-condition-appropriate task topology model for subsequent multi-agent game scheduling and optimal strategy solving.

[0030] In a further embodiment, step S3 specifically includes: S3.1: A three-tiered intelligent agent architecture is constructed. Based on the hierarchical architecture characteristics and division of labor requirements of industrial edge-cloud collaborative scheduling, a three-tiered intelligent agent system with differentiated functions and collaborative linkage is specifically built, namely, the cloud-based overall planning intelligent agent, the edge-based collaborative intelligent agent, and the terminal execution intelligent agent. Among them, the cloud-based overall planning intelligent agent is mainly responsible for global task planning, cross-edge node resource coordination, overall production benefit optimization, and global anomaly control, and has the ability to make decisions across the entire domain. The edge-based collaborative intelligent agent is mainly responsible for the coordination and allocation of atomic tasks within its jurisdiction, the resolution of local resource conflicts, the collaborative interaction of neighboring nodes, and the overall coordination of regional operating conditions. The terminal execution intelligent agent is mainly responsible for the implementation and execution of single-point atomic tasks, real-time on-site status collection, local resource control, and execution feedback reporting. Finally, a three-tiered intelligent agent scheduling architecture with clear hierarchical division of labor, vertical linkage and collaboration, and overall domain coordination and local adaptation is formed.

[0031] S3.2: Construction of a Multi-Agent Game Scheduling Model. Based on the established task hypergraph topology and global state feature vector, a multi-agent non-cooperative game scheduling model adapted to edge-cloud collaborative scenarios is constructed. Cloud, edge, and terminal agents are defined as the main participants in the game, and the allocation schemes, resource configuration methods, and execution timing plans of each atomic task at different levels are set as game strategy variables. Combining the core optimization objectives of industrial production, a multi-dimensional comprehensive benefit function is constructed, covering global production completion benefits, comprehensive utilization rate of software and hardware resources, task execution latency benefits, and equipment energy consumption. At the same time, the maximum capacity limit of equipment resources, network transmission latency threshold, production business priority rules, and process execution logic are used as hard game constraints to complete the construction of the multi-agent game scheduling model.

[0032] S3.3: Solving the optimal scheduling strategy based on Nash equilibrium, the constructed multi-agent game scheduling model is iteratively deduced and precisely solved. The allocation and combination schemes of atomic tasks at different levels of terminals, edge, and cloud are comprehensively traversed, and the task allocation strategy and resource allocation strategy of each level of agent are continuously iteratively optimized. As the iterations converge, the entire game system eventually reaches a stable state of Nash equilibrium. In this equilibrium state, the hierarchical allocation and execution scheduling scheme of all atomic tasks can maximize the overall comprehensive benefits of industrial production, and minimize local contradictions such as resource contention, timing conflicts, and energy consumption exceeding limits across levels and tasks, accurately satisfying the dual scheduling optimization objectives of global optimization and minimization of local conflicts.

[0033] S3.4: Standardized scheduling instruction generation and distribution. The optimal scheduling strategy obtained after the Nash equilibrium solution converges is structurally analyzed and decomposed. Combining the responsibilities, resource permissions, and actual execution capabilities of each level of intelligent agent, the execution requirements, timing requirements, and constraints of each atomic task are matched to generate a standardized scheduling instruction package containing task execution node location, precise execution timing, dedicated resource configuration parameters, latency control standards, energy consumption limit thresholds, and multi-agent collaborative interaction rules. Following the hierarchical division of the edge-cloud three-tier architecture, the standardized scheduling instructions are precisely distributed to the corresponding industrial intelligent agents, completing the implementation of the globally optimal scheduling decision and providing precise instruction basis for subsequent task execution.

[0034] In a further embodiment, the solution for the optimal scheduling strategy of the Nash equilibrium specifically includes: S3.3.1: Game Parameter Initialization Configuration. Based on the established task hypergraph topology and the real-time updated global state feature vector, the basic parameter initialization configuration for multi-agent game solving is completed. The executable policy space, resource scheduling weight, comprehensive payoff calculation coefficient, and various constraint thresholds of the cloud, edge, and terminal agents are initialized and assigned values ​​respectively. The set of schedulable atomic tasks, the upper limit of hardware resource occupation, the maximum allowable network transmission latency, and the benchmark threshold of device energy consumption for each level of agent are clearly defined, and the boundary range, constraint interval, and initial iteration benchmark parameters of the game solution are fully defined, providing initial conditions for subsequent strategy iteration and deduction.

