Intelligent yield management method and system based on multi-agent network

Abstract

The invention provides an intelligent yield management method and system based on a multi-agent network, and relates to the field of quality control and artificial intelligence, and the method comprises the following steps: a tutor agent carries out semantic understanding and task analysis on natural language interaction information, and routes decomposed sub-tasks to corresponding expert agents; the expert agent calls a corresponding tool to execute the task according to the subtask, and returns a task execution result to the tutor agent; and the tutor agent analyzes and integrates the task execution results, integrates the task execution results into final natural language response information, and returns the natural language response information to the user. The intelligent mode of yield management is fundamentally changed, the problems that a traditional yield management system is lack of autonomous execution capacity, low in intelligent degree and limited in expansibility are solved by constructing a multi-agent framework, and meanwhile the limitation of a single agent in the process of automatically executing cross-module and cross-function complex tasks is made up.

Description

Technical Field

[0001] The present invention relates to the fields of quality control in industrial manufacturing and artificial intelligence. Specifically, it relates to an intelligent yield management method and system based on a multi-agent network. Background Art

[0002] A yield management system (YMS) is a core industrial software system designed specifically for high-precision manufacturing industries such as semiconductors and panels. By systematically collecting, integrating, and analyzing a large amount of data in the production process, it helps factories identify yield loss points, locate process fluctuations, and predict product quality trends, thereby providing accurate improvement suggestions and decision-making support. The architecture of a yield management system usually includes a data acquisition layer, a data processing layer, an analysis engine layer, and an application interface layer. These levels need to be seamlessly connected to ensure the timely flow of data and the accuracy of analysis results. Due to the increasingly complex industrial manufacturing environment, the yield management system must also have strong scalability and adaptability in order to flexibly adjust analysis strategies and business processes according to the specific needs of different factories and changing process conditions, so as to continuously improve product yield and production efficiency.

[0003] However, there are some inherent limitations in the intelligent and automated aspects of current yield management systems, which restrict their application effects and efficiency improvement. Traditional yield management systems often adopt a relatively rigid mode in design, with a high degree of coupling between functional modules and a lack of intelligent decision-making and autonomous execution capabilities. This deficiency in architecture directly leads to the following problems: ① Lack of autonomous execution ability: Traditional yield management systems mainly serve as analysis tools, capable of providing data display and analysis results, but lacking the ability to execute operations autonomously. After obtaining system suggestions, users still need to manually perform specific operations such as data query, analysis calculation, parameter adjustment, and result generation. A large amount of manual intervention not only reduces efficiency but also increases operating costs.

[0004] ② High operation complexity: In order to use and manage a yield management system, users need to master complex operation skills and professional knowledge. The interface of traditional yield management systems is complex and the functional modules are scattered. Users need to perform a large number of operations in the cumbersome interface to complete tasks such as data query, analysis, and report generation, which limits the actual application of the system.

[0005] ③ Low intelligence level: Traditional yield management systems mainly rely on preset rules and algorithms for data analysis, lacking the ability of autonomous learning and intelligent decision-making, and are difficult to cope with complex and changing production environments. The system's ability to identify and handle abnormal situations is limited, and it is difficult to actively discover potential problems and automatically take corresponding measures.

[0006] ④Limited scalability and flexibility: Traditional yield management systems have low modularity, making it difficult for the systems to quickly integrate new data sources, analysis tools, or business processes. When the production process or management requirements of a factory change, the system often requires a large amount of customized development, resulting in high upgrade and maintenance costs and difficulty in quickly adapting to new requirements.

[0007] With the rapid development of artificial intelligence technology, autonomous execution systems based on Large Language Model Agents (LLM Agents) have gradually become a research hotspot. Large Language Model Agents can not only understand natural language instructions, but more importantly, can make decisions and automatically execute specific operations on behalf of users through tool calling, significantly improving work efficiency.

[0008] However, yield management involves very complex tasks, such as data processing, product yield management, yield improvement, and manufacturing management. Each type of task requires different tools and methods, and there are complex dependencies between these tasks. The complexity of yield management is mainly reflected in the need to process massive heterogeneous data from MES, equipment, and inspection systems, and that yield problem analysis and solution usually need to span multiple functional modules, such as data collection and cleaning, statistical analysis, process parameter optimization, equipment monitoring, and quality prediction. In an actual production environment, typical cross-module and cross-functional yield problems include multi-site correlation analysis (which requires analyzing data from multiple upstream and downstream process sites simultaneously to find hidden correlation problems), integrated optimization of equipment-process-materials (which requires considering the interactive effects of equipment status, process parameters, and raw material characteristics), yield ramp-up of new products (which requires coordinating data and knowledge from multiple departments such as R & D, process, equipment, and quality), and systematic yield fluctuation analysis (which requires combining information from multiple systems such as production planning, material management, equipment maintenance, and process control). Although a single agent can usually play an effective role when using a small number of tools in a single field, even the most advanced models such as GPT-4o will have greatly reduced effectiveness when multiple tools are required to execute the above complex processes simultaneously. This is because a single agent needs to process a large amount of information, coordinate multiple tools, maintain long-term memory, and perform complex reasoning simultaneously, making it prone to problems such as distracted attention, incorrect tool selection, or broken reasoning chains. Especially when autonomous execution of complex cross-module and cross-functional tasks is required, a single agent is often unable to cope and cannot fully unleash the potential of Large Language Model Agents. Summary of the Invention

