Intelligent Decision-Making and Business Collaboration Methods Based on OAG Ontology and LLM Large Model

Abstract

This invention relates to the field of intelligent discrete manufacturing, and particularly to an intelligent decision-making and business collaboration method based on an OAG ontology and an LLM large model. The method includes the following steps: collecting historical data and business documents; fine-tuning the basic large language model to form an LLM large model; parsing the original data and business documents of the discrete manufacturing scenario to construct an OAG ontology library; receiving real-time field data through the LLM large model, updating the OAG ontology library, and feeding back into the LLM large model to form a two-way feedback; integrating real-time and historical data into the LLM large model, transforming it into business knowledge through the OAG ontology library in a hierarchical manner, and generating primary and backup decision schemes; building an intelligent agent federation hub to issue decision schemes, collecting execution data to feed back into the LLM large model, and achieving cross-scenario collaboration and a closed-loop data flow throughout the entire process. This invention effectively solves the core pain point of existing LLM models—that they can calculate but not understand business—through a closed-loop data flow approach, realizing the full-link value transformation of discrete manufacturing data.

Description

Technical Field

[0001] This invention relates to the field of intelligent discrete manufacturing, and in particular to an intelligent decision-making and business collaboration method based on OAG ontology and LLM large model. Background Technology

[0002] As the discrete manufacturing industry upgrades towards intelligent manufacturing, Large Language Models (LLMs), with their powerful data computation and generation capabilities, are increasingly being applied to core scenarios such as production control, quality control, and logistics control. Meanwhile, traditional technical approaches often employ a "mechanical correspondence between processes and software," attempting to achieve business management through fixed procedures. However, existing technologies have significant shortcomings: LLM models lack semantic understanding and logical adaptation capabilities for discrete manufacturing business scenarios, leading to a disconnect between output results and actual production needs—in other words, "they can calculate but don't understand the business." Furthermore, discrete manufacturing data generally suffers from semantic inconsistencies, poor business collaboration, and fragmented data governance. Traditional technologies cannot achieve the end-to-end transformation of data into business value, making it difficult to support intelligent decision-making and cross-scenario collaboration across the entire production process, severely hindering the intelligent upgrading process of the discrete manufacturing industry. Summary of the Invention

[0003] To address the technical deficiencies in the background technology, this invention proposes an intelligent decision-making and business collaboration method based on OAG ontology and LLM large model, which solves the aforementioned technical problems and meets practical needs. The specific technical solution is as follows:

[0004] The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model includes the following steps:

[0005] Collect historical data and business documents from the discrete manufacturing field, and fine-tune the basic large language model to form an LLM large model adapted to discrete manufacturing business.

[0006] LLM large model is used to parse raw data and business documents of discrete manufacturing scenarios, and an OAG ontology library containing element ontology, business ontology and logical rule ontology is constructed.

[0007] The LLM large model receives real-time field data, mines implicit rules to update the OAG ontology library, and at the same time the OAG ontology library feeds back to optimize the LLM large model, forming a two-way feedback.

[0008] Real-time and historical data are integrated into the LLM large model, and then transformed into business knowledge through the OAG ontology library to generate primary and backup decision-making schemes;

[0009] Establish a federal intelligent agent hub to issue decision-making plans, collect execution data and feed it back to the LLM large model, so as to realize cross-scenario collaboration and full-process data closed loop.

[0010] Furthermore, the specific steps for targeted fine-tuning of the basic large language model are as follows:

[0011] Collect historical production data, quality control records, logistics control logs, business process documents, industry rule manuals, equipment maintenance manuals, order contract data, and bill of materials data in the discrete manufacturing field, and form standardized data after preprocessing;

[0012] Based on standardized data, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio.

[0013] Select a basic large language model, build a supervised fine-tuning framework, associate the training set data with the corresponding business labels and input them into the fine-tuning framework, start model fine-tuning training, and adjust the model's attention mechanism weights and fully connected layer parameters through the backpropagation algorithm;

[0014] During the model fine-tuning training process, after every 10 rounds of parameter iteration, the validation set data is input into the model in the current training state to obtain the semantic parsing results, business rule matching results, and preliminary decision-making schemes output by the model, and to calculate the semantic parsing accuracy, business rule matching accuracy, and the rationality of the decision-making scheme.

[0015] When the semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality all meet the requirements, stop the model fine-tuning training and obtain the preliminary fine-tuned model; otherwise, adjust the parameters and continue iterative training.

[0016] Based on the initial fine-tuning of the model, input the test set data, output the final output results, and determine whether the semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality meet the requirements. If they meet the requirements, complete the targeted fine-tuning and output the LLM large model. If they do not meet the requirements, repeat the above steps until the model performance meets the requirements.

[0017] Furthermore, the specific steps for constructing the OAG ontology library are as follows:

[0018] Based on the LLM model, the original data and business documents of the discrete manufacturing scenario are imported, and four core entities, namely equipment, materials, workstations and personnel, are extracted. The specific attributes of the core entities and the relationships between the core entities are clarified.

[0019] Based on core entities, specific attributes, and relationships, an element ontology is established, and semantic ambiguity from different data sources is eliminated through a unified data semantic definition.

[0020] Based on business documents, we break down three core business actions: production control, quality control, and logistics control. We then analyze the execution process of each action, the logical relationships between actions, and the corresponding business rules. We establish a business ontology, standardize the expression of business processes, and clarify the execution order and constraints of business actions.

[0021] Based on raw data and business documents from discrete manufacturing scenarios, we extract three types of industry experience accumulated in the discrete manufacturing industry: process parameter thresholds, quality anomaly judgment standards, and safe operation specifications. These are then transformed into structured judgment logic that can be directly parsed by the LLM big model, establishing a logical rule ontology and forming a clear basis for business decision-making.

[0022] By associating the element ontology, business ontology, and logical rule ontology, a complete OAG ontology library is formed.

[0023] Furthermore, after uncovering implicit rules, a weighted Bayesian confidence algorithm is used to calculate the confidence level of the implicit rules. Rules that meet the confidence threshold are pushed to the OAG ontology library for automatic updates, while rules with low confidence are manually reviewed and updated. The mathematical expression of the weighted Bayesian confidence algorithm is as follows:

[0024] ;

[0025] in, This represents the confidence score of the latent rule R obtained through mining, with a value ranging from 0 to 1. Let i represent the i-th type of data that supports the implicit rule R. Indicates based on the i-th type of data The conditional probability of determining that the implicit rule R is true. Represents the i-th type of data The weighting coefficients are set based on the reliability of the data source, where n represents the total number of data types that support the implicit rule R. When the value is greater than or equal to a set threshold, the implicit rule R is automatically pushed to the OAG ontology library to complete the update. When the value is less than the set threshold, a manual review process is triggered. Once the implicit rule R is confirmed to be valid after review, it is then updated to the OAG ontology library.

[0026] Furthermore, the specific steps for forming the bidirectional feedback are as follows:

[0027] Configure the data acquisition interface of the LLM large model to receive real-time data from equipment sensor data, order change information, material inventory change data, workstation production status data, AGV running position data and real-time quality control results from the discrete manufacturing site, forming real-time data.

[0028] Activate the multi-dimensional association analysis function of the LLM large model to perform cross-comparison and in-depth mining on the received real-time data, and identify implicit association rules that are not covered by the current rules of the OAG ontology library.

[0029] The confidence score of each implicit association rule is calculated using a weighted Bayes confidence algorithm, and the OAG ontology is updated.

[0030] Enable the feedback optimization channel of the OAG ontology library, transfer the updated ontology content to the semantic parsing module, logical reasoning module and decision generation module of the LLM large model, adjust the parameter configuration of the corresponding modules of the LLM large model, and repeat the above steps to form a continuous iterative bidirectional feedback closed loop.

[0031] Furthermore, before generating the primary and backup decision schemes, business knowledge is combined with uncertainty event quantification analysis. The uncertainty quantification analysis process adopts a Bayesian network algorithm, and the mathematical expression is as follows:

[0032] ;

[0033] in, Indicates equipment fault F in real-time parameters The posterior probability under the combined effect These represent the real-time monitoring parameters related to equipment failure F, specifically: equipment operating temperature, speed, vibration frequency, voltage, current, and operating time. This indicates the real-time parameters assuming equipment failure F occurs. The joint likelihood that appears simultaneously This represents the prior probability of equipment failure F occurring, derived from statistical analysis of historical failure data. Indicates real-time parameters The marginal probability of simultaneous occurrence;

[0034] Based on the Bayesian network algorithm, the LLM large model quantifies the probability of occurrence of four types of uncertain events: equipment failure, material delay, order change, and process adjustment. Combined with the corresponding coping strategies in business knowledge, it generates a master decision plan that takes into account both feasibility and safety. At the same time, it formulates alternative plans for uncertain events with higher probability. The alternative plans clearly define the triggering conditions, adjustment steps, and resource allocation methods to ensure that a quick switch can be made if the master decision plan cannot be executed smoothly.

[0035] Furthermore, the specific steps for generating the primary and backup decision scheme are as follows:

[0036] Real-time data from discrete manufacturing is collected through a data acquisition module and combined with historical data to input into the LLM large model. The real-time data includes real-time operating parameters of equipment sensors, real-time position of AGVs, production status of workstations, real-time results of quality control, order change notifications, and real-time changes in material inventory. The historical data includes production control records, quality control reports, logistics control logs, equipment maintenance records, and order delivery data.

[0037] The feature ontology in the OAG ontology library is invoked, and the entity association mechanism is activated to integrate the scattered real-time data and historical data according to entity attributes and association relationships, and transform them into structured information. The structured information explicitly includes entity name, entity attribute, entity relationship and corresponding data value.

[0038] The business ontology and logical rule ontology in the OAG ontology library are called to perform business logic judgment on structured information. The judgment results are combined with business experience in the discrete manufacturing industry to transform the structured information into business knowledge that can be directly applied.

[0039] Based on business knowledge, the uncertainty event analysis module is launched. Using the Bayesian network algorithm, structured information data related to four types of uncertainty events—equipment failure, material delay, order change, and process adjustment—are input to quantify the probability of occurrence of each type of event, resulting in quantified uncertainty event probabilities. Real-time production demand parameters are obtained through the LLM large model, and multi-dimensional fusion analysis is performed by combining business knowledge and quantified uncertainty event probabilities to generate fusion analysis results.

[0040] Based on the results of the fusion analysis, a master decision-making scheme that conforms to the actual production situation is generated. At the same time, for uncertain events with a probability of occurrence greater than the threshold, corresponding alternative schemes are formulated, and the triggering conditions, adjustment steps and resource allocation methods of the alternative schemes are clarified.

[0041] Furthermore, the cross-scenario collaboration employs a conflict risk quantification model, the mathematical expression of which is as follows:

[0042] ;

[0043] in, This represents the risk value C for cross-scenario conflict, ranging from 0 to 100. The probability of cross-scenario conflict C is represented by the prediction made by the LLM model based on historical conflict data and real-time business data. This represents the weight coefficient for the j-th business objective, set according to the business strategy of discrete manufacturing enterprises. The weight coefficient for the order delivery period objective is set to 0.6, the weight coefficient for the equipment utilization rate objective is set to 0.2, the weight coefficient for the cost control objective is set to 0.15, and the weight coefficient for the product quality objective is set to 0.05. This represents the impact of cross-scenario conflict C on the j-th business objective. The value ranges from 0 to 10. The larger the impact value, the more serious the interference of the conflict on the business objective. j=1 corresponds to the order delivery period objective, j=2 corresponds to the equipment utilization rate objective, j=3 corresponds to the cost control objective, and j=4 corresponds to the product quality objective.

[0044] when When the value is greater than or equal to a preset threshold, the intelligent agent federation center triggers a collaborative adjustment mechanism. Based on the conflict type and scope of impact, it coordinates the intelligent agents in the relevant scenarios to adjust their execution plans, ensuring that each business objective is not severely affected by the conflict. At the same time, the collaborative adjustment results are synchronized to the LLM large model and the OAG ontology library to provide a reference for subsequent decision-making.