[0035] S3.3.2: Layered strategy iterative deduction adopts a distributed parallel iterative deduction mechanism to carry out multi-agent strategy coupling optimization. Each level of agent independently generates initial task scheduling and resource allocation strategies based on its own permission scope and local real-time running status. The cloud agent coordinates and optimizes the global strategy, the edge agent completes the adaptation and coordination of multi-node strategies and conflict pre-resolution within its jurisdiction, and the terminal agent provides real-time feedback on local device execution constraints and operating condition limitations to complete the initial coupling and matching of multi-level scheduling strategies. Finally, multiple sets of alternative scheduling strategy schemes are generated to provide sample support for subsequent strategy evaluation and screening.

[0036] S3.3.3: Comprehensive Benefit and Conflict Quantification Evaluation. For each set of candidate scheduling strategy schemes generated iteratively, a preset multi-dimensional comprehensive benefit function is substituted into the scheme for quantitative calculation. The global production benefit, hardware and software resource utilization efficiency, task latency optimization benefit, and energy consumption control benefit corresponding to each scheme are accurately calculated to complete the quantitative evaluation of the global benefit dimension. At the same time, the conflict problems existing in each scheme are comprehensively detected. The number and intensity of resource preemption conflicts, execution timing conflicts, and energy consumption over-limit conflicts across agents and across atomic tasks are counted. The merits and demerits of all candidate strategies are quantitatively compared and evaluated from the dual dimensions of optimal benefit and minimum conflict.

[0037] S3.3.4: Strategy Iteration Optimization and Convergence Judgment. With the core dual optimization objectives of maximizing global production benefits and minimizing local resource conflicts, the scheduling strategies with low benefits, high conflicts, and low adaptability selected through evaluation are subjected to gradient back correction. The task allocation weights, resource allocation ratios, and timing scheduling schemes of each agent are iteratively updated. The strategy deduction, evaluation, and correction iteration work is continuously carried out in a loop until the scheduling strategies of all levels of agents tend to be stable and no longer actively changed. Furthermore, any single agent's adjustment of its own strategy cannot improve its own benefits and may even reduce the global benefits. At this point, the multi-agent game system is judged to have reached a strict Nash equilibrium convergence state.

[0038] S3.3.5: The optimal scheduling strategy output is solidified. A comprehensive validity verification is performed on the balanced scheduling strategy after Nash equilibrium convergence. Each strategy is checked to see if it meets all hard constraints such as process logic, resource carrying capacity limit, latency constraint standard, energy consumption threshold limit, and business priority rules. Invalid and suboptimal strategies with process violations, resource limits, latency exceeding standards, and energy consumption exceeding thresholds are eliminated. Finally, the scheduling scheme that is fully adapted to the field conditions, meets all constraints, and takes into account both global and local optima is selected. The precise allocation of each atomic task in the three-level architecture of end, edge, and cloud is completed and the optimal scheduling strategy is solidified.

[0039] In a further embodiment, step S4 specifically includes: S4.1: Flexible Reorganization and Task Deployment of Industrial Intelligent Agents. Based on the standardized optimal scheduling instructions issued by the edge-cloud three-level architecture, the system accurately decomposes the task execution responsibilities, resource call permissions, cross-node collaborative interaction rules, and work priorities corresponding to each level of intelligent agents. Combining the execution granularity, process timing logic, resource requirement standards, and collaborative requirements of the current batch of atomic tasks, the system dynamically and flexibly combines and reconstructs the functions of the cloud-based coordination unit, the edge-based collaborative unit, and the terminal execution unit. It adaptively adjusts the working mode and resource configuration status of each intelligent agent, quickly builds a multi-agent collaborative execution architecture that adapts to the current task requirements, and completes the architecture adaptation and deployment preparation work before task execution.