[0009] The present invention provides an intelligent yield management method and system based on a multi-agent network, which solves the deficiencies of existing single agents in automatically executing complex yield management tasks.

[0010] In a first aspect, an intelligent yield management method based on a multi-agent network is provided in an embodiment of the present invention. The method includes the following processes: Receive natural language interaction information proposed by a user through a front-end interaction interface, and transmit the natural language interaction information to a tutor agent; The tutor agent performs semantic understanding and task analysis on the natural language interaction information, decomposes the task into subtasks, and routes the subtasks to corresponding expert agents; The expert agent calls an LLM node and a tool node according to the subtask. The LLM node converts the subtask into structured tool call parameters, and the tool node calls a corresponding tool from the equipped tool set according to the tool call parameters to execute the task, and returns the task execution result to the tutor agent through the expert agent; The tutor agent analyzes and integrates the task execution result, integrates the task execution result into the final natural language response information, and returns the natural language response information to the user through the front-end interaction interface.

[0011] In the above embodiment, the present invention not only realizes the natural language interaction between the user and the yield management system, but also can automatically execute complex yield management tasks through the execution of operations by the expert agent and the intelligent scheduling of the tutor, without the user manually participating in the intermediate steps, greatly improving the work efficiency and providing an effective solution for building a new generation of intelligent yield management systems.

[0012] As some alternative embodiments of the present application, the front-end interaction interface includes a Web chat page, a yield management system, or a chat application software.

[0013] As some alternative embodiments of the present application, the tutor agent and the expert agent are respectively implemented based on a large language model.

[0014] As some alternative embodiments of the present application, the tools called by the expert agent include a standardized encapsulation of the functions of the yield management system.

[0015] As some alternative embodiments of the present application, there is a loop link between the tutor agent and the expert agent, and there is a loop link between the LLM node and the tool node.

[0016] As some alternative embodiments of the present application, the expert agent includes a data management expert, a data analysis expert, and a data visualization expert.

[0017] In some alternative embodiments of the present application, the toolset equipped for the data management expert includes an SQL query tool and a data export tool; the toolset equipped for the data analysis expert includes a statistical calculation tool, a trend analysis tool, and a root cause analysis tool; the toolset equipped for the data visualization expert includes a scatter plot drawing tool, a line chart drawing tool, and a pie chart drawing tool.

[0018] In a second aspect, the present invention provides an intelligent yield management system based on a multi-agent network, the system comprising: A data acquisition unit, which receives natural language interaction information proposed by a user through a front-end interaction interface and transmits the natural language interaction information to a tutor agent; A task decomposition unit, which semantically understands and analyzes the natural language interaction information through the tutor agent, decomposes the task into subtasks, and routes the subtasks to corresponding expert agents; A task execution unit, which, through the expert agent, calls an LLM node and a tool node according to the subtask. The LLM node converts the subtask into structured tool call parameters, and the tool node calls a corresponding tool from the equipped toolset according to the tool call parameters to execute the task, and returns the task execution result to the tutor agent through the expert agent; A data return unit, which analyzes and integrates the task execution result through the tutor agent, integrates the task execution result into final natural language response information, and returns the natural language response information to the user through the front-end interaction interface.

[0019] In a third aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the intelligent yield management method based on a multi-agent network.

[0020] In a fourth aspect, the present invention provides a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, it implements the intelligent yield management method based on a multi-agent network.

[0021] The beneficial effects of the present invention are as follows: 1. The present invention applies the large language model to the human-computer interaction layer of the yield management system, realizing an intuitive and natural interaction method. Different from the complex and cumbersome graphical interfaces, numerous function buttons, nested menus and forms of traditional yield management systems, users no longer need to learn complex operation interfaces and professional terms, nor remember cumbersome operation processes and parameter settings. They only need to express their needs in daily language to understand and execute corresponding operations. Traditional yield management systems often require users to have a professional background to operate correctly. The interface design is centered on functions rather than user experience, resulting in high learning costs and low usage efficiency. The natural language interaction of the present invention not only greatly reduces the usage threshold, but also supports the expression of complex instructions and multi-turn conversations, enabling users to communicate with the system at a deeper level, including clarifying requirements, providing feedback and iteratively optimizing results, greatly enhancing the universality and usability of the system.

