A virtual student classroom simulation system and method based on a personality-knowledge-dynamics three-module framework

Abstract

This invention belongs to, but is not limited to, the field of virtual student classroom simulation technology, and discloses a virtual student classroom simulation system and method based on a three-module framework of personality, knowledge, and dynamics. The system includes: a multimodal data acquisition subsystem for real-time extraction of three types of input signals, aligning them with a unified timestamp to form multimodal feature frames, and sending them to the PDK engine; a PDK virtual student simulation engine, including P, K, and D modules, for virtual student state modeling; and a teacher behavior feedback and evaluation subsystem for generating adaptive teaching ability evaluation reports. This invention can achieve heterogeneous dynamic simulation at the class level, real-time coupling of virtual student states and teacher behaviors, and the behavioral output of this invention has a fully interpretable causal root, without the need for real student participation; the training scenario can be reproduced with infinite precision.

Description

Technical Field

[0001] This invention belongs to, but is not limited to, the field of virtual student classroom simulation technology, and particularly relates to a virtual student classroom simulation system and method based on a three-module framework of personality-knowledge-dynamics. Background Technology

[0002] Pre-service training for teacher trainees is a core component of the teacher education system, directly impacting the quality of the basic education teacher workforce. Traditionally, training teacher trainees' teaching abilities primarily relied on experts observing and evaluating lessons or arranging trial lectures in real classrooms. However, with the development of educational information technology, various technical solutions to support teaching training have emerged, currently forming three main technical approaches:

[0003] The first type is the expert observation and evaluation model, in which the supervising teacher or external expert observes the classroom teaching of the student teacher and provides qualitative evaluations based on teaching experience. This model has significant limitations in practice: the evaluation process is labor-intensive, resulting in student teachers only receiving a limited number of evaluation opportunities throughout the training period; furthermore, the presence of experts can interfere with the naturalness of teacher-student interaction, making it difficult for the measured teacher performance to accurately reflect their actual classroom teaching ability.

[0004] The second category is microteaching and peer role-playing. Some universities use microteaching methods, where peer students play the roles of middle school students to conduct trial lectures with student teachers. While this method allows for repeated training, the role-players already have a certain understanding of the teaching content and cannot accurately reflect the cognitive confusion of real students. Furthermore, peer feedback is mostly based on subjective feelings, lacks objectivity, and is difficult to accurately represent the diverse behavioral characteristics of real student groups.

[0005] The third category is intelligent classroom analysis systems, which have been widely used in the field of smart education in recent years. These systems use real classrooms as the data collection object, deploying cameras and microphone arrays in the classroom and utilizing computer vision and natural language processing technologies to conduct real-time or post-class automated analysis and quantitative evaluation of teacher and student classroom behavior. Representative domestic products integrate facial recognition, posture estimation, and speech recognition technologies to output dual-dimensional evaluation reports on student participation and teacher behavior. Representative international products, based on speech natural language processing technology, automatically detect the ratio of teacher and student speech and teacher questioning patterns, providing teachers with asynchronous post-class feedback. They also support asynchronous viewing and evidence-based teaching research based on video recording and AI annotation, analyzing teacher positioning heatmaps, blackboard writing frequency, and patrol coverage in real time to generate visual diagnostic reports.

[0006] While the aforementioned intelligent classroom analysis systems have made significant progress in the refined observation and quantitative feedback of teacher behavior, their core focus is on post-event analysis of real classrooms or providing quantitative reports on behavioral statistics. This presents two fundamental limitations that prevent their direct application to training teacher trainees' proactive teaching abilities: First, the system evaluates real students, and the acquisition of evaluation data depends on the existence of a real classroom; it lacks the ability to construct a training environment without real students. Second, the student states collected and analyzed by the system (participation, facial expressions, posture) are passive records of overt behavior; the system itself lacks the simulation capability to dynamically update student cognitive states based on teacher input and proactively generate student behavioral responses. Therefore, existing intelligent classroom analysis systems are essentially classroom observation tools rather than classroom simulators, failing to meet the training needs of teacher trainees for repeated practice and causal diagnostic feedback in a controlled environment.

[0007] Based on the above analysis, the urgent technical problems that need to be solved in the existing technology are:

[0008] (1) Lack of simulation ability for dynamic heterogeneous student groups

[0009] In existing virtual systems, the reaction logic of all virtual students in the same class is the same or highly homogeneous. In real classrooms, different students have significant differences in personality traits, prior knowledge levels, and emotional states. These differences profoundly affect teachers' classroom management decisions and teaching strategy choices. Existing systems cannot reproduce this heterogeneity, resulting in a serious lack of representativeness in training scenarios.

[0010] (2) There is a lack of real-time coupling mechanism between the virtual student status and the teacher's behavioral input.

[0011] In existing systems, the virtual student's behavior cannot be perceived and responded to in real time by the teacher's multimodal input, such as vocal emotion, spatial movement, and gaze direction. This results in insufficient simulation realism, making it difficult for teachers to obtain effective contextual feedback during training, and significantly reducing the training value.

[0012] (3) The lack of knowledge state modeling makes it impossible to reflect the dynamics of students' cognition.

[0013] Existing solutions do not incorporate knowledge tracing technology into virtual student modeling. Students' levels of confusion and rate of knowledge acquisition in real classrooms vary from person to person and change dynamically with the teaching content. Current systems cannot capture this dynamic cognitive process, and therefore cannot support the training of teachers' ability to provide differentiated instructional responses to different areas of knowledge weakness.

[0014] (4) The behavioral output lacks an explainable causal root.

[0015] The lack of a causal mechanism makes training feedback unable to accurately target teachers' specific skill gaps, and the evaluation results only remain at the level of behavior counting statistics, failing to support diagnostic improvement guidance.

[0016] (5) The AI ​​classroom analysis system only observes and does not simulate, and cannot support the needs of active training.

