AI collaborative hierarchical multi-agent industrial internet production scheduling method and system

Abstract

The application discloses an AI cooperative hierarchical multi-agent industrial internet production scheduling method and system, relates to the technical field of industrial internet production scheduling, and comprises the following steps: uploading a scheduling task to a meta-controller for pre-evaluation, determining a task hierarchical agent architecture; receiving scheduling task execution low-dimensional embedding coding and scheduling trajectory decision, obtaining a prior task strategy; importing the prior task strategy into a scheduling optimizer, performing hierarchical cascade scheduling configuration based on the task hierarchical agent architecture, and fusing to obtain a task scheduling strategy; and performing automatic production scheduling management on the scheduling task according to the task scheduling strategy. The application solves the technical problems of existing industrial production scheduling, such as fixed hierarchical architecture, single optimization mode, poor hierarchical collaboration and insufficient dynamic adaptation capability, achieves hierarchical multi-agent adaptive scheduling and double-branch parallel optimization, and improves the technical effects of the collaboration and dynamic adaptation of industrial production scheduling.

Description

Technical Field

[0001] This invention relates to the field of industrial internet production scheduling technology, specifically to an AI-collaborative hierarchical multi-agent industrial internet production scheduling method and system. Background Technology

[0002] In existing industrial internet production scheduling technologies, fixed-level scheduling architectures are mostly adopted. The hierarchical structure cannot be dynamically adjusted according to the characteristics of scheduling tasks. At the same time, scheduling decisions rely on traditional rule algorithms, the optimization methods are singular, and there is a lack of efficient collaboration mechanisms among multiple intelligent agents. It is difficult to adjust the scheduling strategy in real time according to changes in working conditions. There are problems such as unreasonable hierarchical coupling control, weak dynamic response capability, and insufficient scheduling accuracy and collaboration efficiency in complex scenarios, which cannot meet the actual scheduling needs of flexible and intelligent industrial production.

[0003] The existing industrial production scheduling suffers from technical problems such as a rigid hierarchical architecture, a single optimization mode, poor hierarchical coordination, and insufficient dynamic adaptability. Summary of the Invention

[0004] This application provides an AI-collaborative hierarchical multi-agent industrial internet production scheduling method and system, which is used to address the technical problems of rigid hierarchical architecture, single optimization mode, poor hierarchical collaboration, and insufficient dynamic adaptation capabilities in existing industrial production scheduling.

[0005] In view of the above problems, this application provides a hierarchical multi-agent industrial internet production scheduling method and system based on AI collaboration.

[0006] The first aspect of this application provides a hierarchical multi-agent industrial internet production scheduling method with AI collaboration, the method comprising: The scheduling task is uploaded to the meta-controller for pre-evaluation. By measuring task scheduling entropy, coupling density, and time urgency, the hierarchical structure and control flow graph of the hierarchical intelligent agent architecture are adaptively adjusted to determine the task hierarchical intelligent agent architecture. Using joint distribution samples and marginal distribution samples as training data, a scheduling decision-maker is constructed in conjunction with the optimization objective. This decision-maker receives the scheduling task, performs low-dimensional embedding encoding and scheduling trajectory decision-making, and obtains the prior task policy. The prior task policy is imported into the scheduling optimizer, and targeted optimization under generative perturbation is performed in parallel. This optimization is then fused with the hierarchical cascaded scheduling configuration based on the task hierarchical intelligent agent architecture to obtain the task scheduling policy. The task scheduling policy is identified by communication configuration parameters. Based on the task scheduling policy, automated production scheduling management is performed on the scheduling task.

[0007] A second aspect of this application provides an AI-coordinated hierarchical multi-agent industrial internet production scheduling system, the system comprising: The task-layered intelligent agent architecture determination module uploads the scheduling tasks to the meta-controller for pre-evaluation. By measuring task scheduling entropy, coupling density, and time urgency, it adaptively adjusts the hierarchical structure and control flow graph of the layered intelligent agent architecture to determine the task-layered intelligent agent architecture. The prior task policy acquisition module uses joint distribution samples and marginal distribution samples as training data, combined with optimization objectives, to construct a scheduling decision-maker. It receives the scheduling tasks, performs low-dimensional embedding encoding and scheduling trajectory decisions, and obtains the prior task policy. The task scheduling policy fusion module imports the prior task policy into the scheduling optimizer. Through parallel execution of targeted optimization under generative perturbation, it fuses the prior task policy with the hierarchical cascaded scheduling configuration based on the task-layered intelligent agent architecture to obtain the task scheduling policy. The task scheduling policy is identified by communication configuration parameters. The production scheduling management module performs automated production scheduling management on the scheduling tasks according to the task scheduling policy.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages: The scheduling task is uploaded to the meta-controller for pre-evaluation, and the hierarchical structure and control flow graph of the hierarchical agent architecture are adaptively adjusted to determine the task hierarchical agent architecture. Using joint distribution samples and marginal distribution samples as training data, the scheduling task is received and low-dimensional embedding encoding and scheduling trajectory decision are performed to obtain the prior task policy. The prior task policy is imported into the scheduling optimizer, and targeted optimization under generative perturbation is performed in parallel. This is then fused with the hierarchical cascaded scheduling configuration based on the task hierarchical agent architecture to obtain the task scheduling policy. According to the task scheduling policy, automated production scheduling management is performed on the scheduling task. This achieves the technical effect of realizing hierarchical multi-agent adaptive scheduling and dual-branch parallel optimization, improving the coordination and dynamic adaptability of industrial production scheduling. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 A schematic diagram of the AI-coordinated hierarchical multi-agent industrial internet production scheduling method provided in this application embodiment; Figure 2 A schematic diagram of the structure of an AI-coordinated hierarchical multi-agent industrial internet production scheduling system provided in this application embodiment.

