Intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement

Abstract

The invention discloses an intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement, and the method comprises the steps: analyzing a natural language instruction of a user through a coordinator agent, and constructing a global context in combination with a historical session and a course portrait; dynamically generating a supervision execution plan in the form of a directed acyclic graph (DAG) by a global planning agent; the parallel actuator agent analyzes the plan, distributes the parallelizable subtasks to a plurality of expert agents configured with different prompt word templates, and performs multi-view evaluation; through an independent state mark synchronization mechanism arranged in a directed acyclic graph (DAG) aggregation node, consistency aggregation is carried out on the output of each agent; and finally, inputting a convergence result into a processing assembly line based on a chain of responsibility mode, and outputting a structured supervision report through evidence matching, improved suggestion generation and compliance filtering. According to the method, the problems of task decomposition, parallel conflict, data competition and reasoning illusion in multi-agent collaboration are effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of digital data processing and artificial intelligence technology, and further relates to an intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement in the context of educational informatization. Background Technology

[0002] With the deepening of educational informatization, the monitoring and evaluation of classroom teaching quality has become a key link in improving the level of education. Traditional classroom supervision mainly relies on experts or academic staff to conduct manual observations. Although this method has a deep evaluation dimension, it is limited by human resources and has inherent defects such as narrow coverage, strong subjectivity, and delayed feedback, making it difficult to meet the needs of large-scale, routine monitoring of teaching quality.

[0003] Existing classroom teaching supervision systems primarily rely on traditional information technology and statistical analysis methods. They typically identify and classify teaching states based on classroom behavior counting, big data statistics, or single deep learning models. The analysis process is output-oriented, lacking intelligent understanding and comprehensive reasoning capabilities regarding the teaching process. Consequently, these systems struggle to conduct multi-perspective collaborative analysis of complex classroom scenarios, and evaluation results often depend on a single model or fixed rules, lacking consistency guarantees and interpretable evidence.

[0004] To address the efficiency issues of manual supervision, automated systems based on big data analytics and rule engines have emerged. These existing technologies typically employ a "data collection—feature engineering—statistical analysis—manual review" model, or a linear, sequential processing flow of "speech transcription—keyword matching—rule scoring." However, this approach has significant limitations when facing complex real-world classroom scenarios: for classroom videos typically lasting 45 minutes, sequentially executing perception and evaluation tasks leads to extremely high system latency. Big data analytics often focuses on post-class batch processing, failing to meet the need for immediate feedback. Statistical and keyword-based big data methods struggle to understand complex teaching logic and teacher-student interaction contexts, often only outputting superficial indicators such as frequency and percentage, lacking qualitative evaluations of the depth of teaching content and the application of teaching methods. Due to the lack of interpretability and evidentiary support for the analysis results, the final supervision report often still requires secondary manual review and polishing, failing to achieve true full automation.

[0005] How to enhance the depth of evaluation by leveraging multiple agents while building upon the breadth of big data analysis, and how to address the issues of efficient parallelism, temporal synchronization, data consistency, and compliance auditing among heterogeneous agents, are the key technical challenges that urgently need to be solved in building a new generation of intelligent teaching supervision systems. Summary of the Invention