[0045] Furthermore, the specific steps of the entire data closed loop are as follows:

[0046] Establish a federal hub for intelligent agents, configure data interaction modules, conflict identification modules, collaborative scheduling modules, and feedback processing modules, and clarify the functional responsibilities and data interaction protocols of each module;

[0047] Deploy scenario-based intelligent agents, including production control intelligent agents, quality control intelligent agents, and logistics control intelligent agents. Configure each intelligent agent with an independent execution module, data receiving interface, and feedback interface, and clarify the functional boundaries and interaction permissions of each intelligent agent.

[0048] Based on the primary and backup decision-making scheme, the intelligent agent federation hub is built to break down the scheme into three business scenarios: production, quality, and logistics. The scheme is then distributed to the corresponding scenario-based intelligent agents through dedicated interfaces. The scenario-based intelligent agents transmit real-time execution data and running status information to the data interaction module of the intelligent agent federation hub through feedback interfaces.

[0049] The conflict identification module extracts the execution data and status information of each agent from the data interaction module. Combined with the business logic conflict risk quantification model in the OAG ontology library, it calculates the risk value of potential conflicts in real time. When the conflict risk value is greater than the threshold, the collaborative scheduling module formulates a special adjustment strategy for the conflict type and scope of impact. The strategy is then sent to the relevant scenario-based agents through the interaction interface. After receiving the adjustment strategy, the relevant agents adjust the execution plan in real time and continue to execute, outputting the final execution data.

[0050] The feedback processing module standardizes and organizes the final execution data and conflict resolution results of each agent to form feedback data, which is then sent back to the LLM large model and the OAG ontology library through dedicated feedback channels. After receiving the feedback data, the LLM large model optimizes the parameter configuration of its own decision generation module, and after receiving the feedback data, the OAG ontology library updates the corresponding business rules and logical judgment criteria.

[0051] Furthermore, after LLM large model optimization and OAG ontology library update, the following steps are also included:

[0052] The updated OAG ontology library is matched with historical business data and real-time data to verify the entity coverage completeness of the feature ontology, the process adaptation accuracy of the business ontology, and the conflict rate of the logical rule ontology. Ontologies that pass the verification are updated, while ontologies that fail the verification remain unchanged.

[0053] Connect the optimized LLM model to the test environment, input feedback data and historical test sets, and evaluate the semantic parsing accuracy, logical reasoning matching accuracy, and decision scheme compliance rate. If all three indicators meet the requirements, the optimization is complete; otherwise, keep the original version of the LLM model.

[0054] Compared with existing technologies, the intelligent decision-making and business collaboration method based on OAG ontology and LLM large model provided by this invention has the following beneficial effects:

[0055] This invention effectively addresses the core pain point of existing LLM models—their inability to perform calculations effectively—by employing a technical approach that includes targeted fine-tuning of the LLM large model, construction of the OAG ontology library, bidirectional feedback loops, and end-to-end data loops. It achieves end-to-end value transformation of discrete manufacturing data from data, information, knowledge, to intelligence. Through the application of weighted Bayesian confidence algorithms, Bayesian network algorithms, and conflict risk quantification models, combined with the cross-scenario collaborative capabilities of the intelligent agent federation hub, it significantly improves decision-making accuracy and business collaboration efficiency, optimizes production control response speed and quality anomaly handling efficiency, and reduces logistics control-related costs. Simultaneously, the scalable design of the OAG ontology library and the adaptation mechanism of the LLM large model enable flexible adaptation to different discrete manufacturing sub-sectors such as automotive parts manufacturing and electronic equipment manufacturing, without requiring a reconstruction of the technical framework. This significantly lowers the industry application threshold and drives a leapfrog upgrade of discrete manufacturing from discrete control to intelligent collaboration. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating the intelligent decision-making and business collaboration method based on OAG ontology and LLM large model in this invention. Detailed Implementation

[0057] In the description of this invention, it should be understood that the terms "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "middle," and "inner," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, features defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, it should be noted that unless otherwise explicitly specified and limited, the terms "installed," "connected," and "joined" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention through specific circumstances.

[0058] The embodiments of the present invention will be described below with reference to the accompanying drawings and related examples. The embodiments of the present invention are not limited to the following examples, and the present invention relates to the relevant necessary components in this technical field, which should be regarded as well-known technology in this technical field and can be known and mastered by those skilled in this technical field.

[0059] See Figure 1 This invention provides an intelligent decision-making and business collaboration method based on OAG ontology and LLM large model, comprising the following steps:

[0060] Step S100: Collect historical data and business documents in the discrete manufacturing field, and fine-tune the basic large language model to form an LLM large model adapted to discrete manufacturing business.

[0061] Historical data refers to accumulated records related to production operations, encompassing time-series or non-time-series information such as scheduling logs, quality control results, and equipment status. It can be used as input for fine-tuning the LLM model, providing background knowledge to enhance the model's understanding of the manufacturing process. In this embodiment, historical data can be exported from enterprise MES, ERP, SCADA, and other information system databases, or extracted and cleaned using ETL tools. Business documents are unstructured text materials describing enterprise operating rules, process standards, and management systems, such as SOP manuals, BOM descriptions, and exception handling guidelines. They can supplement the semantic context missing in historical data, helping the LLM establish a cognitive framework for manufacturing business logic. Furthermore, business documents can be digitized from internal document management systems or paper documents, preprocessed using OCR and NLP, and then used for model learning. The basic large language model is a general-purpose large language model that has not been domain-adapted, possessing basic language generation and understanding capabilities. It can be used as a starting point for fine-tuning, providing initial language processing capabilities, and subsequently achieving specialized transformation through domain data injection. In one specific embodiment, the basic large language model can adopt a publicly released pre-trained model architecture and acquire general semantic representation capabilities through training on massive amounts of internet text. Targeted fine-tuning is the process of adjusting the parameters of the pre-trained model using relevant datasets for a specific application domain. It can be used to enable the basic large language model to acquire terminology recognition, contextual understanding, and logical reasoning capabilities in the discrete manufacturing domain.

[0062] Step S200: Use the LLM large model to parse the original data and business documents of the discrete manufacturing scenario, and build an OAG ontology library containing element ontology, business ontology, and logical rule ontology;

[0063] The LLM large-scale model is a large language model with semantic understanding capabilities in the discrete manufacturing domain after targeted fine-tuning. It can process natural language instructions, parse on-site data, and generate decision suggestions. It can be used to undertake tasks such as data parsing, implicit rule mining, knowledge transformation, and decision scheme generation, and is the core computing engine for realizing intelligent decision-making. The OAG ontology is a structured knowledge representation system used to organize entities, relationships, and logical rules in the discrete manufacturing domain. It contains a three-level ontology structure: element ontology, business ontology, and logical rule ontology. It can be used to transform unstructured manufacturing data into structured knowledge, supporting the reasoning accuracy and decision interpretability of the LLM large-scale model, and serving as a core intermediary carrier for data dimensionality enhancement. The element ontology is the lowest-level constituent unit of the OAG ontology, representing indivisible basic entities in the manufacturing process, such as materials, equipment, work orders, and personnel. It can be used to provide atomic-level semantic support for higher-level ontologies, ensuring the integrity and consistency of subsequent knowledge expression. The business ontology is a knowledge unit describing various processes, tasks, and their interrelationships in manufacturing activities, such as assembly processes, inspection procedures, and mold change operations. It can be used to characterize the behavioral structure of the manufacturing process and support the modeling and optimization of production paths. Logical rule ontology can be a set of knowledge that expresses constraints, causal relationships and decision-making logic in a manufacturing system, such as "processing is suspended if the temperature exceeds the limit" or "high-precision equipment is given priority to critical components". It can be used to provide reasoning basis for decision generation and enhance the compliance and stability of the system.

[0064] Step S300: Receive real-time field data through the LLM large model, mine implicit rules to update the OAG ontology library, and at the same time, the OAG ontology library feeds back to optimize the LLM large model, forming a two-way feedback.

[0065] On-site data refers to real-time data streams collected from the production workshop, including dynamic information such as equipment operating status, sensor readings, and manually entered information. This data can serve as input for the LLM large-scale model to perceive the current production status and to discover abnormal patterns and potential regularities. Implicit rules are empirical practices or potential correlations that are not explicitly written down but are generally followed in actual operation, such as the tendency of certain equipment to deviate under specific temperature and humidity conditions. These rules can supplement operational wisdom missing from formal documentation, improving the system's adaptability to complex situations. Two-way feedback is the interactive mechanism between the LLM large-scale model and the OAG ontology library, where they interact and serve as each other's input and output. This mechanism can be used to maintain the timeliness and accuracy of the knowledge system, forming a self-evolving intelligent system architecture.

[0066] Step S400: Integrate real-time data and historical data into the LLM large model, transform them into business knowledge through the OAG ontology library, and generate primary and backup decision-making schemes;

[0067] Business knowledge is actionable knowledge that has been semantically refined and logically organized. It manifests as specific information units that can guide decision-making and can be used to bridge the semantic gap between data and decisions, enabling machines to output judgments that are more business-reasoned. Primary and backup decision schemes are a set of candidate decision options generated by the system. The primary scheme is the preferred recommendation, while the backup schemes are alternative strategies to deal with unforeseen circumstances. They can be used to improve decision-making flexibility and robustness, supporting rapid switching of response strategies in uncertain environments.

[0068] Step S500: Build an intelligent agent federation hub, issue decision-making schemes, collect execution data and feed it back to the LLM large model to achieve cross-scenario collaboration and full-process data closed loop.

[0069] The intelligent agent federation hub is a centralized coordination node responsible for managing communication, task distribution, and state synchronization among multiple distributed intelligent agents. It supports collaborative scheduling across production scenarios and can be used to achieve decision-making linkage and data feedback between different processes, systems, or devices, ensuring the effective implementation of primary and backup decision-making schemes and full-process tracking. The end-to-end data closed loop is a complete link loop mechanism from data collection, knowledge generation, decision-making to execution feedback. It can be used to ensure that the system has the ability to continuously learn and optimize, avoiding decision-making biases caused by information breakpoints.

[0070] This invention continuously receives field data through LLM and mines implicit rules to update the OAG ontology library. Simultaneously, the OAG ontology library optimizes the LLM's reasoning process through knowledge injection, forming a two-way feedback loop that enhances the system's adaptability to dynamic environments. In the decision-making phase, real-time and historical data are integrated, leveraging the hierarchical structure of the OAG ontology library to progressively upgrade from raw data to business knowledge. A weighted Bayesian confidence algorithm and a conflict risk quantification model are combined to generate primary and backup decision schemes with risk assessment capabilities. By building an intelligent agent federation hub, instruction issuance and execution coordination among multiple production units are achieved, and execution data is collected and fed back to the LLM large model, constructing a complete data loop spanning data collection, knowledge generation, decision-making, and execution feedback. The OAG ontology library adopts a modular design, supporting flexible expansion of ontology sub-libraries according to industry characteristics. Combined with the lightweight adaptation mechanism of LLM, it can be migrated to different manufacturing scenarios such as automotive parts and electronic equipment without reconstructing the system architecture, effectively reducing the technical barriers to cross-industry deployment.

[0071] In one embodiment of the present invention, the specific steps for targeted fine-tuning of the basic large language model are as follows:

[0072] Step S101: Collect historical production data, quality control records, logistics control logs, business process documents, industry rule manuals, equipment maintenance manuals, order contract data, and bill of materials data in the discrete manufacturing field, and preprocess them to form standardized data;

[0073] Historical production data consists of workstation production records and equipment operation logs (such as daily output of stamping stations and lathe speed variation curves) retrieved from the enterprise's MES (Manufacturing Execution System) over the past 1-3 years; quality control records are product dimension inspection reports and surface defect judgment results exported from the quality inspection system (such as "batch B123 product dimension deviation +0.01mm, judged as qualified"); logistics control logs are obtained from the WMS (Warehouse Management System) for AGV delivery path records and material inbound / outbound timestamps; business process documents are structured documents such as "Production Control Management Specifications" and "Quality Anomaly Handling Procedures" provided by the production department; industry rule manuals are referenced from national / industry standards (such as "Automotive Parts Manufacturing Quality Standards") and the enterprise's internal process manuals; equipment maintenance manuals are equipment parameter thresholds (such as "welding equipment current limit 300A") and maintenance cycle records provided by equipment suppliers; order contract data is extracted from the ERP (Enterprise Resource Planning) system for order delivery dates, product models, quantities, and other information; and bill of materials data is obtained from the BOM (Bill of Materials) system for material codes, specifications, product levels, and other data. Preprocessing operations include data cleaning, data deduplication, format standardization, and missing value imputation.