[0017] Existing AI classroom analysis systems are limited to observing and statistically analyzing behavior in real classrooms. They cannot proactively generate student behavioral responses, nor can they construct controllable training scenarios without the participation of real students. Pre-service teacher trainees lack stable real classes for repeated practice and cannot afford the ethical risks of repeated trial and error with real students. None of the aforementioned systems possess the ability to construct reproducible, configurable, and diagnosable virtual training environments, and their analysis reports lack modeling of the causal links between teacher behavior and changes in student cognition, thus failing to directly translate into guidance for teacher trainees' skill improvement.

[0018] Existing technologies still have the following pressing technical problems: lack of simulation capabilities for dynamic and heterogeneous student groups; the reaction logic of all virtual students in the same class is the same or highly homogeneous, failing to reproduce the significant differences in personality traits, prior knowledge levels, and emotional states among students in a real classroom; lack of a real-time coupling mechanism between virtual student states and teacher behavioral input, failing to perceive and respond to multimodal inputs such as teacher's voice and emotion, spatial movement, and gaze direction in real time; lack of knowledge state modeling, failing to introduce knowledge tracking technology into virtual student modeling, failing to capture the dynamic process of student cognition; and lack of explainable causal roots for behavioral output, making it impossible for training feedback to accurately target the teacher's specific skill deficiencies. Summary of the Invention

[0019] To address the problems existing in the prior art, this invention provides a virtual student classroom simulation system and method based on a three-module framework of personality, knowledge, and dynamics.

[0020] This invention is implemented as follows: a virtual student classroom simulation system based on a three-module framework of Personality, Knowledge, and Dynamics, comprising:

[0021] The multimodal data acquisition subsystem is used to extract three types of input signals in real time, align them with a unified timestamp to form a multimodal feature frame, and send it into the PDK engine.

[0022] The PDK virtual student simulation engine includes P, K, and D modules for virtual student state modeling. The P module models the stable personality characteristics of each virtual student agent. The K module dynamically tracks the knowledge mastery status of each virtual student agent on the current course content and outputs a confusion index to drive subsequent behavior generation. The D module combines the internal states output by the P and K modules with real-time input signals and the class environment to generate the external behavior output of each virtual student agent at the current moment through causal decision logic.

[0023] The teacher behavior feedback and evaluation subsystem is used to post-process the teacher behavior data and virtual student status response data recorded throughout the process, and automatically generate an adaptive teaching ability evaluation report in four dimensions.

[0024] Furthermore, the multimodal data acquisition subsystem deploys the following sensing facilities in the physical environment of the virtual classroom:

[0025] (1) Camera in the teacher area (Camera-T):

[0026] Mounted directly above the blackboard, its field of view covers the teacher's standing area and the entire blackboard. It uses a human posture estimation model to analyze the teacher's body movements and standing position coordinates in real time. The network outputs an attention orientation vector by estimating the gaze direction. The system uses a Visual Language Model (VLM) to periodically (every 30 seconds) capture and parse images of the blackboard area, extracting the descriptive text of the knowledge points currently written on the board. ;

[0027] (2) Student area cameras (Camera-S):

[0028] Installed at the rear of the classroom, with a field of view covering the entire seating area, it is used for panoramic recording and assisting in behavioral verification, especially the teacher's position in the classroom and interaction with student terminals.

[0029] (3) Virtual Student Tablet (VST):

[0030] Each virtual student is represented by a tablet terminal, with a total of N tablet terminals arranged in the seating area according to the layout of a real classroom desk. Each tablet terminal is equipped with: a screen to display the virtual student's facial expression animation and body state; a microphone to collect the teacher's voice that can be heard from the seat position; a speaker to play the generated voice response when the virtual student is triggered to raise their hand to answer; and a camera to identify the relative distance to the teacher and to assess the size and clarity of the blackboard writing.

[0031] Furthermore, the multimodal data acquisition subsystem extracts three types of input signals from the aforementioned sensors in real time, the three types of input signals including:

[0032] (1) Signal S1: Teacher speech transcription and emotion annotation

[0033] Each tablet terminal runs a local speech recognition program that transcribes the speech stream captured by the terminal's microphone in real time and outputs the transcribed text fragment within the current time window. The current time window is a 5-second sliding window. The transcribed text is simultaneously fed into a lightweight sentiment classification model running on a tablet terminal, which outputs sentence-by-sentence sentiment polarity labels. Together with the confidence score, they constitute signal S1;

[0034] (2) Signal S2: Teacher's spatial position and attention orientation

[0035] The coordinates of the human body key points output by Camera-T are transformed to calculate the teacher's current position in the classroom coordinate system. Then calculate the relationship between the teacher and the first Euclidean distance between tablet terminals:

[0036] ,

[0037] As a teacher to virtual students The physical proximity metric is input into the PDK engine, and the attention orientation vector is output by the gaze direction estimation network. After projection calculation, determine whether the teacher is currently facing the first... Tablet terminal, outputs Boolean attention markers , and Together they constitute signal S2;

[0038] (3) Signal S3: Blackboard image semantics

[0039] Camera-T periodically captures screenshots of the blackboard area, triggered every 30 seconds or when the system detects significant content changes. The screenshots are then parsed by VLM to extract the descriptive text of the knowledge points currently written on the board. The K module of the PDK engine is used as the basis for updating knowledge content.

[0040] Furthermore, the PDK engine instantiates N virtual student agents at the start of a session: VSA1… Each agent Defined by the internal state vectors of the three modules, the state of each module at each time step (Period 1 second) Updated collaboratively based on input signals S1, S2, and S3.