[0011] Figure labeling: Task hierarchical intelligent agent architecture determination module 10, prior task strategy acquisition module 20, task scheduling strategy fusion module 30, production scheduling management module 40. Detailed Implementation

[0012] This application provides an AI-coordinated hierarchical multi-agent industrial internet production scheduling method and system to address the technical problems in existing industrial production scheduling, such as rigid hierarchical architecture, single optimization mode, poor hierarchical collaboration, and insufficient dynamic adaptation capabilities.

[0013] 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 them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] Example 1, as Figure 1 As shown, this application provides an AI-coordinated hierarchical multi-agent industrial internet production scheduling method, the method comprising: Step S100: Upload the scheduled task to the meta controller for pre-evaluation. By measuring the task scheduling entropy, coupling density, and time urgency, adaptively adjust the hierarchical structure and control flow graph of the hierarchical agent architecture to determine the task hierarchical agent architecture.

[0015] Specifically, the scheduling task is uploaded to the system meta-controller for pre-evaluation, with task scheduling entropy, coupling density, and time urgency as the core evaluation indicators. Based on the indicator calculation results, the hierarchical intelligent agent architecture is adaptively adjusted in terms of hierarchical structure and control flow graph. The hierarchical structure adaptation includes adding, merging, or splitting layers. At the same time, a differentiable temporal abstraction library is inserted between the adjusted architecture layers. High-frequency general scheduling patterns are generated by online clustering of upper and lower layer action sequences, and then encapsulated and stored in registers. Finally, the task hierarchical intelligent agent architecture adapted to the current scheduling task is determined.

[0016] Step S200: Using joint distribution samples and marginal distribution samples as training data, construct a scheduling decision-maker by combining the optimization objective, receive low-dimensional embedding encoding of scheduling task execution and scheduling trajectory decision, and obtain the prior task policy.

[0017] Specifically, the training data is obtained by retrieving joint distribution samples containing task embedding samples and scheduling trajectory samples, as well as randomly paired edge distribution samples. The scheduling decision-maker is constructed with the optimization objective of maximizing the mutual information between task embedding and scheduling trajectory and minimizing the mutual information between task embedding and environmental dynamics. The scheduling decision-maker is deployed with a first decision node for low-dimensional task embedding and a second decision node for task scheduling decision. After receiving the scheduling task, the first decision node extracts the order task features and concatenates them into a feature long vector. The feature long vector is then encoded with low-dimensional embedding to obtain a low-dimensional task embedding vector, which is input into the second decision node to complete the scheduling trajectory decision. Finally, the prior task strategy is output.

[0018] Step S300: Import the prior task strategy into the scheduler optimizer, and integrate it with the hierarchical cascaded scheduling configuration based on the task hierarchical intelligent agent architecture through parallel execution of targeted optimization under generative perturbation. The task scheduling strategy is identified by communication configuration parameters.

[0019] Specifically, the prior task strategy is imported into a scheduler optimizer that includes a generative perturbation branch and a dynamic hierarchical optimization branch. The two branches perform optimization processing in parallel. The generative perturbation branch generates task perturbation data based on the low-dimensional task embedding vector and performs targeted optimization on the prior task strategy under generative perturbation to obtain the first optimized scheduling strategy. The dynamic hierarchical optimization branch performs hierarchical cascaded scheduling configuration association based on the determined task hierarchical intelligent agent architecture to obtain the second optimized scheduling strategy. The first optimized scheduling strategy and the second optimized scheduling strategy are fused to finally obtain the task scheduling strategy carrying communication configuration parameters. The task scheduling entropy is proportional to the frequency of inter-layer communication interaction, and the coupling density is proportional to the communication bandwidth and communication frequency. The high coupling layer adopts a high-frequency broadcast communication mode, and the low coupling layer adopts an event-triggered communication mode.

[0020] Step S400: Perform automated production scheduling management on the scheduled task according to the task scheduling strategy.

[0021] Specifically, based on the generated task scheduling strategy, the hierarchical intelligent agent architecture implements hierarchical task scheduling coordination and management. Taking the smallest process unit in the task scheduling strategy as the benchmark, the scheduling data such as the actual start time, end time and equipment status of each process unit are fed back to the meta controller in real time. The meta controller compares the actual task scheduling data with the strategy progress step by step. When the accumulated deviation exceeds the preset threshold, a feedback scheduling strategy is automatically generated and distributed to the corresponding architecture layer to complete the closed-loop feedback scheduling and full-process automated production scheduling management.

