Intelligent management system and method based on multi-agent cooperation

Abstract

The invention provides an intelligent management system and method based on multi-agent cooperation. The intelligent management system comprises a data acquisition module used for acquiring production field data in real time; the data processing module is used for data cleaning; the large model module is used for performing semantic analysis on a task input by a user based on a large model; the agent module is used for performing task scheduling based on a TSQlearning scheduling algorithm; and the RAG knowledge base module is used for constructing an RAG knowledge base and providing knowledge retrieval based on hierarchical indexing and a mixed retrieval mechanism. According to the method, user requirements are analyzed through the large model module, the task comprehensive scheduling capability is improved through the distributed decision of the agent module and the TSQlearning algorithm, the resource utilization rate is improved, the information retrieval accuracy is improved from 70% to 92% through the hierarchical indexing and mixed retrieval mechanism of the RAG knowledge base, and the management efficiency and the intelligent level of the smart factory are improved.

Description

Technical Field

[0001] The present invention relates to the field of smart factory technology, and in particular to a smart management system and method based on multi-agent collaboration. Background Art

[0002] A smart factory is a factory model that uses high-tech means such as big data and artificial intelligence to optimize production efficiency and quality. In a smart factory, by comprehensively monitoring and coordinating various data indicators in the production process, real-time early warning and decision-making are carried out, thereby maximizing production efficiency and quality. Most companies use a multi-system coordination approach to implement smart factory operations. In this case, although production efficiency can be improved through technologies such as big data, systems such as equipment monitoring (MES), planning and scheduling (APS), and warehouse management (WMS) operate independently, with high data exchange latency. The lack of a unified data bus makes unified scheduling impossible, and various systems are difficult to coordinate and cooperate. Scheduling conflicts make efficient processing and management difficult. In addition, the processing of process parameters and fault anomalies in the factory production process relies on manual experience, and the training cycle for new employees is long and inefficient.

[0003] In the existing technology, MOM systems are mostly used for comprehensive system management. However, the MOM system adopts a vertical layered architecture (equipment layer → control layer → workshop layer → enterprise layer). Data needs to be transmitted layer by layer. It takes minutes or even hours for events to be transmitted from the device end to the decision-making layer, which easily misses the best response window. Data is synchronized between systems through timed batch processing. When orders change suddenly, the planning and scheduling modules cannot obtain inventory changes in real time, resulting in a disconnect between production plans and actual conditions. All decisions rely on central server processing, and high concurrency and latency are high. In addition, traditional scheduling algorithms, such as MRP-based production scheduling, rely on historical data training and cannot cope with real-time disturbances. In emergencies, it takes a long time to reschedule plans, resulting in production capacity loss. Summary of the Invention

[0004] The present invention is made in view of the above-mentioned problems, and its purpose is to provide an intelligent management system and method based on multi-agent collaboration, analyze user needs through a large model module, improve the comprehensive task scheduling capability and resource utilization through the distributed decision-making and TS_Qlearning algorithm of the intelligent agent module, and increase the information retrieval accuracy from 70% to 92% through the hierarchical indexing and hybrid retrieval mechanism of the RAG knowledge base, thereby improving the management efficiency and intelligence level of the smart factory.

[0005] Specifically, the first aspect of the present invention provides an intelligent management system based on multi-agent collaboration, including:

[0006] Data acquisition module: used to collect factory production site data in real time and send it to the data processing module;

[0007] Data processing module: used to clean the received data and send it to the RAG knowledge base module;

[0008] Large model module: performs semantic analysis on the tasks input by the user based on the large model and queries the RAG knowledge base module and / or the scheduling agent module;

[0009] Agent module: includes the system agent and several agents responsible for processing specific tasks of their own, and performs task scheduling based on the TS_Qlearning scheduling algorithm;

[0010] RAG knowledge base module: builds the RAG knowledge base and provides knowledge retrieval based on hierarchical indexing and hybrid retrieval mechanisms.

[0011] This invention implements intelligent manufacturing management through a three-tier architecture: the data layer integrates a variety of factory information, including real-time sensor data, equipment status, inventory status, and production orders, through a data acquisition module; the intermediate processing layer uses an agent module to perform distributed and integrated decision-making, and provides knowledge support in conjunction with the RAG knowledge base; and the application layer uses a large model module to parse user-input tasks based on a large-model-driven natural language interactive interface, enabling intelligent management and control of the entire production process. This architecture effectively addresses the response lag, knowledge fragmentation, and decision-making rigidity issues of traditional MOM systems, significantly improving the adaptability and intelligence of production systems.

