Multi-agent ecological environment big data analysis system based on large model

Abstract

The invention provides a multi-agent ecological environment big data analysis system based on a large model, and solves the problem that ecological environment big data analysis is difficult to master and analyze skillfully due to professional barriers, and the system comprises a main agent and a plurality of sub-agents, each sub-agent is used for executing a task issued by the main agent and reporting respective mastered resources to the main agent, and the main agent is used for receiving a user data analysis request, synchronously enhancing a user intention through an identification module, and according to an output result of the identification module and resource conditions provided in responsibilities of the sub-agents, performing resource management on the sub-agents according to the output result of the identification module and the resource conditions provided in the responsibilities of the sub-agents. Setting a first sub-task to be completed by one of the sub-agents, reporting a task result to the main agent after the corresponding sub-agent completes the first sub-task, synchronously judging whether the task is completed by the main agent according to an original user request, if so, returning a final result, and ending a task process; if not, the task setting process is repeated, and the corresponding sub-agents are issued.

Description

Technical Field

[0001] The present invention relates to the technical field of environmental monitoring and protection, and in particular to a multi-agent ecological environment big data analysis system based on a large model. Background Art

[0002] Ecological environment big data analysis is an important technical guarantee for the digitalization and precision of environmental governance. However, in actual business, front-line ecological environment managers often find it difficult to master data query and analysis skills due to the knowledge barriers between environmental and computer majors, resulting in insufficient business efficiency.

[0003] Traditional solutions include: 1. Cultivate professional technical personnel: This solution is costly and time-consuming; 2. Build an information platform to solidify specific data analysis needs with platform functions: Since platform construction requires program design, bidding and procurement, construction acceptance and other links, the construction cycle is long, which conflicts with the changing environmental analysis needs; 3. Function Calling technology based on large models to achieve natural language understanding and data analysis: This solution requires the pre-writing of relevant data analysis functions for large models to call, which makes it difficult to cope with unplanned data analysis needs.

[0004] To this end, we propose a multi-agent ecological environment big data analysis system based on a large model to solve the above problems. Summary of the invention

[0005] The technical problem to be solved by the present invention is to overcome the defects of the prior art. The present invention proposes a multi-agent ecological environment big data analysis system based on a large model, which solves the practical problem of high professionalism and standardization requirements for business processing in the ecological environment field, and improves the multi-agent system's intention recognition accuracy for users' brief and vague questions.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: a multi-agent ecological environment big data analysis system based on a large model, including: a main agent and sub-agents, The sub-agents include multiple sub-agents, each of which is used to execute tasks assigned by the main agent and report to the main agent the resources they control. The main intelligent agent is used to receive the user data analysis request, and simultaneously enhance the user intention through the recognition module. According to the output result of the recognition module and the resource conditions provided in the responsibilities of all sub-intelligent agents, the first sub-task is set to be completed by one of the sub-intelligent agents. After the corresponding sub-intelligent agent completes the first sub-task, it reports the task result to the main intelligent agent. The main intelligent agent synchronously compares the original user request to determine whether the task is completed. If completed, the final result is returned and the task process ends; if not completed, the aforementioned task setting process is repeated and sent to the corresponding sub-intelligent agent. Among them, the recognition module enhances the user intention process specifically by analyzing the possible M categories of user intentions according to the application scenario requirements, using the gradient descent method to train the intent recognition enhancement method, calculating the probability that the user input belongs to intention j in the M categories, and taking the intention with the highest probability as the final user intention result.

[0007] Furthermore, the identification module recognizes and outputs the steps as follows: According to the application scenario requirements, analyze the user's possible intention M categories, use professional field standard documents, and establish a training and recognition library; Build a recognition model and prepare N samples to perform intensive training on the model. Select the last hidden layer vector output by the recognition model as the user input feature. , a 1×D vector, initializes the intent recognition classification matrix , is a matrix of size D×M, and the matrix elements are initialized according to the uniform distribution of [0,1]; For any sample i in N training texts, calculate the probability of its classification as j, (1) in, represents the jth column of the W matrix, e is the natural logarithm, k is a positive integer from 1, 2, 3 to M, Calculate the classification error loss function, (2) in, is the correct classification label. When the category is j, the value is 1, otherwise the value is 0; The trained recognition model is used to more accurately identify user intent. For user input, the probability that the user input belongs to intent j among M categories can be calculated using formula (1). The intent with the highest probability is taken as the final user intent identified. Read the business template content corresponding to the recognition intent type j, use the big model to summarize the template business logic and process, and enhance the user intent based on this.

