A collaborative method and system for the entire process of a digital factory

Abstract

This invention relates to the field of digital factory technology, and more particularly to a collaborative method and system for the entire process of a digital factory. The method includes the following steps: acquiring production decision data, and analyzing process parameters based on the production decision data to obtain process parameter data; constructing factory resource agents based on the process parameter data using a preset digital factory model to obtain factory resource agent data; conducting multi-agent negotiation based on the factory resource agent data to obtain digital workshop task allocation data; and performing industrial IoT monitoring based on the digital workshop task allocation data to obtain real-time industrial IoT data, thereby enabling collaborative auxiliary operations throughout the entire digital factory process. This invention, by constructing a task allocation mechanism based on semantic understanding and multi-agent collaboration, achieves intelligent matching and flexible scheduling among various resources in the digital factory, effectively improving the system's task execution efficiency and resource utilization.

Description

Technical Field

[0001] This invention relates to the field of digital factory technology, and in particular to a collaborative method and system for the entire process of a digital factory. Background Technology

[0002] A digital factory is a new production organization method that uses data related to the entire product lifecycle as a foundation and, based on the principles of virtual manufacturing, plans, simulates, optimizes, and reorganizes the entire production process in a virtual environment. As industrial manufacturing rapidly evolves towards intelligence and digitalization, traditional production scheduling and resource allocation methods based on static rules and manual decision-making are no longer sufficient to meet the actual needs of flexible manufacturing, customized production, and multi-variety, small-batch tasks. Existing digital factory management systems often employ preset task priorities or linear scheduling algorithms. When faced with dynamic environments such as changes in resource status, task conflicts, and multi-objective optimization, they lack sufficient real-time response and collaborative decision-making capabilities, easily leading to low resource utilization, frequent task delays, and a decline in overall system scheduling efficiency. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a collaborative method and system for the entire process of a digital factory, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a collaborative method for the entire process of a digital factory, including the following steps:

[0005] Step S1: Obtain production decision data and analyze the process parameters based on the production decision data to obtain process parameter data;

[0006] Step S2: Construct a factory resource intelligence agent based on the process parameter data using a preset digital factory model to obtain factory resource intelligence agent data;

[0007] Step S3: Perform multi-agent negotiation based on factory resource agent data to obtain digital workshop task allocation data;

[0008] Step S4: Perform industrial IoT monitoring based on the task allocation data of the digital workshop to obtain real-time industrial IoT data for collaborative auxiliary operations throughout the digital factory process.

[0009] This invention analyzes production requirements based on MES systems, ERP systems, or user-defined inputs to form a structured production task model, encompassing process routes, material requirements, equipment capabilities, and time constraints. Key factory resources (such as CNC machine tools, industrial robots, AGV transport units, and personnel) are abstracted into heterogeneous intelligent agents with perception, reasoning, and communication capabilities. Each agent embeds a parameter modeling module, a real-time status monitoring module, and a scheduling interface module, enabling dynamic resource modeling and task response capabilities. These agents can engage in multiple rounds of negotiation regarding task priorities, resource conflicts, and path avoidance, achieving self-organizing and adaptive task collaborative scheduling. The system utilizes industrial sensors, edge computing units, and PLC controllers deployed on-site to collect real-time task execution status, equipment operating parameters, and environmental anomaly data. Combined with data-driven models or rule engines, it provides anomaly warnings and task adjustments, achieving a closed-loop control process from "task generation → distribution → execution feedback → dynamic adjustment." This realizes a fully closed-loop collaborative process from task generation to task execution in a digital factory. Compared to traditional methods relying on centralized scheduling or rule-based allocation, this approach, which constructs resource agents based on process parameters, enhances the dynamic matching capability between tasks and resources. Utilizing a multi-agent negotiation mechanism significantly improves the scheduling system's adaptability in the face of task conflicts, resource constraints, or multi-objective optimization scenarios. Furthermore, combining industrial IoT monitoring to achieve virtual-real mapping and dynamic feedback enhances the system's execution reliability and real-time response capabilities. Overall, this method exhibits greater flexibility, intelligence, and system stability in dynamic production environments.

[0010] Preferably, step S1 specifically includes:

[0011] Obtain production decision data;

[0012] Process flow data is obtained by extracting process flow data from production decision data;

[0013] Process parameters are analyzed based on the process flow data to obtain process parameter data, which includes the required equipment type, process sequence, material consumption, estimated working hours, energy requirements, and environmental constraints.

[0014] This invention transforms macro-level production plans into executable process parameters by structuring production decision data. Standardized process paths are obtained through process flow extraction, ensuring the continuity and controllability of task execution. The analysis process incorporates the extraction of multi-dimensional information such as equipment type, material usage, process sequence, and time estimation, which not only improves the accuracy of process modeling but also provides a data foundation for resource matching and task collaboration. Compared to traditional manual analysis or template matching methods, this invention offers greater versatility, automation, and engineering adaptability, significantly improving the accuracy of process parameter generation and the early response efficiency of the scheduling system.

[0015] Preferably, step S2 specifically includes:

[0016] Based on the process parameter data, the factory resource type intelligent agent is instantiated through the preset digital factory model to obtain the factory resource type intelligent agent data, which includes production line intelligent agent data, equipment intelligent agent data, personnel intelligent agent data and material intelligent agent data.

[0017] The agent data of factory resource type is configured with attributes to obtain agent attribute data, where attribute configuration includes state information configuration, intention function configuration and policy space configuration.

[0018] Based on the process parameter data, a task collaboration relationship graph is constructed from the agent attribute data to obtain the factory resource agent data.

[0019] This invention combines process parameter data with a pre-set digital factory model to transform factory resources from static information to intelligent agent modeling. At the resource type level, multiple intelligent agents, such as production lines, equipment, personnel, and materials, are instantiated, achieving structured modeling of resource management. At the attribute configuration level, state information, intention functions, and policy spaces are introduced, enabling each type of intelligent agent to possess autonomous perception and response capabilities, breaking the static rule-based model of traditional resource scheduling. Simultaneously, a graph of task collaboration relationships between intelligent agents is constructed based on process parameters, forming a resource network with collaborative logic, providing a data foundation for task game theory and scheduling optimization. Compared to traditional BOM-driven or linear scheduling models, this invention significantly enhances the system's ability to express resource heterogeneity, dynamism, and strategic aspects, exhibiting excellent intelligent scheduling adaptability and scalability.

[0020] Preferably, step S3 specifically includes:

[0021] Initial negotiation data is obtained by performing negotiation initialization based on data from the factory resource intelligence agent.

[0022] Based on the preliminary negotiation data, task game is conducted to obtain negotiation game data;

[0023] Initial local optima are extracted from the negotiation game data to obtain the negotiation local optima data.

[0024] Global resource conflict detection is performed based on negotiated local optimal data to obtain global resource conflict data;

[0025] Based on the global resource conflict data, the negotiation game data is labeled and limited to obtain the negotiation label agreement data;

[0026] Based on the negotiated and labeled data, limited local optimum extraction is performed to obtain digital workshop task allocation data.

[0027] This invention achieves flexible task allocation optimization for complex resource environments by introducing a multi-stage, multi-strategy intelligent negotiation and task game mechanism. Negotiation initialization ensures that various resource agents possess collaborative intentions and behavioral strategies; candidate allocation schemes are dynamically generated through a multi-agent task game mechanism, overcoming the limitations of traditional static scheduling rules; furthermore, intelligent decoupling is achieved under conditions of task complexity and diverse resource conflicts by combining local optimum extraction with global conflict detection; and by labeling and optimizing the negotiation game data, the system achieves optimal resource utilization while ensuring executability. Compared to single-time scheduling or heuristic allocation methods, this approach exhibits significant adaptability, distributed intelligence, and conflict tolerance, more effectively supporting the task scheduling needs of digital factories under multi-objective and dynamic loads.

[0028] Preferably, the negotiation initialization specifically includes:

[0029] Based on the factory resource agent data, an agent semantic graph mapping is performed to obtain cross-agent shared semantic data.

[0030] By performing intention function semantic embedding on cross-agent shared semantic data, we obtain intelligent task context data;

[0031] Based on the factory resource intelligent agent data, a digital twin image is loaded to obtain the intelligent agent digital image data;

[0032] By linking and integrating intelligent task context data and intelligent agent digital mirror data, preliminary negotiation data is obtained.

[0033] This invention constructs an intelligent negotiation foundation with semantic understanding and virtual-real synchronization capabilities by integrating semantic graph mapping, intention function embedding, and digital twin mirroring technologies. It utilizes agent semantic graph mapping to achieve terminological consistency and cognitive alignment among multiple types of resource agents, resolving collaboration barriers between heterogeneous data. By embedding intention functions into shared semantic data, it constructs an intelligent task context with context awareness, making the resource agents' responses to tasks more personalized and strategic. Loading digital twin mirrors of resource agents provides a visualized and predictable virtual environment for strategy deduction and behavioral simulation. Through the fusion of semantic context and digital mirrors, preliminary negotiation data with real-state mapping and semantic cognitive consistency is generated. Compared to traditional negotiation methods based on static attributes or rule-driven approaches, this step achieves cognitive synchronization, behavioral pre-playing, and collaborative perception among resource agents, significantly improving the intelligence, interpretability, and predictability of the early negotiation phase.