[0022] In one possible implementation, step S100 further includes: Step S110: Upload the scheduling task to the meta controller, perform a pre-evaluation of the task scheduling, and determine the evaluation indicators, wherein the evaluation indicators include task scheduling entropy, coupling density, and time urgency.

[0023] Step S120: Based on the evaluation metrics, perform structural adaptation and control flow graph adaptation on the hierarchical agent architecture to determine the task hierarchical agent architecture, wherein structural adaptation includes adding, merging or splitting layers.

[0024] Specifically, the production scheduling tasks to be processed are uploaded to the system's core meta-controller. The meta-controller then conducts a comprehensive pre-assessment of the overall scheduling attributes of the tasks, such as complexity, relationships, execution sequence, and resource requirements. By quantitatively analyzing the task's execution process, resource consumption, process relationships, and delivery deadline, three core evaluation indicators are determined: task scheduling entropy, which characterizes the degree of disorder and decision complexity of task scheduling; coupling density, which characterizes the degree of close relationship between tasks; and time urgency, which characterizes the urgency of task execution. This provides a standardized and quantifiable evaluation basis for the subsequent adaptive adjustment of the intelligent agent architecture.

[0025] Based on the quantitative evaluation results of task scheduling entropy, coupling density, and time urgency, the hierarchical intelligent agent architecture is first dynamically adapted to the task execution at the macro-architectural level. The original tree-like architecture is adaptively transformed into a directed acyclic graph architecture that is more suitable for complex scheduling scenarios. Then, structural adaptation and control flow graph adaptation are carried out separately. The structural adaptation adds, merges, or splits the intelligent agent hierarchy according to the task coupling degree and complexity requirements. The control flow graph adaptation re-plans the node flow, data transmission path, and scheduling execution logic based on the adjusted hierarchical relationship. Finally, the task-hierarchical intelligent agent architecture that is suitable for the current scheduling task is determined.

[0026] In one possible implementation, step S120 further includes: Step S121: The task-layered intelligent agent architecture inserts a differentiable temporal abstraction library between layers.

[0027] Step S122: The differentiable temporal abstraction library generates a high-frequency general scheduling pattern by online clustering of upper and lower layer action sequences.

[0028] Step S123: Encapsulate and store the high-frequency general scheduling mode in registers.

[0029] Specifically, in the task-layered intelligent agent architecture after structural adaptation and control flow graph adaptation adjustment, a differentiable temporal abstraction library is inserted at the interaction interface between each layer. This temporal abstraction library serves as an intermediate component for inter-layer information transmission, action coordination, and scheduling logic reuse. It is used to uniformly receive, parse, and forward the action sequences and scheduling instructions of upper and lower layer intelligent agents, thereby improving the efficiency and consistency of layered collaborative scheduling.

[0030] The differentiable temporal abstraction library collects and caches temporal action sequences generated by upper and lower layer agents in real time during the scheduling process using a sliding window approach. Based on the similarity of temporal features, it performs online clustering on the cached action sequences. By calculating the distance metric between action execution timing, calling relationships, and interaction patterns, it identifies frequently occurring common scheduling sub-processes and generates high-frequency general scheduling patterns that can be reused across layers and trained differentially, providing standardized general execution units for hierarchical cascade scheduling.

[0031] The high-frequency general scheduling pattern generated by online clustering is standardized and encapsulated, and a unified calling interface, input and output parameters and timing constraints are configured for it. It is then written into a register for storage in a structured form that can be directly addressed and quickly invoked, so that the hierarchical intelligent agent can directly read and reuse it when making scheduling decisions, thereby improving the response speed and execution efficiency of hierarchical cascade scheduling.

[0032] In one possible implementation, step S200 further includes: Step S210: Retrieve joint distribution samples and marginal distribution samples, wherein the joint distribution samples include task embedding samples - scheduling trajectory samples, and the marginal distribution samples are randomly paired samples.

[0033] Step S220: Define an optimization objective, wherein the optimization objective is to maximize the mutual information between task embedding and scheduling trajectory, and minimize the mutual information between task embedding and environmental dynamic characteristics.

[0034] Step S230: Deploy the first decision node with low-dimensional task embedding, deploy the second decision node with task scheduling decision, and perform convergence training based on joint distribution samples and marginal distribution samples with the optimization objective as a constraint to obtain the scheduling decision-maker.

[0035] Specifically, sample data for training the scheduling decision-maker is retrieved from the system sample library. Joint distribution samples and marginal distribution samples are obtained respectively. The joint distribution samples are composed of task embedding samples after task features are encoded and matched one-to-one with the scheduling trajectory samples generated by actual execution, so as to maintain the real correlation between tasks and trajectories. The marginal distribution samples are generated by randomly shuffling the task embedding samples and scheduling trajectory samples and pairing them non-correspondingly, and are used to construct the contrast constraint sample set for model training.