[0006] To address the aforementioned shortcomings, this invention proposes an intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration. By introducing a multi-agent division of labor and cooperation mechanism, dynamic graph execution plan, parallel task scheduling, and a chain-of-responsibility-based audit and report generation process, this invention achieves efficient perception, parallel analysis, and consistent evaluation of multimodal classroom data, thereby improving the automation level, analytical accuracy, and system robustness of teaching supervision.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: An intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration includes the following steps: S1. Intent Recognition and Dynamic Programming: Utilizing a coordinator agent to receive natural language input commands from the user. By combining historical conversation memory to construct a global context vector, and inputting it into the global planning agent, a global execution plan containing a directed acyclic graph (DAG) including execution steps, input and output parameters and pre- and post-dependent relationships is generated through large language model reasoning. S2. Parallel Task Parsing and State Initialization: Input the global execution plan generated in step S1 into the parallel executor, automatically identify parallel task groups with no strong dependencies, including data-aware tasks and intelligent evaluation tasks, and reset the corresponding parallel branch completion flags in the global state bus to the incomplete state. Subsequently, the parallel executor triggers the execution entry of the perception subgraph corresponding to the data-aware task and the execution entry of the evaluation subgraph corresponding to the intelligent evaluation task simultaneously through the asynchronous scheduling mechanism. S3. Multimodal Data Perception and Feature Extraction: During the execution phase of the perception subgraph, the content summary node is called to perform ASR speech transcription and text summarization on the classroom audio and video stream to extract the core teaching links; at the same time, the trend analysis node is called to statistically analyze the frequency of teacher-student interaction and sentiment based on the time window, generate visualized classroom trend feature data, and update the completion mark of the perception branch after execution. S4. Multi-perspective evaluation and risk scanning: In the evaluation subgraph execution phase, based on the parallelizable subtasks in step S3, multiple expert agents with different prompt word templates are instantiated to form a multi-perspective expert evaluation module. The supervised object is analyzed from different dimensions, and the conflict of expert opinions is eliminated through the evaluation fusion node to generate a consistent evaluation conclusion. Then, the completion marker of the evaluation subtask is updated. S5. Convergence Synchronization Based on Independent Markers: The convergence synchronization node acts as a logical barrier, polling and checking the independent completion marks written in steps S3 and S4; only when all preset parallel branch marks are detected to be in the "complete" state, the logical block is released and the full converged data is transmitted downstream; if any branch is not completed, the current process is suspended to avoid data race and illusion problems. S6. Chain of Responsibility Audit and Report Generation: The full data collected in step S5 is input into the report generation pipeline built based on the chain of responsibility model. The data goes through data preprocessing, evidence chain matching between evaluation opinions and raw data, retrieval and generation of teaching improvement suggestions, and compliance content filtering in sequence. Finally, a structured supervision report with multimodal evidence support is output.

[0008] For the user's natural language input command in step S1 The coordinator agent first loads a length of Historical session sequence and course profile characteristics ; Utilize large language models to process users' natural language input commands Historical conversation sequence Course profile characteristics Perform joint encoding to construct the current global context vector. Input a global planning agent; the global planning agent is based on the global context vector. The system generates a global execution plan in JSON format, consisting of a directed acyclic graph (DAG) containing execution steps, input / output parameters, and pre- and post-dependencies, through reasoning based on thought chains. Global execution plan , among which, nodes Representing the Each execution step contains an attribute tuple. ,in Let each edge be a directed acyclic graph (DAG), representing the set of preceding dependent nodes for that step.

[0009] The specific method for step S2 is as follows: The parallel executor traverses the global execution plan. Construct a task dependency topology graph; define the set of currently completed steps as... For any step $S_i$, if it satisfies the dependency condition If it has not been executed, add it to the execution queue. .

[0010] The specific method for step S3 is as follows: Global execution plan for a directed acyclic graph (DAG) Input a parallel executor agent, and the parallel executor agent traverses the global execution plan. Construct a task dependency topology graph: Define the set of currently completed steps as... For any execution step If it satisfies the dependency condition If it has not been executed, add it to the execution queue. Automatically identify parallelizable subtasks: perception tasks With evaluation task The classroom audio and video stream data is input into a perception subgraph, which includes a process for generating a classroom content summary based on speech transcription and a process for extracting teaching behavior trend features based on time window statistics. The two processes are executed in parallel or serially to generate structured classroom perception feature data. 3.1 Content Summarization: Converting audio streams into timestamped text sequences. ,in, For text words, For time points; the text sequence with timestamps is processed using a sliding window algorithm. The lesson was divided into several segments, and a large language model was used to generate structured class notes. ; 3.2 Trend Feature Extraction: Divide the classroom timeline into... A time window of equal length For each window Statistics on the density of teacher-student interaction Construct a classroom trend feature matrix based on the proportion of voice usage $V_k$. ; After execution, the perception results are written to the global state, and a synchronization flag is set. Set to 1.