[0040] S4.2: Atomic tasks are executed in a hierarchical and collaborative manner. After flexible reorganization, the various levels of industrial intelligent agents strictly follow the execution sequence, resource allocation standards, latency constraint thresholds, and energy consumption control requirements specified in the scheduling instructions. According to the task process logic, each atomic task is executed collaboratively in a hierarchical parallel or serial combination. Among them, the cloud intelligent agent is responsible for global scheduling and monitoring, overall progress coordination, and cross-regional anomaly control. The edge intelligent agent is responsible for the collaborative connection of multiple atomic tasks within the region, cross-terminal data transfer and processing, and local operating condition control. The terminal intelligent agent is responsible for the precise implementation of atomic tasks on site, real-time operating condition control, and equipment status monitoring. All parties cooperate to ensure the efficient, orderly, and compliant progress of the overall complex production task.

[0041] S4.3: Real-time feedback and collection of execution status throughout the entire lifecycle of atomic tasks. A real-time data feedback link is established across the entire domain to continuously collect the real-time completion progress, actual execution latency, actual energy consumption, real-time hardware and software resource utilization, industrial equipment operating status, edge-cloud network transmission quality, and various abnormal fault alarm information of each atomic task. This data is integrated to form a complete task execution feedback dataset. All real-time feedback data is then synchronously transmitted back to the global status matrix in real time, and historical status data is iteratively updated to ensure that the global status matrix is ​​synchronized and dynamically consistent with the actual operating conditions of the industrial site in real time.

[0042] S4.4: Hypergraph edge weight gradient correction and dynamic topology update. This involves calculating the difference between actual performance metrics such as latency, energy consumption, resource usage, and completion accuracy during task execution and the pre-set standard constraint metrics of the task hypergraph, dimension by dimension, to accurately obtain the multi-dimensional execution feedback error. Based on the quantified feedback error, a gradient descent algorithm is used to adaptively correct the various constraint weights corresponding to the hyperedges in the task hypergraph. These include process dependency constraint weights, resource usage constraint weights, latency constraint weights, energy consumption threshold weights, and business priority weights. This dynamically optimizes the constraint association strength and association ratio of the hypergraph topology, enabling adaptive iterative updates of the task hypergraph topology structure as the on-site working conditions change.

[0043] S4.5: Dynamic iterative optimization of the game scheduling strategy. Combining the real-time updated global state matrix and the new task hypergraph topology after weight correction, the multi-agent game iterative solution process is restarted to solve the Nash equilibrium optimal scheduling strategy again. The payoff function weights and constraint boundary conditions of the multi-agent game model are dynamically updated to effectively resolve the problems of strategy lag and insufficient adaptability caused by historical scheduling deviations and dynamic changes in working conditions. The scheduling accuracy is continuously optimized, and finally a fully closed-loop, adaptive, and self-iterative industrial intelligent agent scheduling mechanism of "scheduling execution - state feedback - topology correction - strategy update" is constructed.

[0044] Example 2:

[0045] An industrial intelligent agent scheduling system based on edge-cloud task graph orchestration is constructed based on the method described above. (See reference...) Figure 2 It includes: a global state perception module, a task hypergraph orchestration module, a multi-agent game scheduling module, and an iterative execution module; The global state perception module is used to collect multi-source heterogeneous data in industrial production scenarios, complete data cleaning, spatiotemporal alignment and feature fusion processing, construct and update the global state matrix in real time, and output a standardized global state feature vector to provide full-dimensional state data support for task orchestration and scheduling decisions. The task hypergraph orchestration module is communicatively connected to the global state perception module. It is used to complete the atomic decomposition of composite production tasks based on global state feature vectors, extract multi-dimensional task constraints, construct a directed acyclic task hypergraph topology containing multi-constraint hyperedges, and realize the visualization and quantitative modeling of production tasks and constraint relationships. The multi-agent game scheduling module is connected to the task hypergraph orchestration module and the global state perception module respectively. It is used to build a three-level intelligent agent architecture of edge cloud, construct a multi-agent game scheduling model, and solve the optimal scheduling strategy with the maximum global benefit and the minimum local conflict based on Nash equilibrium iterative solution, and generate and issue standardized hierarchical scheduling instructions. The iterative execution module, together with the multi-agent game scheduling module and the task hypergraph orchestration module, forms a closed-loop linkage. It is used to receive scheduling instructions to complete the flexible reorganization of industrial agents and the hierarchical collaborative execution of atomic tasks, collect task execution status in real time and update the global state matrix, and complete the hypergraph edge weight gradient correction and game strategy iterative optimization based on the execution feedback error, so as to realize the adaptive dynamic update of the scheduling system.