[0034] Preferably, the agent semantic graph mapping is performed through a pre-defined shared semantic understanding graph model, and the construction steps of the shared semantic understanding graph model include:

[0035] Acquire key terminology data corresponding to the factory resource intelligent agent data;

[0036] Term alignment is performed based on key term data and a pre-defined domain ontology library to obtain term alignment data, which includes concept layer alignment, example layer alignment and attribute layer alignment.

[0037] Positive and negative sample pairs are extracted from the terminology-aligned data to obtain positive and negative sample pair data.

[0038] Cross-agent semantic training is performed on positive and negative sample pairs of data to obtain a shared semantic understanding graph model. The cross-agent semantic training is carried out jointly by a Transformer structure encoding network and a bidirectional gated recurrent unit encoder.

[0039] This invention constructs a shared semantic understanding graph model, achieving semantic alignment and cognitive consistency among various types of factory resource agents, providing a semantic foundation for intelligent collaboration. By extracting key terms from the factory resource agent data and combining them with a pre-defined domain ontology library, multi-level term alignment from the concept layer, instance layer, to the attribute layer is achieved, ensuring the structural comprehensiveness and industry adaptability of the semantic mapping. Based on the term alignment results, positive and negative sample pairs are constructed. During training, a joint modeling strategy combining a Transformer structure encoding network and a bidirectional gated recurrent unit encoder is introduced, balancing long-distance semantic modeling capabilities with temporal contextual expression capabilities, effectively improving the accuracy and generalization ability of semantic representation. The generated shared semantic understanding graph model can achieve unified interpretation and dynamic mapping of task semantics among different resource agents. Compared with traditional methods relying on static keyword matching or rule mapping, this step possesses technical advantages such as adaptability, strong contextual understanding, and trainability / evolution, significantly improving the semantic collaboration efficiency and intelligent task scheduling level in multi-agent systems.

[0040] Preferably, the semantic embedding of the intention function is specifically as follows:

[0041] Attention weight data of the factory resource agents is obtained by performing attention calculation on the semantic data shared across agents;

[0042] The factory resource agent data is weighted and aggregated based on the agent attention weight data to obtain semantically perceptible feature data.

[0043] Semantic prior distribution data is obtained by calculating semantic prior distribution based on semantic data shared across agents;

[0044] Obtain the current state data of the agent corresponding to the shared semantic data across agents, and fit the current state data of the agent to a conditional Gaussian model to obtain the agent observation model;

[0045] Based on semantic prior distribution data and agent observation model, Bayesian posterior inference is performed to obtain task preference probability data;

[0046] Expected value weights are calculated based on semantic perception feature data and task preference probability data to obtain intelligent task context data.

[0047] This invention performs attention calculations on resource agents based on shared semantic data across agents, obtaining their attention weights in the current semantic scenario to achieve semantic-driven feature focusing. Subsequently, weighted aggregation is used to extract semantic-aware features, enhancing the agents' understanding of the task context. A prior distribution of task preferences is generated by combining semantic vectors, and an observation model is fitted using the current resource state as input. A learnable agent state-response mapping is constructed using a conditional Gaussian model. A Bayesian posterior inference method is used to fuse the semantic prior and the observation model, obtaining more stable and interpretable task preference probabilities. Intelligent task context data is generated through expectation weight fusion, providing refined and dynamic preference cognition basis for game strategy selection. Compared with traditional static intention scoring methods, this step has technical advantages such as strong semantic adaptability, superior predictive ability, and high model interpretability, significantly enhancing the multi-agent cognition and collaborative ability in complex task environments within a digital factory scheduling system.

[0048] Preferably, the task game specifically refers to:

[0049] Based on the preliminary negotiation data, candidate task resource mappings are extracted to obtain candidate task resource mapping data;

[0050] Based on the factory resource intelligence agent data, the candidate mapping data of task resources is mapped to intentions to obtain the game response matrix data;

[0051] Contextual semantic resource response is performed based on the game response matrix data to obtain contextual semantic resource response data;

[0052] Based on the contextual semantic resource response data, a task assistance request is made to the factory resource agent data to obtain the agent role response bidding data.

[0053] Based on the bidding data of the agent role response, the collaborative capability assessment is performed to obtain the collaborative capability assessment data, which includes temporal consistency assessment and semantic preference conflict assessment.

[0054] Temporary task groups are generated based on collaborative capability assessment data to obtain negotiation and game data.

[0055] This invention extracts candidate task-resource mapping relationships based on preliminary negotiation data to construct an allocable task range; it integrates the state and semantic preferences of factory resource agents to generate a game response matrix among multiple agents, achieving a quantitative expression of task competition intent; it enhances the adaptability of task allocation to the current environmental context through contextual semantic resource responses; it realizes collaborative bidding for tasks through assistance requests and role-based response bidding for task sub-roles, and verifies the temporal consistency and semantic preference conflicts among responding agents during the collaborative capability evaluation phase; and it outputs the negotiation game results through a temporary task group generation mechanism to construct an executable collaborative relationship between tasks and resources. Unlike traditional one-to-one allocation or static auction mechanisms, this step, based on semantic drive, integrates intelligent technologies such as dynamic alliances and game optimization, possessing stronger scheduling flexibility, resource coordination, and intelligent task adaptation.

[0056] Preferably, the initial local optimum extraction specifically involves:

[0057] Based on the negotiation game data, candidate data for task resource game are extracted to obtain candidate data for task resource game.

[0058] Local optimal target labeling is performed on the candidate data of task resource game to obtain locally optimal target labeled data;

[0059] Based on the locally optimal target annotation data, the task execution of the factory resource agent data is optimally allocated to obtain the preliminary allocation object data;

[0060] Based on the initial allocation object data, task mutual exclusion filtering is performed to obtain negotiated locally optimal data.

[0061] This invention constructs a local task allocation optimization mechanism to achieve high-quality task selection and allocation optimization based on the results of multi-agent task game theory. It extracts candidate pairs of task-resource games based on negotiation game data, clarifying the range of available resources and strategy information, providing a foundation for local optimum extraction. By introducing a local optimum target labeling mechanism, key indicators in the task and resource allocation process (such as willingness, response latency, and energy consumption priority) are transformed into quantifiable labeled data, making local optimization goal-oriented. Subsequently, combining the labeled data with the resource agent state, the optimal matching strategy for task allocation is executed, ensuring that the preliminary results are most suitable for the current resource configuration within a local range. Through a task mutual exclusion filtering mechanism, allocation schemes that cannot be executed simultaneously or have conflicting strategies are excluded, effectively improving the executability and resource utilization efficiency of global scheduling. Compared with the single-round optimal selection or fixed priority strategy in traditional scheduling, this invention has technical advantages such as strong local responsiveness, high strategy adaptability, and strong task conflict avoidance ability, laying a key foundation for the system to achieve high-efficiency, low-interference intelligent scheduling.

[0062] Preferably, the digital factory end-to-end collaborative system for executing the digital factory end-to-end collaborative method as described above includes:

[0063] The process parameter parsing module is used to acquire production decision data and parse process parameters based on the production decision data to obtain process parameter data.

[0064] The resource intelligence agent construction module is used to construct the factory resource intelligence agent based on the process parameter data and a preset digital factory model, and obtain the factory resource intelligence agent data.

[0065] The multi-agent negotiation and task allocation module is used to conduct multi-agent negotiation based on factory resource agent data to obtain digital workshop task allocation data.

[0066] The Industrial IoT Monitoring and Feedback Module is used to monitor the Industrial IoT based on the task allocation data of the digital workshop, obtain real-time Industrial IoT data, and carry out collaborative auxiliary operations throughout the entire process of the digital factory.

[0067] The beneficial effects of this invention are as follows: Step S1 performs structured process parameter analysis on production decision data, achieving rapid transformation from high-level decisions to executable processes and improving the early response efficiency of the scheduling system; Step S2 constructs a factory resource intelligence agent with state awareness, willingness assessment, and strategy response capabilities based on the digital factory model, and establishes collaborative relationships between resources through attribute configuration and collaboration graph construction, enhancing the model's resource expression capabilities and scheduling flexibility; Step S3 realizes adaptive negotiation and alliance-style task allocation among multiple intelligence agents, effectively solving the problem of intelligent matching under task conflicts and resource constraints; Step S4 combines industrial IoT monitoring methods to synchronize digital scheduling results to the physical execution layer, achieving virtual-physical closed-loop feedback and real-time adjustment. Compared with existing systems based on static rules or linear scheduling, this invention possesses stronger intelligence, semantic consistency, resource collaboration capabilities, and system dynamic adaptability, significantly improving the collaborative efficiency and flexible production capabilities of digital factories in complex environments. Attached Figure Description

[0068] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0069] Figure 1 A flowchart illustrating the steps of a digital factory end-to-end collaborative method according to one embodiment is shown.

[0070] Figure 2 A flowchart illustrating the steps of a process parameter analysis method according to one embodiment is shown.

[0071] Figure 3A flowchart illustrating the steps of a resource agent construction method according to an embodiment is shown.

[0072] Figure 4 A flowchart illustrating the steps of a multi-agent negotiation and task allocation method according to an embodiment is shown. Detailed Implementation

[0073] The technical method of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0074] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0075] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0076] The task is obtained through the smart factory management interface and analyzed. The task includes producing 10 high-precision customized parts and delivering them within 48 hours. The process includes cutting, welding, and inspection (each part requires 60 minutes). Resources are queried, including cutting machine A (idle), welding arm B (currently performing a regular task and will be idle in 2 hours), welding arm C (available for overtime), inspection station C (prioritizing urgent orders), personnel P1 (not on duty during the night shift), and P2 (available 24 / 7). Parameter analysis shows that the time required for each product is 60 minutes, for a total of 600 minutes, which can be broken down into 10 task units, each containing 3 steps.