[0011] The specific method for step S4 is as follows: 4.1 Multi-perspective scoring: Instantiating multiple expert agents with different prompt word templates This constitutes a multi-perspective expert evaluation module; each expert evaluation module independently receives the classroom minutes output from step S3. Classroom Trend Feature Matrix Output score vector and text evaluation ; 4.2 Evaluation Fusion: A fusion node is introduced, and a weighted average method is used to aggregate the scores from each expert evaluation module to obtain the final score. Furthermore, a large-scale model is used to deduplicate and resolve conflicts in the text evaluations of each expert evaluation module, generating a comprehensive evaluation report. ; in, For the first The confidence weights of each expert are assigned; after execution, the evaluation results are written to the global state, and a synchronization flag is set. Set to 1.

[0012] Step S5 designs a logical barrier based on independent state tags, which is implemented as follows: When the aggregation synchronization node is triggered by any upstream branch, it executes a status check function. : If the result is WAITING, the current triggering thread will terminate directly with END, without performing any subsequent operations, and the system will be in a suspended state to wait for other branches; If the result is READY, it indicates that the data for all parallel tasks is ready in the global state bus, the system is unblocked, and downstream report generation is activated.

[0013] The specific method for step S6 is as follows: Construct a report generation pipeline based on the chain of responsibility pattern, containing several sequentially connected processors. Full data context Flowing sequentially through each processor: in: Evidence Matching Processor : Calculation of evaluation viewpoints Sentences from class notes semantic similarity between When the similarity exceeds the threshold At that time, establish a reference anchor point; It is recommended to generate a processor For indicator dimensions with scores below preset values, the teaching knowledge base is searched to generate improvement suggestions; Compliance filter processor The output text is validated for compliance using a sensitive word database, and a structured supervisory report containing multimodal evidence is finally output.

[0014] Compared with the prior art, the advantages of the present invention are: First, addressing the problem that existing classroom teaching supervision systems often rely on a single model or centralized analysis process, making it difficult to collaboratively understand complex classroom teaching processes from multiple perspectives, this invention, in step S4, schedules multiple expert agents with different evaluation strategies and analytical focuses in parallel to independently analyze and score the unified classroom perception feature data constructed in step S3. Furthermore, it resolves conflicts and constrains consistency of multi-source evaluation results through evaluation fusion nodes, thereby avoiding evaluation bias caused by a single analysis model in complex teaching scenarios. This allows the supervision results to comprehensively reflect multiple dimensions such as instructional design, teaching pace, and teacher-student interaction, significantly improving the objectivity and reliability of teaching evaluation.

[0015] Secondly, in response to the problem that existing supervision systems generally adopt linear serial processing flow, which makes it difficult to execute multimodal perception analysis and teaching evaluation tasks in parallel and limits the system processing efficiency, this invention generates a global execution plan in the form of a directed acyclic graph through a global planning agent in step S1, and automatically parses the task dependencies in step S2 by a parallel executor, identifying the data perception task and the intelligent evaluation task as parallel branches without strong dependencies and triggering their execution simultaneously. This achieves parallel processing of multiple tasks while ensuring logical correctness, significantly improving the execution efficiency and system throughput of the overall supervision process.

[0016] Third, in order to address the issues of data competition, inconsistent states, and evaluation bias caused by premature reasoning based on incomplete information that are prone to occur in parallel analysis scenarios, this invention sets up a convergence synchronization node based on independent completion markers as a logical barrier in step S5, and performs state polling control on the parallel execution branches of steps S3 and S4. The subsequent report generation process is only triggered after all perception and evaluation tasks have been completed. From the system control level, this effectively avoids the generation of evaluation illusions and erroneous reasoning results, and significantly improves the consistency and stability of supervision results.

[0017] Fourth, in response to the problem that existing classroom teaching supervision results lack evidence support and process traceability, this invention introduces a report generation pipeline based on the chain of responsibility model in step S6. The evaluation conclusions are anchored to the evidence chain with the original classroom perception data in sequence, and combined with the generation of teaching improvement suggestions and compliance content filtering, so that the final output supervision report has evaluation conclusions, evidence basis and improvement suggestions, thereby significantly improving the interpretability and practical application value of teaching supervision results.