[0012] The core technical advantages of this invention are reflected in three aspects: First, the hierarchical indexing mechanism of the RAG knowledge base (abstract index + document block index) is used to achieve efficient retrieval of manufacturing knowledge, which is more than 50% faster than traditional vector retrieval; second, a multi-agent collaborative architecture is adopted, and each professional agent (production scheduling, quality control, equipment maintenance, etc.) realizes parallel processing of complex problems through distributed decision-making, and then makes comprehensive decisions through the system agent; finally, based on the large model, the tasks input by the user are parsed and production tasks are generated. It can dynamically generate professional knowledge such as equipment parameter tuning plans and exception handling plans, significantly reducing dependence on human experts.

[0013] The RAG knowledge base of the present invention includes equipment historical operation data, warehouse information data, technical solution data, customer information data, work log information data, fault condition data, process documents and maintenance manuals.

[0014] The RAG knowledge base of this invention uses a hierarchical indexing architecture, constructing a two-tier index structure: a top-level summary index and a bottom-level document block index. When processing a user query, the system first quickly filters the relevant document range using the summary index, then performs precise document block-level retrieval within the limited range. This mechanism significantly reduces the number of vectors that need to be compared, making it particularly suitable for efficient retrieval of massive amounts of technical documents (such as equipment manuals, process standards, and SOP files) in factory scenarios.

[0015] The hierarchical index includes:

[0016] Primary index: entity classification (including equipment, process, quality standards, customers, and warehousing);

[0017] Through entity classification, we can distinguish knowledge domains, achieve accurate entity positioning, and quickly distinguish query categories.

[0018] Secondary index: document fragment vector library (FAISS clustering index);

[0019] In the process of establishing a secondary index, long documents are first segmented into semantically coherent small segments to facilitate vectorization and efficient retrieval. Then, the text segments are converted into numerical vector representations to capture semantic information. The vector representations are then organized through the FAISS clustering index, and similar vectors are assigned to the same cluster. Each cluster is set with a center vector to speed up index search.

[0020] Level 3 index: fault case decision tree (generated by ID3 algorithm);

[0021] The construction process of the fault case decision tree is as follows: collect fault case data - extract fault characteristics (equipment type, process type, parameter type, parameter range) - generate a tree through the ID3 algorithm - prune and optimize the generated tree - API service.

[0022] In the hybrid retrieval process, the first-level index is traversed to locate the knowledge domain, and then the second-level / third-level index is queried in parallel.

[0023] At the retrieval technology level, the system integrates multimodal retrieval strategies:

[0024] Vector semantic retrieval: uses domain-optimized embedding models (such as bge-large fine-tuned for industrial terminology) to capture deep semantic associations.

[0025] Keyword search: Accurately match structured data such as equipment numbers and material codes;

[0026] Metadata filtering: filter by document type, department, validity period and other attributes;

[0027] HyDE hypothetical document embedding: uses a large model to generate hypothetical responses based on user queries, and uses the response vector to assist retrieval and enhance contextual relevance;

[0028] Retrieval results are optimized using a rerank model, rescoring candidate documents based on the semantic match between the user's question and the query, ensuring that the most relevant technical documents are returned first. In a factory troubleshooting scenario, this technology increased the recall rate of key repair solutions by 35%, significantly reducing downtime.

[0029] Aiming at the complex document structure unique to industrial scenarios, the system integrates advanced document parsing technology:

[0030] Multimodal layout analysis: Identifies text, tables, images, and formula areas in technical documents, particularly enhancing the parsing capabilities of industrial drawings such as device wiring diagrams and PLC ladder diagrams.

[0031] Industrial table recognition: uses the TSR (Table Structure Recognition) model to convert parameter tables and process matrices into structured data;

[0032] Formula detection and recognition: Extract equipment calculation formulas through MFD (Mathematical Formula Detection) and MFR (Mathematical Formula Recognition) technologies;

[0033] Reading order restoration: Reorganize the parsed content based on the "up-down, left-right" principle to ensure the correct order of steps in the device operation process;

[0034] After document parsing, an industrial semantic slicing algorithm is used to avoid semantic gaps caused by traditional equal segmentation. This algorithm divides knowledge blocks according to the title hierarchy (H1-H6), preserving the integrity of documents such as equipment operating manuals and process specifications. High-frequency term extraction technology is also introduced to construct a scenario knowledge base index, establishing a fast search channel for specialized terms such as equipment models and material codes.

[0035] Furthermore, the factory's production site data includes:

[0036] Warehouse material inventory status, production work order data, equipment voltage and current data, equipment temperature data, and equipment vibration data.

[0037] The inventory status of warehouse materials includes material code / ID, material name / description, specification model, batch number / serial number (for traceability), supplier information, unit (piece, kilogram, meter, etc.), safety stock level, minimum order quantity, storage location (warehouse, storage area, shelf, storage location), storage condition requirements (temperature and humidity, light protection, etc.), expiration date / shelf life, current real-time inventory quantity, inventory entry records (time, quantity, source - procurement, production, allocation), outbound records, in-transit inventory and reserved inventory.