[0041] Furthermore, the P module, each agent A personality state vector is assigned during session initialization. It includes the following three parameter dimensions:

[0042] Emotional sensitivity Quantitative Proxy The intensity of response to teachers' emotional cues (including tone of voice, eye contact, and verbal reinforcement);

[0043] Participation tendency Control Agent The basic probability of spontaneously participating in classroom interaction without being called upon by the teacher. Correspondingly, avoid participation. Corresponding to neutral, Corresponding to active participation, Directly participate in the hand-raising behavior triggering judgment in module D;

[0044] Social conformity index Modeling Agent The degree to which the participation status is influenced by the behavior of the surrounding virtual student group, during the generation of behavior in module D, by the agent. The probability distribution of behavior will be based on As the weight, shift towards the average behavior distribution of the current neighboring agents:

[0045]

[0046] Personality state vector Initialize based on the class template at the start of the session. The system has at least 3 preset class templates:

[0047]

[0048] During the conversation, Not completely static, when the teacher acts as an agent The cumulative number of interactions exceeds the threshold (Default 3 times), agent participation tendency Will The range is shifted 0.3 in the positive direction to simulate the educational phenomenon that sustained teacher attention gradually encourages introverted students to participate.

[0049] Furthermore, in the K module, each agent Maintaining the knowledge state vector ,in Indicates agent Knowledge units in the current course knowledge graph The degree of mastery, This is the set of knowledge graph nodes for this session. The knowledge graph is defined by the course configuration file during session initialization, and the node types must include at least:

[0050] Conceptual knowledge nodes: such as definitions, theorems, and concepts;

[0051] Procedural knowledge nodes: such as algorithm steps and operation methods;

[0052] Migrate application nodes: such as application scenarios and cross-domain connections;

[0053] The nodes are labeled with prerequisite relationships at each time step. The K module calls an LLM (Large Language Model) agent based on the RAG (Retrieval Enhancement Generation) role-based prompting framework to process the current input:

[0054] Input construction: The current agent Knowledge state vector Current teacher transcription fragments Blackboard writing semantics and agents Error type archive (Randomly assigned during initialization, describing the typical errors the agent is prone to make on which type of knowledge node) Concatenated into a context and sent to the LLM Agent;

[0055] LLM Agent Inference: The system prompts the agent to set its role as a learning and cognitive simulator, requiring it to update the mastery probability of knowledge units based on the teaching content and return the result in JSON format. This represents the increment in the degree of mastery; a negative value indicates forgetting or confusion.

[0056] Status Update:

[0057]

[0058] The K module also outputs a real-time perplexity index. Defined as the set of knowledge units involved in the current teaching time step. Average level of lack of mastery:

[0059]

[0060] The D module is directly input as the core criterion for triggering behavior, and also serves as the raw data for evaluating the appropriateness of knowledge explanation in the subsystem.

[0061] Furthermore, the D module receives the following inputs at each time step:

[0062]

[0063] The behavior generation logic is as follows:

[0064] Module D employs a priority-based causal decision-making framework (not purely probabilistic triggering), judging decisions in the following order of priority:

[0065] Decision Node 1: Teacher Attention Check (Highest Priority)

[0066] If the teacher is in continuous (Default 120 seconds) without attention Pointing to a proxy This triggers a state of inattention, initiating a logic check for distraction:

[0067] like Furthermore, more than two of the adjacent agents are already distracted / conversing: triggering conversation with the person at the table, duration...

[0068] ;

[0069] Otherwise: Triggers a daydreaming or spacing-out behavior, duration

[0070] .

[0071] Decision Node 2: Perplexity Threshold Check

[0072] If the teacher has sufficient attention ( In the near (Appears within), proceed to confusion level judgment:

[0073] like (Default 0.35): Maintaining focused listening behavior, knowledge status in module K continues to rise;

[0074] like (default ): Displays mild confusion, the facial expression changes from calm to a slightly furrowed brow, without triggering any interaction;

[0075] like : Entering the participation tendency judgment:

[0076] like (Active Participation): Triggers the behavior of raising hands to ask questions, and simultaneously generates a message based on the speaker. Ask virtual voice questions about knowledge points that you are currently confused about;

[0077] like or (Neutral or Avoidant): Displays confused behavior, remains in a state of confusion, and does not actively interact;

[0078] The behavior generation results of module D are presented to the teacher through multi-channel output of VST (tablet terminal):

[0079] Screen channel: Displays 2D facial animations corresponding to the behavior (a total of 5 sets of facial animations, corresponding to five types of behavior) and body status indicators (such as a raised hand icon);

[0080] Voice Channel: When a hand-raising action is triggered, a virtual student voice question generated based on TTS is played through the speaker. The content is determined by LLM based on the current knowledge point of confusion and... Archives are generated dynamically;

[0081] System logs: Record the behavior type, trigger time, and trigger basis (confusion value, attention state, etc.) of each VST in real time, which are used for causal tracing analysis of the post-event evaluation subsystem.

[0082] Furthermore, the teacher behavior feedback and competency evaluation subsystem generates an adaptive teaching competency evaluation report with four dimensions:

[0083] (1) Dimension 1: Classroom management ability

[0084] Calculation metric: Total duration of distraction / conversation behavior Delayed intervention response after teachers detect inattentiveness The rate of resolution of distraction behavior after intervention Evaluation formula:

[0085]

[0086] (2) Dimension Two: Instructional Clarity

[0087] Calculation metric: Average class confusion curve during the conversation The rate of decline (knowledge acquisition efficiency), the average knowledge gain of the class at the end of the course .

[0088] (3) Dimension Three: Emotional Connection Ability

[0089] Calculation metric: Uniformity of teacher attention distribution (based on cumulative attention across agents) (Gini coefficient of frequency) The magnitude of the agent's emotional response to positive emotional signals from the teacher.

[0090] (4) Dimension 4: Differentiated teaching response capability

[0091] Calculation indicators: These include assessing whether the teacher adjusts the pace of explanation promptly after agents at different confusion levels become confused, and whether more spatial attention is given to agents in areas of high confusion, based on changes in speaking speed and blackboard writing frequency; and whether more spatial attention is given to agents in areas of high confusion. (Decrease).