[0074] Step S102: Based on standardized data, divide the data into training set, validation set and test set in a ratio of 8:1:1;

[0075] The training set, comprising 80% standardized data, is used for core training of model parameters and is the primary data source for the model to learn the semantics, business rules, and data patterns of the discrete manufacturing industry. The validation set, comprising 10% standardized data, is used to adjust model hyperparameters (such as learning rate and number of iterations) in real time during fine-tuning, evaluating performance changes during training and preventing overfitting (models that only fit the training data and have poor generalization ability to new data). The test set, comprising 10% standardized data, is used to objectively evaluate the model's final generalization ability after fine-tuning. This dataset is not used in the training process and can simulate "new data processing scenarios" in real-world business scenarios, ensuring the evaluation results are realistic and reliable.

[0076] Step S103: Select a basic large language model, build a supervised fine-tuning framework, associate the training set data with the corresponding business labels and input them into the fine-tuning framework, start model fine-tuning training, and adjust the model's attention mechanism weights and fully connected layer parameters through the backpropagation algorithm.

[0077] Prioritize models that support multimodal input (can process text and numerical data) and allow for custom fine-tuning of parameters, such as deepseek-7B (suitable for fine-tuning in small to medium-sized industries) or Tongyi Qianwen-13B (suitable for complex business scenarios); set up a GPU training environment (e.g., 8 NVIDIA A100 graphics cards), install the PyTorch deep learning framework, configure the necessary libraries for model training (e.g., Transformers, Datasets), and set the training batch size (BatchSize=32) and initial learning rate (LearningRate=2e-5). The design principles of business tags must align with the needs of discrete manufacturing operations, covering three categories of tags: semantic understanding, rule matching, and decision-making prototypes. Semantic understanding tags include: for example, "Data 'Stamping Station A' → Tag 'Core entity in the production scenario, responsible for the metal stamping process'"; rule matching tags include: for example, "Data 'Product size deviation +0.03mm' → Tag 'Does not conform to the size tolerance ±0.02mm rule, deemed unqualified'"; and decision-making prototype tags include: for example, "Data 'Urgent order + Station A load rate 60%' → Tag 'Prioritize scheduling this order for production at Station A'". The association method involves a one-to-one correspondence between "Data ID - Tag ID", integrating standardized data and business tags into "training sample pairs" (e.g., "{'Data content': 'Equipment temperature 320℃, lasting 6 minutes', 'Tag': 'Equipment malfunction, triggering shutdown'}"), forming a training set.

[0078] After training begins, the training set is input into the base model in batches. Supervised fine-tuning is used, and the model parameters are iteratively updated through backpropagation. The attention mechanism is adjusted by enhancing the model's attention weights to key industry terms (such as "dimensional tolerance" and "bottleneck workstation"), so that the model prioritizes core business information when processing data. The fully connected layer parameters are updated by adjusting the weights of the model's output and hidden layers through error backpropagation, so that the model's predicted output gradually approaches the business labels. Each batch of training is set to 100 iterations. After each training iteration, the "prediction error loss value" (using the cross-entropy loss function) is calculated. When the loss value decreases by less than 0.001 for 5 consecutive iterations, the training of the current batch is paused, and it is analyzed whether the hyperparameters need to be adjusted.

[0079] Step S104: During the model fine-tuning training process, after every 10 rounds of parameter iteration, the validation set data is input into the model in the current training state to obtain the semantic parsing results, business rule matching results, and preliminary decision scheme output by the model, and the semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality are calculated.

[0080] After every 10 rounds of parameter iteration, validation set data is batch-input into the model in its current training state. The model outputs three types of results: semantic parsing results, rule matching results, and preliminary decision-making schemes. Semantic parsing accuracy is calculated by comparing the semantic parsing results with semantic understanding labels. Rule matching accuracy is calculated by comparing the rule matching results with rule matching labels. The preliminary decision-making schemes are scored by 3-5 production or quality control experts based on whether they conform to actual production processes and meet resource constraints; the average score is used to determine the rationality of the decision-making scheme. Simultaneously, the semantic parsing accuracy is set to ≥90%, and the business rule matching accuracy to ≥88%. The decision-making scheme has three minimum thresholds: rationality ≥ 85%. If all three thresholds are met, the test proceeds to the next step. If the semantic parsing accuracy is low, the business term annotation samples are increased and the model word vector mapping parameters are adjusted before repeating 10 iterations. If the business rule matching accuracy is low, industry rule annotation data is added, the model's learning of rule logic is strengthened, and the learning rate is reduced to 1e-5 before repeating the iteration. If the decision-making scheme is not rational, decision scenario annotation samples are increased and the model decision layer parameters are optimized before repeating the iteration. An iteration log is also established to record the hyperparameters, performance indicators, and adjustment measures for each 10 iterations in order to trace the cause of performance fluctuations.

[0081] Step S105: When the semantic parsing accuracy, business rule matching accuracy and decision scheme rationality all meet the requirements, stop the model fine-tuning training and obtain the preliminary fine-tuned model; otherwise, adjust the parameters and continue iterative training.

[0082] Semantic parsing accuracy is a core metric measuring the model's correctness in understanding the semantics of discrete manufacturing business data. It is calculated as "number of validation samples whose semantic parsing results match the labels / total number of validation samples × 100%", reflecting the model's depth of understanding of industry language. Business rule matching accuracy measures the model's correctness in matching validation data with discrete manufacturing business rules. It is calculated as "number of validation samples whose rule matching results match the labels / total number of validation samples × 100%", reflecting the model's adaptability to industry rules. Decision-making rationality is assessed through "expert review + business scenario simulation" to determine the degree to which the preliminary decision-making solutions (such as scheduling suggestions and inspection judgments) generated by the model conform to actual production conditions. A 100-point scoring system is used (≥85 points indicates reasonableness), reflecting the practicality of the model's decisions.

[0083] Step S106: Based on the initial fine-tuning model, input the test set data, output the final output results, and determine whether the semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality meet the requirements. If they meet the requirements, complete the targeted fine-tuning and output the LLM large model. If they do not meet the requirements, repeat the above steps until the model performance meets the requirements.

[0084] The test set data does not participate in any training or validation process. Before input, only format validation is performed to ensure consistency with the training set format. No other preprocessing is done to simulate the scenario of new data in actual business. Then, the test set data is input into the model in batches of 16 for preliminary fine-tuning. The semantic parsing results, rule matching results, and preliminary decision schemes are output, and all output results are saved for evaluation. Then, the three core indicators on the test set—semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality—are recalculated according to the indicator calculation method in step 104. If all three indicators meet the following criteria, the semantic parsing accuracy ≥ 90%, the business rule matching accuracy ≥ 90%, and the business rule matching accuracy ≥ 90%, the model will be considered successful. If the thresholds of ≥88% and decision scheme rationality ≥85% are met, the model's generalization ability is deemed to be up to standard. The discrete manufacturing-adapted LLM large model is output and the model parameter file in pth format is saved. If the test set indicators do not meet the standards, the process returns to step 101 to re-optimize the data preprocessing or returns to step 103 to adjust the fine-tuning framework until the test set indicators meet the standards. Finally, the final LLM large model is encapsulated as an API service, providing a standardized calling interface such as RESTful API, specifying the input data format, calling frequency limits, and other interface parameters to facilitate subsequent docking with the OAG ontology library construction module and the intelligent agent federation hub.

[0085] In one embodiment of the present invention, the specific steps for constructing the OAG ontology library are as follows:

[0086] Step S201: Based on the LLM large model, import the original data and business documents of the discrete manufacturing scenario, extract the four core entities of equipment, materials, workstations and personnel, and clarify the specific attributes of the core entities and the relationship between the core entities.

[0087] Based on the already fine-tuned LLM model, raw data and business documents from the discrete manufacturing scenario are imported. The raw data comes from core discrete manufacturing business systems such as MES, ERP, and WMS, including equipment operation logs, material flow records, workstation production data, and personnel scheduling information. The business documents include production management specifications, process manuals, and job operation guides. The LLM model accurately extracts four core entities—equipment, materials, workstations, and personnel—from the above data and documents through semantic parsing and entity recognition.

[0088] For each type of core entity, its specific attributes are clearly defined. For example, the attributes of an equipment entity include equipment model, rated parameters, operating status, and maintenance cycle; the attributes of a material entity include material code, specifications, material type, and inventory quantity; the attributes of a workstation entity include workstation number, workshop, work process, and compatible equipment; and the attributes of a personnel entity include employee number, job type, skill level, and assigned workstation. Simultaneously, the relationships between core entities are clarified. For example, the compatibility relationship between equipment and workstations identifies the types of equipment specific to or compatible with a particular workstation; the correspondence between materials and workstations determines which workstations process a particular type of material; the responsibility relationship between personnel and workstations identifies the dedicated operators for a particular workstation; and the compatibility relationship between equipment and materials determines the types of materials that a particular piece of equipment can process. This forms a complete list of entity-attribute-relationships.

[0089] Step S202: Based on core entities, specific attributes, and relationships, establish an element ontology and eliminate semantic ambiguity from data from different sources by unifying data semantic definitions;

[0090] Based on the core entities, specific attributes, and relationships identified in the first step, the element ontology construction process is initiated. First, a unified semantic definition standard is established to standardize and regulate the attribute names, data formats, and value ranges of entities of the same type. For example, the different expressions of "equipment operating temperature" in different systems are unified into a standard field, temperature values ​​are standardized to degrees Celsius, and the "normal / abnormal" expressions for equipment operating status are uniformly defined. Second, based on the standardized semantic definitions, core entities, attributes, and relationships are entered into the ontology construction tool to establish a hierarchical structure for the element ontology. The top layer is the core entity category, the middle layer is the specific entity attributes, and the bottom layer is the attribute values ​​and relationship rules. Finally, through a semantic mapping mechanism, entity expressions from different data sources are associated with the standard semantics in the element ontology. For example, "stamping machine A" in the MES system is precisely mapped to "equipment-stamping equipment-model XXX" in the element ontology, eliminating semantic ambiguity caused by differences in expression between different data sources and achieving semantic unification of discrete manufacturing data.

[0091] Step S203: Based on business documents, break down the three core business actions of production control, quality control, and logistics control, sort out the execution process of each business action, the relationship logic between actions and the corresponding business rules, establish the business ontology, standardize the standardized expression of business processes, and clarify the execution order and constraints of business actions.

[0092] Based on discrete manufacturing business documents, this study focuses on three core business operations: production control, quality control, and logistics control, and deconstructs the business processes. For production control operations, the complete execution flow from order receipt, capacity calculation, task allocation, process sequencing to schedule generation is outlined. The input conditions for each stage are clearly defined, such as the order receipt stage requiring the product model, quantity, and delivery date; the output results, such as the schedule generation stage outputting the production tasks, execution time, and resource requirements for each workstation; and the interrelationships between actions, such as capacity calculation based on order demand and equipment capacity data, and task allocation based on capacity calculation results and workstation load. Simultaneously, the business rules for production control are clarified, including rules prioritizing urgent orders, balancing bottleneck workstation load, and centralized production of similar products. For quality control operations, the process from initiating inspection tasks, setting inspection parameters, sample collection, inspection execution, to result judgment is broken down. The operational specifications, inspection parameter thresholds, and sample collection ratios for each stage are clearly defined. The criteria for judging whether inspection results are qualified or unqualified are outlined, as well as the subsequent processing logic for rework and scrapping of unqualified products. For logistics management operations, the process is broken down from material demand initiation, inventory query, route planning, AGV scheduling to material receipt, clarifying the triggering conditions, logistics route optimization principles, and AGV scheduling priority rules for each step. Based on the above-described processes, related logic, and business rules, a business ontology is established, standardizing the expression of each business action, clarifying the execution order and constraints of the business actions, and ensuring that the LLM model can orchestrate tasks according to actual business logic.