[0077] Resource analysis reveals that cutting machine A is idle with a response score of 0.92, welding arm B has an estimated idle time of 2 hours, welding arm C is currently busy but has a policy set to "allow night shifts" with a response score of 0.85, and inspection station C has a score of 0.96 and prioritizes urgent orders. P1 is unavailable, while P2 can participate in the welding agent to improve its score.

[0078] The system performs a game-theoretic scoring and sorts the tasks, resulting in cutting machine A > A2 (choose A), welding arm C (because B is still busy), and inspection station C (performing inspection work). The system selects (A, C, C) as the minimum delay execution combination. IoT monitoring shows the cutting execution time slightly exceeds expectations (21 minutes). The system dynamically adjusts the welding start time by +5 minutes. The inspection station experiences a delay of one unit; the system records this and adjusts the task buffer time strategy accordingly.

[0079] Please see Figures 1 to 4 This application provides a collaborative method for the entire process of a digital factory, including the following steps:

[0080] Step S1: Obtain production decision data and analyze the process parameters based on the production decision data to obtain process parameter data;

[0081] Specifically, structured data such as planned production tasks, material information, product specifications, production time windows, and delivery cycles are collected from the enterprise's production management system (such as MES, ERP, or other task scheduling platforms). Data access can be achieved via API interfaces or the OPCUA protocol. Key process parameters, such as processing temperature, pressure, speed, and proportions, are automatically extracted based on a comparison between product specifications and a process standard library. For multi-variety mixed-line production, process templates are matched using preset product classification rules (such as material code identification based on BOM structure). A rule engine (such as Drools) is used to parse process constraint logic, such as "the combination of material A and equipment B can only be used for process type C," ensuring that the generated results conform to technical specifications. Process parameters are organized in JSON format, and the parameter source and parsing path are recorded for traceability.

[0082] Step S2: Construct a factory resource intelligence agent based on the process parameter data using a preset digital factory model to obtain factory resource intelligence agent data;

[0083] Specifically, the model is built based on BIM and simulation models, containing digital twins of virtual equipment, workstations, manpower, and transportation equipment. Model identification uses unique IDs and resource tags (e.g., "laser cutting machine #A01"). Each type of resource is abstracted as an "Agent," containing basic attributes (e.g., type, capability, status, current load) and behavioral functions (e.g., order acceptance, execution, feedback). The system uses parsed process parameter data as input features and performs task allocation judgments on resource agents through built-in matching rules. These rules may include adaptation condition judgments, such as matching equipment types with high-temperature tolerance if "processing temperature > 100℃"; equipment priority strategies, prioritizing equipment with high idle time and good historical success rates among multiple available resources; and production cycle time and capacity constraints, such as requiring a certain type of processing task to simultaneously meet the minimum output capacity per unit time. Each Agent has a state machine that manages idle / preparing / processing / fault states and updates them dynamically. The output data structure is a list of Agent objects, including Agent-ID, a list of compatible processes, resource capabilities (numerical or enumerated), and current status.

[0084] Step S3: Perform multi-agent negotiation based on factory resource agent data to obtain digital workshop task allocation data;

[0085] Specifically, a market bidding simulation mechanism based on multi-agent systems is used to achieve optimal matching between tasks and resources. Each resource agent bids based on its idle time and task fitness. A task-resource fit scoring function is introduced, for example: Fit Score = α × Resource Load + β × Device Compatibility + γ × Historical Success Rate. α, β, and γ can be dynamically adjusted according to the factory strategy. Taking the device candidate set for task T_001 as an example, the current status of device E_23 is as follows: current resource load is 0.60, device compatibility score is 0.90, historical task success rate is 0.85, and the system strategy weights are set to α = 0.4, β = 0.3, and γ = 0.3. When multiple agents respond to bids, the system scheduler will prioritize the allocation of the task to the resource with the highest fit score based on the scoring results, and record this scheduling decision process in the scheduling log to support traceability analysis and subsequent optimization. All resource agents receive task requests and broadcast response information through a unified lightweight communication channel (such as MQTT protocol or ROS node). The main scheduling agent aggregates all response information and comprehensively considers task priority, resource load balancing strategies (such as minimum load priority, round-robin strategy, historical performance backtesting, etc.) to generate a task allocation table. During the negotiation process, the system supports a timeout retry mechanism (automatically resending the task broadcast if no suitable response is received within a set time) and a failure rollback mechanism (if multiple rounds of negotiation fail, a backup resource is introduced or manual approval mode is entered) to ensure the robustness and continuity of the task allocation process. The resulting task allocation data structure includes: a unique task identifier (Task ID), an allocated resource number (Resource ID), a planned start time, an expected completion time, and necessary process parameters (such as processing temperature, cutting speed, robot trajectory, etc.). This structure will be used for subsequent task execution control and integration with the industrial monitoring system to achieve closed-loop collaboration of the production process.

[0086] Step S4: Perform industrial IoT monitoring based on the task allocation data of the digital workshop to obtain real-time industrial IoT data for collaborative auxiliary operations throughout the digital factory process.

[0087] Specifically, edge computing gateways are deployed in the digital workshop to connect various field devices (including programmable logic controllers (PLCs), CNC machine tools, temperature controllers, and industrial robots) to the industrial Internet of Things (IIoT) system, enabling standardized access to equipment status and task execution information. The edge gateways support multiple industrial communication protocols (such as MQTT, OPC UA, and Modbus), and the collected monitoring data includes, but is not limited to, the following indicators: task progress, equipment status, processing temperature, vibration signals, energy consumption, and alarm information. Each indicator is associated with an assigned task, constructing a task-status mapping table. Real-time data is fed back to the scheduling center; if delays, anomalies, or equipment downtime occur, a rescheduling process is triggered. The collaborative operation system dynamically adjusts resource scheduling and task priorities based on real-time data, forming a closed-loop control. All monitoring data is displayed through a visual panel, and a RESTful API is provided for upper-layer business systems to call.

[0088] Preferably, step S1 specifically includes:

[0089] Step S11: Obtain production decision data;

[0090] Specifically, production decision data can be extracted from the enterprise's ERP system, APS system, or order management platform, mainly including order ID, order type, delivery date; product model, specifications; customer customization requirements; inventory status and procurement plan; and safety / quality related constraints. Synchronization is performed periodically via a pre-set API or database intermediate table; if an MES system is used, data can be obtained in real time via a WebService interface; for situations with multiple systems coexisting, unified extraction is performed through a data platform to avoid redundancy and conflicts.

[0091] Step S12: Extract the process flow from the production decision data to obtain process flow data;

[0092] Specifically, a product category-based process template library is constructed. Each template contains a standardized processing flow (e.g., blank preparation → rough machining → heat treatment → finish machining → inspection → packaging). Templates are modeled in flowchart form, using BPMN or Petri net structures to express processing nodes and their dependencies. The closest process flow template is matched based on product specification fields (e.g., model, size, material). For customized products, a classifier trained on historical production data (e.g., random forest) is used to predict the optimal process path. A conditional branching mechanism (e.g., whether heat treatment is required) is introduced to dynamically trim process nodes. Process flow data is represented in a directed graph structure, with each node containing processing step ID, required resource type, and process name.

[0093] Step S13: Analyze the process parameters based on the process flow data to obtain process parameter data, which includes the required equipment type, process sequence, material consumption, estimated working hours, energy requirements, and environmental constraints.

[0094] Specifically, each process node queries the equipment capability library for the type of equipment that supports the process based on its processing method field (such as laser cutting or CNC milling); if the same process is applicable to multiple equipment, the priority rule is "capacity priority > equipment availability > energy consumption optimization".

[0095] Based on the node topology sorting in the flowchart, the dependency relationship (predecessor-successor) is directly mapped to the execution order; special cases (such as multiple steps in parallel) are marked as parallelizable processes and are handled by the task allocation module.

[0096] Extract the unit material consumption from the BOM structure; calculate the total material requirements based on the order quantity and process loss rate; if semi-finished product conversion is involved, add an intermediate material balance calculation module.

[0097] Establish a time model based on empirical formulas or historical data, such as: time = basic processing time + material correction coefficient + difficulty correction coefficient, and use models such as linear regression or XGBoost to compare and dynamically calibrate multiple historical cases.

[0098] Energy consumption is estimated based on equipment type, processing time, and rated power; special processes (such as heat treatment) require the inclusion of auxiliary energy types (gas, steam, etc.) in the calculation; the output is a structured record in kWh / unit or total energy consumption / batch.

[0099] In terms of environmental constraint identification, the system can determine whether each process involves environmental control conditions, such as temperature and humidity requirements, cleanliness level (e.g., electronic packaging requires a 100,000-level clean environment), etc. At the same time, it can detect whether there are process nodes with high pollution, high noise or harmful emissions. If so, it will mark them as tasks that require environmental protection linkage and record their constraint type, target value and constraint duration.