[0092] Another objective of this invention is to provide a virtual student classroom simulation method based on a three-module framework of personality, knowledge, and dynamics, comprising the following steps:

[0093] S1: Multimodal signal extraction. The multimodal data acquisition subsystem extracts three types of input signals in real time, aligns them with a unified timestamp, and forms a multimodal feature frame, which is then sent to the PDK engine.

[0094] S2: PDK three-module virtual student state modeling, P module models the stable personality characteristics of each virtual student agent, K module dynamically tracks the knowledge mastery status of each virtual student agent on the current course content and outputs confusion index to drive subsequent behavior generation, D module combines the internal state output of P module and K module with real-time input signals and class environment, and generates the external behavior output of each virtual student agent at the current moment through causal decision logic.

[0095] S3: The evaluation subsystem performs post-processing on the teacher behavior data and virtual student status response data recorded throughout the process, and automatically generates an adaptive teaching ability evaluation report in four dimensions.

[0096] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0097] First, existing research indicates that multimodal learning analytics helps to more accurately understand the learning process, knowledge tracking can characterize and quantify students' mastery of knowledge points, and virtual classroom training combined with personalized feedback can support teachers' skills practice and improvement. This system, through collaborative modeling of personality, knowledge, and dynamics modules, unifies the stable individual differences, real-time knowledge mastery status, and overt classroom behavior of virtual students into a single decision-making chain. This allows virtual student responses to no longer rely on a single rule trigger but to dynamically change according to the teacher's voice, emotions, spatial attention, and blackboard content.

[0098] Module P enhances the individual differences and realism of group interactions among student agents; Module K utilizes knowledge graphs, RAG, and LLM agents to continuously update mastery levels and output confusion levels, improving the correlation between behavioral triggers and teaching content; Module D generates behaviors such as inattentiveness, confusion, raising hands, and asking questions based on causal priority rules, making the simulation process interpretable and traceable. The evaluation subsystem further transforms behavioral logs, confusion curves, and teacher intervention data into scores for classroom management, knowledge explanation, emotional connection, and differentiated response capabilities, thereby providing teachers with objective, granular, and reviewable training feedback, enhancing the realism, adaptability, and teaching evaluation value of the virtual classroom simulation.

[0099] This invention achieves heterogeneous dynamic simulation at the class level: because each virtual student agent has independent P-module personality parameters (ES (emotional sensitivity), PP (participation tendency), and SC (conformity tendency) three-dimensional independent sampling initialization), different agents produce differentiated responses under the same teacher behavioral input, thus restoring the diversity of student groups in a real class. Teachers need to adopt different interaction strategies for agents with different personality types, which is highly consistent with the needs of real classroom management.

[0100] This invention achieves real-time coupling between virtual student state and teacher behavior: Through multimodal signal acquisition and PDK state update mechanism, the internal state and behavior output of the virtual student can respond to the teacher's immediate behavior within a 1-second time window, making the realism and feedback immediacy of the training scenario significantly better than existing systems based on FSM.

[0101] The behavioral outputs of this invention have fully explainable causal roots: each behavioral output corresponds to a clearly defined triggering condition in module D, and the system log records a precise snapshot of the state at the time of each behavioral trigger. Both the teacher and the system can trace the cause of each behavior, allowing training feedback to directly point to specific skill gaps, rather than simply providing behavior count statistics.

[0102] This invention does not require the participation of real students, and the training scenarios can be reproduced with infinite precision: the parameters of virtual students can be flexibly configured through class templates, and specific difficult scenarios (such as classes with high confusion or classroom management out of control) can be reproduced with the same parameters with precision. It supports teacher trainees to systematically and repeatedly practice the same scenario, overcoming the fundamental limitation that once students in real classrooms come into contact with the knowledge, it cannot be reused.

[0103] This invention is installed directly above the blackboard, with a field of view covering the teacher's standing area and the entire blackboard. It uses a human posture estimation model to analyze the teacher's body movements and standing position coordinates in real time. (in (These represent the teacher's horizontal and vertical coordinates in the classroom coordinate system, respectively), and the network outputs an attention orientation vector by estimating the gaze direction. (A unit vector representing the teacher's current gaze direction) The blackboard area image is periodically (every 30 seconds) captured and parsed using a Visual Language Model (VLM) to extract the descriptive text of the knowledge points currently written on the board. (This refers to the text describing the current knowledge point on the whiteboard after being parsed by VLM).

[0104] Secondly, as supporting evidence of the inventiveness of this invention, it is also reflected in the following important aspects:

[0105] (1) The expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0106] The virtual student classroom simulation system of this invention has significant commercial value and social benefits. In the field of teacher education, more than one million teacher trainees in my country need pre-service training every year. The traditional model relies on real classrooms and expert evaluation, which is costly and difficult to scale. This system, through AI-driven virtual student simulation, can build a high-fidelity training environment without the participation of real students, significantly reducing training costs. It is estimated that it can save each teacher training institution an average of 500,000 to 1 million yuan annually in expert evaluation fees. At the same time, the system supports accurate scenario reproduction and repeated practice, improving the training efficiency of teacher trainees by 3-5 times. In terms of commercial applications, this system can be extended to multiple fields such as corporate training, online education, and AI teaching assistants, with an estimated market size of several billion yuan. In addition, the multi-dimensional teaching ability evaluation report generated by the system can provide data support for teachers' professional development and promote the transformation of educational evaluation from subjective experience to data-driven scientific methods.

[0107] (2) The technical solution of this invention fills a technical gap in the industry both domestically and internationally:

[0108] This invention is the first to integrate personality psychology models, knowledge tracking technology, and multimodal real-time perception into the field of virtual student classroom simulation, filling a technological gap both domestically and internationally in the generation of explainable causal virtual student behavior. Existing virtual teaching systems are mainly based on finite state machines (FSMs) or simple probability models, unable to simulate the cognitive dynamics and personality heterogeneity of real students; while existing AI classroom analysis systems only possess observation capabilities and lack simulation generation capabilities. This invention, through a PDK three-module framework, achieves causal correlation modeling between the internal cognitive state and external behavior of virtual students, ensuring that each behavioral output has a clear triggering condition and an explainable causal root cause—something unachieved in existing technologies. Particularly in knowledge state modeling, this invention is the first to introduce LLMAgent based on RAG role cues into virtual student knowledge tracking, achieving dynamic quantification of students' cognitive confusion and filling a gap in deep cognitive modeling of virtual educational agents.