[0018] In summary, the multi-agent collaborative execution process constructed by the present invention through steps S1 to S6 enables efficient multimodal perception, parallel analysis, and consistent evaluation of the classroom teaching process. It is significantly superior to the prior art in terms of system execution efficiency, evaluation stability, and result interpretability, and has good engineering feasibility and application value. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the implementation of the present invention.

[0020] Figure 2 This is the interface for displaying the overall supervision status of the system under conditions of parallel operation in multiple classrooms.

[0021] Figure 3 This is a comprehensive analysis interface for supervising single-session classroom teaching.

[0022] Figure 4 In a simulation environment, this invention enables the automatic generation of multi-dimensional teaching quality evaluation results for a single classroom teaching process.

[0023] The invention will now be further described with reference to the accompanying drawings.

[0024] To make the objectives, technical solutions, and beneficial effects of the present invention clearer and more complete, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only a part of the embodiments of the present invention, and not all of them. Other implementation methods obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0025] As attached Figure 1 As shown, this embodiment provides a classroom teaching intelligent supervision method based on multi-agent collaboration and dynamic graph arrangement. The overall process includes the following steps: S1. Intent Recognition and Dynamic Programming: The Coordinator Agent receives the user's natural language guidance instructions, combines historical conversation memory to construct a global context, and inputs it into the Global Planner Agent. Through large language model reasoning, a global execution plan containing a directed acyclic graph (DAG) containing execution steps, input and output parameters, and pre- and post-dependencies is generated. Specifically, step S1 involves processing the natural language supervision instructions input by the user. ,in Indicates the first Each word element, Given the instruction length, the coordinator agent first loads an instruction of length from the session memory module. Historical session sequence And at the same time, obtain the course profile features corresponding to the current classroom. The course profile features are used to describe the course's basic attributes, teaching objectives, and historical teaching characteristics. Subsequently, the coordinator agent transmits the user instructions... Historical conversation sequence and course profile characteristics Joint encoding is performed, and the input is fed into the context understanding module based on a large language model. Semantic modeling is then used to obtain the global context representation corresponding to the current supervision task. In obtaining the global context representation Then, it is input into the global planning agent, which performs task decomposition and process planning for the current supervised task based on the reasoning mechanism, generating a global execution plan in the form of a directed acyclic graph: In any execution step Represented as attribute tuples: in, A unique identifier representing a step; Indicates the task type to which the step belongs, used to distinguish between data-aware tasks and intelligent evaluation tasks; This represents the set of preceding dependent steps for this step, used to describe the execution constraints between steps.

[0026] S2. Parallel Task Parsing and State Initialization: Input the global execution plan generated in step S1 into the Parallel Executor, automatically identify parallel task groups without strong dependencies (i.e., data-aware tasks and intelligent evaluation tasks), reset the corresponding parallel branch completion flags in the global state bus to the incomplete state, and then simultaneously activate the execution entry points of the awareness subgraph and the evaluation subgraph. In step S2, the system takes the global execution plan generated in step S1 as input and parses the plan using a parallel executor. The parallel executor analyzes the pre- and post-deployment dependencies between execution steps, automatically identifying execution steps without strong dependencies, thus determining the set of tasks that can be executed in parallel. Before identifying parallel tasks, the system first initializes and resets the synchronization flags in the global state to ensure that each parallel branch is in a unified incomplete state at startup. Subsequently, the parallel executor distributes the identified tasks without strong dependencies to the execution entry points of the data perception module and the intelligent evaluation module, and simultaneously triggers the execution of these modules through an asynchronous scheduling mechanism. This step achieves decoupling and parallel scheduling of different analysis tasks, avoiding the efficiency bottleneck caused by the fixed task order in traditional integrated linear processes, and providing a foundation for parallel execution of subsequent multimodal perception analysis and multi-perspective evaluation.