[0093] Step S204: Based on the original data and business documents of the discrete manufacturing scenario, extract the three types of industry experience accumulated in the discrete manufacturing industry: process parameter thresholds, quality anomaly judgment standards, and safe operation specifications. Transform them into structured judgment logic that can be directly parsed by the LLM big model, establish the logical rule ontology, and form a clear basis for business decision-making.

[0094] From raw data and business documents in discrete manufacturing scenarios, three categories of industry experience were extracted: process parameter thresholds, quality anomaly judgment criteria, and safe operation procedures. Process parameter thresholds include the range of key process parameters for each process, such as the stamping pressure threshold for stamping, the welding current and voltage thresholds for welding, and the cutting speed and feed rate thresholds for machining. This data primarily originates from equipment maintenance manuals, process specification documents, and records of optimal parameters from historical production. Quality anomaly judgment criteria include product dimensional tolerance ranges, surface defect judgment criteria, and acceptable performance index ranges, derived from quality control specifications, industry quality standards, and historical quality control data. Safe operation procedures include safety thresholds for equipment operation, such as safe distances during equipment operation, operator protection requirements, and operational restrictions for hazardous processes, derived from safety operation manuals, industry safety standards, and internal enterprise safety management systems. The extracted industry experience is transformed into structured decision logic that can be directly parsed by the LLM model. This logic uses a "condition-result" rule format, such as: "If the stamping pressure in the stamping process is >300MPa and the duration is >5 seconds, then the process parameter is considered abnormal, triggering a equipment shutdown command"; "If the product size deviation exceeds ±0.02mm, then the quality is considered unqualified, triggering a rework process"; "If the operator is not wearing protective equipment, then starting the equipment is prohibited, triggering a safety alarm." These structured decision logics are then entered into an ontology construction tool to establish a logical rule ontology, forming a clear basis for business decision-making.

[0095] Step S205: Associate the element ontology, business ontology, and logical rule ontology to form a complete OAG ontology library.

[0096] Establish the association mapping relationship between the element ontology, business ontology, and logical rule ontology: Each business action in the business ontology is associated with the corresponding core entity in the element ontology, clarifying the entities involved in the business action, such as equipment, materials, workstations, and personnel; simultaneously, it is associated with the corresponding judgment logic in the logical rule ontology, clarifying the process parameter thresholds, quality standards, safety specifications, etc., that the business action must follow. For example, the production control business action is associated with the equipment, material, and workstation entities in the element ontology, clarifying the entity attributes such as equipment capacity, material inventory, and workstation load that need to be considered in scheduling; it is also associated with the judgment logic such as emergency order priority rules and load balancing rules in the logical rule ontology, ensuring that the scheduling results conform to the business rules. Through association mapping, the three ontology modules are integrated into a complete OAG ontology library. Subsequently, the integrity of the ontology library is verified to check the coverage of core entities, ensuring that no key entities in production, quality, and logistics scenarios are omitted; the integrity of business processes is verified to confirm that there are no missing links in the execution process of each core business action; and the comprehensiveness of logical rules is verified to ensure that the main process parameters, quality standards, and safety specifications have been converted into structured rules. After the verification is passed, the final OAG ontology library is formed.

[0097] This invention establishes a standardized element ontology by clearly defining the attributes and relationships of core entities, thus completely resolving the problem of inconsistent semantics in discrete manufacturing data and eliminating misunderstandings caused by differences in data representation across different systems and documents. Previously isolated data scattered across various systems are effectively linked through the element ontology's association mechanism, significantly improving data interpretability and usability. Data utilization is increased by more than 40% compared to traditional methods, laying a solid foundation for the subsequent formation of structured information.

[0098] It should be noted that after uncovering implicit rules, a weighted Bayesian confidence algorithm is used to calculate the confidence level of the implicit rules. Rules that meet the confidence threshold are pushed to the OAG ontology library for automatic updates, while rules with low confidence are manually reviewed and updated. The mathematical expression of the weighted Bayesian confidence algorithm is as follows:

[0099] ;

[0100] in, This represents the confidence score of the latent rule R obtained through mining, with a value ranging from 0 to 1. Let i represent the i-th type of data that supports the implicit rule R. Indicates based on the i-th type of data The conditional probability of determining that the implicit rule R is true. Represents the i-th type of data The weighting coefficients are set based on the reliability of the data source, where n represents the total number of data types that support the implicit rule R. When the value is greater than or equal to a set threshold, the implicit rule R is automatically pushed to the OAG ontology library to complete the update. When the value is less than the set threshold, a manual review process is triggered. Once the implicit rule R is confirmed to be valid after review, it is then updated to the OAG ontology library.

[0101] The differences Depending on the data source and business significance, i takes values ​​from 1 to 3, where This represents real-time data from equipment sensors, such as the "temperature data" and "speed data" of the stamping station equipment that support rule R. These data come directly from field sensors and have the highest real-time performance and accuracy. This represents quality control record data, such as "product size deviation detection results" and "surface defect judgment records" that support rule R. These data come from the quality inspection system and reflect the impact of the rule on the quality results. This represents order execution and delivery data, such as "order delivery delay records" and "rework order percentage data" that support rule R. These data come from the ERP system and reflect the indirect impact of the rule on production progress. The weighting values ​​are set based on the reliability of the data source; the higher the reliability, the higher the weighting. The larger the value, the higher the weight of high-reliability data in the confidence calculation, thus avoiding interference from low-reliability data in the rule determination.

[0102] Among them, real-time data from the equipment sensors ( The data is directly collected from the production site, with small errors, strong real-time performance, and the highest reliability; its weighting coefficient is set to 0.9. Quality control record data ( The data is generated using professional equipment and has a small error margin, but there may be delays due to manual data entry. The weighting factor is set to 0.8. (Order execution and delivery data...) Due to the influence of multiple factors (such as logistics and supply chain), its direct correlation with rule R is relatively weak, and the weighting coefficient is set to 0.7. The weighting setting needs to be combined with the actual data quality of the enterprise. If the accuracy of a company's quality control equipment is insufficient, the weighting coefficient can be adjusted accordingly. The weight was lowered to 0.75 to ensure that the weight matches the actual reliability.

[0103] In discrete manufacturing scenarios, implicit rules mined by large LLM models (such as "the pass rate of a certain material decreases by 12% when processed in an environment with humidity > 65%) often lack clear documentation support, and their effectiveness needs to be verified by multiple types of data. Furthermore, the reliability of data from different sources (such as real-time data from equipment sensors and manually recorded quality results) varies. Simply using "data volume ratio" to judge rule effectiveness can easily lead to misjudgments due to interference from low-reliability data. The core function of the weighted Bayes confidence algorithm is to quantify the correlation strength between rules and data through Bayesian conditional probability calculation, while highlighting the influence of high-reliability data through data weight allocation, ultimately outputting a confidence score in the 0-1 range, achieving an objective and quantitative determination of the effectiveness of implicit rules.

[0104] In one embodiment of the present invention, the specific steps for forming bidirectional feedback are as follows:

[0105] Step S301: Configure the data acquisition interface of the LLM large model. The data acquisition interface is used to receive real-time data from equipment sensor data, order change information, material inventory change data, workstation production status data, AGV running position data and real-time quality control results from the discrete manufacturing site, forming real-time data.

[0106] The data acquisition interface of the LLM large-scale model is the technical channel for data interaction between the LLM and the discrete manufacturing site. It needs to support real-time access to multi-source data, covering various dynamic information such as equipment, orders, and inventory, and is the foundation for the model to perceive the site status. When configuring the data acquisition interface of the LLM large-scale model, the technical specifications of the interface and the interface objects must be clearly defined first: the interface must support the TCP / IP protocol to ensure data transmission stability, use JSON format for data encapsulation to ensure format uniformity, and configure a data filtering module to remove invalid and redundant data. The interface objects cover the entire data source of the discrete manufacturing site, including equipment sensors (real-time acquisition of operating parameters such as temperature, speed, and vibration frequency), order management system (pushing order change information, such as product model adjustment and early delivery date), material inventory system (synchronizing material inventory change data, such as inbound quantity and outbound workstation), workstation status monitoring module (feedback on the current production status of the workstation, such as busy / idle or waiting for materials), AGV scheduling system (transmitting AGV real-time location and task progress), and quality control equipment (uploading real-time detection results, such as dimensional deviation and surface defect judgment). After the interface is deployed, connectivity testing is required to ensure that the transmission delay of various types of data is controlled within 1 second and the data packet loss rate is less than 0.1%, so as to ensure that real-time data can be accurately and timely input into the LLM large model.

[0107] Step S302: Activate the multi-dimensional association analysis function of the LLM large model, perform cross-comparison and in-depth mining on the received real-time data, and identify implicit association rules that are not covered by the current rules of the OAG ontology library.

[0108] Multi-dimensional correlation analysis is the core capability of the LLM (Local Management Model) to uncover potential business relationships from real-time data. By cross-referencing the inherent logic of different data types, it identifies implicit relationships not covered by existing rules. Implicit correlation rules are business logic relationships in discrete manufacturing scenarios that are not currently included in the OAG (Ontology Aggregator) ontology but exist in actual production; these rules must undergo credibility verification before being included in the ontology. When activating the multi-dimensional correlation analysis function of the LLM, the analysis dimensions must be divided according to the business scenario to ensure that the direction of data mining aligns with the actual needs of discrete manufacturing. For production scenarios, cross-referencing equipment operating parameters (such as temperature and speed) with quality control results (such as pass / fail and defect type) analyzes the potential correlation between parameter anomalies and quality problems. For logistics scenarios, linking AGV real-time location, material inventory data, and workstation demand information uncovers the linkage between path congestion and insufficient inventory. For order scenarios, combining order change information with production control execution data identifies the interference patterns of order adjustments on scheduling progress. During the analysis, the LLM large model needs to call the existing OAG ontology library rules as a reference. By comparing the difference between "actual data association" and "expected existing rules", implicit association rules not covered by the OAG ontology library are screened out. For example, the association relationship "when the welding station temperature is >280℃ and the humidity is >65%, the welding defect rate increases by 15%" is found, which is not included in the existing rules.

[0109] Step S303: Calculate the confidence score of each implicit association rule using the weighted Bayes confidence algorithm and update the OAG ontology library;

[0110] First, each discovered implicit association rule is matched with its corresponding supporting data category to determine the weight coefficient for each data category (e.g., 0.9 for equipment sensor data, 0.8 for quality control data, and 0.7 for order data). Then, based on historical data, the conditional probability of each rule under different data types is calculated, and the result is substituted into the formula to obtain a confidence score. If the score is ≥0.85, the system automatically triggers the OAG ontology update process, converting the implicit rule into structured judgment logic and adding it to the logical rule ontology. At the same time, the system updates the relevant entity associations or process constraints in the element ontology or business ontology. If the score is <0.85, the system generates a manual review work order, which includes the supporting data and calculation process for the rule. Process engineers or production managers verify the actual validity of the rule. After the review is passed, the rule is manually entered into the OAG ontology to ensure the reliability of the rules.

[0111] Step S304: Open the feedback optimization channel of the OAG ontology library, transmit the updated ontology content to the semantic parsing module, logical reasoning module and decision generation module of the LLM large model, adjust the parameter configuration of the corresponding modules of the LLM large model, repeat the above steps to form a continuous iterative bidirectional feedback closed loop.

[0112] The feedback optimization channel is the technical path for transferring updated knowledge from the OAG ontology library to the LLM large model. This channel transforms structured ontology content into parameter adjustment criteria that the model can absorb, enabling iterative improvement in the model's business understanding capabilities. The continuous iterative two-way feedback loop is a cyclical mechanism that links data collection, rule mining, ontology updates, and model optimization, ensuring that the LLM and OAG ontology library can continuously evolve based on actual on-site conditions, preventing knowledge and model capabilities from lagging behind business changes. Enabling the OAG ontology library's feedback optimization channel and adjusting the LLM large model parameters requires precise transmission of updated content across modules. For example, newly added entity attributes or relationships in the element ontology are transmitted to the LLM's semantic parsing module, adjusting the entity recognition weight parameters to enable the model to more accurately identify new entities or relationships. Updated process steps or constraints in the business ontology are transmitted to the logic reasoning module, optimizing the reasoning path of the business process to ensure the model generates decisions according to the updated process logic. Supplemented judgment criteria in the logical rule ontology are transmitted to the decision generation module, adjusting the priority parameters of rule calls to ensure the model prioritizes the latest rules when making decisions. After parameter adjustments, verification using small batches of real-time data is required to observe changes in the model's semantic parsing accuracy and rule matching accuracy. If the metrics improve without abnormal fluctuations, this round of feedback optimization is complete. The above four steps are then repeated, with data collection and analysis triggered hourly and a feedback optimization report generated every 24 hours, forming a continuously iterative two-way feedback loop.