[0100] Preferably, step S2 specifically includes:

[0101] Step S21: Based on the process parameter data, instantiate the factory resource type intelligent agent through the preset digital factory model to obtain the factory resource type intelligent agent data, which includes production line intelligent agent data, equipment intelligent agent data, personnel intelligent agent data and material intelligent agent data;

[0102] Specifically, a unified resource modeling framework is established, which includes four types of resource metaclasses (production line, equipment, personnel, and materials). Each type contains basic attribute templates (such as resource ID, resource capabilities, status interfaces, communication interfaces, etc.). All templates are defined based on JSON Schema or XML, which facilitates subsequent instantiation and parsing.

[0103] Iterate through the processing parameters and resource type requirements (e.g., "Required equipment type: laser cutting machine", "Required production line is automatic packaging line") in the process parameter data; map the matched resource types to the model template one by one, generate intelligent agent instances as needed, and assign them a unique identifier ID (e.g., AGENT_DEVICE_LASER_001); if there are multiple candidate entities for a certain resource (e.g., multiple devices of the same model), instantiate them in batches, and record the corresponding physical resource number and its initial state for each intelligent agent.

[0104] Step S22: Configure the attributes of the intelligent agent data of the factory resource type to obtain the intelligent agent attribute data, wherein the attribute configuration includes state information configuration, intention function configuration and policy space configuration;

[0105] Specifically, each agent maintains a set of dynamic state attributes to describe its real-time availability and current task status. These attributes mainly include: status, which represents the current running state and can take values ​​such as idle, running, and maintenance; load, which represents the current task load as a percentage, such as 0% to 100%; and energy_state, which represents the remaining working capacity, specifically the remaining battery power, remaining executable time, or working hours budget. The initial state is read from the MES or edge gateway, and the update cycle is dynamically configured from 5 to 30 seconds.

[0106] Each agent possesses a "willingness function" to determine its willingness to accept orders under different task requests. This function is composed of the following factors: current load (the lower the load, the higher the willingness); task matching degree (the degree of fit with its capability model); historical execution success rate; equipment depreciation priority (high-value equipment is given priority to undertake key tasks). An example willingness function is as follows (value range 0-1): w = w1·(1-current load) + w2·task matching degree + w3·success rate-w4·equipment depreciation index, where w1 is the load weight coefficient with a value of 0.30, w2 is the matching degree weight coefficient with a value of 0.30, w3 is the success rate weight coefficient with a value of 0.20, and w4 is the depreciation penalty coefficient with a value of 0.10.

[0107] Each agent defines an executable set of policies, such as order acceptance, adjustment, and rejection policies, and possesses autonomous decision-making capabilities. For example, the device agent's policy space includes `accept_task_if(load<70% and match_score>0.8)`, indicating that the task is accepted when the load is low and the match rate is high; `delay_task_if(temperature>threshold)`, indicating that the task is delayed when the device temperature exceeds the safety threshold; and `reject_task_if(predictive_failure=true)`, indicating that the task is rejected if the device state is predicted to be a potential failure. The policy set is expressed using a DSL or rule engine and supports hot updates.

[0108] Step S23: Construct a task collaboration relationship graph based on the process parameter data and the agent attribute data to obtain the factory resource agent data.

[0109] Specifically, the process flow diagram (steps, sequence, dependencies) in the process parameters is used as input; each processing step is treated as a task node, with node attributes including required resource type, process ID, estimated time, etc.; connections between nodes represent dependencies (e.g., OP20 needs to wait for OP10 to complete), forming a directed task graph. A set of candidate agents is assigned to each task node (matched according to resource type); a many-to-many mapping table between tasks and resources is established; the nodes record the set of collaborating resources and the logical order between resources (e.g., constraints requiring tasks to be executed consecutively on the same production line). This is done using adjacency lists and resource mapping tables, or using a graph database (e.g., Neo4j); each graph node contains the bound resource type, the set of candidate agents, and a status flag (assigned / unassigned / conflicted).

[0110] Preferably, step S3 specifically includes:

[0111] Step S31: Perform negotiation initialization based on the factory resource agent data to obtain preliminary negotiation data;

[0112] Specifically, all agent entities with task candidate qualifications are extracted from the output factory resource agent data. Task candidate qualification is determined based on the following conditions: the resource type must match the task requirements (e.g., matching processing type, resource capacity covering required process parameters), the current state must be idle or schedulable, and it must not be bound to other tasks. The initial negotiation state includes generating a candidate resource agent set for each task; initializing the intention function (as mentioned above) and state attributes of each agent; constructing an initial task-resource mapping table for all tasks and their corresponding candidate resource sets; registering unique communication identifiers for all participating agents and establishing communication channels (e.g., ROS nodes, MQTT topics) for information broadcasting and feedback collection between agents. Task adaptability analysis is performed on all agent objects in the factory resource agent data. Based on their resource type, capability matching degree, and state parameters, a task candidate set is initialized, a task-resource matching mapping relationship is constructed, and the agent negotiation state is set to unlocked, resulting in preliminary negotiation data.

[0113] Step S32: Conduct task game based on the preliminary negotiation data to obtain negotiation game data;

[0114] Specifically, a task-oriented multi-agent game mechanism (such as discrete-time round-robin game) is adopted. In each round, each resource agent actively participates in the competition based on its willingness function, submitting its task acceptance proposal. The game rounds are defined, and in each round, agents bid on acceptable tasks (calculated based on load, willingness score, etc.). After receiving a proposal, the task node evaluates the candidate bids according to the scoring function. After each round, the agent's bidding status and task allocation intention table are updated. The task response score = α × matching degree + β × availability - γ × current load + δ × priority. α is the matching degree weight, with a value of 0.4, emphasizing matching degree, representing the task needing to "find the right person / equipment." β is the availability weight, with a value of 0.2, indicating that availability is important but not the only decision point. γ is the load suppression coefficient, with a value of 0.3, moderately suppressing high-load resources to avoid overload. δ is the priority enhancement weight, with a value of 0.1, which can be increased to 0.2–0.3 if the system is sensitive to task urgency. After each round of the game, the system updates the agents' bidding status, task response status, and some resource conflicts, maintaining a task allocation intention table and a task game status table. Then, it proceeds to the next round until a preset number of rounds is reached or task allocation converges. Based on the initial negotiation data, the system constructs a task game model, integrates the candidate resource responses received by each task node, combines task requirements with current resource capabilities, generates a game mapping relationship between tasks and resources, and outputs the negotiated game data.

[0115] Step S33: Extract the initial local optimum based on the negotiation game data to obtain the negotiation local optimum data;

[0116] Specifically, after completing the initial round of task negotiation and game among multiple agents, the system obtains the willingness score data between all tasks and their candidate resource agents. The game score is a scoring index generated by weighting the aforementioned multiple factors, reflecting the suitability of each resource for a specific task. The system will perform the following processing flow for each task: sorting candidate agents and ranking the candidate resource set corresponding to the task by score; extracting the preferred agent and selecting the resource agent with the highest score as the current preferred execution candidate for the task; recording the initial mapping relationship, binding the task ID with its preferred resource agent identifier and recording it as the preferred resource field; and constructing a local optimal table, summarizing the preferred matching relationships of all tasks and constructing an initial local optimal task allocation table.

[0117] Step S34: Perform global resource conflict detection based on the negotiated local optimal data to obtain global resource conflict data;

[0118] Specifically, the system detects whether multiple tasks are assigned to the same agent (resource conflict) or whether resources are scheduled for tasks with overlapping timeframes. It checks for duplicate resource IDs; if the same resource is occupied by multiple tasks simultaneously, it records this as a static resource conflict event, and the conflict data is categorized and summarized by conflicting resource ID and conflicting task group. The system further compares the time intervals of all resource tasks; if it finds that the same resource has overlapping time intervals across multiple tasks (i.e., the task execution time ranges intersect), it records this as a dynamic time conflict event. It performs resource reuse detection on negotiated local optimal data, identifies conflicts such as duplicate scheduling of resource IDs or overlapping task execution times, and generates global resource conflict data.

[0119] Step S35: Label and limit the negotiation game data based on the global resource conflict data to obtain the negotiation label agreement data;

[0120] Specifically, conflicting resources are marked as conflicting in the negotiation game data, and acceptable scheduling conditions or restrictions are set, such as marking conflicting tasks as requiring further negotiation; the policy space of resources is updated, for example, setting a minimum interval between tasks; a penalty term is added to the willingness function to reduce the priority of conflicting resources in the game, such as the willingness value equals the original willingness value minus the conflict penalty factor, which is calculated based on the conflict level and historical conflict frequency, and the regression coefficient is fitted based on historical data. The global resource conflict data is fed back into the negotiation game model, and the relevant task agents are labeled and restricted according to the conflict type, including execution order restrictions, resource reallocation priority adjustments, policy space shrinkage, etc., generating negotiation label agreement data.

[0121] Step S36: Perform restricted local optimum extraction based on the negotiated annotation agreement data to obtain digital workshop task allocation data.

[0122] Specifically, under conflict constraints, the local optimal solution is recalculated to form the least conflict-free task allocation table: unavailable resources are excluded or new scheduling windows are set; the optimal allocation is reconstructed based on the updated intention function and policy rules, that is, the weight calculation of the resource state, task priority, and policy matching degree is minimized, with the corresponding weight of resource state being 0.5, the corresponding weight of task priority being 0.3, and the corresponding weight of policy matching degree being 0.2; the system uses heuristic matching based on task priority ranking, maximum weight matching algorithm, or local graph search algorithm to combine and screen candidate task-resource pairs, ensuring that the results retain the original task intention as much as possible while satisfying scheduling constraints, taking into account executability and resource balance, and outputting conflict-free and time-reasonable task → resource binding relationships. Based on the negotiated and labeled agreement data, a task reallocation optimization operation is performed under constrained conditions, and the local optimal task undertaking scheme is re-extracted based on resource usage constraints and time constraints to obtain digital workshop task allocation data.