[0008] Furthermore, the sub-agents include data scientists, information officers, programmers, tool libraries and inquirers. The three agents, data scientists, information officers and tool libraries, report their respective resources to the main agent, which are data set information, knowledge base information and tool library information respectively. The programmer agent is responsible for writing data analysis python code.

[0009] Furthermore, the inquirer agent is used to interact with the user and confirm the user's intention enhancement result with the user. When the user confirms that it is correct, it is handed over to the main agent for execution. Otherwise, the intention enhancement result is regenerated according to the user's suggestion.

[0010] Furthermore, the data engineer agent is responsible for accessing the data set and selecting a suitable data set for data query according to the data query task assigned by the master agent. The specific steps include: Access a data set, which includes multiple data tables, each of which includes a data field name, field description, and field data format description; Write SQL statements based on assigned tasks; According to the execution of the SQL statement, determine whether the assigned task is completed. If not, repeat the previous step; If the task is judged to be completed, the result will be returned; if there is a problem in the execution of the task or additional information is required, such as the data to be queried does not exist or the data query scope is unclear, the inquirer will further confirm the relevant information with the user; if the task is not completed after multiple cycles, an error message will be returned.

[0011] Furthermore, the information officer agent is responsible for querying professional domain knowledge, mainly accessing the ecological environment knowledge base and the network retrieval interface. The specific steps include: Access to the ecological environment knowledge base and network search interface, the ecological environment knowledge base includes the knowledge base name, knowledge base ID, and knowledge base description; Based on the assigned information query task and the accessed knowledge base content, determine whether to use knowledge base query or network search; Obtain knowledge base query or network search results to determine whether the assigned task is completed; if the task is not completed, repeat the previous step; If the task is judged to be completed, the result will be returned; if there are problems in the execution of the task or additional information is required, the inquirer will further confirm the relevant information with the user.

[0012] Furthermore, the programmer agent performs the following steps: Write Python code according to the assigned code writing tasks; Execute the Python code and determine whether the task is completed based on the returned result; if the task is not completed, re-execute the previous step; If the task is judged to be completed, the result is returned. If there is a problem with the task execution or additional information is needed, such as unclear program scope results, the inquirer will further confirm the relevant information with the user. If the task is not completed after multiple cycles, an error message is returned.

[0013] Furthermore, the tool library agent is responsible for calling the industry professional tool API, and the specific steps include: Access to industry professional analysis tool APIs, including API name, function description, and input parameter description information; According to the assigned subtask, select the appropriate API, extract valid input parameters based on the task information, and call the tool API; Determine whether the task is completed based on the result returned by the API. If not, re-execute the previous step. When the task is judged to be completed, the result is returned. If there is a problem with the task execution or additional information is needed, such as unclear key parameters required for tool calling, the inquirer will further confirm the relevant information with the user. If the task is not completed after multiple cycles, an error message is returned.

[0014] Compared with the prior art, the beneficial effects of the present invention include: 1. The present invention realizes the comprehensive integration and utilization of five types of data, namely structured database, unstructured knowledge base, network information, professional tools and historical business documents, by designing six intelligent agents with different functions, thus avoiding the problem that traditional large-model language processing technology is difficult to cope with unplanned data analysis needs; 2. This invention solves the practical problem of high requirements for professionalism and standardization of business processing in the field of ecological environment by designing intention enhancement based on template reference, and improves the intention recognition accuracy of multi-agent system for users' brief and fuzzy questions 3. The present invention realizes the convenience of continuous interaction with users during the execution of complex tasks by designing the inquirer intelligent agent module, so that users can supplement key information of the task during the execution of the task, or assist in processing anomalies in the intelligent agent's call information, thereby improving the accuracy and stability of the intelligent agent's task completion. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The disclosure of the present invention is described with reference to the accompanying drawings. It should be understood that the drawings are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. In the accompanying drawings, the same reference numerals are used to refer to the same components. Among them: Figure 1 The figure schematically shows the composition of an analysis system according to one embodiment of the present invention. DETAILED DESCRIPTION

[0016] It is easy to understand that according to the technical solution of the present invention, without changing the essential spirit of the present invention, a person skilled in the art can propose a variety of interchangeable structural modes and implementation modes. Therefore, the following specific implementation modes and drawings are only exemplary descriptions of the technical solution of the present invention, and should not be regarded as the whole of the present invention or as a limitation or restriction to the technical solution of the present invention.

[0017] According to one embodiment of the present invention, Figure 1 Shown.