[0113] It should be noted that before generating the primary and backup decision-making schemes, business knowledge is combined with quantitative analysis of uncertain events. The process of uncertainty quantification analysis adopts the Bayesian network algorithm, and the mathematical expression is as follows:

[0114] ;

[0115] in, Indicates equipment fault F in real-time parameters Posterior probability under combined effects, such as welding equipment at temperature =300℃, rotation speed The probability of failure occurring under parameters such as 1500 r / min; These represent the real-time monitoring parameters related to equipment failure F, specifically: equipment operating temperature, speed, vibration frequency, voltage, current, and operating time. This indicates the real-time parameters assuming equipment failure F occurs. The joint likelihood that appears simultaneously This represents the prior probability of equipment failure F occurring, derived from historical failure data statistics. It is calculated as the ratio of the number of historical failures to the total operating time. For example, if the welding equipment operated for 8760 hours in the past year, with a failure duration of 262.8 hours, then... =262.8 / 8760=0.03; Indicates real-time parameters The marginal probability of simultaneous occurrence;

[0116] Uncertainty event quantification analysis refers to the process of using mathematical algorithms to calculate the probability of occurrence of events in discrete manufacturing that are uncertain in terms of whether they will occur but may affect production (such as equipment failure and material delays), transforming them into quantifiable values ​​to provide objective basis for decision-making. This differs from traditional subjective judgment and enhances the scientific rigor of decision-making. Bayesian network algorithms are graphical models based on probabilistic reasoning. They construct networks through nodes (events) and edges (relationships), combining prior probabilities with real-time data to calculate posterior probabilities. They are suitable for handling uncertainties with multivariate dependencies and can accurately quantify the probability of events occurring under the combined influence of multiple parameters in discrete manufacturing. The master decision scheme is the optimal execution method generated based on current business knowledge and low-risk uncertainty events. The primary plan is the preferred path for production and operation, and must meet the requirements of feasibility and efficiency. Alternative plans are contingency plans developed for uncertain events with a high probability of occurrence, clearly defining triggering conditions and adjustment measures to ensure rapid switching when the primary plan is disrupted, thus guaranteeing production continuity. Real-time monitoring parameters refer to equipment or environmental data related to uncertain events collected in real time at the discrete manufacturing site, such as equipment operating temperature and speed; these are the core inputs for calculating the probability of event occurrence. Prior probability is the probability of an uncertain event occurring naturally, derived from historical data statistics, reflecting the inherent trend of event occurrence. Posterior probability is the probability of event occurrence calculated by combining real-time monitoring parameters and prior probability, dynamically reflecting the actual risk of event occurrence under current operating conditions.

[0117] Based on the Bayesian network algorithm, the LLM large model quantifies the probability of occurrence of four types of uncertain events: equipment failure, material delay, order change, and process adjustment. Combined with the corresponding coping strategies in business knowledge, it generates a master decision plan that takes into account both feasibility and safety. At the same time, it formulates alternative plans for uncertain events with higher probability. The alternative plans clearly define the triggering conditions, adjustment steps, and resource allocation methods to ensure that a quick switch can be made if the master decision plan cannot be executed smoothly.

[0118] The first step requires preparing the basic data and business knowledge for the quantitative analysis of uncertainty events: First, identify the four types of uncertainty events that need to be quantified in discrete manufacturing scenarios—equipment failure, material delays, order changes, and process adjustments. For each type of event, determine the corresponding real-time monitoring parameters and historical data sources. Real-time monitoring parameters for equipment failure include equipment operating temperature. Rotation speed Vibration frequency ,Voltage Current runtime Data is collected in real time from equipment sensors; parameters corresponding to material delays, including material transit time, supplier capacity utilization, and inventory warning status, are obtained from the logistics management system and supplier management system; parameters corresponding to order changes, including order urgency and customer demand adjustment frequency, are extracted from the order management system; parameters corresponding to process adjustments, including process parameter deviations and product model switching frequency, are obtained from the production process system. Simultaneously, business knowledge in the OAG ontology library is reviewed, including response strategies for various events (such as backup equipment scheduling rules in case of equipment failure and alternative material lists for material delays) and production demand constraints (such as order delivery dates and resource load limits), to ensure that subsequent decision-making schemes conform to business logic.

[0119] The second step is to use the Bayesian network algorithm to calculate the probability of the occurrence of uncertain events. The core of this step is to substitute the data into the algorithm formula to complete the calculation.

[0120] The third step combines business knowledge with quantified probabilities to generate primary and backup decision-making schemes: First, based on production demand constraints in business knowledge (such as only 3 days left for order delivery and 60% load rate of workstation A) and low-probability uncertainty events (5% for order changes and 8% for process adjustments), a primary decision-making scheme is generated, such as "prioritize the stamping task of batch B123 to workstation A, call equipment number M01, and retrieve materials from inventory area B to ensure that production is completed within 3 days". This scheme must meet the matching of equipment capacity, material inventory, and time constraints. For uncertain events with a high probability of occurrence (equipment failure 15%, material delay 20%), alternative plans are formulated: The trigger condition for the alternative plan for equipment failure is "equipment M01 failure probability ≥ 15% or real-time parameters exceed the threshold", the adjustment steps are "switch to backup equipment M02, and coordinate material delivery delay for 1 hour", and the resource allocation method is "prioritize occupying the idle time of equipment M02 and notify the logistics department to prioritize the delivery of materials to workstation A"; The trigger condition for the alternative plan for material delay is "material delivery delay probability ≥ 20% or inventory warning", the adjustment steps are "activate alternative material C123 and modify process parameters to adapt to the alternative material", and the resource allocation method is "urgently retrieve alternative materials from inventory area C and notify the process department to provide a parameter adjustment list".

[0121] In one embodiment of the present invention, the specific steps for generating the primary / backup decision scheme are as follows:

[0122] Step S401: Collect real-time discrete manufacturing data through the data acquisition module and input it into the LLM large model in combination with historical data. The real-time data includes real-time operating parameters of equipment sensors, real-time position of AGV, production status of workstations, real-time results of quality control, order change notifications and real-time changes in material inventory. The historical data includes production control records, quality control reports, logistics control logs, equipment maintenance records and order delivery data.

[0123] The data acquisition module is a hardware and software combination that integrates real-time and historical data from discrete manufacturing. It is responsible for connecting to various data sources to achieve unified collection and transmission of multiple types of data, serving as the core carrier of decision-making data input. Real-time data consists of various types of information dynamically generated on the production site, directly reflecting the current production status and providing timely support for decision-making. Historical data is a record of data accumulated in past production operations, including information on past production control, quality control, and equipment maintenance, providing experience reference for decision-making. The data acquisition module connects to equipment sensors, AGV scheduling systems, workstation status monitoring modules, quality control equipment, order management systems, and inventory management systems, collecting real-time data such as real-time operating parameters of equipment sensors, real-time AGV positions, workstation production status, real-time results of quality control, order change notifications, and real-time changes in material inventory. At the same time, it retrieves historical data such as production control records, quality control reports, logistics control logs, equipment maintenance records, and order delivery data from the enterprise's MES, ERP, and WMS systems. After collection, the data is categorized by data type. Real-time data is transmitted through a high-speed transmission interface with a latency controlled within 1 second. Historical data, after integrity verification, is input into the LLM large model through a standardized interface to ensure that the data format is consistent, meets the model processing requirements, and avoids parsing deviations caused by format differences.

[0124] Step S402: Call the feature ontology in the OAG ontology library, start the entity association mechanism, integrate the scattered real-time data and historical data according to entity attributes and association relationships, and transform them into structured information. The structured information explicitly includes entity name, entity attribute, entity relationship and corresponding data value.

[0125] Structured information is a standardized data format formed after integration and association of OAG element ontology. It clearly includes entity names, attributes, relationships, and corresponding values, eliminating the semantic ambiguity and dispersion of the original data. The LLM large model calls the OAG element ontology, initiates the preset entity association mechanism, and integrates the scattered real-time and historical data according to the entity dimension based on the mapping relationship and association rules of the core entity attributes defined in the ontology. For example, the inventory data of batch B123 material, the corresponding stamping station A information, the operating speed and temperature data of equipment M01, and the information of the person in charge of the operation are associated to form structured information containing "material B123 - attribute (specification Φ10mm, inventory 50 pieces) - associated station (stamping A) - associated equipment (M01, speed 1500r / min, temperature 280℃) - associated personnel (employee 003)". At the same time, the data type is clearly marked as real-time or historical to avoid confusion of data with different time periods and ensure the interpretability of information.

[0126] Step S403: Call the business ontology and logical rule ontology in the OAG ontology library to make business logic judgments on the structured information. Combine the judgment results with business experience in the discrete manufacturing industry to transform the structured information into business knowledge that can be directly applied.

[0127] Business knowledge is actionable knowledge formed by integrating structured information with business logic and industry experience. It serves as a key intermediary connecting data and decision-making. The LLM large model calls upon the OAG business ontology and logical rule ontology to compare structured information with process specifications (such as production control sequence and logistics control path) in the business ontology and judgment criteria (such as equipment parameter thresholds and quality acceptance range) in the logical rule ontology. For example, it verifies whether the temperature of equipment M01 (280℃) is lower than the 300℃ threshold set in the logical rule ontology, and whether the inventory of 50 units of material B123 meets the minimum inventory requirements for scheduling in the business ontology. Combining operational experience in discrete manufacturing, such as "preferential processing of similar materials at suitable workstations," invalid judgment results are eliminated, and valid information is integrated to form business knowledge that "material B123 has sufficient inventory, equipment M01 is operating normally, stamping station A has a load rate of 60%, and this batch of materials can be prioritized for processing," ensuring that the knowledge has the practicality to directly guide decision-making.

[0128] Step S404: Based on business knowledge, start the uncertainty event analysis module, use the Bayesian network algorithm, input structured information data related to four types of uncertainty events: equipment failure, material delay, order change, and process adjustment, quantify the probability of occurrence of each type of event, obtain the quantified uncertainty event probability, obtain real-time production demand parameters through the LLM large model, and perform multi-dimensional fusion analysis by combining business knowledge and the quantified uncertainty event probability to generate fusion analysis results.

[0129] The Uncertainty Event Analysis module is a functional unit that executes Bayesian network algorithms and quantifies the probability of various random events, providing a quantitative risk basis for decision-making. The fusion analysis result is the output of a comprehensive evaluation of business knowledge, uncertainty event probabilities, and real-time production demand parameters, serving as the direct basis for generating decision-making solutions. Upon activating the Uncertainty Event Analysis module, using the Bayesian network algorithm, structured information data related to four types of events—equipment failure, material delays, order changes, and process adjustments—is input. For example, data such as the runtime of equipment M01 and past failure records are used to calculate the equipment failure probability, while the transit time of material B123 and supplier capacity data are used to calculate the material delay probability. The LLM large model synchronously acquires real-time production demand parameters, including the remaining 3 days for the delivery of order B123, the current load rate of workstation A at 60%, and the cost control target of no more than 5000 yuan per batch. It performs multi-dimensional fusion analysis of business knowledge, quantified event probabilities, and demand parameters to evaluate the effectiveness of "prioritizing the scheduling of material B123 to workstation A," predict potential risks and corresponding countermeasures, and generate a fusion analysis result including execution priority, risk level, and resource requirements.

[0130] Step S405: Based on the fusion analysis results, generate a master decision scheme that conforms to the actual production situation. At the same time, for uncertain events with a probability of occurrence greater than the threshold, formulate corresponding alternative schemes and clarify the triggering conditions, adjustment steps and resource allocation methods of the alternative schemes.