[0123] Preferably, the negotiation initialization specifically includes:

[0124] Based on the factory resource agent data, an agent semantic graph mapping is performed to obtain cross-agent shared semantic data.

[0125] Specifically, the system defines a factory resource ontology, including classes such as equipment, personnel, production lines, and materials, as well as class attributes such as capabilities, status, region, and communication interface. For each resource agent (e.g., equipment, personnel), the system performs a triplet transformation, mapping its basic attributes (e.g., resource ID, capability set, current status, region) to nodes in a graph, and constructs the graph structure through these attribute relationships. For example, RDF / OWL semantic structures are used to describe the type relationships and upstream / downstream dependencies between agents; each agent's data (e.g., equipment ID, capability set, current status) is mapped to nodes in the graph; and the following edge connections are established: hasCapability (possesses capability), locatedIn (location), connectedTo (resource interface communication association), and isCandidateFor (candidate for a task). Cross-agent shared semantic data is generated: stored and queried using a graph database (e.g., Neo4j, GraphDB); and SPARQL semantic queries are supported to identify combinations of agents with collaborative conditions. A semantic graph is constructed based on factory resource agent data, defining semantic class relationships including equipment, personnel, production lines and materials. Semantic mapping is performed through attributes such as resource capabilities, spatial distribution, and communication interfaces to obtain cross-agent shared semantic data to support multi-agent understanding and collaboration.

[0126] By performing intention function semantic embedding on cross-agent shared semantic data, we obtain intelligent task context data;

[0127] Specifically, knowledge representation learning methods (such as TransE and R-GCN) are used to vectorize entities and relations in the graph. The embedding goal is to construct a low-dimensional vector space to represent the collaborative intention state of each agent. Each agent's intention function (e.g., weighted calculation based on load, matching degree, energy consumption, and task priority) is mapped to function term embeddings. The system embeds this intention function as an "intention expression vector" into the semantic vector space, enabling each agent to form an intention intensity point and resource adaptation tendency vector in the semantic space, expressing its potential scheduling tendency in multi-task collaborative scenarios. To achieve contextual matching between resources and tasks, the system also embeds task requirements (including processing accuracy requirements, energy consumption limitations, execution order dependencies, and special process constraints) into the same semantic vector space through feature extraction and label encoding, forming a task semantic embedding representation. The contextual matching degree between the agent and the task is obtained by calculating cosine similarity or Euclidean distance. Based on cross-agent shared semantic data, a semantic embedding mechanism is used to vectorize the intention functions of each agent, construct a low-dimensional feature space representing the ability to understand task context, and form intelligent task context data that can be used by agents to judge task priority and willingness to accept.

[0128] Based on the factory resource intelligent agent data, a digital twin image is loaded to obtain the intelligent agent digital image data;

[0129] Specifically, a corresponding digital twin mirror model is loaded for each physical resource agent, including real-time operating status (state machine / sensor values); historical usage records (failure rate, task success rate); and predictive information (estimated remaining lifetime, energy consumption model). Data can be pulled from edge devices or historical databases using OPCUA and MQTT protocols. Each mirror object adopts a unified structure: {entity ID, current state, historical trajectory, predictable parameters}; the twin state can be optionally displayed using a visualization interface (such as Unity or Cesium).

[0130] By linking and integrating intelligent task context data and intelligent agent digital mirror data, preliminary negotiation data is obtained.

[0131] Specifically, the semantic embedding (behavioral tendencies) of the same agent is merged with its digital twin mirror image (real-time capabilities); a hybrid structure is established, including the current state; collaborative context scoring; capability-task matching degree; and scheduling suggestion weights. A weighted fusion formula is used to generate an initial negotiation score: This is the willingness weighting coefficient, with a value of 0.4. This is the state weight coefficient, with a value of 0.3. The reliability weight coefficient, set to 0.3, is used to generate an initial collaborative candidate matrix for each task agent, labeling the candidate levels. The intelligent task context data and the agent's digital mirror data are integrated at the feature and policy levels, fusing their task understanding capabilities, current resource status, and behavior prediction capabilities to construct preliminary negotiation data suitable for the game negotiation process, supporting local optimum calculations and conflict detection. Based on the above scoring results, the system constructs a task-resource scoring matrix, where each element represents the collaborative fit between the task and the resource agent. According to this matrix, the resource agents that can be matched for each task are ranked, and candidate levels are divided according to their scores. For example, Level A: scores greater than 0.85, high-priority candidates; Level B: scores between 0.7 and 0.85, optional candidates; Level C: scores below 0.7, secondary candidates or alternative resources.

[0132] Preferably, the agent semantic graph mapping is performed through a pre-defined shared semantic understanding graph model, and the construction steps of the shared semantic understanding graph model include:

[0133] Acquire key terminology data corresponding to the factory resource intelligent agent data;

[0134] Specifically, key information fields are extracted from the structured and semi-structured data of the factory resource agent, such as type fields (e.g., injection molding machine, process section); capability fields (e.g., maximum load, supported material type); and status fields (e.g., running, pre-maintenance). TF-IDF sorting is performed on the structured data field names and values; NER (Named Entity Recognition) is used to extract terminology phrases from descriptive fields (e.g., equipment description, task annotation); low-frequency and generic words are filtered out, retaining specialized terms. Chinese and English terms are uniformly encoded (e.g., converted to Pinyin-underscore style: chu_kou_leng_que_xi_tong); the original context, location, and entity ID mapping of terms are preserved. Terminology information involved in the factory resource agent data is obtained, including resource type, capability parameters, status labels, and attribute descriptions. Keyword extraction and named entity recognition methods are used to extract key terminology data, providing basic input for subsequent semantic alignment.

[0135] Term alignment is performed based on key term data and a pre-defined domain ontology library to obtain term alignment data, which includes concept layer alignment, example layer alignment and attribute layer alignment.

[0136] Specifically, an ontology knowledge base for the intelligent manufacturing field is constructed, comprising: a concept layer (e.g., equipment, process, personnel); an attribute layer (e.g., power, precision, number of workstations, displacement range); and an instance layer (e.g., specific equipment models, personnel roles, material batches). Concept layer alignment uses word embedding similarity (Word2Vec, FastText) and edit distance to determine if concepts belong to the same category; for example, CNC machining equipment and CNC machine tools will be aligned to the equipment category concept node. Instance layer alignment clusters based on task context and equipment task history behavior similarity. Feature encoding is performed based on task context semantic vectors (e.g., task description, instruction path) and equipment historical behavior data. Unsupervised clustering algorithms (e.g., KMeans or DBSCAN) are used to analyze the behavioral similarity of multiple devices / personnel to determine if they belong to the same instance category. Attribute layer alignment compares the semantic similarity between terms and ontology attribute names (e.g., aligning "maximum force" with "rated thrust"). Semantic similarity calculations are performed between attribute items in terms and standard attributes in the ontology (using models such as BERT and SBERT), and it is determined whether an attribute alignment relationship exists. For example, maximum force and rated thrust will be identified as semantically equivalent attributes; if the attribute units differ (e.g., N and kN), they will be uniformly converted to standard units and normalized (e.g., maximum value normalization or standard deviation normalization). Key terminology data will be aligned with a pre-defined intelligent manufacturing domain ontology library. This alignment includes concept-level alignment, example-level alignment, and attribute-level alignment, using semantic similarity calculation and attribute normalization to generate structured terminology alignment data.

[0137] Positive and negative sample pairs are extracted from the terminology-aligned data to obtain positive and negative sample pair data.

[0138] Specifically, the positive sample generation rules are as follows: if the matching degree between a term and an ontology item is greater than the threshold (e.g., 0.85), it is marked as a positive sample; term groups that are successfully aligned with concepts are considered as positive pairs; attribute names and units that are consistent or completely matched after conversion are considered as attribute-aligned positive pairs.

[0139] Negative sample generation rules: extract from terms with different semantic categories or mismatched contexts; control the positive-to-negative ratio (e.g., 1:3) to improve training generalization ability; exclude words with similar spellings but contradictory meanings (e.g., internal cooling system vs. cooling shell);

[0140] A semantic alignment sample set is constructed based on term alignment data. The sample includes matching term pairs as positive sample pairs and term pairs that are conceptually inconsistent or context-detached as negative sample pairs, forming positive and negative sample pair data that can be used for semantic relation learning.

[0141] Cross-agent semantic training is performed on positive and negative sample pairs of data to obtain a shared semantic understanding graph model. The cross-agent semantic training is carried out jointly by a Transformer structure encoding network and a bidirectional gated recurrent unit encoder.