[0027] S3. Multimodal Data Perception and Feature Extraction: During the execution phase of the perception subgraph, the content summary node is called to perform ASR speech transcription and text summarization on the classroom audio and video stream to extract the core teaching links; at the same time, the trend analysis node is called to statistically analyze the frequency of teacher-student interaction and sentiment based on the time window, generate visualized classroom trend feature data, and update the completion mark of the perception branch after execution. In step S3, the system enters the data perception module to analyze and process the audio and video stream data generated during classroom teaching. The data perception module includes a content summarization node and a trend analysis node, used to perceive and extract features from classroom data from different dimensions. Specifically, the content summarization node performs speech-to-text transcription on the classroom audio stream and extracts summaries from the transcribed text to obtain the core content and key teaching segments of the classroom teaching process. The trend analysis node uses a time window as the basic unit to statistically analyze teacher-student interaction behavior during classroom teaching, calculating indicators such as interaction frequency and speech ratio, thereby depicting the changing trends of classroom teaching activities over time. Through the above processing, the original unstructured classroom audio and video data is transformed into structured teaching behavior feature data and written into the global state for use by the subsequent intelligent evaluation module. This step provides a unified and reusable data foundation for teaching evaluation.

[0028] S4. Multi-perspective evaluation and risk scanning: During the evaluation subgraph execution phase, multiple expert evaluation agents with different personalities (such as supportive, strict, and data-driven) are scheduled in parallel. Based on the perception data (or preprocessed data) generated in step S3, they are independently scored and risk points are scanned against the teaching syllabus. The expert opinion conflicts are eliminated through the evaluation fusion node to generate a consistent evaluation conclusion, and then the completion mark of the evaluation branch is updated. In step S4, the system enters the intelligent evaluation module to evaluate the teaching quality of the classroom teaching process from multiple perspectives and dimensions. This intelligent evaluation module schedules multiple expert agents with different evaluation strategies to analyze the classroom teaching process from different emphases. Each expert agent, based on the teaching behavior characteristic data output in step S3 and combined with the teaching syllabus or preset evaluation standards, performs risk scanning and multi-dimensional scoring of the classroom teaching process, forming its own independent evaluation results. Subsequently, the system comprehensively processes the multi-source evaluation results through an evaluation fusion node, resolving conflicts between differing evaluation opinions and generating a consistent teaching quality evaluation result. By introducing a multi-expert collaborative evaluation and fusion mechanism, the system effectively reduces the bias that may arise from a single evaluation strategy, improving the stability and reliability of the evaluation results.

[0029] S5. Convergence Synchronization Based on Independent Markers: The convergence synchronization node (Graph Join Node) is used as a logical barrier to poll and check the independent completion marks written in steps S3 and S4; the logical block is lifted and the full converged data is passed downstream only when all preset parallel branch marks are detected to be in the "complete" state; if any branch is not completed, the current process is suspended to avoid data race and illusion problems. In step S5, the system uses a convergence synchronization node as a logical barrier to synchronize the parallel execution of each analysis branch. The convergence synchronization node uses a polling method to detect the independent completion flags written by the data perception branch and the intelligent evaluation branch to determine the execution status of each parallel task. Only when all preset parallel branches are detected to be complete does the convergence synchronization node release the logical block and transmit the full data, including the perception and evaluation results, to the downstream process; if any branch is not yet complete, the current process remains in a waiting state and does not trigger subsequent processing. This logical barrier mechanism avoids data race problems caused by incomplete data during parallel processing, and also prevents the evaluation module from generating inference bias or evaluation illusions when the perceived data is incomplete.