[0131] The primary decision plan is the preferred execution plan generated based on optimal efficiency and feasibility, suitable for normal production scenarios. Alternative plans are contingency plans developed for high-probability uncertainties, ensuring rapid switching to the primary plan if its execution is hindered, thus guaranteeing production continuity. Based on the fusion analysis results, combined with low-probability uncertainties (such as a 5% probability of order changes and an 8% probability of process adjustments) and production demand constraints, the primary decision plan is generated, specifying that "batch B123 materials will be prioritized for processing at stamping station A, using equipment M01, with execution time from 14:00 to 17:00 today, operated by employee 003, and materials retrieved from inventory area B," ensuring optimal efficiency and feasibility. For uncertainties with a probability greater than a preset threshold (such as 30%), such as a 15% probability of equipment M01 failure and a 20% probability of material B123 delivery delay, alternative plans are developed: the trigger condition for the alternative plan in case of equipment failure is "equipment M01 failure probability..." If the error is ≥15% or the real-time parameter exceeds the threshold, the adjustment steps are: "Switch to backup equipment M02, adjust the processing time to 15:00-18:00 today," and the resource allocation method is: "Prioritize occupying the idle time of equipment M02, and notify the logistics department to delay material delivery by 1 hour." The alternative solution for material delay is triggered by "material B123 delivery delay warning or inventory less than 30 pieces." The adjustment steps are: "Activate alternative material C123, modify the stamping process parameters to adapt to the alternative material," and the resource allocation method is: "Emergency transfer of alternative material C123 from inventory area C, and notify the process department to provide a parameter adjustment list." This ensures that the main solution can be quickly switched when it is blocked, thus ensuring production continuity.

[0132] It should be noted that the cross-scenario collaboration employs a conflict risk quantification model, the mathematical expression of which is as follows:

[0133] ;

[0134] in, This represents the risk value of cross-scenario conflict C, ranging from 0 to 100. A higher risk value indicates a more severe overall interference with business objectives by the conflict. For example... =60 indicates that conflict C poses a comprehensive risk of 60 points to the overall business objectives, requiring urgent handling; This represents the probability of a cross-scenario conflict C occurring, predicted by the LLM model based on historical conflict data and real-time business data. For example, the LLM model might count 20 instances of conflicts where "two orders simultaneously require workstation B" in the past year, with a total of 100 scheduling attempts. Combining this with current real-time data showing workstation B's load rate at 90% and emergency orders accounting for 30%, the probability of this conflict occurring in this instance is predicted. =0.3; This represents the weight coefficient for the j-th business objective, set according to the business strategy of discrete manufacturing enterprises. The weight coefficient for the order delivery period objective is set to 0.6, the weight coefficient for the equipment utilization rate objective is set to 0.2, the weight coefficient for the cost control objective is set to 0.15, and the weight coefficient for the product quality objective is set to 0.05. This represents the impact of cross-scenario conflict C on the j-th business objective. The value ranges from 0 to 10. The larger the impact value, the more serious the interference of the conflict on the business objective. j=1 corresponds to the order delivery period objective, j=2 corresponds to the equipment utilization rate objective, j=3 corresponds to the cost control objective, and j=4 corresponds to the product quality objective.

[0135] when When the value is greater than or equal to a preset threshold, the intelligent agent federation center triggers a collaborative adjustment mechanism. Based on the conflict type and scope of impact, it coordinates the intelligent agents in the relevant scenarios to adjust their execution plans, ensuring that each business objective is not severely affected by the conflict. At the same time, the collaborative adjustment results are synchronized to the LLM large model and the OAG ontology library to provide a reference for subsequent decision-making.

[0136] Cross-scenario collaboration refers to breaking down the boundaries of single business scenarios such as production, quality, and logistics in discrete manufacturing. It involves coordinating the actions of intelligent agents across scenarios through a centralized coordination node (the intelligent agent federation hub), avoiding conflicts between scenarios (such as production control adjustments not being synchronized with logistics capacity), and achieving optimal overall business objectives. The conflict risk quantification model uses mathematical formulas to transform the "probability of occurrence" and "business impact" of cross-scenario conflicts into calculable risk values. Its core is to quantify the comprehensive interference of conflicts on various business objectives, distinguishing it from traditional subjective judgment and making conflict handling more objective. The intelligent agent federation hub is the core control unit for cross-scenario collaboration, possessing data interaction, conflict identification, and... The collaborative scheduling and feedback processing functions are responsible for connecting the LLM large model with intelligent agents in various scenarios, and are the main body for triggering collaborative adjustments. The business target weight coefficient is the proportion of importance of each business target set according to the business strategy of discrete manufacturing enterprises, reflecting the enterprise's priority ranking of order delivery, equipment utilization, cost control, and product quality. The conflict impact degree is a quantitative value of the severity of the interference caused by cross-scenario conflicts to a certain business target, directly reflecting the degree of obstruction of the conflict to the achievement of the business target. The preset threshold is the critical standard for determining whether to trigger collaborative adjustments. It needs to be set in combination with the enterprise's tolerance for conflict to ensure that adjustments are only initiated for high-risk conflicts, balancing adjustment costs and business stability.

[0137] The first step is to calculate the risk value of cross-scenario conflicts: The conflict identification module of the intelligent agent federation hub extracts the execution data (such as the order scheduling plan of the production control intelligent agent and the AGV capacity data of the logistics control intelligent agent) and status information (such as workstation load rate and material inventory status) of each scenario-based intelligent agent from the data interaction module, and inputs this data into the LLM large model; the LLM large model calls the business logic (such as scheduling priority rules and logistics path constraints) in the OAG ontology library, and combines it with historical conflict data (such as the frequency of occurrence and handling results of similar conflicts in the past year) to predict the probability of occurrence of cross-scenario conflict C. Meanwhile, the conflict identification module assesses the impact of conflict C on each business objective based on the type of conflict C (such as resource contention or process conflict) and its relevance to each business objective. For example, the conflict of "insufficient AGV capacity leading to material delivery delays" has a high impact on order delivery time targets but a low impact on quality targets; finally, by substituting into the conflict risk quantification model formula, the following can be calculated: If substituted =0.3、 =0.6、 =8、 =0.2、 =6、 =0.15、 =0.05、 =0d =0.3×(0.6×8+0.2×6+0.15×4+0.05×0)=0.3×(4.8+1.2+0.6+0)=0.3×6.6=1.98.

[0138] The second step requires determining the risk value and triggering collaborative adjustments: The intelligent agent federation central control presets a risk value threshold (usually set to 5, but can be adjusted according to enterprise needs), and calculates the... Compare with the threshold; if ≥5, such as a conflict =7.2, triggering the collaborative adjustment mechanism; the collaborative scheduling module, based on the conflict type (e.g., resource contention, process conflict) and scope of impact (e.g., involving production and logistics scenarios), calls the collaborative rules in the OAG ontology library (e.g., "allocate resources according to order priority when there is a resource conflict" and "optimize execution order when there is a process conflict") to formulate specific adjustment strategies. For example, in the conflict of "two orders simultaneously needing workstation B", the adjustment strategy is "prioritize arranging urgent orders to workstation B, adjust regular orders to backup workstation C, and simultaneously notify the logistics management intelligent agent to adjust the material delivery route to workstation C"; if <5, as mentioned above =1.98, indicating that the impact of the conflict is small and there is no need to trigger coordinated adjustments. Only the conflict information is recorded for subsequent analysis.

[0139] The third step requires implementing adjustments and synchronizing the results: The collaborative scheduling module distributes adjustment strategies to relevant scenario-based intelligent agents through interactive interfaces, such as the production control intelligent agent adjusting order workstation allocation and the logistics control intelligent agent optimizing AGV paths. After receiving the adjustment strategies, the relevant intelligent agents update their execution plans in real time and execute them. For example, the production control intelligent agent adjusts regular orders from workstation B to workstation C, and the logistics control intelligent agent changes the delivery target of AGV-05 from workstation B to workstation C. At the same time, the intelligent agent federation hub synchronizes the collaborative adjustment results (such as the adjusted scheduling plan and logistics path) to the LLM large model and OAG ontology library through a dedicated channel. The LLM large model uses the adjustment results as a reference for subsequent decisions and optimizes the decision generation logic. The OAG ontology library transforms effective adjustment strategies into new collaborative rules (such as "urgent orders prioritize occupying bottleneck workstations"), supplementing them to the business ontology or logical rule ontology and improving cross-scenario collaborative knowledge.

[0140] In one embodiment of the present invention, the specific steps of the end-to-end data closed loop are as follows:

[0141] Step S501: Build the intelligent agent federation hub, configure the data interaction module, conflict identification module, collaborative scheduling module, and feedback processing module, and clarify the functional responsibilities and data interaction protocols of each module;

[0142] The Intelligent Agent Federation Hub is the core control unit for cross-scenario collaboration in discrete manufacturing. It integrates four functional modules: data interaction, conflict identification, collaborative scheduling, and feedback processing. It connects the LLM (Limited Least Model) big data model with scenario-based intelligent agents, enabling the issuance of decision commands, collection of execution data, collaborative conflict handling, and feedback data transmission. It serves as the "central brain" ensuring seamless data flow and business collaboration throughout the entire process. The data interaction module acts as the "data entry and exit point" of the Intelligent Agent Federation Hub, receiving decision schemes output by the LLM big data model and collecting execution data and status information from various scenario-based intelligent agents. Standardized protocols ensure the stability and format consistency of data transmission. The conflict identification module acts as a "conflict early warning system," identifying potential cross-scenario conflicts from execution data based on the business logic and conflict risk quantification model of the OAG (Ontology Aggregator) library, preventing conflicts from escalating and impacting business objectives. The collaborative scheduling module acts as a "conflict resolver," developing specific adjustment strategies for identified conflicts and coordinating relevant intelligent agents to adjust their execution plans, ensuring smooth business processes. The feedback processing module acts as a "data processor and transmitter," standardizing execution data and conflict handling results before transmitting them back to the LLM big data model and the OAG library, providing data support for system optimization. In building the intelligent agent federation hub, it is necessary to first clarify the functional responsibilities and data interaction protocols of each module: The data interaction module must support bidirectional communication with the LLM large model and scenario-based intelligent agents, use the TCP / IP protocol to ensure data transmission stability, unify the data format as JSON, and define data fields including scheme ID, intelligent agent ID, data type, content, and timestamp to ensure that data can be parsed between different systems; The conflict identification module must preset the data extraction frequency (e.g., extract execution data from the data interaction module once every 30 seconds), clarify the OAG ontology logic to be combined (e.g., production control priority, logistics path constraints) and conflict risk quantification model parameters (e.g., risk value threshold 5); The collaborative scheduling module must establish a mapping relationship between conflict types and adjustment strategies, such as resource contention conflicts corresponding to the "allocate resources according to order priority" strategy, and process conflict conflicts corresponding to the "optimize execution order" strategy; The feedback processing module must formulate data standardization rules, such as uniformly using "normal / abnormal / idle" to describe the "equipment status" in the execution data, and the conflict handling results must include conflict ID, handling measures, handling time, and scope of impact. After the module configuration is completed, joint debugging tests are required to verify whether the data flow between the modules is smooth. For example, whether the decision scheme output by the LLM can be accurately sent to the agent through the data interaction module to ensure that the central function is normal.

[0143] Step S502: Deploy scenario-based intelligent agents, including production control intelligent agents, quality control intelligent agents, and logistics control intelligent agents. Configure each intelligent agent with an independent execution module, data receiving interface, and feedback interface, and clarify the functional boundaries and interaction permissions of each intelligent agent.