[0022] 2. The present invention realizes the technological progress from analysis and suggestion to automatic execution. Traditional question-and-answer systems can only provide information query and improvement suggestions, and still require users to manually execute specific operations; while the present invention enables the large language model agent to directly execute specific operations such as SQL queries, data analysis, and visualization chart generation through tool invocation technology, truly replacing users to complete complex tasks. Moreover, it not only can execute clear instructions, but also has the ability of autonomous decision-making. It can intelligently judge the best execution path and required resources according to users' vague or incomplete instructions, automatically plan the execution process, and dynamically adjust the strategy according to the intermediate results during the execution process. This autonomous decision-making and execution ability enables the system to have problem-solving ability close to that of human experts, greatly improving work efficiency.

[0023] 3. The present invention applies a multi-agent collaboration architecture of "tutor-expert", adopting the idea of "divide and conquer" to effectively solve the problem of the ability limitation of a single agent in dealing with complex tasks. This architecture forms an intelligent collaboration network, and the multi-turn interaction between agents constructs an effective decision-making-execution feedback mechanism, enabling the system to handle tasks that are highly complex and span multiple functional modules. This multi-agent architecture significantly enhances the system's ability to handle complex business scenarios and achieves a collaborative processing effect that is difficult to achieve by a single agent.

[0024] 4. The present invention adopts a highly modular design, decomposing the system into a tutor node and expert agent modules. Each module has clear responsibilities, standardized interfaces, and low coupling. This design brings good scalability and portability: users can flexibly add expert agents and tool sets without modifying the existing architecture; at the same time, the system separates the business logic from the underlying implementation, enabling it to easily adapt to different yield management systems and data environments. This greatly reduces the system maintenance cost and technical debt, enabling it to continuously adapt to changing production environments and business requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as limiting the scope. For those skilled in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0026] Figure 1 is the flowchart of the intelligent yield management method based on a multi-agent network according to the embodiments of the present invention; Figure 2 is a comparison diagram of the interaction levels of the multi-agent network and the traditional interaction level based on a graphical user interface (GUI) according to the embodiments of the present invention; Figure 3 is the architecture diagram of the multi-agent network according to the embodiments of the present invention; Figure 4 is the architecture diagram of the expert agent according to the embodiments of the present invention; Figure 5 is the dialogue diagram of Scenario 1: The user requests to draw the yield curve for the most recent month according to the embodiments of the present invention; Figure 6 is the dialogue diagram of Scenario 1: The tutor agent assigns a task to the data management expert according to the embodiments of the present invention; Figure 7 is the dialogue diagram of Scenario 1: The data management expert feedbacks data element information according to the embodiments of the present invention; Figure 8 is the dialogue diagram of Scenario 1: The tutor agent continues to route the task to the data management expert according to the embodiments of the present invention; Figure 9 is the dialogue diagram of Scenario 1: The tutor agent returns the final result according to the embodiments of the present invention; Figure 10 is the complete process schematic diagram of Scenario 1 according to the embodiments of the present invention; Figure 11 is the dialogue diagram of Scenario 2: The user requests to analyze the yield data of different production lines according to the embodiments of the present invention; Figure 12 is the dialogue diagram of Scenario 2: The tutor agent assigns a task to the data management expert according to the embodiments of the present invention; Figure 13 is the dialogue diagram of Scenario 2: The data management expert performs tool calls according to the embodiments of the present invention; Figure 14 is the dialogue diagram of Scenario 2: The tutor agent continues to route the task to the data analysis expert according to the embodiments of the present invention; Figure 15This is the dialogue diagram of scenario two described in the embodiments of the present invention: the data analysis expert executes tool calls; Figure 16 This is the dialogue diagram of scenario two described in the embodiments of the present invention: the tutor agent continues to route the task to the data visualization expert; Figure 17 This is the dialogue diagram of scenario two described in the embodiments of the present invention: the tutor agent returns the final result; Figure 18 This is the complete flowchart of scenario two described in the embodiments of the present invention. Detailed implementation manners

[0027] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0028] To solve the deficiencies of existing single agents in automatically performing complex yield management. Embodiments of the present invention provide an intelligent yield management method and system based on a multi-agent network. Before introducing the specific technical solutions of the present invention, first introduce the intelligent yield management system based on a multi-agent network. The system includes a front-end layer, an agent layer, and a back-end layer; wherein, the agent layer includes a tutor agent and an expert agent, and the expert agent includes an LLM node and a tool node. The system corresponds one-to-one with the intelligent yield management method based on a multi-agent architecture.

