CN-121998510-A - Intelligent supervision method based on multi-agent cooperation and dynamic graph arrangement

CN121998510ACN 121998510 ACN121998510 ACN 121998510ACN-121998510-A

Abstract

An intelligent supervision method based on multi-agent cooperation and dynamic graph arrangement utilizes a coordinator agent to analyze user natural language instructions, combines history session and course portrait to construct global context, dynamically generates a supervision execution plan in a Directed Acyclic Graph (DAG) form by the global planning agent, analyzes the plan by a parallel executor agent, distributes parallelizable subtasks to expert agents with different prompt word templates for multi-view evaluation, carries out consistency convergence on the output of each agent through an independent state mark synchronization mechanism arranged on a convergence node of the Directed Acyclic Graph (DAG), finally inputs a convergence result into a processing pipeline based on a responsibility chain mode, carries out evidence matching, improved suggestion generation and compliance filtration, and outputs a structured supervision report.

Inventors

MIAO QIGUANG
CHEN CHI
LIU RUYI
ZHAO PEIPEI
WANG XIANGYANG
WANG QUAN

Assignees

西安电子科技大学

Dates

Publication Date: 20260508
Application Date: 20260209

Claims (7)

1. An intelligent supervision method based on multi-agent cooperation and dynamic graph arrangement is characterized by comprising the following steps: s1, intention recognition and dynamic programming, namely receiving natural language input instructions of a user by utilizing a coordinator intelligent body The method comprises the steps of constructing a global context vector by combining history session memory, inputting the global context vector into a global planning agent, and generating a directed acyclic graph DAG global execution plan comprising execution steps, input and output parameters and front-rear dependency relationships through large language model reasoning; S2, inputting the global execution plan generated in the step S1 into a parallel executor, automatically identifying that a parallel task group without strong dependency relationship comprises a data sensing task and an intelligent evaluation task, resetting a corresponding parallel branch completion flag in a global state bus to be in an unfinished state, and then triggering a sensing sub-graph execution inlet corresponding to the data sensing task and an evaluation sub-graph execution inlet corresponding to the intelligent evaluation task by the parallel executor through an asynchronous scheduling mechanism; S3, multi-mode data perception and feature extraction, namely, in the execution stage of a perception subgraph, invoking a content abstract node to perform ASR speech transcription and text abstract on a classroom audio-video stream, and extracting a core teaching link; S4, in the execution stage of the evaluation sub-graph, according to the parallelizable subtasks in the step S3, instantiating a plurality of expert intelligent agents with different prompt word templates to form a multi-view expert evaluation module, analyzing the supervision objects from different dimensions, eliminating expert opinion conflicts through evaluation fusion nodes to generate consistency evaluation conclusion, and then updating the completion mark of the evaluation subtask; S5, aggregation synchronization based on independent marks, namely using an aggregation synchronization node as a logic barrier, polling and checking the independent completion marks written in the step S3 and the step S4, and releasing logic blocking and transmitting the aggregated full data to the downstream only when all preset parallel branch marks are detected to be in a 'completion' state; s6, inputting the total data collected in the step S5 into a report generation pipeline constructed based on a responsibility chain mode, sequentially carrying out data preprocessing, evidence chain matching of evaluation views and original data, teaching improvement suggestion retrieval generation and compliance content filtering, and finally outputting a structured supervision report containing multi-mode evidence support.
2. The intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement according to claim 1, wherein the natural language input instruction for the user in step S1 The coordinator agent first loads a length of History of session sequences of (a) Course portrait feature Natural language input instruction to user by large language model Historical session sequence And course portrait features Performing joint coding to construct current global context vector Inputting global planning agent based on global context vector Generating a directed acyclic graph DAG global execution plan in a JSON format comprising execution steps, input and output parameters and front-rear dependency relations through thinking chain reasoning Global execution plan Wherein, the node Represents the first Each step includes an attribute tuple Wherein The set of pre-dependency nodes for this step is represented as edges of the directed acyclic graph DAG.
3. The intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement according to claim 1, wherein the specific method in step S2 is as follows: Parallel executors traversing global execution plans Constructing task dependent topology graph, defining the current completed step set as For any step If it satisfies the dependency condition And is not executed, then add it to the queue to be executed 。
4. The intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement according to claim 1, wherein the specific method in step S3 is as follows: DAG global execution plan for directed acyclic graph Inputting parallel executor agent, and traversing global execution plan by parallel executor agent Constructing a task dependent topology graph, namely defining a current completed step set as For any execution step If it satisfies the dependency condition And is not executed, then add it to the queue to be executed Automatically identifying parallelizable subtasks, namely sensing tasks Task of evaluation Inputting classroom audio and video stream data into a perception subgraph, wherein the perception subgraph comprises a classroom content abstract generating process based on voice transcription and a teaching behavior trend feature extracting process based on time window statistics, and the two processes are executed in a parallel or serial mode to generate structured classroom perception feature data: 3.1 Content digest generation of converting an audio stream into a sequence of text with time stamps , wherein, In the case of a text word, Time-stamped text sequences by sliding window algorithm Splitting into a plurality of fragments, and calling a large language model to generate a structured class summary ; 3.2 Trend feature extraction, namely dividing a classroom time axis into Time windows of equal length For each window Counting interaction density of teachers and students Constructing a classroom trend feature matrix by the voice ratio of $V_k $ ; After execution is completed, the perception result is written into the global state, and the synchronous mark is marked Set to 1.
5. The intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement according to claim 1, wherein the specific method of step S4 is as follows: 4.1 Multiple view scoring instantiating multiple expert agents with different prompt word templates Forming multi-view expert evaluation modules, wherein each expert evaluation module independently receives the class summary output in the step S3 And classroom trend feature matrix Outputting the scoring vector And text evaluation ; 4.2 The evaluation fusion is to introduce fusion nodes and aggregate the scores of all expert evaluation modules by adopting a weighted average method to obtain final scores Performing de-duplication and conflict resolution on the text evaluation of each expert evaluation module through a large model to generate a comprehensive evaluation report ; Wherein, the Is the first Writing the evaluation result into the global state after the execution is completed, and marking the synchronization Set to 1.
6. The intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement according to claim 1, wherein the step S5 designs a logic barrier based on independent state markers, and is specifically implemented as follows: executing a state check function when the aggregation synchronous node is triggered by any upstream branch : If the result is WAITING, the current trigger thread directly terminates END without any subsequent operation, and the system is in a suspended state to wait for the rest branches; if READY is the result, the system unblocks and activates downstream report generation indicating that all parallel tasks are READY in the global status bus.
7. The intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement according to claim 1, wherein the specific method in step S6 is as follows: Constructing a report generating pipeline based on a responsibility chain mode, comprising a plurality of orderly connected processors Full data context Sequentially flowing through the processors: Wherein: Evidence matching processor Calculating an evaluation viewpoint Class summary sentence Semantic similarity between When the similarity exceeds a threshold When the reference anchor point is established; Suggestion generation processor Aiming at index dimension with score lower than preset value, searching teaching knowledge base to generate improvement suggestion; compliance filter processor And (3) carrying out compliance verification on the output text by using the sensitive word stock, and finally outputting a structured supervision report containing multi-mode evidence support.