[0038] Production work order data includes work order number / ID, product code / name / specifications, planned production quantity, actual production quantity, work order status (created, issued, in production, suspended, completed, closed, canceled), planned start time / planned completion time, actual start time / actual completion time, priority, and associated sales orders / customer needs.

[0039] The device voltage and current data, device temperature data, and device vibration data are provided by corresponding sensors. The device vibration data is collected by the vibration sensor. The specific vibration data collected includes displacement data, velocity data, and acceleration data in three-dimensional coordinates.

[0040] Furthermore, the data cleaning includes supplementing missing values ​​with a spatiotemporal collaborative filling method, marking outliers with an isolation forest algorithm, and performing secondary verification by matching historical fault thresholds with the RAG knowledge base.

[0041] The present invention uses a spatiotemporal collaborative filling method to supplement missing values. In order to supplement missing values, the missing values ​​are supplemented by combining the weight of 0.6 for equipment neighboring data and the weight of 0.4 for production line similar data. The isolation forest algorithm is used to mark outliers to eliminate outliers. Data greater than the isolation forest threshold set for the corresponding data is preliminarily marked as suspected anomalies. For data marked as suspected anomalies, the context information of the data is converted into a natural language query. The historical fault case library in the RAG knowledge base and the domain expert experience rules are used to find the sensor data threshold / pattern features that are highly correlated with the context information before / when the fault occurs, and a secondary threshold matching check is performed based on the found data.

[0042] Furthermore, the task processing of the large model module includes the following steps:

[0043] Step 4.1: Perform preliminary analysis of the user's input task based on ASR and NLP technologies;

[0044] ASR technology is speech recognition, which is used to convert user language input into text instructions. NLP is used for semantic analysis. After obtaining the instruction text, it first classifies the task type, then extracts keywords and outputs a structured instruction object.

[0045] Step 4.2: Decompose the task after preliminary analysis into atomic operation chains;

[0046] The purpose of this step is to decompose complex tasks into executable basic operation units. Atomic operations are the lowest-level, indecomposable operation units. Atomic operation chains are generated through the target decomposition engine.

[0047] Step 4.3: Initiate multiple rounds of searches to the RAG knowledge base based on the atomic operation chain. The search trigger condition is that the large model confidence or task complexity level meets the requirements;

[0048] When the confidence of the large model is less than 0.7 or the complexity of the task is greater than 3, that is, when the confidence of the large model for the currently decomposed atomic operation chain is lower than the preset threshold, it is necessary to search the RAG knowledge base to add relevant atomic units to enhance the context; or when the complexity of the decomposed atomic task is still large, it is necessary to search the RAG knowledge base.

[0049] Step 4.4: The large model verifies the feasibility of the task through the reinforcement learning reward function based on the search results of the RAG knowledge base;

[0050] The reward function is designed as follows:

[0051] R = αR s +βR c +γR e ;

[0052] Where: R is the total reward;

[0053] α is the safety reward weight coefficient;

[0054] R s Reward for safety;

[0055] β is the cost-reward weight coefficient;

[0056] R c for cost incentives;

[0057] γ is the efficiency reward weight coefficient;

[0058] R e Reward for efficiency.

[0059] The safety reward is -10×the number of dangerous operations, the cost reward is 1 / (estimated energy consumption×electricity price), and the efficiency reward is log(standard working hours / estimated working hours). The feasibility of the task is verified by calculating the comprehensive reward.

[0060] Step 4.5: If the feasibility of the task is greater than the threshold, the task is assigned to the agent module for processing.

[0061] Furthermore, the structure of the intelligent agent module is as follows: the system intelligent agent connects the intelligent agents of each production stage task through a bus to form a hierarchical control structure.

[0062] Furthermore, the task processing of the agent module includes the following steps:

[0063] Step 5.1: After receiving the task, the system agent assigns the task to the corresponding agent according to the task domain and production stage;

[0064] After receiving a new task, the system agent first analyzes the attributes of the task through the task context information, including:

[0065] Field: The professional field to which the task belongs (including: mechanical processing, circuit welding, software compilation, logistics and transportation line A).

[0066] Production stage: the stage at which the task is located in the overall process (including: preprocessing, processing, assembly, testing, packaging).

[0067] The system agent then assigns the task to the agent responsible for that specific area and / or that specific production stage based on predefined rules.

[0068] Step 5.2: Each agent makes distributed decisions based on the local state of the responsible part;

[0069] Distributed decision-making may include: inserting a new task into a certain position in the queue, estimating the completion time of the new task, judging whether it can accept the new task based on its own status (if the allocation is negotiated), and adjusting its own resource parameters to adapt to the new task.

[0070] Step 5.3: The system agent optimizes the task sequence of each agent based on the TS_Qlearning scheduling algorithm.