[0109] (3) The technical solution of the present invention solves a technical problem that people have long wanted to solve but have never been able to solve successfully:

[0110] For a long time, teacher education has faced the challenge of how to conduct effective teaching skills training without real students. Issues such as the unreproducibility of real classrooms, unpredictable student responses, and limited expert resources have constrained the quality and scale of teacher training. While virtual reality and artificial intelligence technologies offer possible solutions, previous approaches have failed to overcome three core bottlenecks: first, virtual student behavior lacks cognitive roots, only achieving superficial behavioral simulation; second, virtual student groups are highly homogeneous, failing to replicate the heterogeneity of real classes; and third, the system lacks an interpretable causal feedback mechanism, unable to provide teachers with diagnostic improvement guidance. This invention, through a PDK three-module collaborative architecture, organically integrates personality traits, knowledge status, and dynamic decision-making for the first time, achieving a closed-loop simulation of cognition-driven behavior and behavior-driven cognition feedback. This allows virtual students to ask questions due to cognitive confusion, become distracted due to being ignored, and change their attitudes due to teacher attention, just like real students, fundamentally solving the long-standing technical problem of virtual teaching training environments being similar in form but lacking in essence.

[0111] (4) The technical solution of this invention overcomes technological bias: There is a technological bias in the traditional field of educational technology, namely, that the application of AI in education should focus on replacing teachers or assisting teaching, while ignoring the potential of AI in training teachers. This bias has led to a large amount of R&D resources being invested in intelligent teacher systems and automatic grading tools, while simulation technology for teacher training has long been neglected. This invention breaks through this bias, repositioning AI technology from a teaching aid tool as an infrastructure for teacher training, proving that AI can not only teach students, but also play the role of students to train teachers. This shift in technical approach opens up new directions for the application of AI in education and overcomes the industry's narrow understanding of the scope of AI applications in education. Attached Figure Description

[0112] Figure 1 This is an overall architecture diagram of the virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework provided in this embodiment of the invention;

[0113] Figure 2 This is a flowchart of the three-layer state coupling update process of PDK provided in the embodiment of the present invention (single time step).

[0114] Figure 3 This is a schematic diagram of the physical deployment of a virtual classroom provided in an embodiment of the present invention;

[0115] Figure 4 This is a schematic diagram illustrating the dynamic changes of the knowledge state vector and confusion level provided in an embodiment of the present invention;

[0116] Figure 5 This is a schematic diagram of a D-layer behavior generation causal decision tree provided in an embodiment of the present invention;

[0117] Figure 6 This is a flowchart of a virtual student classroom simulation method based on a three-module framework of personality, knowledge, and dynamics, provided in an embodiment of the present invention. Detailed Implementation

[0118] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0119] This invention provides a virtual student classroom simulation system based on a three-module framework of personality, knowledge, and dynamics. The system collects classroom behavior data of teachers through multimodal sensing, drives N virtual student agents (VSAs) with heterogeneous internal states to update in real time and generate visible behavior outputs, thereby constructing a high-fidelity virtual classroom training environment.

[0120] like Figure 1 As shown, the overall system architecture consists of three subsystems: a multimodal data acquisition subsystem, a PDK virtual student simulation engine, and a teacher behavior feedback and evaluation subsystem. The following sections provide a detailed explanation of each module in conjunction with the system's processing flow.

[0121] The multimodal data acquisition subsystem, with physical sensor deployment as follows: Figure 3 As shown, the following sensing facilities are deployed in the physical environment of the virtual classroom:

[0122] (1) Camera in the teacher area (Camera-T):

[0123] Mounted directly above the blackboard, its field of view covers the teacher's standing area and the entire blackboard. It uses a human posture estimation model to analyze the teacher's body movements and standing position coordinates in real time. The network outputs an attention orientation vector by estimating the gaze direction. The visual language model (VLM) is used to periodically capture and parse the blackboard area image every 30 seconds to extract the descriptive text of the knowledge points currently written on the board. .

[0124] (2) Student area cameras (Camera-S):

[0125] Installed at the rear of the classroom, with a field of view covering the entire seating area, it is used for panoramic recording and assisting in behavioral verification, especially the teacher's position in the classroom and interaction with student terminals.

[0126] (3) Virtual Student Tablet (VST):

[0127] Each virtual student is represented by a tablet terminal, with a total of N terminals. The default N=20, and the configurable range is 20–40. They are arranged in the seating area according to the layout of a real classroom desk. Each tablet terminal is equipped with: a screen to display the virtual student's facial expressions and body language animations; a microphone to capture the teacher's voice heard from a distance; a speaker to play the generated voice response when the virtual student raises their hand to answer; and a camera to identify the relative distance to the teacher and to assess the size and clarity of the blackboard writing.

[0128] The multimodal data acquisition subsystem extracts three types of input signals from the aforementioned sensors in real time, aligns them with a unified timestamp to form a multimodal feature frame, and sends it to the PDK engine. The unified timestamp has an accuracy of 100ms. The three types of input signals include:

[0129] (1) Signal S1: Teacher speech transcription and emotion annotation

[0130] Each tablet terminal runs a local speech recognition program that transcribes the speech stream captured by the terminal's microphone in real time and outputs the transcribed text fragment within the current time window. The current time window is a 5-second sliding window. This is due to the physical distance between each tablet and the teacher. The difference in volume attenuation leads to variations in speech recognition confidence. This is a design feature intentionally retained by the system to reflect the physical reality of different hearing levels between students in the front and back rows in a real classroom.