Description

Intelligent supervision method based on multi-agent cooperation and dynamic graph arrangement Technical Field The invention belongs to the technical field of digital data processing and artificial intelligence, and further relates to an intelligent supervision method based on multi-agent cooperation and dynamic graph arrangement in an educational informatization scene. Background With the deep advancement of education informatization, the monitoring and evaluation of classroom teaching quality has become a key link for improving education level. Traditional classroom supervision mainly relies on specialists or educational staff to conduct manual lectures. The method has the advantages of deep evaluation dimension, limited labor cost, narrow coverage, strong subjectivity, feedback lag and other inherent defects, and is difficult to meet the requirement of large-scale normalized teaching quality monitoring. Aiming at the existing classroom teaching supervision system, the traditional informatization and statistical analysis method is mainly adopted, the teaching state identification and classification are generally completed based on classroom behavior counting, big data statistics or a single deep learning model, the analysis process takes result output as a guide, and the intelligent understanding and comprehensive reasoning capability of the teaching process is lacked. Under the background, the system is difficult to carry out multi-view collaborative analysis on complex classroom scenes, and the evaluation result often depends on a single model or fixed rules and lacks consistency guarantee and interpretable evidence support. To solve the efficiency problem of manual supervision, automated systems based on big data analysis and rule engines have been developed. Such prior art typically employs a "data acquisition-feature engineering-statistical analysis-manual review" model, or a "speech transcription-keyword matching-rule scoring" linear serial processing flow. However, this technical route has a significant disadvantage in facing complex real classroom scenarios, in that serial execution of the perception and evaluation tasks results in extremely high system delays for classroom videos of typically 45 minutes in duration. Big data analysis often focuses on post-class batch processing, and the demand of instant feedback cannot be met. The big data method based on statistics and keywords is difficult to understand complex teaching logic and interaction contexts of teachers and students, often only shallow indexes such as frequency, duty ratio and the like can be output, and qualitative evaluation on teaching content depth and teaching method application is lacking. Because the analysis result lacks the interpretability and the evidence support, the final supervision report always needs to be manually subjected to secondary review and color rendering, and the real full automation cannot be realized. How to utilize multiple intelligent agents to promote evaluation depth on the basis of inheritance of big data analysis breadth and solve the problems of efficient parallelism, time sequence synchronization, data consistency and compliance audit among heterogeneous intelligent agents is a technical problem to be solved urgently in constructing a new-generation intelligent teaching supervision system. Disclosure of Invention Aiming at the defects, the intelligent supervision method based on multi-agent cooperation and dynamic diagram arrangement is provided, and the efficient sensing, parallel analysis and consistency evaluation of multi-modal data in a classroom are realized by introducing a multi-agent division cooperation mechanism, a dynamic diagram execution plan, parallel task scheduling and an audit and report generation flow based on a responsibility chain, so that the automation level, analysis accuracy and system robustness of teaching supervision are improved. In order to achieve the above purpose, the present invention adopts the following technical scheme: An intelligent supervision method based on multi-agent cooperation and dynamic graph arrangement comprises the following steps: s1, intention recognition and dynamic programming, namely receiving natural language input instructions of a user by utilizing a coordinator intelligent body The method comprises the steps of constructing a global context vector by combining history session memory, inputting the global context vector into a global planning agent, and generating a directed acyclic graph DAG global execution plan comprising execution steps, input and output parameters and front-rear dependency relationships through large language model reasoning; S2, inputting the global execution plan generated in the step S1 into a parallel executor, automatically identifying that a parallel task group without strong dependency relationship comprises a data sensing task and an intelligent evaluation task, resetting a corresponding parallel branch completi