[0018] In order to solve the problem that traditional information platforms or digital analysis methods based on large model function calling need to define possible analysis types in advance, it is difficult to cope with flexible analysis needs, such as Figure 1 As shown, in this embodiment, the multi-agent system is mainly composed of six agents, namely, planner, data engineer, information officer, programmer, tool library, and inquirer. Among them, the planner is the main agent; the other five agents are sub-agents. Each sub-agent is used to execute the tasks issued by the main agent and report to the main agent the resources they have. The main agent is used to receive user data analysis requests and simultaneously enhance user intentions through the recognition module. According to the output results of the recognition module and the resource conditions provided in the responsibilities of all sub-agents, the first sub-task is set to be completed by one of the sub-agents. After the corresponding sub-agent completes the first sub-task, it reports the task results to the main agent. The main agent simultaneously compares the original user request to determine whether the task is completed. If completed, the final result is returned and the task process ends; if not completed, the aforementioned task setting process is repeated and issued to the corresponding sub-agent. Among them, the recognition module enhances the user intention process specifically by analyzing the possible M categories of user intentions according to the application scenario requirements, using the gradient descent method to train the intent recognition enhancement method, calculating the probability that the user input belongs to intention j in the M categories, and taking the intention with the highest probability as the final user intention result.

[0019] Combined with the above system composition, the system analysis process can be summarized as follows: 1) The system accesses external resources, including database, knowledge base, Internet search, and professional tool library; 2) The planner agent receives the user request and enhances the user's intention based on the user request and existing resources, especially based on the business template reference; 3) Interact with the user by asking the agent to confirm the intended enhancement results; 4) Based on the intention enhancement results, the tasks are broken down and assigned to members of the agent team; 5) Other agent team members receive tasks from the planner, complete the tasks according to their own functions and give feedback. During this process, they can interact with the user at any time by asking the agent to solve problems in the task execution process; 6) The planner determines whether the task is completed or needs further execution based on the feedback from each agent. If the task is completed, the result is returned; if further execution is required, step 5 is re-executed); 7) The front-end system provides real-time feedback to users on the progress of the task and the final result of the task based on the task log.

[0020] The following is a detailed description of the system in combination with each intelligent agent: 1. Multi-agent control process 1) The three agents, data engineer, information officer and tool library, report their respective resources to the planner, which are data set information, knowledge base information and tool library information respectively; 2) The planner receives the user data analysis request, enhances the user's intention, and sets subtask 1 based on the user's intention enhancement result, combined with the member's responsibilities and the resources provided in step 1), and assigns it to one of the sub-agents to complete subtask 1; 3) The sub-agent completes subtask 1 and reports the task completion result to the planner (main agent). The planner determines whether the task is completed based on the task completion result and the user's original request. If the task is not completed, subtask 2 is set and step 2 is repeated); 4) When the planner determines that the task is completed, the final result is directly returned and the task process ends.

[0021] 2. Intention Enhancement 2.1 Intent Enhancement Based on Template Reference In order to meet the professional and normative requirements of the ecological environment industry, the user intention is enhanced based on the template reference. The specific method is: 1) Analyze the user's possible intentions M according to the application scenario requirements, use professional field standard documents, and establish a training and recognition library; 2) Prepare N samples to conduct intensive training on the model. Select the last hidden layer vector output by the large model as the user input feature , a vector of size 1×D. Initialize the intent recognition classification matrix , is a matrix of size D×M, and the matrix elements are initialized according to the uniform distribution of [0,1]; 3) Use the gradient descent method to train the intent recognition enhancement method. For any sample i in the N training texts, calculate the probability of its classification as j: (1) in, represents the jth column of the W matrix, e is the natural logarithm, k is a positive integer from 1, 2, 3 to M, Calculate the classification error loss function: (2) in, is the correct classification label. When the category is j, the value is 1, otherwise the value is 0.

[0022] 4) Use the trained model to more accurately identify user intent. For user input, use formula (1) to calculate the probability that the user input belongs to intent j among M categories, and take the intent with the highest probability as the final user intent identified; 5) Read the business template content corresponding to the recognition intent type j, use the big model to summarize the template business logic and process, and enhance the user intent based on this; Example: User input: Help me analyze the air quality in Nanjing today; Intent recognition: Daily air quality analysis.