[0131] The transcribed text is simultaneously fed into a lightweight sentiment classification model running on a tablet terminal, which outputs sentence-by-sentence sentiment polarity labels. Together with the confidence score, they constitute the signal S1.

[0132] (2) Signal S2: Teacher's spatial position and attention orientation

[0133] The coordinates of the human body key points output by Camera-T are transformed to calculate the teacher's current position in the classroom coordinate system. Then calculate the relationship between the teacher and the first Euclidean distance between tablet terminals:

[0134] ,

[0135] As a teacher to virtual students The physical proximity metric is input into the PDK engine.

[0136] The attention orientation vector output by the gaze direction estimation network After projection calculation, determine whether the teacher is currently facing the first... Tablet terminal, outputs Boolean attention markers . and Together they constitute signal S2.

[0137] (3) Signal S3: Blackboard image semantics

[0138] Camera-T periodically captures screenshots of the blackboard area, triggered every 30 seconds or when the system detects significant content changes. The screenshots are then parsed by VLM to extract the descriptive text of the knowledge points currently written on the board. This is expressed as a list of knowledge points on the blackboard at this moment and their corresponding course node identifiers, which is sent to the K module of the PDK engine as the basis for updating knowledge content.

[0139] like Figure 2 As shown, the PDK engine instantiates N virtual student agents at the start of a session: Each agent Defined by the internal state vectors of the three modules, the state of each module at each time step The system updates collaboratively based on input signals S1, S2, and S3. The three modules include:

[0140] (1) P module: Personality module

[0141] The P module is responsible for modeling the stable personality traits of each virtual student agent, providing the basis for individualized parameters for the dynamic updates of the K and D modules. Each agent... A personality state vector is assigned during session initialization. It includes the following three parameter dimensions:

[0142] Emotional sensitivity Quantitative Proxy The intensity of response to teachers' emotional cues. In mathematics, As an amplification factor for the intensity of emotional signals: when teachers emit positive emotional signals ( When ), the agent's sentiment vector update magnitude is ; The higher the value, the more significant the mood fluctuations.

[0143] Participation tendency Control Agent The basic probability of spontaneously participating in classroom interaction without being called upon by the teacher. Correspondingly, avoid participation. Corresponding to neutral, Active participation is required. The hand-raising behavior directly triggers the judgment in module D.

[0144] Social conformity index Modeling Agent The degree to which the participant's participation status is influenced by the behavior of the surrounding virtual student group. During behavior generation in module D, the agent... The probability distribution of behavior will be based on As the weight, shift towards the average behavior distribution of the current neighboring agents:

[0145]

[0146] This mechanism is used to recreate the social diffusion effect of one student's inattention in the classroom, which then distracts the students around them.

[0147] Personality state vector Initialize based on the class template at the start of the session. The system has at least 3 preset class templates:

[0148]

[0149] During the conversation, Not completely static, when the teacher acts as an agent The cumulative number of interactions exceeds the threshold (Default 3 times), agent participation tendency Will The range is shifted 0.3 in the positive direction to simulate the educational phenomenon that sustained teacher attention gradually encourages introverted students to participate.

[0150] (2) K module: Knowledge module

[0151] like Figure 4 As shown, module K is responsible for dynamically tracking the knowledge mastery status of each virtual student agent on the current course content and outputting a confusion index to drive subsequent behavior generation.

[0152] Each agent Maintaining the knowledge state vector ,in Indicates agent Knowledge units in the current course knowledge graph The degree of mastery, This is the set of knowledge graph nodes for this session. The knowledge graph is defined by the course configuration file during session initialization, and the node types must include at least:

[0153] Conceptual knowledge nodes: such as definitions, theorems, and concepts;

[0154] Procedural knowledge nodes: such as algorithm steps and operation methods;

[0155] Migrate application nodes: such as application scenarios and cross-domain connections;

[0156] The nodes are labeled with prerequisite relationships, that is, if If the data is not fully understood, i.e., below the threshold of 0.4, then the dependency is considered insufficient. nodes The rate at which one can master the skill will be suppressed.

[0157] At each time step The K module calls an LLM Agent based on the RAG role-based prompting framework to process the current input:

[0158] Input construction: The current agent Knowledge state vector Current teacher transcription fragments Blackboard writing semantics and agents Error type archive The context is concatenated and sent to the LLM Agent.

[0159] LLM Agent Inference: The system prompts the agent to set its role as a learning and cognitive simulator, requiring it to update the mastery probability of knowledge units based on the teaching content and return the result in JSON format. , represents the increment of mastery, and can be negative to indicate forgetting or confusion.

[0160] Status Update:

[0161]

[0162] The K module also outputs a real-time perplexity index. Defined as the set of knowledge units involved in the current teaching time step. Average level of lack of mastery:

[0163]

[0164] The higher the value, the greater the level of confusion. The D module is directly input as the core criterion for triggering behavior, and also serves as the raw data for evaluating the appropriateness of knowledge explanation in the subsystem.

[0165] (3) Module D: Dynamic Module

[0166] like Figure 5 As shown, module D is the core output module of the PDK engine. It is responsible for combining the internal states output by modules P and K with real-time input signals and the class environment, and generating the external behavior output of each virtual student agent at the current moment through causal decision logic.

[0167] Module D receives the following input at each time step:

[0168]

[0169] The behavior generation logic is as follows:

[0170] Module D employs a priority-based causal decision-making framework (not purely probabilistic triggering), judging decisions in the following order of priority:

[0171] Decision Node 1: Teacher Attention Check (Highest Priority)

[0172] If the teacher is in continuous (Default 120 seconds) without attention Pointing to a proxy This triggers a state of inattention, initiating a logic check for distraction:

[0173] like Furthermore, if more than two people in the adjacent seats are distracted or talking: triggering a conversation with the person at the same table, duration...

[0174] ;

[0175] Otherwise: Triggers a daydreaming or spacing-out behavior, duration

[0176] .