[0036] To ensure the accuracy and robustness of scheduling decisions, the system constructs a dual-constraint optimization objective. By maximizing the mutual information between task embedding and scheduling trajectory, the task embedding can fully retain the identifiable and effective information of the scheduling trajectory. At the same time, by minimizing the mutual information between task embedding and dynamic environmental characteristics, transient noise and irrelevant interference caused by environmental disturbances are filtered out. This allows the final embedded spatial distance to accurately reflect the true scheduling similarity between different order combinations, providing a stable and reliable optimization guide for subsequent decisions.

[0037] In the model structure of the scheduling decision-maker, a first decision node and a second decision node are set up respectively. The first decision node is used to implement low-dimensional embedding encoding of task features and output a standardized low-dimensional task embedding vector. The second decision node is used to complete scheduling trajectory inference and decision output based on the low-dimensional task embedding vector. During the model training process, maximizing the mutual information between task embedding and scheduling trajectory and minimizing the mutual information between task embedding and environmental dynamic characteristics are taken as the core optimization objectives. Jointly distributed samples are used as positive samples with real associations and marginally distributed samples are used as negative samples with random pairings to form the training dataset. Supervised convergence training is performed on the two-node model. In each iteration, the model calculates the current embedding and decision results through forward propagation, calculates the loss value based on mutual information constraints, and performs backpropagation to continuously update the node parameters until the model converges and stabilizes. Finally, a scheduling decision-maker that can simultaneously complete low-dimensional embedding encoding and scheduling trajectory decision is trained.

[0038] In one possible implementation, step S200 further includes: Step S240: Upload the scheduling task, and based on the first decision node, extract the order task features and concatenate them into a long feature vector.

[0039] Step S250: Perform low-dimensional embedding encoding on the feature long vector to determine the low-dimensional task embedding vector.

[0040] Step S260: Embed the low-dimensional task into a vector flow and transfer it to the second decision node, then output the prior task strategy.

[0041] Specifically, the actual order scheduling tasks to be scheduled are input into the scheduling decision-maker. The first decision node in the scheduling decision-maker parses and processes the scheduling tasks, extracting multi-dimensional order task features such as task type, process sequence, equipment requirements, delivery time limit, resource constraints, and task coupling relationship. The extracted features are then concatenated and integrated in a preset order to generate a complete feature long vector, providing standardized input data for subsequent low-dimensional embedding encoding.

[0042] After the first decision node performs standardized preprocessing on the concatenated feature long vector, it carries out low-dimensional embedding encoding through nonlinear feature mapping and dimensionality compression operations to remove redundant noise information in the features, and maps the high-dimensional feature long vector to a unified low-dimensional feature space to generate a low-dimensional task embedding vector that can fully represent the core attributes of the order scheduling task, thus realizing a concise expression of task features.

[0043] The low-dimensional task embedding vector output by the first decision node is directly input into the second decision node. The second decision node, based on the converged decision network after training, combines task hierarchical constraints, process timing rules and resource occupation relationships to perform time-series deduction and scheduling scheme solution on the low-dimensional embedding vector, and outputs a standardized prior task strategy containing process sequence, equipment allocation and time window, providing an initial scheduling scheme for subsequent hierarchical cascade optimization.

[0044] In one possible implementation, step S300 further includes: Step S310: Construct a scheduling optimizer, wherein the scheduling optimizer includes a generative perturbation branch and a dynamic hierarchical optimization branch, wherein the dynamic hierarchical optimization branch embeds a dynamically updated task hierarchical agent architecture.

[0045] Step S320: Import the prior task strategy into the scheduler optimizer, perform dual-path branch parallel optimization and coordination, and generate a task scheduling strategy.

[0046] Specifically, a scheduling optimizer is constructed using a fusion algorithm of deep reinforcement learning and generative adversarial networks. This optimizer is designed as a dual-branch collaborative structure. The generative perturbation branch is implemented based on a conditional variational autoencoder (CVAE), which generates diverse candidate scheduling schemes by applying controllable structured perturbations to the prior task policy. The dynamic hierarchical optimization branch is implemented based on the deep deterministic policy gradient (DDPG) algorithm. Each branch embeds a dynamically updated task-hierarchical agent architecture. During training, the completion time, resource utilization, and cost of production scheduling are used as the comprehensive reward function. The prior task policy is used as the initial input. The generative perturbation branch first samples candidate schemes, and then the agents in the dynamic hierarchical optimization branch interact with the production environment in real time for trial and error. The agent network parameters are iteratively updated using the actor-critic mechanism of DDPG, while simultaneously updating the hierarchical relationships and interaction rules of the task-hierarchical agent architecture. The two branches perform end-to-end collaborative training through joint reward feedback until the scheduling optimizer converges, ultimately obtaining a scheduling optimization model with global optimization and dynamic adaptation capabilities.