[0029] Please refer to Figure 1 , Figure 1 This is the flowchart of the intelligent yield management method based on a multi-agent network. The process of the intelligent yield management method based on a multi-agent network is as follows: Step 1: Receive the natural language interaction information proposed by the user through the front-end interaction interface of the front-end layer, and transmit the natural language interaction information to the tutor agent of the agent layer.

[0030] In the embodiments of the present invention, the user uses the front-end interaction interface to pose questions or instructions to the yield management system in natural language. For example: "Query the yield data of a certain product line today", "Generate a yield trend chart for the last week", "Analyze the yield differences of different devices", etc.

[0031] Specifically, the front-end interaction interface can be an independently developed Web chat page, can also be integrated into the existing yield management system, or directly embedded in a chat application software (such as enterprise WeChat). The embodiments of the present invention do not limit this. It should be emphasized that the intelligent yield management method based on a multi-agent architecture provided by the embodiments of the present invention does not require redeveloping the yield management system and will not conflict with the interaction method of the existing yield management system. The user can select a suitable front-end interaction interface according to actual needs. Please refer to Figure 2 ,Figure 2 This is a comparison chart of the interaction levels of the multi-agent network involved in the present invention and the traditional interaction levels based on the graphical user interface (GUI). Compared with the traditional GUI-based interaction, the natural language interaction adopted in the embodiments of the present invention has significant advantages: ① More efficient interaction experience: Users do not need to learn complex menu structures and operation processes. They only need to express their needs in natural language like having a conversation with a person to get a quick response.

[0032] ② Lower learning threshold: Natural language is the most familiar interaction method for users, reducing the threshold for users to use and eliminating the need for professional computer operation skills.

[0033] ③ More flexible interaction method: Natural language can express more complex and flexible needs, breaking through the limitations of the GUI-based interaction method in terms of expression ability.

[0034] ④ More intelligent human-machine collaboration: Natural language interaction lays the foundation for achieving deeper human-machine collaboration, facilitating better understanding of user intentions and actively providing assistance and suggestions.

[0035] Step 2: The tutor agent in the proxy layer performs semantic understanding and task analysis on the natural language interaction information, decomposes the natural language interaction information into subtasks, and routes the subtasks to the corresponding expert agents.

[0036] In the embodiments of the present invention, the tutor agent is the core control component of the multi-agent network, implemented based on a large language model. It receives the user's natural language questions or instructions, performs semantic understanding and task parsing, analyzes the user's intentions, and identifies the task types. Subsequently, the tutor agent formulates a task execution plan, decomposes complex tasks into subtasks, and routes the subtasks to the corresponding expert agents. Among them, the expert agents include, but are not limited to, data management experts, data analysis experts, and data visualization experts, etc.

[0037] For example: For the task of "analyzing the possible reasons for the decline in the yield of XX process", the tutor agent may decompose it into subtasks such as "data query", "statistical calculation", "root cause analysis", etc., and route them to the data management expert, data analysis expert, etc. respectively.

[0038] Please refer to Figure 3 , Figure 3 which is the architecture diagram of the multi-agent network, and its key points are as follows: (1) Tutor agent: The tutor agent is the core control component of the multi-agent network, implemented based on a large language model, and mainly undertakes the following tasks: ① User interaction: Receive and understand the user's natural language input.

[0039] ② Task analysis: Analyze the user's intention and formulate a task execution plan.

[0040] ③ Task routing: Assign the decomposed subtasks to the corresponding expert agents. That is, according to the nature and requirements of the subtasks, select the most suitable expert agent to handle the task. This precise task routing ensures that each subtask is handled by the most professional agent, improving the overall execution efficiency and result quality.

[0041] ④ Result integration: Aggregate the execution results of each expert agent to generate the final response.

[0042] Specifically, the technical implementation of the tutor agent mainly includes two key elements: model selection and prompt engineering. In terms of model selection, the embodiments of the present invention comprehensively evaluate multiple dimensions such as performance, tool call ability, and reasoning planning, and recommend using OpenAI GPT-4o, Claude 3.7 Sonnet, or OpenAI o3-mini as the base model. It should be noted that the embodiments of the present invention do not impose mandatory restrictions on model selection, and appropriate alternative solutions can be selected according to specific application scenarios, including open-source models, closed-source commercial models, or customized and fine-tuned dedicated models. In terms of prompt engineering, the embodiments of the present invention integrate a variety of advanced technologies, including: role-playing to enhance professionalism, chain of thought to enhance reasoning ability, control flow to ensure execution logic, output limit to standardize the response format, and few-shot prompting to improve task adaptability, etc. Based on practical experience in the industrial production environment, a tutor agent prompt template can be designed.

[0043] Expert agent: The expert agent is responsible for understanding the instructions of the tutor agent, calling the corresponding tools according to the instructions, and after completing the tool call, returning the result to the tutor agent.