[0023] Reference template: Nanjing Air Quality Daily.docx. Intention enhancement results: Analyze the air quality in Nanjing on December 31, 2024, including the following steps: Calculate the mean and maximum concentrations of sulfur dioxide, nitrogen oxides, PM10, PM2.5 and ozone at each national and provincial control station; Compare the mean concentration of each pollutant with the control target to find out whether there are any factors exceeding the standard; Conduct year-on-year and month-on-month analysis of mean pollutant concentrations to identify abnormally high values ​​(increases of more than 30% year-on-year and month-on-month); For factors exceeding the standard and abnormally high-value factors, the pollution source analysis is carried out using the pollution source tracing model; Use air quality prediction models to conduct forecast analysis for future developments; Integrate the above information and write an analysis report; 5) Use the inquirer agent to confirm the user's intention enhancement results with the user. When the user confirms that it is correct, it will be handed over to the planner for execution. Otherwise, the intention enhancement results will be regenerated according to the user's suggestions.

[0024] 3. Sub-agents 3.1 Data Scientist The data engineer is responsible for accessing the data set and selecting the appropriate data set for data query according to the data query task assigned by the main agent. The specific steps include: 1) Access a data set, which includes multiple data tables, each of which includes a data field name, field description, and field data format description; 2) Write SQL statements according to the assigned tasks; 3) According to the execution of the SQL statement, determine whether the assigned task is completed. If not, repeat step 2); 4) If the task is judged to be completed, the result is returned; 5) If there are problems with the task execution or additional information is needed, such as the data to be queried does not exist or the data query scope is unclear, the inquirer will further confirm the relevant information with the user; 6) If the task is not completed after multiple cycles, an error message is returned.

[0025] 3.2 Information Officer Information officers are responsible for searching professional knowledge, mainly accessing the ecological environment knowledge base and network retrieval interface. The specific steps include: 1) Access to the ecological environment knowledge base and network retrieval interface. The ecological environment knowledge base includes information such as the knowledge base name, knowledge base ID, and knowledge base description; 2) Determine whether to use knowledge base query or network search based on the assigned information query task and the accessed knowledge base content; 3) Obtain the knowledge base query or network search results to determine whether the assigned task has been completed; if the task has not been completed, repeat step 2); 4) If the task is judged to be completed, the result is returned; 5) If there are problems with the task execution or additional information is needed, such as the required information does not exist or the information query content is unclear, the inquirer will further confirm the relevant information with the user; 6) If the task is not completed after multiple cycles, an error message is returned.

[0026] 3.3 Programmer The programmer is responsible for writing the data analysis python code, and the specific steps include: 1) Write Python code according to the assigned code writing tasks; 2) Execute the Python code and determine whether the task is completed based on the returned result; if the task is not completed, re-execute step 1); 3) If the task is judged to be completed, the result is returned; 4) If there are problems with the execution of the task or additional information is needed, such as unclear program scope results, the inquirer will further confirm the relevant information with the user; 5) If the task is not completed after multiple cycles, an error message is returned.

[0027] 3.4 Tool Library The tool library is responsible for calling the industry professional tool API. The specific steps include: 1) Access to industry professional analysis tool APIs, including API name, function description, and input parameter description information; 2) Select the appropriate API according to the assigned subtask, extract valid input parameters based on the task information, and call the tool API; 3) Determine whether the task is completed based on the result returned by the API. If not, re-execute step 2); 4) When the task is judged to be completed, the result is returned; 5) If there are problems with task execution or additional information is needed, such as unclear key parameters required for tool invocation, the inquirer will further confirm the relevant information with the user; 6) If the task is not completed after multiple cycles, an error message is returned.

[0028] Similarly, the sub-agents in the above scheme can be screened according to actual needs, and the alternative scheme may not completely include the above five sub-agents, or only include the functions of one of the knowledge base and network retrieval in the information officer agent.

[0029] In addition, the above-mentioned method steps and system can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0030] The technical scope of the present invention is not limited to the contents in the above description. Those skilled in the art can make various deformations and modifications to the above embodiments without departing from the technical idea of ​​the present invention, and these deformations and modifications should all fall within the protection scope of the present invention.

Claims

1. A multi-agent ecological environment big data analysis system based on a large model, characterized by: include: The main agent and sub-agents, The sub-agents include multiple sub-agents, each of which is used to execute tasks assigned by the main agent and report to the main agent the resources they control. The main intelligent agent is used to receive the user data analysis request, and simultaneously enhance the user intention through the recognition module. According to the output result of the recognition module and the resource conditions provided in the responsibilities of all sub-intelligent agents, the first sub-task is set to be completed by one of the sub-intelligent agents. After the corresponding sub-intelligent agent completes the first sub-task, it reports the task result to the main intelligent agent. The main intelligent agent synchronously compares the original user request to determine whether the task is completed. If completed, the final result is returned and the task process ends; if not completed, the aforementioned task setting process is repeated and sent to the corresponding sub-intelligent agent. Among them, the recognition module enhances the user intention process specifically by analyzing the possible M categories of user intentions according to the application scenario requirements, using the gradient descent method to train the intent recognition enhancement method, calculating the probability that the user input belongs to intention j in the M categories, and taking the intention with the highest probability as the final user intention result.