[0030] S6. Chain of Responsibility Audit and Report Generation: The full data aggregated in step S5 is input into a report generation pipeline built on the Chain of Responsibility model. This pipeline sequentially undergoes data preprocessing, evidence chain matching between evaluation viewpoints and original data, retrieval and generation of teaching improvement suggestions, and compliance content filtering, ultimately outputting a structured supervision report supported by multimodal evidence. In step S6, the system inputs the full data aggregated in step S5 into the processing pipeline built on the Chain of Responsibility model for further auditing and organizing the supervision results. The processing pipeline consists of multiple processors with clearly defined functions and sequential connections. These processors process the data in a predetermined order. Specifically, firstly, the data preprocessing processor standardizes the format and organizes the content of the aggregated data; then, the evidence chain anchoring processor establishes the correlation between evaluation viewpoints and original classroom data to form traceable evidence support; next, for identified teaching problems, the teaching improvement suggestion processor retrieves and generates corresponding improvement suggestions; finally, the compliance content filtering processor verifies the output content to ensure that the generated results meet relevant regulatory requirements. Through the above chain-of-responsibility process, the system ultimately outputs a structured classroom teaching supervision report, which organically combines evaluation conclusions, evidence, and improvement suggestions, thereby enhancing the interpretability and practicality of the teaching supervision results.

[0031] Simulation experiment Table 1.1 Hardware Configuration Table Hardware projects Configuration parameters Processor (CPU) QEMU Virtual CPU (based on KVM virtualization), 32 cores, clock speed approximately 2.10 GHz Processor architecture x86_64 Core Structure Single-threaded architecture, 8 sockets × 4 cores RAM 62 GB Disk storage 1 TB (Logical Volume LVM Mount) GPU acceleration Not configured (CPU inference environment) Virtualization KVM Full Virtualization The simulation environment is shown in Table 1.1. In the simulation verification environment, the system uses real-time audio and video streams collected in actual classrooms as the input data source. On the server side, a multimodal data perception module, a multi-agent collaborative evaluation module, a parallel scheduling and synchronization control module, and a chain-of-responsibility report generation module are deployed to construct a complete intelligent classroom teaching supervision and operation system. The video stream is collected using high-definition camera equipment at a fixed frame rate and transmitted in real time through a streaming media protocol. The audio signal is collected synchronously and processed for speech recognition. The running end aligns the audio and video data based on a unified timestamp mechanism and processes it in segments according to preset time windows. Within each time window, student behavior feature extraction, classroom participation statistics, and teaching semantic information extraction are completed in parallel. The multimodal results are aggregated to form window-level classroom fact objects. Subsequently, the parallel scheduling mechanism triggers multiple evaluation agents to perform collaborative reasoning and conflict resolution based on the classroom facts and corresponding evidence indexes, outputting stable and consistent teaching quality evaluation results. Finally, the chain-of-responsibility processing flow completes evidence verification, logical consistency review, and structured report generation and storage archiving, thereby realizing the full-process operation verification of the intelligent classroom teaching supervision method under real data conditions.

[0032] By simulating the continuous operation of multiple classes, the overall processing capability of the system, the stability of evaluation results, and the effectiveness of the supervision report generation were comprehensively verified.

[0033] like Figure 2 The image shows the overall supervision status display interface of the system under the condition of multiple classrooms operating in parallel. The system can perform real-time statistical analysis on a large number of classroom teaching processes, and automatically summarize key indicators such as the number of evaluations, classroom data scale, evaluation result distribution, and teaching behavior trends.

[0034] Simulation results show that, with the support of a parallel scheduling mechanism based on dynamic graph execution plans, the system can simultaneously complete multimodal data perception and analysis and multi-agent teaching evaluation tasks, maintain stable operation, and significantly improve the processing capacity and system throughput efficiency of the classroom supervision process in high-concurrency scenarios.

[0035] like Figure 3 The image shows the comprehensive analysis interface for supervising a single classroom lesson. The system can simultaneously display classroom audio and video streams, key time points in the teaching process, statistical results of multimodal perception features, and teaching evaluation scores from multiple perspectives. During the data perception phase, the system automatically completes classroom speech transcription, extracts teaching content summaries, and analyzes trends in teacher-student interaction. In the intelligent evaluation phase, multiple expert agents independently analyze and score the classroom teaching process from different dimensions based on the aforementioned perception results, and generate consistent teaching quality evaluation results through a fusion mechanism.

[0036] Simulation results show that the present invention can achieve comprehensive perception and multi-perspective collaborative evaluation of the classroom teaching process without human intervention, verifying the effectiveness and stability of the proposed multi-agent collaborative mechanism in complex teaching scenarios.