[0177] Decision Node 2: Perplexity Threshold Check

[0178] If the teacher has sufficient attention, proceed to the confusion assessment:

[0179] like (Default 0.35): Maintaining focused listening behavior, knowledge status in module K continues to rise;

[0180] like (default ): Displays mild confusion, the facial expression changes from calm to a slightly furrowed brow, without triggering any interaction;

[0181] like : Entering the participation tendency judgment:

[0182] like (Active Participation): Triggers the behavior of raising hands to ask questions, and simultaneously generates a message based on the speaker. Ask virtual voice questions about knowledge points that you are currently confused about;

[0183] like or (Neutral or Avoidant): Displays confused behavior, remains in a state of confusion, and does not actively interact;

[0184] The behavior generation results of module D are presented to the teacher through VST's multi-channel output:

[0185] Screen channel: Displays 2D facial animations and body status indicators corresponding to the behavior;

[0186] Voice Channel: When a hand-raising action is triggered, a virtual student voice question generated based on TTS is played through the speaker. The content is determined by LLM based on the current knowledge point of confusion and... Archives are generated dynamically;

[0187] System logs: Record the behavior type, trigger time, and trigger basis of each VST in real time, which is used for causal tracing analysis of the post-event evaluation subsystem.

[0188] The teacher behavior feedback and competency evaluation subsystem, after the session ends, performs post-processing on the teacher behavior data and virtual student status response data recorded throughout the process, and automatically generates an adaptive teaching competency evaluation report with four dimensions:

[0189] Dimension 1: Classroom Management Ability

[0190] Calculation metric: Total duration of distraction / conversation behavior Delayed intervention response after teachers detect inattentiveness The rate of resolution of distraction behavior after intervention Evaluation formula:

[0191]

[0192] Dimension Two: Instructional Clarity

[0193] Calculation metric: Average class confusion curve during the conversation The rate of decline, the average knowledge gain of the class at the end of the course .

[0194] Dimension Three: Emotional Connection Ability

[0195] Calculation indicators: uniformity of teacher attention distribution, high The magnitude of the agent's emotional response to positive emotional signals from the teacher.

[0196] Dimension Four: Ability to Respond to Differentiated Instruction

[0197] Calculation indicators: Judging whether the teacher adjusts the pace of explanation in a timely manner after the agent becomes confused at different levels of confusion by using changes in speech rate and blackboard writing frequency; and whether more spatial attention is given to the agent in areas of high confusion.

[0198] The evaluation results, combined with the LLM-driven reflection prompt module, generate personalized diagnostic reports for teachers, including: an interpretable description of specific skill gaps with timestamps of specific classroom events, targeted improvement suggestions, and a recommended class template configuration for the next training session.

[0199] like Figure 6 As shown, the virtual student classroom simulation method based on a three-module framework of personality-knowledge-dynamics provided by this embodiment of the invention specifically includes:

[0200] S1: Multimodal signal extraction. The multimodal data acquisition subsystem extracts three types of input signals in real time, aligns them with a unified timestamp, and forms a multimodal feature frame, which is then sent to the PDK engine.

[0201] S2: PDK three-module virtual student state modeling, P module models the stable personality characteristics of each virtual student agent, K module dynamically tracks the knowledge mastery status of each virtual student agent on the current course content and outputs confusion index to drive subsequent behavior generation, D module combines the internal state output of P module and K module with real-time input signals and class environment, and generates the external behavior output of each virtual student agent at the current moment through causal decision logic.

[0202] S3: The evaluation subsystem performs post-processing on the teacher behavior data and virtual student status response data recorded throughout the process, and automatically generates an adaptive teaching ability evaluation report in four dimensions.

[0203] I. Specific application areas or related products of this invention.

[0204] The virtual student classroom simulation system of this invention can be widely applied in the following fields and related products:

[0205] (1) Pre-service training system for teacher trainees

[0206] As a core application, this system can be deployed in the educational training centers of teacher training colleges, providing teacher trainees with an immersive virtual classroom training environment. The system supports course content configuration for various subjects (mathematics, Chinese, English, physics, etc.), allowing teacher trainees to practice teaching in virtual classes. The system generates real-time feedback on virtual student behavior and provides a four-dimensional teaching ability evaluation report after class. This product can directly replace traditional micro-teaching classrooms and peer role-playing models, becoming a standard tool for teacher trainees' teaching skills training.

[0207] (2) In-service teacher continuing education platform

[0208] This system can serve as a training platform for the professional development of in-service teachers, supporting them in conducting specialized training for specific teaching skills (such as classroom management, differentiated instruction, and emotional communication). Teachers can choose virtual class templates of varying difficulty and type to repeatedly practice specific scenarios (such as handling unexpected classroom events, guiding introverted students to participate, and dealing with highly confused classes). The system records training data and generates competency growth curves, providing objective evidence for teacher professional title evaluation and continuing education.

[0209] (3) Educational AI Research and Evaluation Benchmarks

[0210] This system can serve as an experimental platform and benchmark for educational artificial intelligence research. Researchers can test the effectiveness of different teaching strategies, classroom management methods, or AI teaching assistant algorithms in a controlled virtual environment. The standardized virtual student groups and quantitative evaluation indicators provided by the system ensure comparability between different research approaches. Furthermore, the multimodal datasets generated by the system (teacher behavior data, student status data, and causal decision logs) can be used for educational data mining and learning analytics research.

[0211] (4) Corporate training and online education

[0212] The core technology of this system can be transferred to the corporate training field, used to train corporate trainers, customer service trainers, and other positions that require face-to-face explanations and interactions with audiences. In the online education field, the system can serve as the core engine of an AI teaching assistant, simulating student questions and feedback to help online course instructors optimize teaching content and interaction strategies.

[0213] II. Evidence related to the technical effects obtained by the embodiments of the present invention.