[0046] Example 3:

[0047] An electronic device, see Figure 3The system includes a processor, a memory, and a communication bus that are coupled together. The processor and the memory communicate with each other through the communication bus. The memory stores a computer program. When the computer program is executed by the processor, it implements the industrial intelligent agent scheduling method based on edge-cloud task graph orchestration as described above. The electronic device specifically includes an industrial server, an edge computing gateway, or a cloud data center node. The processor is specifically a CPU, GPU, or FPGA chip to adapt to the computing power requirements in different scenarios.

[0048] Example 4:

[0049] A computer storage medium stores computer program instructions that can be loaded and executed by a processor. When the instructions are executed by the processor, the industrial intelligent agent scheduling method based on edge-cloud task graph orchestration is performed. The computer storage medium specifically includes ROM, RAM, disk or optical disk. The computer program instructions include data acquisition instructions, hypergraph construction instructions, game scheduling instructions and execution control instructions, thereby constructing a dedicated industrial scheduling system on general-purpose hardware.

[0050] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An industrial intelligent agent scheduling method based on edge-cloud task graph orchestration, characterized in that, Includes the following steps: S1: Collect multi-source heterogeneous data from industrial production scenarios, construct a global state matrix covering production operations, equipment status, resources, network environment, and historical iteration dimensions, and generate a global state feature vector through spatiotemporal alignment algorithm and feature fusion processing. S2: Based on the global state matrix, the composite production task is decomposed into atomic task units, and a task hypergraph topology is constructed that includes process dependence, resource constraints, latency determinism, energy consumption threshold and business priority association; the task hypergraph topology is a directed acyclic network structure with atomic tasks as nodes and various constraints as hyperedges. S3: Build a cloud-based overall planning intelligent agent, an edge-based collaborative intelligent agent, and a terminal execution intelligent agent. Based on Nash equilibrium, solve for the optimal scheduling strategy of each atomic task in the task hypergraph at the terminal, edge, and cloud levels and generate scheduling instructions. The optimal scheduling strategy simultaneously satisfies the dual objectives of maximizing global benefits and minimizing local resource conflicts. S4: Based on the scheduling instructions, the industrial intelligent agent is flexibly reorganized into atomic units and tasks are executed. The task execution status is collected and fed back to the global state matrix to drive the dynamic update of the task hypergraph topology and game strategy. The dynamic update includes the correction of the hypergraph edge weight gradient based on the feedback error.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S1.1: Multi-source heterogeneous data full-domain collection. Industrial production data is collected at multiple levels through industrial terminal sensors, edge gateways, and cloud production management platforms. Data types include production business data, equipment status data, resource usage data, network environment data, and historical iterative scheduling data. The raw data covers structured, semi-structured, and unstructured data, completing the full-dimensional data collection. S1.2: Construction of the global state matrix. The collected raw data is cleaned, completed and anomaly removed. The data is preprocessed by time as the row and equipment space nodes as the column. The five major dimension attribute parameters are associated to construct a real-time updated full-coverage global state matrix to realize the quantitative representation of the full state of industrial scenarios. S1.3: Multi-source data spatiotemporal alignment processing, using a timestamp synchronization algorithm to unify the time base of multi-source data, eliminating acquisition delay and asynchronous deviation; combining device topology location and task space region to complete spatial registration, realizing spatiotemporal dimension unification of cross-end, edge, and cloud data; S1.4: Multi-dimensional feature fusion and feature vector generation. Deep and shallow features are extracted and coupled and fused into spatiotemporally aligned data to mine core features of business, equipment, resources, network and iterative optimization. Redundant features are removed and normalized to generate a global state feature vector that can be directly used for task decomposition and scheduling decisions.