[0047] The output prior task strategy is used as initial input and fully imported into the trained and converged scheduler optimizer, initiating a parallel collaborative optimization process involving a generative perturbation branch and a dynamic hierarchical optimization branch. The generative perturbation branch, based on the Conditional Variational Autoencoder (CVAE) algorithm, applies controllable structured perturbations to key parameters of the prior task strategy, such as process timing and resource allocation, generating multiple sets of differentiated candidate scheduling schemes while filtering out invalid and conflicting schemes. The dynamic hierarchical optimization branch, relying on the Deep Deterministic Policy Gradient (DDPG) algorithm and its embedded dynamic task hierarchical agent architecture, performs hierarchical cascade optimization of each set of candidate schemes, taking into account real-time production environment status and resource occupancy. The upper-level agent is responsible for global resource management and task priority adjustment, while the lower-level agent is responsible for refining single-task processes and correcting local conflicts. Through real-time data interaction and coordination mechanisms, the two branches screen and iteratively optimize candidate schemes with the core objectives of shortest completion time, highest resource utilization, and lowest cost, ultimately outputting a task scheduling strategy that balances feasibility, optimality, and dynamic adaptability.

[0048] In one possible implementation, step S320 further includes: Step S321: The generative perturbation branch receives the low-dimensional task embedding vector and generates task perturbation data.

[0049] Step S322: Based on the task disturbance data, perform targeted optimization on the prior task strategy to obtain a first optimized scheduling strategy.

[0050] Step S323: The dynamic hierarchical optimization branch receives the prior task strategy, and based on the task hierarchical intelligent agent architecture, executes hierarchical cascaded scheduling logic configuration association to determine the second optimized scheduling strategy.

[0051] Step S324: Combine the first optimized scheduling strategy and the second optimized scheduling strategy as the task scheduling strategy, wherein the task scheduling strategy includes communication configuration parameters based on a task-layered intelligent agent architecture.

[0052] Specifically, the generative perturbation branch takes the received low-dimensional task embedding vector as input and, relying on the trained Conditional Variational Autoencoder (CVAE) network, performs controlled sampling of the mean and variance of the hidden feature space corresponding to the embedding vector. While maintaining the core constraints of the task and the rationality of the scheduling, it applies structured Gaussian perturbations to key scheduling features such as process sequence, resource allocation weights, execution time windows, and equipment matching preferences. Finally, it generates standardized task perturbation data containing temporal offsets, resource adjustment coefficients, and priority perturbation terms, providing diversified and conflict-free adjustment basis for subsequent prior strategy optimization.

[0053] The task perturbation data output by the generative perturbation branch is mapped to the process layout, equipment allocation, time constraints, and resource ratio dimensions corresponding to the prior task strategy. Based on the time offset, resource adjustment coefficient, and priority perturbation items in the perturbation data, targeted corrections and fine-tuning are performed on issues such as local resource idleness, poor process connection, and time window redundancy in the prior task strategy. At the same time, scheduling constraint conflicts are checked and unreasonable adjustment items are eliminated. After iterative fine-tuning, a first optimized scheduling strategy that takes into account the diversity of solutions and local optimality is generated.

[0054] The dynamic hierarchical optimization branch receives the original prior task strategy and uses a dynamically updated task hierarchical intelligent agent architecture as the decision framework. It divides the structure into a hierarchical structure of global coordination intelligent agent, resource scheduling intelligent agent, and process execution intelligent agent, and carries out hierarchical cascade optimization from top to bottom. The upper-level global intelligent agent coordinates task priority and overall resource allocation, the middle-level resource intelligent agent matches equipment capacity and task load, and the lower-level process intelligent agent refines the process connection timing. Through the configuration association and constraint verification of the scheduling logic, the intelligent agents at each level correct hierarchical conflicts, resource imbalances and timing deviations in the prior task strategy, and finally outputs a second optimized scheduling strategy that adapts to the multi-agent collaborative constraints.

[0055] A weighted fusion and conflict verification mechanism is adopted to integrate the local disturbance optimization results of the first optimization scheduling strategy and the hierarchical collaborative optimization results of the second optimization scheduling strategy. Parameter normalization processing is performed on dimensions such as process sequence, resource allocation, and task priority to eliminate scheduling conflicts between the two types of strategies. At the same time, based on the task hierarchical intelligent agent architecture, communication configuration parameters such as communication frequency, instruction interaction format, and status feedback cycle between levels are configured, and intelligent agent collaborative constraints are embedded into the scheduling scheme. Finally, a task scheduling strategy with global optimality, local flexibility, and multi-agent collaborative adaptability is generated.

[0056] In one possible implementation, step S110 further includes: Step S111: The task scheduling entropy is positively correlated with the frequency of inter-layer communication interactions.

[0057] Step S112: The coupling density is positively correlated with the communication bandwidth and communication frequency. According to the preset coupling density distribution, a high coupling layer and a low coupling layer are divided. The high coupling layer adopts a high-frequency broadcast communication mode, and the low coupling layer adopts an event-triggered communication mode.