[0044] (3) There is a loop link between the tutor agent and the expert agent: The tutor agent can continue to route the subtasks to the expert agent according to the feedback of the expert agent, and this process can be repeated multiple times.

[0045] Step 3: Through the expert agent in the proxy layer, call the LLM node and the tool node according to the subtask. The LLM node converts the subtask into structured tool call parameters, and the tool node calls the corresponding tool from the equipped tool set according to the tool call parameters, combines the data interaction in the backend layer to execute the task, and returns the task execution result to the tutor agent through the expert agent.

[0046] In the embodiments of the present invention, the expert agent that receives the subtask, according to the instructions of the tutor agent, calls the corresponding tool from its own tool set to perform operations. After the expert agent completes the tool call, it returns the execution result (such as: queried data, analysis results, visualization charts, etc.) to the tutor agent.

[0047] Please refer to Figure 4 , Figure 4 which is the architecture diagram of the expert agent. The expert agent is implemented based on a large language model and is constructed using the ReAct (Reasoning and Acting) framework. The core points are as follows: (1) LLM node: The LLM node is the core component of the expert agent, responsible for processing instructions from the tutor agent. It converts the instructions into structured tool call parameters through natural language understanding and passes these parameters to the tool node to perform specific operations. After the tool node completes the operation, the LLM node will analyze and summarize the execution result to form clear feedback information and return it to the tutor agent.

[0048] Specifically, similar to the tutor agent, the construction of the LLM node in the ReAct framework involves two key elements: model selection and prompt engineering. In terms of model selection, in the embodiments of the present invention, multiple dimensions such as performance, deployment, tool call ability, and reasoning planning are comprehensively evaluated, and it is recommended to use OpenAI GPT-4o, Claude 3.7 Sonnet, or OpenAI o3-mini as the base model. The embodiments of the present invention do not impose mandatory restrictions on model selection, and appropriate alternative solutions can be selected according to specific application scenarios, including open-source models, closed-source commercial models, or specially customized and fine-tuned dedicated models. In terms of prompt engineering, the embodiments of the present invention integrate a variety of advanced technologies, including: role-playing to enhance professionalism, chain of thought to enhance reasoning ability, control flow to ensure execution logic, output limitation to standardize the response format, and few-shot prompting to improve task adaptability, etc. Based on practical experience in the industrial production environment, a prompt template for the expert agent can be designed.

[0049] (2) Tool Node: The tool node is the execution component of the expert agent, responsible for calling the relevant functions of the yield management system according to the parameters provided by the LLM node and returning the execution results to the LLM node. The tools here are standardized encapsulations of the functions of the yield management system, including but not limited to the call interfaces for functions such as SQL queries, root cause analysis, and data visualization. It should be noted that the construction of the tools described in the embodiments of the present invention does not involve the development of specific business functions, but only refers to the encapsulation of the calls to these business functions. The specific business functions are provided by the yield management system. For example, "root cause analysis" may involve algorithms such as machine learning, and the specific algorithm implementation is provided by the yield management system. The embodiments of the present invention only develop the interfaces for calling these functions and do not involve the implementation of these functions. This design is similar to the relationship between the API layer and the business layer, ensuring that the expert agent can easily call system functions, while not affecting the complexity of the underlying business logic, being easy to maintain and expand, and having good portability.

[0050] (3) There is a loop link between the LLM node and the tool node. The LLM node will analyze and judge according to the return results of the tool node calls. If the current task has not been completed, it will continue to call the relevant tools. This action-feedback loop process continues until the LLM node determines that it can return the relevant results. This iterative working mode enables the expert agent to handle complex multi-step tasks.

[0051] Step Four: The tutor agent in the proxy layer analyzes and integrates the task execution results, integrates the task execution results into the final natural language response information, and returns the natural language response information to the user through the front-end interaction interface.

[0052] In the embodiments of the present invention, the tutor agent receives the results returned by the expert agent, conducts result analysis and sorting. The tutor agent judges whether the current task has been completed, whether it is necessary to further instruct the expert agent to execute the task, or whether iterative processing is required. If the task has not been completed, the tutor agent may return to Step Three, continue to instruct the expert agent to execute the task and receive the results. This process may be iterated multiple times until the tutor agent determines that the task has been completed. When the task is completed, the tutor agent will integrate all the return results of the expert agent, generate the final natural language response, and return it to the user through the front-end interaction interface. This iterative optimization mechanism can ensure the accuracy and integrity of the output results and better meet the user's needs.

[0053] The following takes two scenarios as examples to elaborate on the detailed steps of the embodiments of the present invention.

[0054] Scenario 1: To better understand the technical solution of the embodiments of the present invention, taking the yield problem as an example based on a certain yield database, a specific embodiment is used for detailed description.