[0037] like Figure 4 As shown, in a simulation environment, this invention can automatically generate multi-dimensional teaching quality evaluation results for a single classroom teaching process, and simultaneously output diagnostic analysis and improvement suggestions corresponding to each evaluation dimension. The system uses a multi-agent collaborative evaluation mechanism to independently analyze multiple dimensions such as teaching content quality, teaching structure design, classroom pace control, and student participation, and forms consistent scoring results through evaluation fusion nodes, avoiding bias problems caused by a single evaluation model. Simultaneously, relying on a chain-of-responsibility audit mechanism, each evaluation conclusion can be linked to corresponding classroom behavior characteristics and time window evidence, achieving traceable expression of evaluation results and automatically generating targeted teaching improvement suggestions.

[0038] Simulation results show that, compared with the traditional single-model evaluation method, the present invention has significantly improved the stability, consistency and interpretability of evaluation results, verifying the effectiveness of multi-agent collaborative analysis and evidence chain closed-loop mechanism in teaching supervision scenarios.

[0039] Through the process described in the above embodiments, the present invention realizes multimodal data perception of the classroom teaching process, multi-agent collaborative evaluation, and automatic generation of supervision reports based on evidence chains, effectively improving the intelligence level, evaluation consistency, and interpretability of classroom teaching supervision.

Claims

1. An intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration, characterized in that, Includes the following steps: S1. Intent Recognition and Dynamic Programming: Utilizing a coordinator agent to receive natural language input commands from the user. By combining historical conversation memory to construct a global context vector, and inputting it into the global planning agent, a global execution plan containing a directed acyclic graph (DAG) including execution steps, input and output parameters and pre- and post-dependent relationships is generated through large language model reasoning. S2. Parallel Task Parsing and State Initialization: Input the global execution plan generated in step S1 into the parallel executor, automatically identify parallel task groups with no strong dependencies, including data-aware tasks and intelligent evaluation tasks, and reset the corresponding parallel branch completion flags in the global state bus to the incomplete state. Subsequently, the parallel executor triggers the execution entry of the perception subgraph corresponding to the data-aware task and the execution entry of the evaluation subgraph corresponding to the intelligent evaluation task simultaneously through the asynchronous scheduling mechanism. S3. Multimodal Data Perception and Feature Extraction: During the execution phase of the perception subgraph, the content summary node is called to perform ASR speech transcription and text summarization on the classroom audio and video stream to extract the core teaching links; at the same time, the trend analysis node is called to statistically analyze the frequency of teacher-student interaction and sentiment based on the time window, generate visualized classroom trend feature data, and update the completion mark of the perception branch after execution. S4. Multi-perspective evaluation and risk scanning: In the evaluation subgraph execution phase, based on the parallelizable subtasks in step S3, multiple expert agents with different prompt word templates are instantiated to form a multi-perspective expert evaluation module. The supervised object is analyzed from different dimensions, and the conflict of expert opinions is eliminated through the evaluation fusion node to generate a consistent evaluation conclusion. Then, the completion marker of the evaluation subtask is updated. S5. Convergence Synchronization Based on Independent Markers: The convergence synchronization node acts as a logical barrier, polling and checking the independent completion marks written in steps S3 and S4; only when all preset parallel branch marks are detected to be in the "complete" state, the logical block is released and the full converged data is transmitted downstream; if any branch is not completed, the current process is suspended to avoid data race and illusion problems. S6. Chain of Responsibility Audit and Report Generation: The full data collected in step S5 is input into the report generation pipeline built based on the chain of responsibility model. The data goes through data preprocessing, evidence chain matching between evaluation opinions and raw data, retrieval and generation of teaching improvement suggestions, and compliance content filtering in sequence. Finally, a structured supervision report with multimodal evidence support is output.