[0142] Specifically, a Transformer-structured encoder is used to extract semantic features of term context; a parallel Bi-GRU (bidirectional gated recurrent unit) structure is used to capture word sequence order features; the model training adopts a triplet loss function: Loss = max(0, margin + d(anchor, positive) - d(anchor, negative)), where Loss is the triplet loss value, max is the maximization function, margin is the set minimum semantic discrimination threshold, d(x, y) is the distance between vectors x and y in the embedding space (e.g., Euclidean distance), anchor is the encoding vector of the current target term, positive is a synonymous term with a similar meaning to anchor, and negative is a term with a meaning unrelated to anchor or with semantic differences. A minimum discrimination threshold margin is set (e.g., 0.5) to achieve the effect of semantic contrastive learning; the term text is converted into a token sequence, input into the Transformer to extract semantic feature vectors; simultaneously, it is input into the Bi-GRU encoder to extract contextual relevance; the joint output is concatenated and fed into a fully connected layer for training and classification. After optimization, synonyms are made as close as possible in the embedding space, while the distance between synonyms is increased, resulting in strong semantic discrimination ability. After training, the model can be used for automatic classification of new terms and semantic label recommendation. Cross-agent semantic training is performed on positive and negative sample pairs of data. A joint semantic representation model composed of a Transformer-structured encoding network and a bidirectional gated recurrent unit (Bi-GRU) is adopted. The model is optimized and trained using a triple loss function to generate a shared semantic understanding graph model. This model is used to establish a unified semantic understanding and contextual expression capability among multiple agents.

[0143] Preferably, the semantic embedding of the intention function is specifically as follows:

[0144] Attention weight data of the factory resource agents is obtained by performing attention calculation on the semantic data shared across agents;

[0145] Specifically, input A: factory resource agent data (attributes of each agent, task adaptation labels, etc.); input B: shared semantic understanding graph embedding (term-level semantic vectors, upstream and downstream dependency information). For each task objective, the attention score of all candidate agents in the semantic graph context is calculated using the following scoring function: Where A t t eniFor, Softmax is, Q i Let be the attribute vector of the i-th agent, K be the set of key vectors from the semantic graph embedding, T be the transpose symbol, and d be the attribute vector of the i-th agent. k The vector dimension scaling factor is used; the output is the agent attention weight matrix for each task. The attention score represents the priority of each agent in the current task and semantic context. The system can realize an agent matching and evaluation mechanism based on semantic enhancement, which can improve the dynamic adaptation and collaborative efficiency between tasks and resources.

[0146] The factory resource agent data is weighted and aggregated based on the agent attention weight data to obtain semantically perceptible feature data.

[0147] Specifically, attention weights are used as weighting coefficients to linearly combine the original attribute vectors of each agent, resulting in new feature representations that incorporate semantic awareness. Feature dimensions include capability dimensions (productivity, accuracy, etc.); state dimensions (availability, health); and semantic dimensions (task relevance score). Based on the agent attention weight data, the factory resource agent attribute vectors are weighted and aggregated, and task context semantic attention information is integrated to generate semantically aware feature data with contextual understanding capabilities.

[0148] Semantic prior distribution data is obtained by calculating semantic prior distribution based on semantic data shared across agents;

[0149] Specifically, joint frequency modeling is performed on the adaptation of task type and historical task agent; multiple distribution priors of task type → agent category are constructed; the system collects the matching relationship between task type and participating resource agents in historical task allocation records; and joint modeling is performed on the collaboration frequency between task semantic tags and resource semantic tags to obtain the empirical adaptation degree of task to resource category. Based on the above frequency statistics, a conditional probability distribution model from task type to resource category is constructed, forming a multinomial distribution structure: P(RT│TT)=f(TT,HC), where P(RT│TT) is the conditional probability distribution, i.e., semantic prior distribution data, RT is the resource type, TT is the task type, and f(TT,HC) is the collaboration frequency mapping function, which maps the task type TT and historical collaboration data HC into a function structure of conditional probability (e.g., normalized frequency, Laplace smoothing, etc.), and HC is the historical collaboration frequency data. To enhance the stability and robustness of the model and prevent the zero probability problem caused by sample sparsity, the system introduces Dirichlet prior estimation into the above multinomial distribution. By assigning a non-zero prior smoothing factor to each resource category, the model has a certain generalization ability for unseen task-resource combinations. The system adopts a hierarchical Bayesian modeling structure, with task type as the upper-level topic distribution and agent capability category as the lower-level resource response distribution, forming the following two-level modeling framework: upper-level modeling: semantic topic distribution of task categories; lower-level modeling: resource capability response distribution under various tasks. Combined with Dirichlet distribution, a conditional resource propensity model is formed, thereby outputting semantic prior distribution data, which is the set of resource category probability mappings corresponding to each task type.

[0150] Obtain the current state data of the agent corresponding to the shared semantic data across agents, and fit the current state data of the agent to a conditional Gaussian model to obtain the agent observation model;

[0151] Specifically, the state characteristic data of the factory resource agent includes, but is not limited to, the following indicators: load rate, remaining energy, reliability score, and fault prediction value; and is conditionally combined with semantic vectors (e.g., task type and state variables form a joint distribution). The system uses a Conditional Gaussian Mixture Model (CGMM) for modeling, with the following structure: conditional variable (X), such as the task semantic embedding vector; response variable (Y), such as a vector composed of resource state variables; and model objective, such as setting the learning P(Y|X), which is the probability distribution of resource state Y under the condition of task semantic X. The modeling process uses the classic EM algorithm for parameter estimation, including the following steps: E-step (expectation step), which estimates the posterior probability of a sample belonging to each Gaussian component under the current parameters; M-step (maximization step): based on the E-step results, updates the mean vector, covariance matrix, and mixing coefficients of each Gaussian distribution; and iterates until the convergence condition is met to obtain the agent observation model.

[0152] Based on semantic prior distribution data and agent observation model, Bayesian posterior inference is performed to obtain task preference probability data;

[0153] Specifically, based on the semantic prior distribution data and the agent observation model, a Bayesian posterior inference mechanism is used to estimate the task preference probability data of each agent for different task types, which is used to express the contextual adaptation tendency of resources to tasks. The semantic prior distribution data represents the task preference trend of various tasks in the semantic space, denoted as P(T). k ),in This represents the set of task types; the agent observation model is a state observation model based on conditional Gaussian distribution modeling, used to estimate the state under given task conditions T. k The current state O of the lower agent j Let the likelihood probability be P(O) j |T k ), where O j This represents the agent's state vector; based on this, the system applies Bayes' theorem to calculate the agent's state for task type T. k The posterior probability is expressed as follows: P(T k |O j ) represents the agent's response to task type T k The preference probability (posterior), P(O) j |T k Let P(T) be the likelihood of the agent's state given a task type. k ) represents the semantic prior probability of the task type, l represents the index of the candidate task type, and O represents the semantic prior probability of the task type. j T represents the current state of the agent. k The task type is defined; this inference process jointly models the agent's current state with the prior features of the task, outputs the preference probability distribution of each agent relative to various tasks, and generates task preference probability data.

[0154] Expected value weights are calculated based on semantic perception feature data and task preference probability data to obtain intelligent task context data.

[0155] Specifically, semantically aware feature vectors (extracted by the pre-encoding module, representing the semantic feature vectors of each resource agent in the semantic embedding space) are used as feature weights; task preference probabilities (generated by the semantic prior modeling module, expressing the probability distribution of the current task type's preference for each resource type) are used as importance factors; and a weighted combination of expected feature vectors is performed for each agent. The semantically aware feature data and task preference probability data are jointly weighted and calculated to construct the expected task response function, i.e., intelligent task context = semantically aware feature data × task preference probability data, thus obtaining intelligent task context data that expresses the agent's response capability in the current task context.

[0156] Preferably, the task game specifically refers to:

[0157] Based on the preliminary negotiation data, candidate task resource mappings are extracted to obtain candidate task resource mapping data;

[0158] Specifically, the task context vector (representing the semantic requirements of the task in the current negotiation scenario) and agent semantic features (representing the capability expression and semantic features of each resource agent) are derived from the initial negotiation data. For each task, matching is performed based on the cosine similarity between the task context vector and the semantic vectors of each agent; a threshold (e.g., similarity > 0.7) is set as the admission condition for candidate resources. The task ID → candidate resource list is recorded to form a task resource mapping table.

[0159] Based on the factory resource intelligence agent data, the candidate mapping data of task resources is mapped to intentions to obtain the game response matrix data;

[0160] Specifically, each agent calculates its own task acceptance willingness value based on its state (load, health, etc.) and preference parameters. A willingness scoring table is constructed for all task-candidate resource combinations, forming a task × agent matrix; for each candidate task-agent pair (T... i A j The system constructs a willingness score function based on the following three types of indicators: M(i,j)=S cap (i,j)+S sem (i,j)―D load (j), where M(i,j) is the willingness score between the candidate task and the resource agent, i is the task number, j is the agent number, and S cap (i,j) represents the matching score between the process required for the i-th task and the resource capability for the j-th task, S sem (i,j) represents the similarity score (e.g., cosine similarity) between the semantic label of the i-th task and the semantic embedding of the j-th resource, D load(j) represents the willingness discount term generated by the current load rate of the j-th agent, which can be obtained by directly setting the load reduction coefficient or by setting a weight for the load coefficient; the system integrates the score tables of all tasks into a two-dimensional response matrix. Based on the factory resource agent data, the willingness function is matched on the candidate mapping data of task resources to construct the game response matrix data between task and agent, which is used to describe the response strength of the agent to each task.