[0144] Scenario-based intelligent agents are execution units corresponding to specific business scenarios, including three categories: production control, quality control, and logistics control. They possess independent execution and data interaction capabilities, responsible for translating decision-making plans into actual business actions and providing feedback on execution status. When deploying scenario-based intelligent agents, separate configurations are required for the three scenarios: production control, quality control, and logistics. The execution module of the production control intelligent agent needs to interface with the enterprise's MES system, possessing permissions to read equipment capacity data, material supply data, and order priority information, and be able to generate detailed scheduling plans including workstations, tasks, time, and resources. The execution module of the quality control intelligent agent needs to interface with quality inspection equipment (such as visual inspection cameras and dimensional measuring instruments), capable of acquiring inspection data and determining pass / fail status, while also linking to the exception handling process in the OAG ontology library. The execution module of the logistics control intelligent agent needs to interface with the AGV scheduling system and WMS system, capable of reading AGV locations, shelf inventory, and workstation requirements, and generating delivery routes and scheduling instructions. Each agent is configured with an independent data receiving interface and feedback interface, and the interface protocol is consistent with the data interaction module of the federation center to avoid communication compatibility issues; the functional boundaries of each agent are clearly defined, such as the production control agent being only responsible for generating and adjusting the scheduling plan and not interfering with the quality control process; interaction permissions are set, such as agents can only upload their own execution data to the federation center and cannot read the core data of other agents (such as the original images of quality control), to ensure data security and operational standards.

[0145] Step S503: Based on the primary / backup decision-making scheme, a federated intelligence agent hub is built, which decomposes the scheme according to three business scenarios: production, quality, and logistics. The schemes are then distributed to the corresponding scenario-based intelligence agents through dedicated interfaces. The scenario-based intelligence agents transmit real-time execution data and operational status information to the data interaction module of the federated intelligence agent hub via feedback interfaces. The execution module is the "action execution core" of the scenario-based intelligence agent, generating specific operation instructions based on the decision scheme. For example, the execution module of the production control intelligence agent generates production task lists for each workstation. The data receiving interface and feedback interface are the "communication channels" between the intelligence agent and the federated hub, receiving decision schemes and adjustment strategies from the hub and providing feedback on real-time execution data and status information. Functional boundaries and interaction permissions are the "constraints" that regulate the operational scope of the intelligence agents, preventing overlapping functions or unauthorized operations by different intelligence agents. For example, the production control intelligence agent cannot directly modify the logistics control path. During the solution distribution and data collection phase, the intelligent agent federation hub first breaks down the primary and backup decision solutions output by the LLM into scenarios: the production scenario solution includes scheduling plans, equipment allocation, and order priorities, and is distributed to the production control intelligent agent; the quality scenario solution includes inspection standards, inspection frequency, and anomaly handling thresholds, and is distributed to the quality control intelligent agent; the logistics scenario solution includes delivery routes, time nodes, and AGV allocation, and is distributed to the logistics control intelligent agent. During the breakdown, a dedicated interface must be used for distribution, with each interface corresponding to a unique intelligent agent to avoid incorrect solution distribution. After receiving the solution, the scenario-based intelligent agent executes the corresponding business actions and simultaneously uploads execution data and status information in real time through the feedback interface. For example, the production control intelligent agent uploads workstation load rate and task completion progress every 10 minutes, and the logistics control intelligent agent uploads AGV location and material delivery status every 5 minutes, ensuring that the federation hub has real-time control over the execution dynamics.

[0146] Step S504: The conflict identification module extracts the execution data and status information of each agent from the data interaction module, and calculates the risk value of potential conflicts in real time by combining the business logic conflict risk quantification model in the OAG ontology library. When the conflict risk value is greater than the threshold, the collaborative scheduling module formulates a special adjustment strategy for the conflict type and scope of impact, and sends it to the relevant scenario-based agents through the interaction interface. After receiving the adjustment strategy, the relevant agents adjust the execution plan in real time and continue to execute, and output the final execution data.

[0147] In the conflict resolution phase, the conflict identification module extracts the execution data and status information of each agent from the data interaction module at a preset frequency. Combining this with the business logic of the OAG ontology (e.g., "AGVs must deliver materials 30 minutes before the start of the production task") and the conflict risk quantification model, it calculates the risk value of potential conflicts. If the risk value exceeds a threshold (e.g., 5), the collaborative scheduling module, based on the conflict type (e.g., "AGV delays causing materials not to be delivered on time" falls under logistics-production conflict) and the scope of impact (involving stamping station A and AGV-05), retrieves a preset adjustment strategy (e.g., "prioritize scheduling idle AGV-06 for material delivery, and simultaneously notify the production control agent to delay the production task at this station by 15 minutes"), and distributes it to the relevant agents through the interaction interface. After receiving the adjustment strategy, the relevant agents update the execution plan in real time. For example, the AGV scheduling agent replans the path of AGV-06, and the production control agent adjusts the task time of stamping station A. After execution, the final execution data (e.g., actual material delivery time, and the adjusted completion time of the production task) is output.

[0148] Step S505: The feedback processing module standardizes and organizes the final execution data and conflict resolution results of each agent to form feedback data, which is then sent back to the LLM large model and the OAG ontology library through dedicated feedback channels. After receiving the feedback data, the LLM large model optimizes the parameter configuration of its own decision generation module. After receiving the feedback data, the OAG ontology library updates the corresponding business rules and logical judgment criteria.

[0149] Feedback data consists of standardized execution results and conflict resolution records, including information such as execution progress, anomalies, and adjustment measures. It serves as the core basis for LLM large-scale model optimization and OAG ontology library updates. During the feedback optimization phase, the feedback processing module standardizes and organizes the final execution data and conflict resolution results of each agent: unifying data formats (e.g., organizing information such as "production task completion rate 95%" and "AGV delivery delay 10 minutes" into "agent ID-scheme ID-execution indicator-value-timestamp" fields), removing invalid data (such as records with incorrect formats), and forming structured feedback data. This data is then transmitted back through dedicated feedback channels: one channel connects to the LLM large-scale model, where feedback data is used to optimize the parameter configuration of the decision generation module (e.g., adjusting the model's predictive weight for AGV capacity in logistics management decisions based on data showing frequent AGV delays); the other channel connects to the OAG ontology library, where feedback data is used to update corresponding business rules and logical judgment standards (e.g., adding "avoid peak hours on path 1" to the business ontology process constraints of logistics management based on the handling result of "AGV-05 frequently congested on path 1"). After receiving feedback data, the LLM large model and OAG ontology library automatically trigger the optimization and update process to ensure that the next round of decisions can be generated based on the latest business experience and data patterns, forming a closed-loop data process of "solution issuance - execution - conflict handling - feedback optimization".

[0150] Based on feedback data, the OAG ontology library implements layered updates. The element ontology update specifically involves: filtering new entities not currently covered by the element ontology (such as newly added equipment models or special material types), entity attribute change records (such as equipment parameter threshold adjustments), and entity relationship correction data from the structured feedback data; calculating the confidence level of newly added information using a weighted Bayes confidence algorithm; automatically adding information with a confidence level ≥ 0.85 to the element ontology; and updating the entity list, attribute definitions, and association rules. The business ontology update specifically involves: analyzing business process execution deviation records in the feedback data, collaboratively adjusting effective paths, and identifying missing process steps and unreasonable execution sequences in the current business ontology. For incomplete constraints, based on the actual needs of discrete manufacturing operations, the definitions of business actions are revised, process connection logic is supplemented, and the expression of constraints is optimized to ensure that the business entity is consistent with the actual execution process. The update of the logical rule entity specifically involves: extracting rule conflict cases, effective solutions for exception handling, and reasons for decision failure from feedback data; identifying loopholes in existing logical rules (such as unreasonable threshold settings or missing judgment conditions); transforming new rules (such as new logic for exception handling) and rule correction solutions (such as threshold calibration values) that have been proven effective in practice into structured judgment logic; updating the logical rule entity after passing confidence verification (confidence ≥ 0.9); and simultaneously deleting old rules that have been proven invalid.

[0151] The optimization of the LLM large-scale model is carried out at the module level, specifically including the optimization of the semantic parsing module, the logical reasoning module, and the decision generation module. The semantic parsing module optimization involves: extracting semantic parsing error cases (such as business terminology recognition deviations and entity association errors) from the preprocessed feedback data, constructing a dedicated optimization dataset, inputting this dataset into the semantic parsing module of the LLM large-scale model, adjusting the module's word vector mapping parameters and entity recognition weights, and retraining the semantic understanding model. The logical reasoning module optimization involves: filtering cases of decision logic deviations and rule matching errors from the feedback data, combining them with the updated OAG logical rule ontology to generate inference optimization training data, fine-tuning the inference layer parameters of the LLM large-scale model, strengthening the model's ability to prioritize business rules, and reducing the bias of data statistical inference. The decision generation module optimization involves: collecting feedback data on decision solutions that did not meet expectations (such as scheduling delays, cost overruns, and repeated collaboration conflicts), analyzing problems such as unreasonable allocation of decision target weights and insufficient prediction of uncertain events, constructing a decision optimization training set based on real-time production demand parameters, adjusting the objective function weights and probability fusion algorithm parameters of the model's decision generation module, and improving the feasibility and adaptability of the primary and backup solutions.

[0152] It should be noted that after LLM large model optimization and OAG ontology library update, the following steps are also included:

[0153] The updated OAG ontology library is matched with historical business data and real-time data to verify the entity coverage completeness of the feature ontology, the process adaptation accuracy of the business ontology, and the conflict rate of the logical rule ontology. Ontologies that pass the verification are updated, while ontologies that fail the verification remain unchanged.

[0154] The first step, verification and updating of the OAG ontology, needs to be carried out in modules: First, the updated OAG ontology is matched with historical business data and real-time data in all dimensions to verify the entity coverage completeness of the element ontology—extract all entities (such as equipment model, material type, workstation number) from historical business data and real-time data, compare them with the entity list in the element ontology, and count the proportion of uncovered entities. If the proportion is ≤2%, the entity coverage completeness is considered to meet the standard. Second, verify the process adaptation accuracy of the business ontology—select typical business processes (such as order scheduling-production-quality inspection-logistics) from historical business data, simulate the execution of the updated business ontology process, and count the proportion of process steps that are consistent with the actual execution. If the proportion is ≥95%, the process adaptation accuracy is considered to meet the standard. Third, verify the conflict occurrence rate of the logical rule ontology—input real-time data into the logical rule ontology and count the number of times contradictions (such as two rules making opposite judgments on the same event) or inapplicability (such as no corresponding rule to judge a certain type of anomaly) occur during rule application. If the conflict occurrence rate is ≤1%, it is considered to meet the standard. After verification, each ontology module is processed separately: for the parts of the element ontology, business ontology, and logical rule ontology that have passed verification, the old version is officially replaced and the update is completed; for the parts that have not passed verification (such as the element ontology missing a new type of material entity, or a process step in the business ontology not matching the actual situation), the original version is kept unchanged, and the reasons for not passing verification are recorded (such as insufficient feedback data, deviation in rule extraction), and the verification is readjusted after more valid data is collected in the future.

[0155] Connect the optimized LLM model to the test environment, input feedback data and historical test sets, and evaluate the semantic parsing accuracy, logical reasoning matching accuracy, and decision scheme compliance rate. If all three indicators meet the requirements, the optimization is complete; otherwise, keep the original version of the LLM model.

[0156] The second step, validation and optimization of the LLM large model, requires setting up a dedicated testing environment and conducting performance evaluation: First, set up a testing environment with the same configuration as the production environment, including the same hardware computing power (such as GPU model and memory capacity) and software dependencies (such as deep learning framework version and data processing library), to ensure that the test results are consistent with the actual application scenario; input the above-generated feedback data (execution data of each agent and conflict handling results) and historical test sets (standardized data used for model evaluation in the past, including semantic parsing samples, logical reasoning cases, and decision scheme evaluation samples) into the optimized LLM large model; start the model performance evaluation process, and calculate the semantic parsing accuracy (number of samples whose semantic parsing results are consistent with the standard results / total number of samples × 100%), logical reasoning matching accuracy (number of samples whose logical reasoning conclusions are consistent with the business rules / total number of samples × 100%), and decision scheme compliance rate (number of samples whose decision schemes meet production needs, resource constraints, and risk control / total number of samples × 100%). If all three indicators meet the preset requirements (usually semantic parsing accuracy ≥92%, logical reasoning matching accuracy ≥90%, and decision solution compliance rate ≥88%), the model optimization is deemed effective, and the old version is officially replaced and put into production. If any indicator fails to meet the requirements (e.g., the decision solution compliance rate is only 85%), the original version of the LLM model is maintained, and the reasons for failure are analyzed (e.g., insufficient data for a certain type of scenario in the feedback data, or excessive parameter adjustment). The process returns to the "Feedback Data Optimize Model" step, and after readjusting the parameters or supplementing the data, this step is executed again for verification until the indicators meet the requirements.