[0055] For the convenience of demonstration, in Scenario 1, Python and the Lang Chain framework are used to develop a simple multi-agent network architecture, and both the tutor agent and the expert agent use Open AI GPT-4o.

[0056] Specifically, the expert agent includes: ① Data management expert: The data management expert is responsible for database query and data acquisition, and can generate accurate SQL statements according to requirements to query relevant information in the yield database, including yield data, process parameters, equipment status, etc. This expert agent has the ability to understand the database structure and can efficiently obtain and process structured data.

[0057] ② Data analysis expert: The data analysis expert focuses on data statistical analysis and mining, and can perform trend analysis, correlation analysis, anomaly detection, etc. on yield data. This data analysis expert has statistical knowledge and data mining capabilities, and can extract valuable information from the original data to provide data support for yield improvement.

[0058] ③ Data visualization expert: The data visualization expert is good at converting data into intuitive charts, and can generate various visualization charts according to analysis requirements, such as trend charts, pie charts, box plots, etc. This data visualization expert has data visualization capabilities and can present complex data in a clear and intuitive way to assist decision-making analysis.

[0059] In this embodiment scenario, the user requests to draw the yield curve for the most recent month. This scenario demonstrates how the tutor agent parses the user's requirements, plans the task path, and coordinates the two module's expert agents (data management expert and data visualization expert) to work together to complete the whole process from data acquisition to final visualization presentation through multiple rounds of iteration.

[0060] Step S10: The user asks a question to the yield management system in natural language through the front-end interaction interface.

[0061] Please refer to Figure 5 , Figure 5 For Scenario 1: The dialogue diagram of the user requesting to draw the yield curve for the most recent month. It can be seen from the message history that the user needs to draw the yield curve for the most recent month. After receiving the user's question, the tutor agent first confirms the specific requirements with the user.

[0062] Step S20: The tutor agent parses the question, plans the task, and routes it to the expert agent.

[0063] Please refer to Figure 6 , Figure 6 Figure 1 is a dialogue diagram of the tutor agent assigning tasks to the data management expert for Scenario 1. The tutor agent assigns tasks to the data management expert, asking him / her to provide relevant data element information. Among them, the messages starting with "[System]" are system logs, including tutor instructions, expert feedback, tool call logs, etc., rather than the information fed back by the tutor agent to the user.

[0064] Step S30: The expert agent executes the tool call and returns the result to the tutor agent.

[0065] Please refer to Figure 7 , Figure 7 Figure 2 is a dialogue diagram of the data management expert providing feedback on data element information for Scenario 1. After receiving the tutor agent's instruction, the data management expert calls relevant tools to obtain data element information and feeds back the relevant information to the tutor agent. From the message history, it can be seen that the expert agent can intelligently generate accurate SQL query statements according to the instruction, independently execute relevant query operations, and clearly feed back the results to the tutor agent. This mode of replacing humans with agents to perform professional operations not only significantly improves work efficiency but also eliminates the rigid requirement for business personnel to master technical languages such as SQL, achieving seamless connection between technology and business.

[0066] Step S21: The tutor agent continues to route the task to other expert agents based on the feedback from the expert agent.

[0067] Please refer to Figure 8 , Figure 8 Figure 3 is a dialogue diagram of the tutor agent continuing to route the task to the data management expert for Scenario 1. The tutor agent continues to route the task to the data management expert based on the feedback from the data management expert, asking him / her to obtain specific data. The tutor agent continues to route the task to the data visualization expert based on the further feedback from the data management expert. This process of "tutor scheduling - expert feedback" repeats multiple times. From the dialogue history, it can be seen that: ① The tutor agent asks the data management expert to obtain specific data.

[0068] ② The data management expert calls the tool to successfully obtain the yield information and feeds back the relevant results and intermediate variables to the tutor agent.

[0069] ③ After receiving the feedback, the tutor agent continues to route the task to the data visualization expert, asking him / her to draw a yield curve.

[0070] ④ After receiving the tutor's instruction, the data visualization expert calls relevant tools to draw a yield curve and returns the relevant results to the tutor agent.

[0071] Step S40: The tutor agent integrates the results, determines whether to iterate, and generates a final response.

[0072] Please refer to Figure 9 , Figure 9 , for Scenario 1: The dialogue diagram of the tutor agent returning the final result. After receiving the feedback from the data visualization expert, the tutor agent performs result integration and analysis. The tutor agent determines that the current task is completed and returns the integrated result to the user. It can be seen from the message history that the tutor node, based on the feedback result of the expert agent, transmits the required picture to the user and then determines that the task is completed. Please refer to Figure 10 , Figure 10 , for Scenario 1: The schematic diagram of the complete process. It should be emphasized that the entire process from the user initiating a request to the tutor agent providing a complete solution takes only about one minute; in contrast, if the same task is performed by a human operator, it may take one minute just to write the necessary SQL query statements, and it is estimated that it will take more than ten minutes to complete the entire workflow, and it also poses a relatively high requirement for the operator's computer professional skills. This significant comparison fully demonstrates the value of the embodiments of the present invention in improving work efficiency, reducing technical thresholds and learning costs, and achieving seamless integration of technology and business.