3. The method according to claim 2, characterized in that, Step S2 specifically includes: S2.1: Composite production task analysis and atomic task decomposition. Based on the business rules and process standards corresponding to the global state matrix, composite production tasks are decomposed hierarchically, redundant processes are eliminated, and coupled modules are split to obtain the smallest atomic task execution unit with uniform granularity and independent function. The type, execution requirements and basic attribute parameters of each atomic task are marked. S2.2: Multi-dimensional task constraint parameter extraction. Based on the global state matrix, extract the process dependence, hardware resources, transmission latency, equipment energy consumption and business priority constraints of each atomic task, quantify various constraint indicators, and construct the atomic task constraint parameter set; S2.3: Construction of task hypergraph nodes and hyperedges. Using atomic tasks as topological nodes and various quantitative constraints as associated hyperedges, a multi-connection hyperedge structure is constructed for scenarios of multi-task resource sharing and process coupling to achieve accurate representation of multi-constraint relationships. S2.4: Directed Acyclic Task Hypergraph Topology Formation. Based on the process flow logic and constraint relationships, the hypergraph is labeled in a directed manner, and invalid topological structures with loops and reverses are eliminated to construct a directed acyclic task hypergraph mesh topology that is acyclic, fully constrained, and traversable.

4. The method according to claim 1, characterized in that, Step S3 specifically includes: S3.1: A three-tiered intelligent agent architecture is built, consisting of a cloud-based overall planning intelligent agent, an edge-based collaborative intelligent agent, and a terminal-based execution intelligent agent. The cloud-based overall planning intelligent agent is responsible for global planning, cross-regional resource coordination, and benefit optimization; the edge-based collaborative intelligent agent is responsible for regional task coordination and resource conflict resolution; and the terminal-based execution intelligent agent is responsible for local task execution, status collection, and resource regulation, forming a three-tiered scheduling architecture with division of labor and collaboration. S3.2: Construction of a multi-agent game scheduling model. Based on the task hypergraph topology, with three-level agents as game participants and atomic task allocation schemes as policy variables, a comprehensive revenue function including production revenue, resource utilization, latency, and energy consumption is constructed. Combining resource, latency, business priority, and process logic constraints, a multi-agent non-cooperative game scheduling model is established. S3.3: Solving the optimal scheduling strategy based on Nash equilibrium, iteratively deducing the task allocation combination of edge-cloud in the game model, continuously optimizing the agent scheduling strategy until Nash equilibrium is reached, minimizing cross-level and cross-task resource conflicts while maximizing global benefits. S3.4: Standardized scheduling instruction generation and issuance, parsing the optimal balanced scheduling strategy, matching the rights and responsibilities and execution capabilities of each intelligent agent, generating standardized scheduling instructions that include execution nodes, timing, resource configuration, latency and energy consumption limits and coordination rules, and issuing them layer by layer to the corresponding industrial intelligent agents.

5. The industrial intelligent agent scheduling method based on edge-cloud task graph orchestration according to claim 4, characterized in that, The solution to the optimal scheduling strategy for Nash equilibrium specifically includes: S3.3.1: Initialization and configuration of game parameters. Based on the task hypergraph and global state feature vector, initialize the policy space, resource weights, payoff coefficients and constraint thresholds of the three-level agent, define the set of schedulable tasks, resource upper limit, latency and energy consumption benchmark, and determine the game solution boundary and initial iteration parameters. S3.3.2: Layered strategy iterative deduction. Each level of intelligent agent generates an initial scheduling strategy based on its own permissions and local state. Through cloud coordination, edge adaptation, and terminal constraint feedback, the multi-level strategy coupling and matching is completed, and multiple sets of candidate scheduling schemes are output. S3.3.3: Comprehensive benefit and conflict quantitative evaluation. Calculate the global comprehensive benefit for each group of candidate strategies, and count the number and intensity of resource, time series, and energy consumption conflicts to complete a two-way quantitative evaluation of the advantages and disadvantages of the strategies. S3.3.4: Policy Iteration Optimization and Convergence Judgment. With the goal of maximizing global benefits and minimizing local conflicts, high-conflict and low-return policies are corrected by gradient, and task allocation and resource configuration schemes are iteratively updated until the policies of each agent are stable and unilateral policy changes can no longer improve benefits, and the Nash equilibrium convergence state is determined. S3.3.5: The optimal scheduling strategy output is solidified, the compliance of the strategy after convergence is verified, invalid schemes that violate process, resource, latency and energy consumption constraints are eliminated, the globally optimal scheduling strategy is selected, and the precise allocation and solidification of atomic tasks at the edge-cloud level is completed.