[0071] The system agent receives local task queue status (including the preliminary sequence after the insertion of new tasks) and current status information fed back by all domain / stage agents. The global goal is to minimize the completion time while meeting any task deadline.

[0072] The system agent uses the TS_Qlearning scheduling algorithm to perform local searches in the solution space (the permutations and combinations of all agent task sequences) through TS. It explores the neighborhood by defining "moves" (such as exchanging the positions of two tasks in different agent queues or moving a task to another agent queue). It uses a taboo list to record recent moves to prevent loops and force exploration of new areas. It also sets a contempt criterion to allow breaking taboos when a significantly better solution is found.

[0073] Through Q-learning, the learned experience is used to evaluate the potential value of movements, and movements with high Q values ​​are prioritized for trial, thereby accelerating convergence to a better solution. The decision-making of the system agent is continuously iterated through the TS_Qlearning scheduling algorithm until the stopping condition is met. This optimized sequence of instructions is then issued to the agents in each field / stage for execution.

[0074] Furthermore, the step 5.2 includes: the corresponding intelligent agent makes local decisions on the task based on its own relevant environmental data.

[0075] The agent responsible for the corresponding task makes preliminary decisions on how to handle new tasks and how to adjust its existing task queue based on the current state of the task queue (including tasks previously assigned to it but not completed) and its own current local state (including: the current load, health status, current location, skill set, and progress of the machines / resources it controls).

[0076] Furthermore, the step 5.3 includes: the system agent learns the long-term optimal strategy from a global perspective based on TS_Qlearning, and optimizes the task sequence by avoiding repeated operations through a taboo table.

[0077] The taboo table records the action combinations that caused resource conflicts during the most recent k scheduling processes, where k is the number of corresponding devices × 0.2. The system agent makes global decisions based on TS_Qlearning. TS provides a search mechanism, and QL provides intelligent guidance. It can handle complex combinatorial optimization problems and dynamic disturbances, and is very suitable for complex scheduling scenarios that need to balance efficiency, flexibility, and global optimization goals.

[0078] A second aspect of the present invention provides an intelligent management method based on multi-agent collaboration, comprising the following steps:

[0079] Step 1: Build a RAG knowledge base based on the factory's production site data, equipment manuals, process parameters, and quality standards;

[0080] Step 2: After receiving the task input by the user, it is initially parsed by the large model and then decomposed into a chain of atomic operations;

[0081] Step 3: Initiate multiple rounds of searches to the RAG knowledge base based on the atomic operation chain and assign tasks to the agent module;

[0082] Step 4: The agent module optimizes task scheduling based on the TS_Qlearning scheduling algorithm and then assigns the corresponding agent to complete the task. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present drawings or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present drawings. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0084] Figure 1 A schematic structural diagram of the system provided by the present invention;

[0085] The purpose, features and advantages of this drawing will be further described with reference to the accompanying drawings in conjunction with the embodiments. DETAILED DESCRIPTION

[0086] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention. Based on the embodiments provided by the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work are within the scope of protection of the present invention.

[0087] Obviously, the drawings described below are merely examples or embodiments of the present invention. Those skilled in the art can apply the present invention to other similar scenarios based on these drawings without inventive effort. Furthermore, it is understood that while the effort involved in such a development process may be complex and lengthy, for those skilled in the art related to the disclosure of the present invention, any design, manufacturing, or production changes based on the technical content disclosed in the present invention are merely conventional technical means and should not be construed as an inadequacy of the disclosure of the present invention.

[0088] Unless otherwise specified, all embodiments and optional embodiments of the present invention can be combined with each other to form new technical solutions.

[0089] Unless otherwise specified, all technical features and optional technical features of the present invention can be combined with each other to form a new technical solution.

[0090] Unless otherwise specified, all steps of the present invention may be performed sequentially or randomly, preferably sequentially. For example, the method includes steps (a) and (b), which means that the method may include steps (a) and (b) performed sequentially, or may include steps (b) and (a) performed sequentially. For example, the method may further include step (c), which means that step (c) may be added to the method in any order, for example, the method may include steps (a), (b) and (c), or may include steps (a), (c) and (b), or may include steps (c), (a) and (b), etc.

[0091] Unless otherwise specified, the terms "include" and "comprising" used in the present invention may be open-ended or closed-ended. For example, "include" and "comprising" may mean that other components not listed may also be included or that only the listed components are included.

[0092] Unless otherwise specified, the term "or" is inclusive in this disclosure. For example, the phrase "A or B" means "A, B, or both A and B." More specifically, the condition "A or B" is satisfied by any of the following conditions: A is true (or exists) and B is false (or does not exist); A is false (or does not exist) and B is true (or exists); or both A and B are true (or exist).