[0073] Scenario 2: In this scenario, the user requests an analysis of the yield data of different production lines. This scenario demonstrates the ability of the tutor agent to understand the user's vague request and coordinate all experts to solve the task.

[0074] Step S10: The user poses a question to the yield management system in natural language through the front-end interaction interface. Please refer to Figure 11 , Figure 11 , for Scenario 2: The dialogue diagram of the user requesting an analysis of the yield data of different production lines.

[0075] Step S20: The tutor agent parses the question, plans the task, and routes it to the expert agent. Please refer to Figure 12 , Figure 12 , for Scenario 2: The dialogue diagram of the tutor agent assigning a task to the data management expert.

[0076] Step S30: The expert agent performs tool invocation and returns the result to the tutor agent.

[0077] Please refer to Figure 13 , Figure 13 , for Scenario 2: The dialogue diagram of the data management expert performing tool invocation. After receiving the instruction from the tutor agent, the data management expert first invokes the tool to obtain the database metadata, then generates an SQL query statement, and invokes the tool to execute the SQL query to obtain the relevant data, and finally feeds back the result to the tutor agent.

[0078] Step S21: The tutor agent continues to route the task to the data analysis expert based on the feedback from the expert agent.

[0079] Please refer to Figure 14 , Figure 14 for Scenario 2: The conversation graph where the tutor agent continues to route the task to the data analysis expert. After receiving the instruction from the tutor agent, the data analysis expert first calls a tool to obtain the column names of the data table provided by the tutor agent, then calls a tool for hypothesis testing and calculates other statistics, and finally returns the results to the tutor agent.

[0080] Step S31: The data analysis expert executes the tool call and returns the results to the tutor agent.

[0081] Please refer to Figure 15 , Figure 15 for Occasion 2: The conversation graph where the data analysis expert executes the tool call. After receiving the instruction from the tutor agent, the data analysis expert first calls a tool to obtain the column names of the data table provided by the tutor agent, then calls a tool for hypothesis testing and calculates other statistics, and finally returns the results to the tutor agent.

[0082] Step S22: The tutor agent continues to route the task to the visualization expert based on the feedback from the expert agent. Please refer to Figure 16 , Figure 16 for Scenario 2: The conversation graph where the tutor agent continues to route the task to the data visualization expert. As can be seen from the figure, the tutor agent calls the data visualization expert and provides specific requirements; after receiving the instruction from the tutor agent, the data visualization expert calls relevant tools to draw bar charts and box plots of the yield data of different production lines and returns the results to the tutor agent.

[0083] Step S40: The tutor agent integrates the results and judgment iterations and generates a final response.

[0084] Please refer to Figure 17 , Figure 17 for Scenario 2: The conversation graph where the tutor agent returns the final result. After receiving the feedback from the data visualization expert, the tutor agent conducts result integration and analysis. The tutor agent determines that the current task is completed and returns the integrated result to the user. As can be seen from the conversation history, the tutor agent determines that the task is completed based on the feedback from the expert agent, calls a tool to send the required pictures to the user, and gives relevant suggestions. Please refer to Figure 18 , Figure 18 for the complete flowchart of Scenario 2.

[0085] In summary, the embodiments of the present invention automatically complete complex tasks through the collaboration among multiple agents. In this architecture, multiple agents can communicate, coordinate, and cooperate with each other to jointly complete tasks that are difficult for a single agent to accomplish. This architecture can not only overcome the problems of the traditional yield management system, such as lack of autonomous execution ability, low intelligence level, and limited scalability, but also make up for the limitations of a single agent in dealing with cross-module and cross-functional complex tasks. Among them, the multi-agent architecture is a typical implementation, which solves the problem of automatic execution of complex tasks through the "divide and conquer" strategy: this architecture creates dedicated agents for each type of task or module, and introduces a tutor agent to intelligently decompose tasks and route them to the corresponding expert agents for automatic execution.

[0086] It should be noted that the intelligent yield management system based on the multi-agent network in this embodiment corresponds one by one to the steps in the intelligent yield management method based on the multi-agent network in the foregoing embodiment. Therefore, for the functions, implementation manners, and achieved technical effects of the intelligent yield management system, reference may be made to the implementation manners of the intelligent yield management method based on the multi-agent network, which will not be elaborated here.