[0058] Specifically, task scheduling entropy is used to quantify the uncertainty and complexity of scheduling schemes under a hierarchical intelligent agent architecture. The higher the frequency of communication and interaction between intelligent agents at different levels, the greater the amount of information exchanged and the more complex the scheduling constraints, the higher the value of task scheduling entropy. Therefore, task scheduling entropy is positively correlated with the frequency of communication and interaction between layers.

[0059] The coupling density between layers characterizes the degree of connection between agents at each layer. Its magnitude is positively correlated with communication bandwidth and communication frequency. The larger the communication bandwidth, the higher the communication frequency, and the denser the data interaction between layers, the higher the coupling density. The system divides the task-layered agent architecture into high-coupling layers and low-coupling layers according to the preset coupling density distribution threshold. For high-coupling layers with strong interaction requirements and close connections, a high-frequency broadcast communication mode is used to ensure real-time synchronization of scheduling instructions. For low-coupling layers with low interaction frequency and weak dependencies, an event-triggered communication mode is used, transmitting data only when a scheduling event occurs, thereby reducing communication redundancy overhead.

[0060] In one possible implementation, step S400 further includes: Step S410: Based on the inter-layer relationships of the task-layered intelligent agent architecture, perform hierarchical driving management for task scheduling and coordination.

[0061] Step S420: Using the smallest process unit based on the task scheduling strategy, feed back the task scheduling data of each process unit to the meta controller in real time, wherein the task scheduling data includes the actual start time, end time and equipment status.

[0062] Step S430: The meta controller compares the task scheduling data with the strategy progress step by step. When the cumulative deviation exceeds a preset threshold, a feedback scheduling strategy is generated.

[0063] Step S440: The feedback scheduling strategy is sent to the associated architecture layer for feedback scheduling management.

[0064] Specifically, based on the hierarchical affiliation and association relationships built by the task-layered intelligent agent architecture, a top-down hierarchical driving management mechanism is established. The upper-level coordinating intelligent agent issues overall scheduling goals and constraint instructions, while the middle and lower-level intelligent agents take over and execute tasks layer by layer. Through the transmission of instructions and status feedback between levels, the orderly coordination and hierarchical control of the entire process task scheduling are realized.

[0065] The task scheduling strategy is broken down into the smallest indivisible process execution unit. The actual start time, actual end time, and corresponding equipment operating status and load conditions of each process unit are collected in real time. The above data is uploaded and fed back to the meta controller in real time, providing real-time operating condition basis for subsequent scheduling progress verification and deviation judgment.

[0066] The meta controller receives real-time task scheduling data uploaded by each process unit, compares and calculates the actual process progress with the planned progress preset by the task scheduling strategy for each process and item, continuously calculates and accumulates the time deviation and resource usage deviation of process execution, and automatically generates a feedback scheduling strategy to correct the scheduling deviation when the overall accumulated deviation exceeds the system's preset deviation threshold, based on the current actual working conditions.

[0067] The meta-controller accurately distributes the generated feedback scheduling strategy to the corresponding level in the task-layered intelligent agent architecture where execution deviations occur. The intelligent agent at that level receives the scheduling correction instruction, adjusts the process sequence, resource allocation, and execution priority, completes dynamic correction, and realizes closed-loop feedback scheduling management.

[0068] Example 2, based on the same inventive concept as the AI-coordinated hierarchical multi-agent industrial internet production scheduling method in the aforementioned examples, such as... Figure 2 As shown, this application provides an AI-coordinated hierarchical multi-agent industrial internet production scheduling system. The system and method embodiments in this application are based on the same inventive concept. The system includes: The task-layered intelligent agent architecture determination module 10 is used to upload the scheduled tasks to the meta controller for pre-evaluation. By measuring the task scheduling entropy, coupling density and time urgency, the hierarchical structure and control flow graph of the layered intelligent agent architecture are adaptively adjusted to determine the task-layered intelligent agent architecture.

[0069] The prior task policy acquisition module 20 is used to construct a scheduling decision-maker by combining joint distribution samples and marginal distribution samples as training data and optimization objectives. It receives low-dimensional embedding encoding of scheduling task execution and scheduling trajectory decision to obtain the prior task policy.

[0070] The task scheduling strategy fusion module 30 is used to import the prior task strategy into the scheduling optimizer, and through parallel execution of targeted optimization under generative perturbation, fuse it with the hierarchical cascaded scheduling configuration based on the task hierarchical intelligent agent architecture to obtain the task scheduling strategy, wherein the task scheduling strategy is identified by communication configuration parameters.

[0071] The production scheduling management module 40 is used to perform automated production scheduling management on the scheduled tasks according to the task scheduling strategy.

[0072] Furthermore, the system is also used to implement the following functions: The scheduling task is uploaded to the meta controller, and the task scheduling is pre-evaluated to determine the evaluation indicators, which include task scheduling entropy, coupling density, and time urgency. Based on the evaluation indicators, the hierarchical agent architecture is subjected to structural adaptation and control flow graph adaptation to determine the task hierarchical agent architecture, wherein structural adaptation includes adding, merging, or splitting layers.