2. The intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration according to claim 1, characterized in that, For the user's natural language input command in step S1 The coordinator agent first loads a length of Historical session sequence and course profile characteristics ; Utilizing large language models to process users' natural language input commands Historical conversation sequence Course profile characteristics Perform joint encoding to construct the current global context vector. Input a global planning agent; the global planning agent is based on the global context vector. The system generates a global execution plan in JSON format, consisting of a directed acyclic graph (DAG) containing execution steps, input / output parameters, and pre- and post-dependencies, through reasoning based on thought chains. Global execution plan , among which, nodes Representing the Each execution step contains an attribute tuple. ,in Let each edge be a directed acyclic graph (DAG), representing the set of preceding dependent nodes for that step.

3. The intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration according to claim 1, characterized in that, The specific method for step S2 is as follows: The parallel executor traverses the global execution plan. Construct a task dependency topology graph; Define the set of currently completed steps as For any step If it satisfies the dependency condition If it has not been executed, it is added to the execution queue. .

4. The intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration according to claim 1, characterized in that, The specific method for step S3 is as follows: Global execution plan for a directed acyclic graph (DAG) Input a parallel executor agent, and the parallel executor agent traverses the global execution plan. Construct a task dependency topology graph: Define the set of currently completed steps as... For any execution step If it satisfies the dependency condition If it has not been executed, it is added to the execution queue. Automatically identify parallelizable subtasks: perception tasks With evaluation task The classroom audio and video stream data is input into a perception subgraph, which includes a process for generating a classroom content summary based on speech transcription and a process for extracting teaching behavior trend features based on time window statistics. The two processes are executed in parallel or serially to generate structured classroom perception feature data. 3.1 Content Summarization: Converting audio streams into timestamped text sequences. ,in, For text words, For time points; the text sequence with timestamps is processed using a sliding window algorithm. The lesson was divided into several segments, and a large language model was used to generate structured class notes. ; 3.2 Trend Feature Extraction: Divide the classroom timeline into... A time window of equal length For each window Statistics on the density of teacher-student interaction Construct a classroom trend feature matrix based on the proportion of voice usage $V_k$. ; After execution, the perception results are written to the global state, and a synchronization flag is set. Set to 1.

5. The intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration according to claim 1, characterized in that, The specific method for step S4 is as follows: 4.1 Multi-perspective scoring: Instantiating multiple expert agents with different prompt word templates This constitutes a multi-perspective expert evaluation module; each expert evaluation module independently receives the classroom minutes output from step S3. Classroom Trend Feature Matrix Output score vector and text evaluation ; 4.2 Evaluation Fusion: A fusion node is introduced, and a weighted average method is used to aggregate the scores from each expert evaluation module to obtain the final score. Furthermore, a large-scale model is used to deduplicate and resolve conflicts in the text evaluations of each expert evaluation module, generating a comprehensive evaluation report. ; in, For the first The confidence weights of each expert are assigned; after execution, the evaluation results are written to the global state, and a synchronization flag is set. Set to 1.

6. The intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration according to claim 1, characterized in that, Step S5 designs a logical barrier based on independent state tags, which is implemented as follows: When the aggregation synchronization node is triggered by any upstream branch, it executes a status check function. : If the result is WAITING, the current triggering thread will terminate directly with END, without performing any subsequent operations, and the system will be in a suspended state to wait for other branches; If the result is READY, it indicates that the data for all parallel tasks is ready in the global state bus, the system is unblocked, and downstream report generation is activated.

7. The intelligent supervision method based on multi-agent collaboration and dynamic graph orchestration according to claim 1, characterized in that, The specific method for step S6 is as follows: Construct a report generation pipeline based on the chain of responsibility pattern, containing several sequentially connected processors. Full data context Flowing sequentially through each processor: in: Evidence Matching Processor : Calculation of evaluation viewpoints Sentences from class notes semantic similarity between When the similarity exceeds the threshold At that time, establish a reference anchor point; It is recommended to generate a processor For indicator dimensions with scores below preset values, the teaching knowledge base is searched to generate improvement suggestions; Compliance filter processor The output text is validated for compliance using a sensitive word database, and a structured supervisory report containing multimodal evidence is finally output.

Images