[0087] In addition, in one embodiment, the present application further provides a computer device, which includes a processor, a memory, and a computer program stored in the memory. When the computer program is run by the processor, it implements the intelligent yield management method based on the multi-agent network in the foregoing embodiment.

[0088] In addition, in one embodiment, the present application further provides a computer storage medium, on which a computer program is stored. When the computer program is run by the processor, it implements the intelligent yield management method based on the multi-agent network in the foregoing embodiment.

[0089] In some embodiments, the computer-readable storage medium may be a memory such as FRAM, ROM, PROM, EPROM, EEPROM, flash memory, magnetic surface memory, optical disc, or CD-ROM; or it may be various devices including one or any combination of the above memories. The computer may be various computing devices including intelligent terminals and servers.

[0090] In some embodiments, the executable instructions may be in the form of a program, software, software module, script, or code, and may be written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including being deployed as an independent program or being deployed as a module, component, subroutine, or other unit suitable for use in a computing environment.

[0091] As an example, executable instructions may, but need not, correspond to a file in a file system, may be stored as part of a file storing other programs or data, such as in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files storing one or more modules, subroutines, or code portions).

[0092] By way of example, executable instructions may be deployed to be executed on one computing device, or on multiple computing devices located at one site, or on multiple computing devices distributed across multiple sites and interconnected by a communication network.

[0093] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or system. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the existence of other identical elements in the process, method, article or system including the element.

[0094] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0095] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as a read-only memory / random access memory, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a multimedia terminal device (which can be a mobile phone, a computer, a television receiver, or a network device, etc.) to execute the methods described in each embodiment of the present application.

[0096] The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the present application specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An intelligent yield management method based on a multi-agent network, characterized in that, The method includes the following processes: Receive the natural language interaction information proposed by the user through the front-end interaction interface, and transmit the natural language interaction information to the tutor agent; The tutor agent performs semantic understanding and task analysis on the natural language interaction information, decomposes the natural language interaction information into subtasks, and routes the subtasks to the corresponding expert agents; The expert agent calls the LLM node and the tool node according to the subtask. The LLM node converts the subtask into structured tool call parameters. The tool node calls the corresponding tool from the equipped tool set according to the tool call parameters to execute the task, and returns the task execution result to the tutor agent through the expert agent; The tutor agent analyzes and integrates the task execution result, integrates the task execution result into the final natural language response information, and returns the natural language response information to the user through the front-end interaction interface.

2. The intelligent yield management method based on a multi-agent network according to claim 1, characterized in that The front-end interaction interface includes a Web chat page, a yield management system, or a chat application software.

3. The intelligent yield management method based on a multi-agent network according to claim 1, wherein The tutor agent and the expert agent are respectively implemented based on a large language model.

4. The intelligent yield management method based on a multi-agent network according to claim 1, characterized in that The tools called by the expert agent include a standardized encapsulation of the functions of the yield management system.

5. The intelligent yield management method based on a multi-agent network according to claim 1, wherein There is a loop link between the tutor agent and the expert agent, and there is a loop link between the LLM node and the tool node.

6. The intelligent yield management method based on a multi-agent network according to claim 1, wherein The expert agent includes a data management expert, a data analysis expert, and a data visualization expert.

7. The intelligent yield management method based on a multi-agent network according to claim 6, characterized in that The tool set equipped by the data management expert includes an SQL query tool and a data export tool; the tool set equipped by the data analysis expert includes a statistical calculation tool, a trend analysis tool, and a root cause analysis tool; The tool set equipped by the data visualization expert includes a scatter plot drawing tool, a line chart drawing tool, and a pie chart drawing tool.

8. An intelligent yield management system based on a multi-agent network, characterized in that, The system includes a front-end layer, an agent layer, and a back-end layer. Among them, the agent layer includes a tutor agent and an expert agent, and the expert agent includes an LLM node and a tool node; The front-end interaction interface of the front-end layer receives the natural language interaction information proposed by the user, and transmits the natural language interaction information to the tutor agent; The tutor agent of the agent layer performs semantic understanding and task analysis on the natural language interaction information, decomposes the task into subtasks, and routes the subtasks to the corresponding expert agents; The expert agent of the agent layer calls the LLM node and the tool node according to the subtask. The LLM node converts the instruction of the subtask into structured tool call parameters. The tool node calls the corresponding tool from the equipped tool set according to the tool call parameters, combines the data interaction of the back-end layer to execute the task, and returns the task execution result to the tutor agent through the expert agent; The tutor agent of the agent layer analyzes and integrates the task execution result, integrates the task execution result into the final natural language response information, and returns the natural language response information to the user through the front-end interaction interface.

9. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the intelligent yield management method based on a multi-agent network described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, it implements the method for intelligent yield management based on a multi-agent network according to any one of claims 1-7.

Images