[0093] In order to better understand the solutions of the embodiments of the present invention, some relevant terms and concepts that may be involved in the embodiments of the present invention are first introduced below.

[0094] (1) RAG knowledge base (Retrieval-Augmented Generation) is a technical architecture that combines information retrieval technology with generative AI models. Its core principle is divided into two steps: retrieval phase: when a user asks a query, the system retrieves relevant information fragments from structured or unstructured knowledge bases (such as documents, databases); generation phase: the retrieved information is used as context input into a generative model (such as a large language model) to generate natural language answers, ensuring the accuracy and traceability of the answers.

[0095] (2) TS_Qlearning scheduling algorithm (Tabu Search-Q Learning), a hybrid optimization algorithm that combines tabu search (TS) and Q learning, is designed specifically for multi-agent system task scheduling. It generates a set of high-quality initial solutions and stores them in the tabu table to avoid repeated searches for inefficient solutions.

[0096] (3) AI Agent is an AI system with autonomous perception, decision-making and execution capabilities. Its core features include: being driven by a large language model, being able to independently plan tasks, call tools (such as APIs), and remember historical interactions.

[0097] In this embodiment, Figure 1 As shown in the figure, the intelligent management system based on multi-agent collaboration includes:

[0098] Data acquisition module: used to collect factory production site data in real time and send it to the data processing module;

[0099] Data processing module: used to clean the received data and send it to the RAG knowledge base module;

[0100] Large model module: performs semantic analysis on the tasks input by the user based on the large model and queries the RAG knowledge base module and / or the scheduling agent module;

[0101] Agent module: includes the system agent and several agents responsible for processing specific tasks of their own, and performs task scheduling based on the TS_Qlearning scheduling algorithm;

[0102] RAG knowledge base module: builds the RAG knowledge base and provides knowledge retrieval based on hierarchical indexing and hybrid retrieval mechanisms.

[0103] This invention implements intelligent manufacturing management through a three-tier architecture: the data layer integrates a variety of factory information, including real-time sensor data, equipment status, inventory status, and production orders, through a data acquisition module; the intermediate processing layer uses an agent module to perform distributed and integrated decision-making, and provides knowledge support in conjunction with the RAG knowledge base; and the application layer uses a large model module to parse user-input tasks based on a large-model-driven natural language interactive interface, enabling intelligent management and control of the entire production process. This architecture effectively addresses the response lag, knowledge fragmentation, and decision-making rigidity issues of traditional MOM systems, significantly improving the adaptability and intelligence of production systems.

[0104] The core technical advantages of this invention are reflected in three aspects: First, the hierarchical indexing mechanism of the RAG knowledge base (abstract index + document block index) is used to achieve efficient retrieval of manufacturing knowledge, which is more than 50% faster than traditional vector retrieval; second, a multi-agent collaborative architecture is adopted, and each professional agent (production scheduling, quality control, equipment maintenance, etc.) realizes parallel processing of complex problems through distributed decision-making, and then makes comprehensive decisions through the system agent; finally, based on the large model, the tasks input by the user are parsed and production tasks are generated. It can dynamically generate professional knowledge such as equipment parameter tuning plans and exception handling plans, significantly reducing dependence on human experts.

[0105] The RAG knowledge base of the present invention includes equipment historical operation data, warehouse information data, technical solution data, customer information data, work log information data, fault condition data, process documents and maintenance manuals.

[0106] The RAG knowledge base of this invention uses a hierarchical indexing architecture, constructing a two-tier index structure: a top-level summary index and a bottom-level document block index. When processing a user query, the system first quickly filters the relevant document range using the summary index, then performs precise document block-level retrieval within the limited range. This mechanism significantly reduces the number of vectors that need to be compared, making it particularly suitable for efficient retrieval of massive amounts of technical documents (such as equipment manuals, process standards, and SOP files) in factory scenarios.

[0107] The hierarchical index includes:

[0108] Primary index: entity classification (including equipment, process, quality standards, customers, and warehousing);

[0109] Through entity classification, we can distinguish knowledge domains, achieve accurate entity positioning, and quickly distinguish query categories.

[0110] Secondary index: document fragment vector library (FAISS clustering index);

[0111] In the process of establishing a secondary index, long documents are first segmented into semantically coherent small segments to facilitate vectorization and efficient retrieval. Then, the text segments are converted into numerical vector representations to capture semantic information. The vector representations are then organized through the FAISS clustering index, and similar vectors are assigned to the same cluster. Each cluster is set with a center vector to speed up index search.

[0112] In this embodiment, a motor maintenance manual is segmented according to fault codes, and the number of blocks after segmentation is 120, with an average length of 342 words; a process specification manual is segmented according to process steps, and the number of blocks after segmentation is 85, with an average length of 210 words.