[0161] Contextual semantic resource response is performed based on the game response matrix data to obtain contextual semantic resource response data;

[0162] Specifically, task context information (such as execution time, material location, and dependent tasks) is incorporated into agent response analysis; a multi-dimensional semantic context feature representation is constructed. If the task context conflicts with the physical environment of the resources (such as remote scheduling versus distant resources), the response value is adjusted. If the physical environment of the agent is inconsistent with the task context (such as long resource distance, response time delay, or incomplete preceding tasks), the system introduces two scenario correction factors: an environmental consistency factor, used to measure the accessibility of resources in terms of spatial geographic location and job unit matching; and a temporal accessibility factor, used to determine whether the current resource can respond to and complete the task within the required time. A scenario weight correction term is added, for example, response value = willingness value × environmental consistency factor × temporal accessibility factor. Based on the game response matrix, the agent's response behavior is semantically corrected and weighted by combining task context scenario information, forming scenario semantic resource response data that integrates environmental constraints and execution feasibility.

[0163] Based on the contextual semantic resource response data, a task assistance request is made to the factory resource agent data to obtain the agent role response bidding data.

[0164] Specifically, the system assigns roles to each task based on process characteristics and workflow requirements, defining multiple roles for each task: main executor (responsible for the main processing tasks), assistant (performing auxiliary tasks such as clamping and supporting), transporter (responsible for workpiece handling and transfer), and quality inspector (responsible for quality inspection of task results). Each role type is mapped to a specific resource category, for example: main executor → processing equipment or core operators; transporter → logistics robots or AGVs; quality inspector → vision inspection equipment or quality engineers. Each task broadcasts a task execution request (based on semantic response ranking) to its candidate resources; each agent returns whether to bid and its bidding intention value according to its own role strategy. For contextual semantic resource response data, the semantic vector of the task role is matched with the semantic response score of the candidate resources to select the current set of responsive candidate resources. For each task role, the system sends a task assistance request (broadcast request) to its candidate resource agents. Upon receiving a request, the resource agent will make a bidding decision based on the following information: its current state (idle, busy, under maintenance); the role-based order-acceptance rules in the current policy space (e.g., role priority, scheduling window); and its willingness function score (representing the resource's responsiveness in the current semantic task). The response result consists of two parts: a bidding response flag indicating whether the resource participates in the role's task; and a bidding willingness value representing the score output by its willingness function. The system records the mapping relationship between all task roles and responding resources as structured agent role response bidding data, in the form of Task ID – Role Type – Candidate Resource ID – Bidding Willingness Value.

[0165] Based on the bidding data of the agent role response, the collaborative capability assessment is performed to obtain the collaborative capability assessment data, which includes temporal consistency assessment and semantic preference conflict assessment.

[0166] Specifically, the system determines whether task overlap exists based on the planned execution window of the tasks responded to by the bidding agents; it also performs time interval overlap detection based on the task schedule table and resource availability window. If there is overlap, it is considered a time-conflicting task pair, and the system will reconcile them based on task priority or game score, or temporarily remove some bids. For each resource agent, its willingness to bid on different tasks may be high simultaneously. If the resource cannot physically execute multiple high-preference tasks concurrently, it is necessary to determine whether its bids have semantic preference conflicts. The system introduces a mutual exclusion coefficient for evaluation, defined as: EC(A k )=∑ i,j,i<j (w ki ·w kj ·δ ij ), of which EC(A k ) represents the mutual exclusion coefficient of the agents, i is the task number, j is the task number different from i, and w kiFor the agent to perform task T i The bid intention value, w kj For the agent to perform task T j The bid intention value, δ ij As a task conflict indicator factor, if task T i With T j If there is time overlap or exclusive resource conflict, then δ ij =1, otherwise 0. The system evaluates the collaborative capabilities of agents responding to bidding data. This evaluation includes assessing the temporal consistency of task execution and the semantic conflict between task preferences, used to determine the feasibility and priority of each resource in participating in the current task game. The system will then base its decisions on EC(A). k Task allocation is reconciled based on the threshold of EC(A), for example, if EC(A) k If EC(A) > θ, where θ is the harmonic threshold configured between [0.2, 0.6], depending on the system's tolerance for resource concurrency, it can be set to 0.2. In this case, the system can prioritize, reduce scores, or partially eliminate conflicting bidding tasks. k If )≈0, it means that the agent's current task combination has no obvious conflict and can directly participate in collaborative optimization.

[0167] Temporary task groups are generated based on collaborative capability assessment data to obtain negotiation and game data.

[0168] Specifically, to construct a resource collaboration scheme that meets the task execution requirements, the system identifies resource role combinations that have passed feasibility verification based on the obtained collaboration capability assessment data, and generates temporary task execution groups accordingly. These temporary task groups provide a set of resources with execution and coordination capabilities during the task negotiation and game phase. Each combination must include at least one primary execution resource; each task involves several supporting roles (such as handling, assembly, and inspection), and the system requires that the actual number of auxiliary roles allocated in the task group is not less than a set coverage threshold (e.g., 80%) of the total required number of roles. If multiple combinations exist, they are ranked by evaluation scores (weighted by indicators of resource willingness, temporal consistency, and load balancing, or directly obtained based on resource willingness), and the Top-N candidate schemes are selected as the negotiation output, including the task ID, a list of resource combinations (labeling the role type of each resource), collaboration capability assessment scores, and role coverage and executable status markers. Temporary task groups that meet the task requirements are constructed based on the collaboration capability assessment data. These task groups include primary execution resources and multiple auxiliary role resources, forming negotiation and game data for negotiation decisions.

[0169] Preferably, the initial local optimum extraction specifically involves:

[0170] Based on the negotiation game data, candidate data for task resource game are extracted to obtain candidate data for task resource game.

[0171] Specifically, the data comes from negotiation and game theory, with each task corresponding to multiple resource combinations (temporary task groups). For each task, the top N high-scoring candidate groups (e.g., N=5) are retained; each candidate group includes resource composition, role configuration, and evaluation score. Based on the negotiation and game theory data, resource response candidate schemes for each task are extracted and ranked, retaining high-scoring task resource combinations that meet the threshold requirements, resulting in task resource game theory candidate data, which is used for optimal target selection.

[0172] Local optimal target labeling is performed on the candidate data of task resource game to obtain locally optimal target labeled data;

[0173] Specifically, during the candidate phase of the game, the system has established multiple resource combination schemes (teams) and matching candidate relationships with each task. To select the optimal combination, the system pre-defines a multi-factor weighted objective function, scores each resource combination scheme, and identifies the optimal candidate accordingly, calculating... Where f(team) is the resource combination score, representing the overall score of the resource combination scheme under the current task; a higher value indicates a better performance. The collaborative scoring weight determines the system's emphasis on the collaborative capabilities between resources (such as role complementarity and parallelism). A value of 0.4 indicates that the team's collaborative capabilities directly affect the overall scheduling stability and execution efficiency. In multi-resource collaborative tasks, the quality of collaborative relationships takes precedence over individual performance. tea m is the synergy score, representing the degree of synergy within the combination in terms of semantics, function, and scheduling relationships. This represents the load balancing weight, indicating the degree of importance placed on load distribution balance. A value of 0.2 indicates that load balancing helps avoid resource bottlenecks, but balance can be appropriately sacrificed for certain high-priority tasks, hence the lower weight. Var load Load variance represents the variance of the current load of each agent in the portfolio; the smaller the value, the more balanced the portfolio. This represents the weighting of energy utilization efficiency, indicating the system's emphasis on combined average energy efficiency. It has a value of 0.2, and energy conservation considerations are a secondary indicator. This weighting should only be increased in green manufacturing scenarios. eff Energy efficiency is the inverse ratio of the energy required to perform the current task; the higher the value, the more energy-efficient. Historical success rate weighting represents the weight of each resource in the portfolio based on its historical successful experience in executing this type of task. The value is 0.2. Success rate reflects reliability, while importance is on par with energy efficiency, making it suitable for experience-driven scheduling scenarios. R historyHistorical success rate refers to the average success rate of each resource in the combination when performing similar tasks historically. Each task is labeled as a local optimum candidate; if the scores of multiple groups are close to a preset tolerance threshold (±ε), they are marked as tied optima. The candidate data of task resource game are scored and sorted according to a preset objective function, and the resource combination scheme with the best score or tied optima is extracted and labeled as a local optimum objective, thus obtaining the locally optimum objective labeled data.

[0174] Based on the locally optimal target annotation data, the task execution of the factory resource agent data is optimally allocated to obtain the preliminary allocation object data;

[0175] Specifically, agents in locally optimal combinations are marked as task assignment candidates; the scheduling status of these agents is updated to pending assignment, and the task ID is recorded. The system supports the temporary participation of the same resource in multiple tasks, but conflict filtering is required; a pre-assigned task list is maintained for each resource. The system supports temporary resource reuse, allowing agents with the same resource to participate in multiple task combinations in the initial stage, but these must be filtered out or optimized in subsequent conflict detection and resource exclusivity assessment. Therefore, each resource agent will maintain the following allocation-related fields: Agent ID: unique identifier of the resource; PreAssignmentList: a set of multiple task IDs currently pre-assigned to the resource; CurrentDispatchStatus: set to "Pending"; based on the locally optimal target annotation data, agents in the optimal task resource combination are marked as pending scheduling objects, and their pre-assigned tasks are recorded, forming preliminary allocation object data to prepare for task conflict handling.