2. The multi-agent ecological environment big data analysis system based on a large model according to claim 1 is characterized by: The identification module identification output steps are as follows: According to the application scenario requirements, analyze the user's possible intention M categories, use professional field standard documents, and establish a training and recognition library; Build a recognition model and prepare N samples to perform intensive training on the model. Select the last hidden layer vector output by the recognition model as the user input feature. , a 1×D vector, initializes the intent recognition classification matrix , is a matrix of size D×M, and the matrix elements are initialized according to the uniform distribution of [0,1]; For any sample i in N training texts, calculate the probability of its classification as j, （1） in, represents the jth column of the W matrix, e is the natural logarithm, k is a positive integer from 1, 2, 3 to M, Calculate the classification error loss function, （2） in, is the correct classification label. When the category is j, the value is 1, otherwise the value is 0; The trained recognition model is used to more accurately identify user intent. For user input, the probability that the user input belongs to intent j among M categories can be calculated using formula (1). The intent with the highest probability is taken as the final user intent identified. Read the business template content corresponding to the recognition intent type j, use the big model to summarize the template business logic and process, and enhance the user intent based on this.

3. The multi-agent ecological environment big data analysis system based on a large model according to claim 1 is characterized by: The sub-agents include data scientists, information officers, programmers, tool libraries and inquirers. The three agents, data scientists, information officers and tool libraries, report their respective resources to the main agent, which are data set information, knowledge base information and tool library information respectively. The programmer agent is responsible for writing data analysis python code.

4. The multi-agent ecological environment big data analysis system based on a large model according to claim 3 is characterized by: The inquirer agent is used to interact with the user and confirm the user's intention enhancement result with the user. When the user confirms that it is correct, it is handed over to the main agent for execution. Otherwise, the intention enhancement result is regenerated according to the user's suggestion.

5. The multi-agent ecological environment big data analysis system based on a large model according to claim 3 is characterized by: The data engineer agent is responsible for accessing the data set and selecting a suitable data set for data query according to the data query task assigned by the master agent. The specific steps include: Access a data set, which includes multiple data tables, each of which includes a data field name, field description, and field data format description; Write SQL statements according to assigned tasks; According to the execution of the SQL statement, determine whether the assigned task is completed. If not, repeat the previous step; If the task is judged to be completed, the result will be returned; if there are problems in the execution of the task or additional information is needed, such as the data to be queried does not exist or the data query scope is unclear, the inquirer will further confirm the relevant information with the user; if the task is not completed after multiple cycles, an error message will be returned.

6. The multi-agent ecological environment big data analysis system based on a large model according to claim 3 is characterized by: The information agent is responsible for querying professional domain knowledge, mainly accessing the ecological environment knowledge base and network retrieval interface. The specific steps include: Access to the ecological environment knowledge base and network search interface, the ecological environment knowledge base includes the knowledge base name, knowledge base ID, and knowledge base description; Based on the assigned information query task and the accessed knowledge base content, determine whether to use knowledge base query or network search; Obtain knowledge base query or network search results to determine whether the assigned task is completed; if the task is not completed, repeat the previous step; If the task is judged to be completed, the result will be returned; if there are problems in the execution of the task or additional information is required, the inquirer will further confirm the relevant information with the user.

7. The multi-agent ecological environment big data analysis system based on a large model according to claim 3 is characterized by: The programmer agent performs the following steps: Write Python code according to the assigned code writing tasks; Execute the Python code and determine whether the task is completed based on the returned result; if the task is not completed, re-execute the previous step; If the task is judged to be completed, the result will be returned. If there are problems in the execution of the task or additional information is needed, such as unclear program scope and results, the inquirer will further confirm the relevant information with the user. If the task is not completed after multiple cycles, an error message will be returned.

8. The multi-agent ecological environment big data analysis system based on a large model according to claim 3 is characterized by: The tool library agent is responsible for calling the industry professional tool API. The specific steps include: Access to industry professional analysis tool APIs, including API name, function description, and input parameter description information; According to the assigned subtask, select the appropriate API, extract valid input parameters based on the task information, and call the tool API; Determine whether the task is completed based on the result returned by the API. If not, re-execute the previous step. When the task is judged to be completed, the result is returned. If there are problems in the execution of the task or additional information is needed, such as unclear key parameters required for tool calls, the inquirer will further confirm the relevant information with the user. If the task is not completed after multiple cycles, an error message is returned.

Images