[0113] Level 3 index: fault case decision tree (generated by ID3 algorithm);

[0114] The construction process of the fault case decision tree is as follows: collect fault case data - extract fault characteristics (equipment type, process type, parameter type, parameter range) - generate a tree through the ID3 algorithm - prune and optimize the generated tree - API service.

[0115] In the hybrid retrieval process, the first-level index is traversed to locate the knowledge domain, and then the second-level / third-level index is queried in parallel.

[0116] At the retrieval technology level, the system integrates multimodal retrieval strategies:

[0117] Vector semantic retrieval: uses domain-optimized embedding models (such as bge-large fine-tuned for industrial terminology) to capture deep semantic associations.

[0118] Keyword search: Accurately match structured data such as equipment numbers and material codes;

[0119] Metadata filtering: filter by document type, department, validity period and other attributes;

[0120] HyDE hypothetical document embedding: uses a large model to generate hypothetical responses based on user queries, and uses the response vector to assist retrieval and enhance contextual relevance;

[0121] Retrieval results are optimized using a rerank model, rescoring candidate documents based on the semantic match between the user's question and the query, ensuring that the most relevant technical documents are returned first. In a factory troubleshooting scenario, this technology increased the recall rate of key repair solutions by 35%, significantly reducing downtime.

[0122] Aiming at the complex document structure unique to industrial scenarios, the system integrates advanced document parsing technology:

[0123] Multimodal layout analysis: Identifies text, tables, images, and formula areas in technical documents, particularly enhancing the parsing capabilities of industrial drawings such as device wiring diagrams and PLC ladder diagrams.

[0124] Industrial table recognition: uses the TSR (Table Structure Recognition) model to convert parameter tables and process matrices into structured data;

[0125] Formula detection and recognition: Extract equipment calculation formulas through MFD (Mathematical Formula Detection) and MFR (Mathematical Formula Recognition) technologies;

[0126] Reading order restoration: Reorganize the parsed content based on the "up-down, left-right" principle to ensure the correct order of steps in the device operation process;

[0127] After document parsing, an industrial semantic slicing algorithm is used to avoid semantic gaps caused by traditional equal segmentation. This algorithm divides knowledge blocks according to the title hierarchy (H1-H6), preserving the integrity of documents such as equipment operating manuals and process specifications. High-frequency term extraction technology is also introduced to construct a scenario knowledge base index, establishing a fast search channel for specialized terms such as equipment models and material codes.

[0128] Furthermore, the factory's production site data includes:

[0129] Warehouse material inventory status, production work order data, equipment voltage and current data, equipment temperature data, and equipment vibration data.

[0130] The inventory status of warehouse materials includes material code / ID, material name / description, specification model, batch number / serial number (for traceability), supplier information, unit (piece, kilogram, meter, etc.), safety stock level, minimum order quantity, storage location (warehouse, storage area, shelf, storage location), storage condition requirements (temperature and humidity, light protection, etc.), expiration date / shelf life, current real-time inventory quantity, inventory entry records (time, quantity, source - procurement, production, allocation), outbound records, in-transit inventory and reserved inventory.

[0131] Production work order data includes work order number / ID, product code / name / specifications, planned production quantity, actual production quantity, work order status (created, issued, in production, suspended, completed, closed, canceled), planned start time / planned completion time, actual start time / actual completion time, priority, and associated sales orders / customer needs.

[0132] The device voltage and current data, device temperature data, and device vibration data are provided by corresponding sensors. The device vibration data is collected by the vibration sensor. The specific vibration data collected includes displacement data, velocity data, and acceleration data in three-dimensional coordinates.

[0133] Furthermore, the data cleaning includes supplementing missing values ​​with a spatiotemporal collaborative filling method, marking outliers with an isolation forest algorithm, and performing secondary verification by matching historical fault thresholds with the RAG knowledge base.

[0134] The present invention uses a spatiotemporal collaborative filling method to supplement missing values. In order to supplement missing values, the missing values ​​are supplemented by combining the weight of 0.6 for equipment neighboring data and the weight of 0.4 for production line similar data. The isolation forest algorithm is used to mark outliers to eliminate outliers. Data greater than the isolation forest threshold set for the corresponding data is preliminarily marked as suspected anomalies. For data marked as suspected anomalies, the context information of the data is converted into a natural language query. The historical fault case library in the RAG knowledge base and the domain expert experience rules are used to find the sensor data threshold / pattern features that are highly correlated with the context information before / when the fault occurs, and a secondary threshold matching check is performed based on the found data.

[0135] Furthermore, the task processing of the large model module includes the following steps:

[0136] Step 4.1: Perform preliminary analysis of the user's input task based on ASR and NLP technologies;

[0137] ASR technology is speech recognition, which is used to convert user language input into text instructions. NLP is used for semantic analysis. After obtaining the instruction text, it first classifies the task type, then extracts keywords and outputs a structured instruction object.