[0073] Furthermore, the system is also used to implement the following functions: The hierarchical intelligent agent architecture for tasks inserts a differentiable temporal abstraction library between layers; the differentiable temporal abstraction library generates a high-frequency general scheduling pattern by online clustering of action sequences between upper and lower layers; and the high-frequency general scheduling pattern is encapsulated and stored in registers.

[0074] Furthermore, the system is also used to implement the following functions: Retrieve joint distribution samples and marginal distribution samples, wherein the joint distribution samples include task embedding samples and scheduling trajectory samples, and the marginal distribution samples are randomly paired samples; define an optimization objective, wherein the optimization objective is to maximize the mutual information between task embeddings and scheduling trajectories, and minimize the mutual information between task embeddings and environmental dynamic characteristics; deploy a first decision node with low-dimensional task embeddings, deploy a second decision node with task scheduling decisions, and perform convergence training based on joint distribution samples and marginal distribution samples with the optimization objective as a constraint to obtain a scheduling decision-maker.

[0075] Furthermore, the system is also used to implement the following functions: The scheduling task is uploaded, and based on the first decision node, the order task features are extracted and concatenated into a long feature vector; the long feature vector is subjected to low-dimensional embedding encoding to determine a low-dimensional task embedding vector; the low-dimensional task embedding vector is transferred to the second decision node, and the prior task strategy is output.

[0076] Furthermore, the system is also used to implement the following functions: A scheduling optimizer is constructed, comprising a generative perturbation branch and a dynamic hierarchical optimization branch, wherein the dynamic hierarchical optimization branch embeds a dynamically updated task hierarchical agent architecture; the prior task strategy is imported into the scheduling optimizer, and parallel optimization and coordination of the two branches are performed to generate a task scheduling strategy.

[0077] Furthermore, the system is also used to implement the following functions: The generative perturbation branch receives the low-dimensional task embedding vector and generates task perturbation data; based on the task perturbation data, it performs targeted optimization on the prior task strategy to obtain a first optimized scheduling strategy; the dynamic hierarchical optimization branch receives the prior task strategy, and based on the task hierarchical intelligent agent architecture, executes hierarchical cascaded scheduling logic configuration association to determine a second optimized scheduling strategy; the first optimized scheduling strategy and the second optimized scheduling strategy are fused to form the task scheduling strategy, wherein the task scheduling strategy includes communication configuration parameters based on the task hierarchical intelligent agent architecture.

[0078] Furthermore, the system is also used to implement the following functions: The task scheduling entropy is positively correlated with the frequency of inter-layer communication interaction; the coupling density is positively correlated with the communication bandwidth and communication frequency. According to the preset coupling density distribution, the high coupling layer and the low coupling layer are divided into a high coupling layer and a low coupling layer. The high coupling layer adopts a high-frequency broadcast communication mode, and the low coupling layer adopts an event-triggered communication mode.

[0079] Furthermore, the system is also used to implement the following functions: Based on the hierarchical intelligent agent architecture, task scheduling and coordination are managed by hierarchical drivers. The task scheduling data of each work unit, based on the task scheduling strategy, is fed back to the meta-controller in real time. This task scheduling data includes the actual start time, end time, and device status. The meta-controller compares the task scheduling data with the strategy progress step-by-step. When the accumulated deviation exceeds a preset threshold, a feedback scheduling strategy is generated. This feedback scheduling strategy is then distributed to the associated architecture layer for feedback scheduling management.

[0080] It should be noted that the order of the embodiments described above is for descriptive purposes only and does not represent the superiority or inferiority of the embodiments. Specific embodiments of this specification have been described above. Furthermore, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0081] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0082] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application intends to include such modifications and variations.

Claims

1. An AI-collaborative hierarchical multi-agent industrial internet production scheduling method, characterized in that, The method includes: The scheduling task is uploaded to the meta controller for pre-evaluation. By measuring the task scheduling entropy, coupling density and time urgency, the hierarchical structure and control flow graph of the hierarchical agent architecture are adaptively adjusted to determine the task hierarchical agent architecture. Using joint distribution samples and marginal distribution samples as training data, a scheduling decision-maker is constructed by combining the optimization objective. It receives low-dimensional embedding encoding of scheduling task execution and scheduling trajectory decision to obtain the prior task policy. The prior task strategy is imported into the scheduler optimizer, and targeted optimization under generative perturbation is performed in parallel. This optimization is then combined with the hierarchical cascaded scheduling configuration based on the task hierarchical intelligent agent architecture to obtain the task scheduling strategy. The task scheduling strategy is identified by communication configuration parameters. According to the task scheduling strategy, automated production scheduling management is performed on the scheduled tasks.

2. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 1, characterized in that, Determine the hierarchical intelligent agent architecture for the task, including: The scheduling task is uploaded to the meta controller to perform a pre-evaluation of the task scheduling and determine the evaluation indicators, which include task scheduling entropy, coupling density and time urgency. Based on the evaluation metrics, the hierarchical agent architecture is subjected to structural adaptation and control flow graph adaptation to determine the task hierarchical agent architecture. The structural adaptation includes adding, merging, or splitting layers.

3. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 2, characterized in that, The hierarchical intelligent agent architecture for the task has a differentiable temporal abstraction library inserted between the layers; The differentiable temporal abstraction library generates a high-frequency general scheduling pattern by online clustering of upper and lower layer action sequences; The high-frequency general scheduling mode is encapsulated and stored in registers.

4. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 1, characterized in that, Using jointly distributed and marginally distributed samples as training data, a scheduling decision-maker is constructed based on the optimization objective, including: Retrieve joint distribution samples and marginal distribution samples, wherein the joint distribution samples include task embedding samples-scheduling trajectory samples, and the marginal distribution samples are randomly paired samples; Define an optimization objective, wherein the optimization objective is to maximize the mutual information between task embedding and scheduling trajectory, and minimize the mutual information between task embedding and dynamic environmental characteristics; The first decision node is deployed by embedding low-dimensional tasks, and the second decision node is deployed by task scheduling decisions. With the optimization objective as a constraint, convergence training based on joint distribution samples and marginal distribution samples is performed to obtain the scheduling decision-maker.

5. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 4, characterized in that, The received scheduling task performs low-dimensional embedding encoding and scheduling trajectory decision-making to obtain the prior task policy, including: The scheduling task is uploaded, and based on the first decision node, the order task features are extracted and concatenated into a long feature vector; The feature long vector is subjected to low-dimensional embedding encoding to determine the low-dimensional task embedding vector; The low-dimensional task is embedded into a vector flow and transferred to the second decision node, whereby the prior task strategy is output.

6. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 5, characterized in that, The task scheduling strategy is obtained, including: Construct a scheduling optimizer, wherein the scheduling optimizer includes a generative perturbation branch and a dynamic hierarchical optimization branch, wherein the dynamic hierarchical optimization branch embeds a dynamically updated task hierarchical agent architecture; The prior task strategy is imported into the scheduler optimizer, and dual-path branch parallel optimization and coordination are performed to generate a task scheduling strategy.

7. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 6, characterized in that, Perform dual-branch parallel optimization and coordination to generate task scheduling strategies, including: The generative perturbation branch receives the low-dimensional task embedding vector and generates task perturbation data; Based on the task disturbance data, the prior task strategy is optimized in a targeted manner to obtain a first optimized scheduling strategy; The dynamic hierarchical optimization branch receives the prior task strategy, and based on the task hierarchical intelligent agent architecture, executes hierarchical cascaded scheduling logic configuration association to determine the second optimized scheduling strategy. The first optimized scheduling strategy and the second optimized scheduling strategy are combined to form the task scheduling strategy, wherein the task scheduling strategy includes communication configuration parameters based on a task-layered intelligent agent architecture.

8. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 7, characterized in that, The task scheduling entropy is positively correlated with the frequency of inter-layer communication interactions; The coupling density is positively correlated with the communication bandwidth and communication frequency. According to the preset coupling density distribution, a high coupling layer and a low coupling layer are divided. The high coupling layer adopts a high-frequency broadcast communication mode, and the low coupling layer adopts an event-triggered communication mode.

9. The AI-coordinated hierarchical multi-agent industrial internet production scheduling method as described in claim 1, characterized in that, Automated production scheduling management is performed on the scheduled tasks, including: Based on the inter-layer relationships of a task-layered intelligent agent architecture, task scheduling and coordination are driven by hierarchical management. Based on the task scheduling strategy, the task scheduling data of each work unit is fed back to the meta controller in real time. The task scheduling data includes the actual start time, end time and equipment status. The meta controller compares the task scheduling data with the strategy progress step by step. When the cumulative deviation exceeds a preset threshold, a feedback scheduling strategy is generated. The feedback scheduling strategy is then distributed to the associated architecture layer for feedback scheduling management.

10. An AI-collaborative hierarchical multi-agent industrial internet production scheduling system, characterized in that: The system is used to implement the AI-coordinated hierarchical multi-agent industrial internet production scheduling method according to any one of claims 1-9, the system comprising: The task-layered intelligent agent architecture determination module is used to upload the scheduled tasks to the meta controller for pre-evaluation. By measuring the task scheduling entropy, coupling density and time urgency, the hierarchical structure and control flow graph of the layered intelligent agent architecture are adaptively adjusted to determine the task-layered intelligent agent architecture. The prior task policy acquisition module is used to construct a scheduling decision-maker by using joint distribution samples and marginal distribution samples as training data and combining them with the optimization objective. It receives low-dimensional embedding encoding of scheduling task execution and scheduling trajectory decision to obtain the prior task policy. The task scheduling strategy fusion module is used to import the prior task strategy into the scheduling optimizer, and through parallel execution of targeted optimization under generative perturbation, fuse it with the hierarchical cascaded scheduling configuration based on the task hierarchical intelligent agent architecture to obtain the task scheduling strategy. The task scheduling strategy is identified by communication configuration parameters. The production scheduling management module is used to perform automated production scheduling management on the scheduled tasks according to the task scheduling strategy.

Images