6. The method according to claim 1, characterized in that, Step S4 specifically includes: S4.1: Flexible reorganization and task deployment of industrial intelligent agents. Based on hierarchical scheduling instructions, the responsibilities, resource permissions and collaborative relationships of each intelligent agent are decomposed, and the edge-cloud intelligent agent units are dynamically and flexibly reconstructed to adapt to the execution granularity, process timing and resource requirements of the current atomic task. S4.2: Atomic tasks are executed in a hierarchical and collaborative manner. Each level of intelligent agent executes atomic tasks in parallel or serially according to the timing, resource, latency and energy consumption constraints of the scheduling instructions. The cloud is responsible for global monitoring and anomaly fallback, the edge is responsible for regional coordination and data transfer, and the terminal is responsible for on-site execution and working condition management. S4.3: Real-time feedback and collection of the entire process execution status. The task progress, latency, energy consumption, resource usage, device status, network quality and abnormal information are collected in real time throughout the entire task execution process, forming a feedback dataset and updating it synchronously to the global status matrix. S4.4: Hypergraph edge weight gradient correction and topology dynamic update. Calculate the multi-dimensional feedback error between the actual task indicators and preset constraints. Correct the hypergraph process, resources, latency, energy consumption and priority weights through gradient descent algorithm to achieve adaptive update of task topology. S4.5: Dynamic iterative optimization of game scheduling strategy. Based on the updated global state matrix and the corrected hypergraph topology, the optimal Nash equilibrium scheduling strategy is re-solved, the game weights and constraint boundaries are dynamically updated, scheduling lag bias is eliminated, and a closed-loop adaptive scheduling mechanism is formed.

7. An industrial intelligent agent scheduling system based on edge-cloud task graph orchestration, characterized in that, The system is constructed based on the method described in any one of claims 1-6, comprising: a global state perception module, a task hypergraph orchestration module, a multi-agent game scheduling module, and an iterative execution module; The global state perception module is used to collect multi-source heterogeneous industrial data, complete data cleaning, spatiotemporal alignment and feature fusion, construct and update the global state matrix in real time, and output the global state feature vector to provide data support for task orchestration and scheduling decisions. The task hypergraph orchestration module is communicatively connected to the global state perception module. It is used to decompose composite tasks based on global state feature vectors, extract multi-dimensional constraints, construct a directed acyclic task hypergraph topology with multi-constraint hyperedges, and realize quantitative modeling of the relationship between tasks and constraints. The multi-agent game scheduling module is connected to the task hypergraph orchestration module and the global state perception module respectively. It is used to build a three-level intelligent agent architecture of edge cloud, construct a game scheduling model and solve the Nash equilibrium optimal scheduling strategy, and generate and issue standardized hierarchical scheduling instructions. The iterative execution module, together with the multi-agent game scheduling module and the task hypergraph orchestration module, forms a closed-loop linkage to perform flexible reorganization of agents and atomic task scheduling, provide real-time feedback on task status to update the global matrix, complete hypergraph weight correction and iterative optimization of game strategy, and realize system adaptive dynamic updates.

8. An electronic device, characterized in that, It includes a processor, a memory, and a communication bus that are coupled together, wherein the processor and the memory communicate with each other through the communication bus; the memory stores a computer program, and when the computer program is executed by the processor, it implements the scheduling method as described in any one of claims 1-6.

9. A computer storage medium, characterized in that, The system stores computer program instructions that can be loaded and executed by a processor, wherein the program instructions, when executed by the processor, implement the scheduling method as described in any one of claims 1-6.

Images