[0138] Step 4.2: Decompose the task after preliminary analysis into atomic operation chains;

[0139] The purpose of this step is to decompose complex tasks into executable basic operation units. Atomic operations are the lowest-level, indecomposable operation units. Atomic operation chains are generated through the target decomposition engine.

[0140] Step 4.3: Initiate multiple rounds of searches to the RAG knowledge base based on the atomic operation chain. The search trigger condition is that the large model confidence or task complexity level meets the requirements;

[0141] When the confidence of the large model is less than 0.7 or the complexity of the task is greater than 3, that is, when the confidence of the large model for the currently decomposed atomic operation chain is lower than the preset threshold, it is necessary to search the RAG knowledge base to add relevant atomic units to enhance the context; or when the complexity of the decomposed atomic task is still large, it is necessary to search the RAG knowledge base.

[0142] Step 4.4: The large model verifies the feasibility of the task through the reinforcement learning reward function based on the search results of the RAG knowledge base;

[0143] The reward function is designed as follows:

[0144] R = αR s +βR c +γR e ;

[0145] Where: R is the total reward;

[0146] α is the safety reward weight coefficient;

[0147] R s Reward for safety;

[0148] β is the cost-reward weight coefficient;

[0149] R c for cost incentives;

[0150] γ is the efficiency reward weight coefficient;

[0151] R e Reward for efficiency.

[0152] The safety reward is -10×the number of dangerous operations, the cost reward is 1 / (estimated energy consumption×electricity price), and the efficiency reward is log(standard working hours / estimated working hours). The feasibility of the task is verified by calculating the comprehensive reward.

[0153] Step 4.5: If the feasibility of the task is greater than the threshold, the task is assigned to the agent module for processing.

[0154] Furthermore, the structure of the intelligent agent module is as follows: the system intelligent agent connects the intelligent agents of each production stage task through a bus to form a hierarchical control structure.

[0155] Furthermore, the task processing of the agent module includes the following steps:

[0156] Step 5.1: After receiving the task, the system agent assigns the task to the corresponding agent according to the task domain and production stage;

[0157] After receiving a new task, the system agent first analyzes the attributes of the task through the task context information, including:

[0158] Field: The professional field to which the task belongs (including: mechanical processing, circuit welding, software compilation, logistics and transportation line A).

[0159] Production stage: the stage at which the task is located in the overall process (including: preprocessing, processing, assembly, testing, packaging).

[0160] The system agent then assigns the task to the agent responsible for that specific area and / or that specific production stage based on predefined rules.

[0161] Step 5.2: Each agent makes distributed decisions based on the local state of the responsible part;

[0162] Distributed decision-making may include: inserting a new task into a certain position in the queue, estimating the completion time of the new task, judging whether it can accept the new task based on its own status (if the allocation is negotiated), and adjusting its own resource parameters to adapt to the new task.

[0163] Step 5.3: The system agent optimizes the task sequence of each agent based on the TS_Qlearning scheduling algorithm.

[0164] The system agent receives local task queue status (including the preliminary sequence after the insertion of new tasks) and current status information fed back by all domain / stage agents. The global goal is to minimize the completion time while meeting any task deadline.

[0165] The system agent uses the TS_Qlearning scheduling algorithm to perform local searches in the solution space (the permutations and combinations of all agent task sequences) through TS. It explores the neighborhood by defining "moves" (such as exchanging the positions of two tasks in different agent queues or moving a task to another agent queue). It uses a taboo list to record recent moves to prevent loops and force exploration of new areas. It also sets a contempt criterion to allow breaking taboos when a significantly better solution is found.

[0166] Through Q-learning, the learned experience is used to evaluate the potential value of movements, and movements with high Q values ​​are prioritized for trial, thereby accelerating convergence to a better solution. The decision-making of the system agent is continuously iterated through the TS_Qlearning scheduling algorithm until the stopping condition is met. This optimized sequence of instructions is then issued to the agents in each field / stage for execution.

[0167] Furthermore, the step 5.2 includes: the corresponding intelligent agent makes local decisions on the task based on its own relevant environmental data.

[0168] The agent responsible for the corresponding task makes preliminary decisions on how to handle new tasks and how to adjust its existing task queue based on the current state of the task queue (including tasks previously assigned to it but not completed) and its own current local state (including: the current load, health status, current location, skill set, and progress of the machines / resources it controls).

[0169] Furthermore, the step 5.3 includes: the system agent learns the long-term optimal strategy from a global perspective based on TS_Qlearning, and optimizes the task sequence by avoiding repeated operations through a taboo table.

[0170] The taboo table records the action combinations that caused resource conflicts during the most recent k scheduling processes, where k is the number of corresponding devices × 0.2. The system agent makes global decisions based on TS_Qlearning. TS provides a search mechanism, and QL provides intelligent guidance. It can handle complex combinatorial optimization problems and dynamic disturbances, and is very suitable for complex scheduling scenarios that need to balance efficiency, flexibility, and global optimization goals.