[0157] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for intelligent decision-making and business collaboration based on OAG ontology and LLM large model, characterized in that, Includes the following steps: Collect historical data and business documents from the discrete manufacturing field, and fine-tune the basic large language model to form an LLM large model adapted to discrete manufacturing business. LLM large model is used to parse raw data and business documents of discrete manufacturing scenarios, and an OAG ontology library containing element ontology, business ontology and logical rule ontology is constructed. The LLM large model receives real-time field data, mines implicit rules to update the OAG ontology library, and at the same time the OAG ontology library feeds back to optimize the LLM large model, forming a two-way feedback. Real-time and historical data are integrated into the LLM large model, and then transformed into business knowledge through the OAG ontology library to generate primary and backup decision-making schemes; Establish a federal intelligent agent hub to issue decision-making schemes, collect execution data and feed it back to the LLM large model to achieve cross-scenario collaboration and full-process data closed loop; After uncovering implicit rules, a weighted Bayesian confidence algorithm is used to calculate the confidence level of the implicit rules. Rules that meet the confidence threshold are pushed to the OAG ontology library for automatic updates, while rules with low confidence are manually reviewed and updated. The mathematical expression of the weighted Bayesian confidence algorithm is as follows: ， in, This represents the confidence score of the latent rule R obtained through mining, with a value ranging from 0 to 1. Let i represent the i-th type of data that supports the implicit rule R. Indicates based on the i-th type of data The conditional probability of determining that the implicit rule R is true. Represents the i-th type of data The weighting coefficients are set based on the reliability of the data source, where n represents the total number of data types that support the implicit rule R. When the value is greater than or equal to a set threshold, the implicit rule R is automatically pushed to the OAG ontology library to complete the update. When the value is less than the set threshold, a manual review process is triggered. Once the implicit rule R is confirmed to be valid after review, it is then updated to the OAG ontology library. The specific steps for forming bidirectional feedback are as follows: Configure the data acquisition interface of the LLM large model to receive real-time data from equipment sensor data, order change information, material inventory change data, workstation production status data, AGV running position data and real-time quality control results from the discrete manufacturing site, forming real-time data. Activate the multi-dimensional association analysis function of the LLM large model to perform cross-comparison and in-depth mining on the received real-time data, and identify implicit association rules that are not covered by the current rules of the OAG ontology library. The confidence score of each implicit association rule is calculated using a weighted Bayes confidence algorithm, and the OAG ontology is updated. Enable the feedback optimization channel of the OAG ontology library, transfer the updated ontology content to the semantic parsing module, logical reasoning module and decision generation module of the LLM large model, adjust the parameter configuration of the corresponding modules of the LLM large model, repeat the above steps, and form a continuous iterative two-way feedback closed loop.

2. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 1, characterized in that, The specific steps for targeted fine-tuning of the basic large language model are as follows: Collect historical production data, quality control records, logistics control logs, business process documents, industry rule manuals, equipment maintenance manuals, order contract data, and bill of materials data in the discrete manufacturing field, and form standardized data after preprocessing; Based on standardized data, the dataset is divided into training, validation, and test sets in an 8:1:1 ratio. Select a basic large language model, build a supervised fine-tuning framework, associate the training set data with the corresponding business labels and input them into the fine-tuning framework, start model fine-tuning training, and adjust the model's attention mechanism weights and fully connected layer parameters through the backpropagation algorithm; During the model fine-tuning training process, after every 10 rounds of parameter iteration, the validation set data is input into the model in the current training state to obtain the semantic parsing results, business rule matching results, and preliminary decision-making schemes output by the model, and to calculate the semantic parsing accuracy, business rule matching accuracy, and the rationality of the decision-making scheme. When the semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality all meet the requirements, stop the model fine-tuning training and obtain the preliminary fine-tuned model; otherwise, adjust the parameters and continue iterative training. Based on the initial fine-tuning of the model, input test set data, output the final output results, and determine whether the semantic parsing accuracy, business rule matching accuracy, and decision scheme rationality meet the requirements. If they meet the requirements, complete the targeted fine-tuning and output the LLM large model. If they do not meet the requirements, repeat the above steps until the model performance meets the requirements.

3. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 1, characterized in that, The specific steps for constructing the OAG ontology library are as follows: Based on the LLM model, the original data and business documents of the discrete manufacturing scenario are imported, and four core entities, namely equipment, materials, workstations and personnel, are extracted. The specific attributes of the core entities and the relationships between the core entities are clarified. Based on core entities, specific attributes, and relationships, an element ontology is established, and semantic ambiguity from different data sources is eliminated through a unified data semantic definition. Based on business documents, we break down three core business actions: production control, quality control, and logistics control. We then analyze the execution process of each action, the logical relationships between actions, and the corresponding business rules. We establish a business ontology, standardize the expression of business processes, and clarify the execution order and constraints of business actions. Based on raw data and business documents from discrete manufacturing scenarios, we extract three types of industry experience accumulated in the discrete manufacturing industry: process parameter thresholds, quality anomaly judgment standards, and safe operation specifications. These are then transformed into structured judgment logic that can be directly parsed by the LLM big model, establishing a logical rule ontology and forming a clear basis for business decision-making. By associating the element ontology, business ontology, and logical rule ontology, a complete OAG ontology library is formed.

4. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 1, characterized in that, Before generating the primary / backup decision scheme, business knowledge is combined with quantitative analysis of uncertain events. The process of uncertainty quantification analysis adopts a Bayesian network algorithm, and the mathematical expression is as follows: ， in, Indicates equipment fault F in real-time parameters The posterior probability under the combined effect These represent the real-time monitoring parameters related to equipment failure F, specifically: equipment operating temperature, speed, vibration frequency, voltage, current, and operating time. This indicates the real-time parameters assuming equipment failure F occurs. The joint likelihood that appears simultaneously This represents the prior probability of equipment failure F occurring, derived from statistical analysis of historical failure data. Indicates real-time parameters The marginal probability of simultaneous occurrence; Based on the Bayesian network algorithm, the LLM large model quantifies the probability of occurrence of four types of uncertain events: equipment failure, material delay, order change, and process adjustment. Combined with the corresponding coping strategies in business knowledge, it generates a master decision plan that takes into account both feasibility and safety. At the same time, it formulates alternative plans for uncertain events with higher probability. The alternative plans clearly define the triggering conditions, adjustment steps, and resource allocation methods to ensure that a quick switch can be made if the master decision plan cannot be executed smoothly.

5. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 4, characterized in that, The specific steps for generating the primary and backup decision schemes are as follows: Real-time data from discrete manufacturing is collected through a data acquisition module and combined with historical data to input into the LLM large model. The real-time data includes real-time operating parameters of equipment sensors, real-time position of AGVs, production status of workstations, real-time results of quality control, order change notifications, and real-time changes in material inventory. The historical data includes production control records, quality control reports, logistics control logs, equipment maintenance records, and order delivery data. The feature ontology in the OAG ontology library is invoked, and the entity association mechanism is activated to integrate the scattered real-time data and historical data according to entity attributes and association relationships, and transform them into structured information. The structured information explicitly includes entity name, entity attribute, entity relationship and corresponding data value. The business ontology and logical rule ontology in the OAG ontology library are called to perform business logic judgment on structured information. The judgment results are combined with business experience in the discrete manufacturing industry to transform the structured information into business knowledge that can be directly applied. Based on business knowledge, the uncertainty event analysis module is launched. Using the Bayesian network algorithm, structured information data related to four types of uncertainty events—equipment failure, material delay, order change, and process adjustment—are input to quantify the probability of occurrence of each type of event, resulting in quantified uncertainty event probabilities. Real-time production demand parameters are obtained through the LLM large model, and multi-dimensional fusion analysis is performed by combining business knowledge and quantified uncertainty event probabilities to generate fusion analysis results. Based on the results of the fusion analysis, a master decision-making scheme that conforms to the actual production situation is generated. At the same time, for uncertain events with a probability of occurrence greater than the threshold, corresponding alternative schemes are formulated, and the triggering conditions, adjustment steps and resource allocation methods of the alternative schemes are clarified.

6. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 1, characterized in that, The cross-scenario collaboration employs a conflict risk quantification model, the mathematical expression of which is as follows: ， in, This represents the risk value C for cross-scenario conflict, ranging from 0 to 100. The probability of cross-scenario conflict C is represented by the prediction made by the LLM model based on historical conflict data and real-time business data. This represents the weight coefficient for the j-th business objective, set according to the business strategy of discrete manufacturing enterprises. The weight coefficients are: order delivery time objective (0.6), equipment utilization rate objective (0.2), cost control objective (0.15), and product quality objective (0.05). This represents the impact of cross-scenario conflict C on the j-th business objective. The value ranges from 0 to 10. The larger the impact value, the more serious the interference of the conflict on the business objective. j=1 corresponds to the order delivery period objective, j=2 corresponds to the equipment utilization rate objective, j=3 corresponds to the cost control objective, and j=4 corresponds to the product quality objective. when When the value is greater than or equal to a preset threshold, the intelligent agent federation center triggers a collaborative adjustment mechanism. Based on the conflict type and scope of impact, it coordinates the intelligent agents in the relevant scenarios to adjust their execution plans, ensuring that each business objective is not severely affected by the conflict. At the same time, the collaborative adjustment results are synchronized to the LLM large model and the OAG ontology library to provide a reference for subsequent decision-making.

7. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 6, characterized in that, The specific steps of the full-process data closed loop are as follows: Establish a federal hub for intelligent agents, configure data interaction modules, conflict identification modules, collaborative scheduling modules, and feedback processing modules, and clarify the functional responsibilities and data interaction protocols of each module; Deploy scenario-based intelligent agents, including production control intelligent agents, quality control intelligent agents, and logistics control intelligent agents. Configure each intelligent agent with an independent execution module, data receiving interface, and feedback interface, and clarify the functional boundaries and interaction permissions of each intelligent agent. Based on the primary and backup decision-making scheme, the intelligent agent federation hub is built to break down the scheme into three business scenarios: production, quality, and logistics. The scheme is then distributed to the corresponding scenario-based intelligent agents through dedicated interfaces. The scenario-based intelligent agents transmit real-time execution data and running status information to the data interaction module of the intelligent agent federation hub through feedback interfaces. The conflict identification module extracts the execution data and status information of each agent from the data interaction module. Combined with the business logic conflict risk quantification model in the OAG ontology library, it calculates the risk value of potential conflicts in real time. When the conflict risk value is greater than the threshold, the collaborative scheduling module formulates a special adjustment strategy for the conflict type and scope of impact. The strategy is then sent to the relevant scenario-based agents through the interaction interface. After receiving the adjustment strategy, the relevant agents adjust the execution plan in real time and continue to execute, outputting the final execution data. The feedback processing module standardizes and organizes the final execution data and conflict resolution results of each agent to form feedback data, which is then sent back to the LLM large model and the OAG ontology library through dedicated feedback channels. After receiving the feedback data, the LLM large model optimizes the parameter configuration of its own decision generation module, and after receiving the feedback data, the OAG ontology library updates the corresponding business rules and logical judgment criteria.

8. The intelligent decision-making and business collaboration method based on OAG ontology and LLM large model according to claim 7, characterized in that, Following LLM large model optimization and OAG ontology update, the following steps are also included: The updated OAG ontology library is matched with historical business data and real-time data to verify the entity coverage completeness of the feature ontology, the process adaptation accuracy of the business ontology, and the conflict rate of the logical rule ontology. Ontologies that pass the verification are updated, while ontologies that fail the verification remain unchanged. Connect the optimized LLM model to the test environment, input feedback data and historical test sets, and evaluate the semantic parsing accuracy, logical reasoning matching accuracy, and decision scheme compliance rate. If all three indicators meet the requirements, the optimization is complete; otherwise, keep the original version of the LLM model.

Images