[0176] Based on the initial allocation object data, task mutual exclusion filtering is performed to obtain negotiated locally optimal data.

[0177] Specifically, the following two types of conflicts are examined: resource-time conflicts, which include agents participating in multiple tasks simultaneously; and functional-role conflicts, which include the inability to perform the same role repeatedly (e.g., one person cannot simultaneously control two processes). Conflicting task groups are prioritized, retaining high-priority tasks; conflicting tasks are returned to the candidate group, triggering a rollback game indicator. Based on the initial allocation object data, task mutual exclusion detection is performed to identify scheduling conflicts caused by resource overlap or role conflicts. Conflict filtering is performed using task priority and resource exclusivity rules, retaining conflict-free task resource combinations and generating negotiated locally optimal data.

[0178] Specifically, the system performs scheduling conflict detection based on the task resource binding information recorded in the preliminary allocation object data. Conflict types mainly include the following two categories: resource time conflict, which refers to the same resource agent being occupied by multiple tasks simultaneously within overlapping time periods (e.g., the same device being called by multiple tasks in the same time period); and functional role conflict, which refers to the same agent simultaneously undertaking multiple mutually exclusive role tasks (e.g., the same operator cannot simultaneously act as the master controller to perform two different procedures). To resolve these conflicts, the system calculates the urgency level of each conflicting task based on its deadline or remaining schedulable time, and initially sorts them from highest to lowest urgency. Tasks with high urgency are prioritized. Deadlines ≤ 2 hours are considered high urgency, deadlines ≤ 4 hours are medium urgency, and deadlines > 4 hours are low urgency. Based on the first round of sorting, the system checks the scheduling priority level of tasks (e.g., level 1 to level 5) and moves higher-priority tasks to the front of the same urgency level to ensure critical tasks are executed first. Among high-urgency tasks, level 1 tasks are ranked first, level 3 tasks are ranked in the middle, and level 5 tasks are ranked last. After the first two levels of sorting, the system further fine-tunes the time window flexibility of tasks within the same priority level. Tasks with tight time windows (low time flexibility) are prioritized before tasks with more flexible windows to reduce the risk of scheduling failure. Delay time ≤ 30 minutes → tight time window → ranked higher; delay time ≥ 2 hours → more flexible time window → ranked lower. Priority is given to retaining task resource combinations with higher scheduling importance and less conflict impact. For conflicting tasks with relatively low scores or non-critical tasks, the system removes them from the current scheduling set and returns them to the original task resource candidate set. Returned task combinations will be set to a game-rollback state, allowing the system to re-enter the candidate pool for task reallocation in subsequent scheduling rounds. This process is recorded internally by a game-rollback flag and linked to the game state management module. The system outputs a set of conflict-free task resource bindings, representing the negotiated locally optimal data, which serves as the basis for the next step of resource conflict detection or final allocation scheme confirmation.

[0179] Preferably, the digital factory end-to-end collaborative system for executing the digital factory end-to-end collaborative method as described above includes:

[0180] The process parameter parsing module is used to acquire production decision data and parse process parameters based on the production decision data to obtain process parameter data.

[0181] The resource intelligence agent construction module is used to construct the factory resource intelligence agent based on the process parameter data and a preset digital factory model, and obtain the factory resource intelligence agent data.

[0182] The multi-agent negotiation and task allocation module is used to conduct multi-agent negotiation based on factory resource agent data to obtain digital workshop task allocation data.

[0183] The Industrial IoT Monitoring and Feedback Module is used to monitor the Industrial IoT based on the task allocation data of the digital workshop, obtain real-time Industrial IoT data, and carry out collaborative auxiliary operations throughout the entire process of the digital factory.

[0184] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0185] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A collaborative method for the entire process of a digital factory, characterized in that, Includes the following steps: Step S1: Obtain production decision data and analyze the process parameters based on the production decision data to obtain process parameter data; Step S2: Construct a factory resource intelligence agent based on the process parameter data using a preset digital factory model to obtain factory resource intelligence agent data; Step S3: Perform negotiation initialization based on the factory resource agent data to obtain preliminary negotiation data; perform task game based on the preliminary negotiation data to obtain negotiation game data; extract initial local optima based on the negotiation game data to obtain negotiation local optima data; perform global resource conflict detection based on the negotiation local optima data to obtain global resource conflict data; and label and limit the negotiation game data based on the global resource conflict data to obtain negotiation label agreement data. Based on the negotiated annotation agreement data, limited local optimum extraction is performed to obtain digital workshop task allocation data; Step S4: Perform industrial IoT monitoring based on digital workshop task allocation data to obtain real-time industrial IoT data for collaborative auxiliary operations throughout the digital factory process; The negotiation initialization specifically involves: Based on the factory resource agent data, an agent semantic graph mapping is performed to obtain cross-agent shared semantic data; the intention function semantic embedding is performed on the cross-agent shared semantic data to obtain intelligent task context data; Based on the factory resource agent data, a digital twin image is loaded to obtain agent digital image data; the intelligent task context data and agent digital image data are correlated and integrated to obtain preliminary negotiation data; The semantic embedding of the intention function is specifically as follows: Attention is calculated on the factory resource agent data based on the cross-agent shared semantic data to obtain agent attention weight data; the factory resource agent data is then weighted and aggregated based on the agent attention weight data to obtain semantic perception feature data. Semantic prior distribution data is obtained by calculating the semantic prior distribution based on the cross-agent shared semantic data; the current state data of the agent corresponding to the cross-agent shared semantic data is obtained, and a conditional Gaussian model is fitted to the current state data to obtain the agent observation model; Bayesian posterior inference is performed based on the semantic prior distribution data and the agent observation model to obtain the task preference probability data; and expected value weights are calculated based on the semantic perception feature data and the task preference probability data to obtain the intelligent task context data.

2. The method according to claim 1, characterized in that, Step S1 is as follows: Obtain production decision data; Process flow data is obtained by extracting process flow data from production decision data; Process parameters are analyzed based on the process flow data to obtain process parameter data, which includes the required equipment type, process sequence, material consumption, estimated working hours, energy requirements, and environmental constraints.

3. The method according to claim 1, characterized in that, Step S2 is as follows: Based on the process parameter data, the factory resource type intelligent agent is instantiated through the preset digital factory model to obtain the factory resource type intelligent agent data, which includes production line intelligent agent data, equipment intelligent agent data, personnel intelligent agent data and material intelligent agent data. The agent data of factory resource type is configured with attributes to obtain agent attribute data, where attribute configuration includes state information configuration, intention function configuration and policy space configuration. Based on the process parameter data, a task collaboration relationship graph is constructed from the agent attribute data to obtain the factory resource agent data.

4. The method according to claim 1, characterized in that, The semantic graph mapping of the intelligent agent is carried out through a pre-defined shared semantic understanding graph model. The construction steps of the shared semantic understanding graph model include: Acquire key terminology data corresponding to the factory resource intelligent agent data; Term alignment is performed based on key term data and a pre-defined domain ontology library to obtain term alignment data, which includes concept layer alignment, example layer alignment and attribute layer alignment. Positive and negative sample pairs are extracted from the terminology-aligned data to obtain positive and negative sample pair data. Cross-agent semantic training is performed on positive and negative sample pairs of data to obtain a shared semantic understanding graph model. The cross-agent semantic training is carried out jointly by a Transformer structure encoding network and a bidirectional gated recurrent unit encoder.

5. The method according to claim 1, characterized in that, The specific task-based game is as follows: Based on the preliminary negotiation data, candidate task resource mappings are extracted to obtain candidate task resource mapping data; Based on the factory resource intelligence agent data, the candidate mapping data of task resources is mapped to intentions to obtain the game response matrix data; Contextual semantic resource response is performed based on the game response matrix data to obtain contextual semantic resource response data; Based on the contextual semantic resource response data, a task assistance request is made to the factory resource agent data to obtain the agent role response bidding data. Based on the bidding data of the agent role response, the collaborative capability assessment is performed to obtain the collaborative capability assessment data, which includes temporal consistency assessment and semantic preference conflict assessment. Temporary task groups are generated based on collaborative capability assessment data to obtain negotiation and game data.

6. The method according to claim 1, characterized in that, The initial local optimum extraction is specifically as follows: Based on the negotiation game data, candidate data for task resource game are extracted to obtain candidate data for task resource game. Local optimal target labeling is performed on the candidate data of task resource game to obtain locally optimal target labeled data; Based on the locally optimal target annotation data, the task execution of the factory resource agent data is optimally allocated to obtain the preliminary allocation object data; Based on the initial allocation object data, task mutual exclusion filtering is performed to obtain negotiated locally optimal data.

7. A digital factory end-to-end collaborative system, characterized in that, For executing the digital factory end-to-end collaborative method as described in claim 1, the digital factory end-to-end collaborative system includes: The process parameter parsing module is used to acquire production decision data and parse process parameters based on the production decision data to obtain process parameter data. The resource intelligence agent construction module is used to construct factory resource intelligence agents based on process parameter data and a preset digital factory model to obtain factory resource intelligence agent data. The multi-agent negotiation and task allocation module is used to conduct multi-agent negotiation based on factory resource agent data to obtain digital workshop task allocation data. The Industrial IoT Monitoring and Feedback Module is used to monitor the Industrial IoT based on the task allocation data of the digital workshop, obtain real-time Industrial IoT data, and carry out collaborative auxiliary operations throughout the entire process of the digital factory.

Images