[0171] The intelligent management method based on multi-agent collaboration includes the following steps:

[0172] Step 1: Build a RAG knowledge base based on the factory's production site data, equipment manuals, process parameters, and quality standards;

[0173] Step 2: After receiving the task input by the user, it is initially parsed by the large model and then decomposed into a chain of atomic operations;

[0174] Step 3: Initiate multiple rounds of searches to the RAG knowledge base based on the atomic operation chain and assign tasks to the agent module;

[0175] Step 4: The agent module optimizes task scheduling based on the TS_Qlearning scheduling algorithm and then assigns the corresponding agent to complete the task.

[0176] It should be noted that the present invention is not limited to the above-mentioned embodiments. The above-mentioned embodiments are merely examples, and any embodiments having substantially the same structure and effect as the technical concept within the scope of the technical solution of the present invention are all included in the technical scope of the present invention. In addition, without departing from the scope of the present invention, other embodiments that can be conceived by those skilled in the art and that combine some of the constituent elements in the embodiments are also included in the scope of the present invention.

Claims

1. An intelligent management system based on multi-agent collaboration, characterized by: include: Data acquisition module: used to collect factory production site data in real time and send it to the data processing module; Data processing module: used to clean the received data and send it to the RAG knowledge base module; Large model module: performs semantic analysis on the tasks input by the user based on the large model and queries the RAG knowledge base module and / or the scheduling agent module; Agent module: includes the system agent and several agents responsible for processing specific tasks of their own, and performs task scheduling based on the TS_Qlearning scheduling algorithm; RAG knowledge base module: builds the RAG knowledge base and provides knowledge retrieval based on hierarchical indexing and hybrid retrieval mechanisms.

2. The intelligent management system based on multi-agent collaboration according to claim 1 is characterized in that: The production site data of the factory includes: Warehouse material inventory status, production work order data, equipment voltage and current data, equipment temperature data, and equipment vibration data.

3. The intelligent management system based on multi-agent collaboration according to claim 1 is characterized in that: The data cleaning includes supplementing missing values ​​with a spatiotemporal collaborative filling method, marking outliers with an isolation forest algorithm, and performing secondary verification by matching historical fault thresholds with a RAG knowledge base.

4. The intelligent management system based on multi-agent collaboration according to claim 1 is characterized in that: The task processing of the large model module includes the following steps: Step 4.1: Perform preliminary analysis of the user's input task based on ASR and NLP technologies; Step 4.2: Decompose the task after preliminary analysis into atomic operation chains; Step 4.3: Initiate multiple rounds of searches to the RAG knowledge base based on the atomic operation chain. The search trigger condition is that the large model confidence or task complexity level meets the requirements; Step 4.4: The large model verifies the feasibility of the task through the reinforcement learning reward function based on the search results of the RAG knowledge base; Step 4.5: If the feasibility of the task is greater than the threshold, the task is assigned to the agent module for processing.

5. The intelligent management system based on multi-agent collaboration according to claim 1 is characterized in that: The structure of the intelligent agent module is as follows: the system intelligent agent is connected to the intelligent agents of each production stage task through a bus to form a hierarchical control structure.

6. The intelligent management system based on multi-agent collaboration according to claim 5 is characterized in that: The task processing of the intelligent agent module includes the following steps: Step 5.1: After receiving the task, the system agent assigns the task to the corresponding agent according to the task domain and production stage; Step 5.2: Each agent makes distributed decisions based on the local state of the responsible part; Step 5.3: The system agent optimizes the task sequence of each agent based on the TS_Qlearning scheduling algorithm.

7. The intelligent management system based on multi-agent collaboration according to claim 6 is characterized in that: The step 5.2 includes: the corresponding intelligent agent makes local decisions on the task based on its own relevant environmental data.

8. The intelligent management system based on multi-agent collaboration according to claim 6 is characterized in that: The step 5.3 includes: the system agent learns the long-term optimal strategy from a global perspective based on TS_Qlearning, and optimizes the task sequence by avoiding repeated operations through a taboo table.

9. The intelligent management method based on multi-agent collaboration is characterized by: The following steps are involved: Step 1: Build a RAG knowledge base based on the factory's production site data, equipment manuals, process parameters, and quality standards; Step 2: After receiving the task input by the user, it is initially parsed by the large model and then decomposed into a chain of atomic operations; Step 3: Initiate multiple rounds of searches to the RAG knowledge base based on the atomic operation chain and assign tasks to the agent module; Step 4: The agent module optimizes task scheduling based on the TS_Qlearning scheduling algorithm and then assigns the corresponding agent to complete the task.

Images