[0214] (1) Implement dynamic simulation of heterogeneity at the class level

[0215] Because each virtual student agent has independent P-module personality parameters (ES, PP, SC three-dimensional independent sampling initialization), different agents produce differentiated responses under the same teacher behavioral input, thus replicating the diversity of student groups in a real classroom. Teachers need to adopt different interaction strategies for agents with different personality types, which is highly consistent with the needs of real classroom management.

[0216] (2) Real-time coupling of virtual student status and teacher behavior

[0217] Through multimodal signal acquisition and PDK state update mechanism, the internal state and behavior output of virtual students can respond to the teacher's real-time behavior within a 1-second time window, making the realism and real-time feedback of the training scene significantly better than the existing system based on FSM.

[0218] (3) The behavioral output has a fully explainable causal root.

[0219] Each behavioral output corresponds to a clearly defined trigger condition in Module D (confusion threshold, attention duration, social conformity index threshold, etc.), and the system log records a precise snapshot of the state at the time of each behavioral trigger. Both teachers and the system can trace the cause of each behavior, allowing training feedback to directly target specific skill gaps, rather than simply providing behavior count statistics.

[0220] (4) No real students are required, and the training scenario can be reproduced with infinite precision.

[0221] Virtual student parameters can be flexibly configured through class templates. Specific difficult scenarios (such as classes with high levels of confusion or classroom management out of control) can be accurately reproduced with the same parameters, supporting teacher trainees to systematically and repeatedly practice the same scenario, overcoming the fundamental limitation that once students in real classrooms come into contact with the knowledge, it cannot be reused.

[0222] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0223] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A virtual student classroom simulation system based on a three-module framework of personality, knowledge, and dynamics, characterized in that: The system includes: The multimodal data acquisition subsystem is used to extract three types of input signals in real time, align them with a unified timestamp to form a multimodal feature frame, and send it into the PDK engine. The PDK virtual student simulation engine includes P, K, and D modules for modeling virtual student states. The P module models the stable personality traits of each virtual student agent. The K module dynamically tracks the knowledge mastery status of each virtual student agent in the current course content and outputs a confusion index to drive subsequent behavior generation. The D module combines the internal states output by the P and K modules with real-time input signals and the class environment to generate the external behavior output of each virtual student agent at the current moment through causal decision logic. The teacher behavior feedback and evaluation subsystem is used to post-process the teacher behavior data and virtual student status response data recorded throughout the process, and automatically generate an adaptive teaching ability evaluation report in four dimensions.

2. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The multimodal data acquisition subsystem deploys the following sensing facilities in the physical environment of the virtual classroom: (1) Cameras in the teachers' area: Mounted directly above the blackboard, its field of view covers the teacher's standing area and the entire blackboard. It uses a human posture estimation model to analyze the teacher's body movements and standing position coordinates in real time. The network outputs an attention orientation vector by estimating the gaze direction. The system uses a visual language model to periodically (every 30 seconds) capture and parse images of the blackboard area, extracting the descriptive text of the knowledge points currently written on the board. ; (2) Cameras in the student area: Installed at the rear of the classroom, with a field of view covering the entire seating area, it is used for panoramic recording and assisting in behavioral verification, especially the teacher's position in the classroom and interaction with student terminals. (3) Virtual student tablet terminal: Each virtual student is represented by a tablet terminal, with a total of N tablet terminals arranged in the seating area according to the layout of a real classroom desk. Each tablet terminal is equipped with: a screen to display the virtual student's facial expression animation and body state; a microphone to collect the teacher's voice that can be heard from the seat position; a speaker to play the generated voice response when the virtual student is triggered to raise their hand to answer; and a camera to identify the relative distance to the teacher and to assess the size and clarity of the blackboard writing.

3. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The multimodal data acquisition subsystem extracts three types of input signals from the aforementioned sensors in real time, including: (1) Signal S1: Teacher speech transcription and emotion annotation Each tablet terminal runs a local speech recognition program that transcribes the speech stream captured by the terminal's microphone in real time and outputs the transcribed text fragment within the current time window. The current time window is a 5-second sliding window. The transcribed text is simultaneously fed into a lightweight sentiment classification model running on a tablet terminal, which outputs sentence-by-sentence sentiment polarity labels. Together with the confidence score, they constitute signal S1; (2) Signal S2: Teacher's spatial position and attention orientation The coordinates of the human body key points output by Camera-T are transformed to calculate the teacher's current position in the classroom coordinate system. Then calculate the relationship between the teacher and the first Euclidean distance between tablet terminals: ， As a teacher to virtual students The physical proximity metric is input into the PDK engine, and the attention orientation vector is output by the gaze direction estimation network. After projection calculation, determine whether the teacher is currently facing the first... Tablet terminal, outputs Boolean attention markers , and Together they constitute signal S2; (3) Signal S3: Blackboard image semantics Camera-T periodically captures screenshots of the blackboard area, triggered every 30 seconds or when the system detects significant content changes. The screenshots are then parsed by VLM to extract the descriptive text of the knowledge points currently written on the board. The K module of the PDK engine is used as the basis for updating knowledge content.

4. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The PDK engine instantiates N virtual student agents at the start of a session: Each agent Defined by the internal state vectors of the three modules, the state of each module at each time step Update in tandem based on input signals S1, S2, and S3.

5. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The P module, each agent A personality state vector is assigned during session initialization. It includes the following three parameter dimensions: Emotional sensitivity Quantitative Proxy The intensity of response to teachers' emotional cues; Participation tendency Control Agent The basic probability of spontaneously participating in classroom interaction without being called upon by the teacher. Correspondingly, avoid participation. Corresponding to neutral, Corresponding to active participation, Directly participate in the hand-raising behavior triggering judgment in module D; Social conformity index Modeling Agent The degree to which the participation status is influenced by the behavior of the surrounding virtual student group, during the generation of behavior in module D, by the agent. The probability distribution of behavior will be based on As the weight, shift towards the average behavior distribution of the current neighboring agents: ； Personality state vector Initialize based on the class template at the start of the session, and during the session... Not completely static, when the teacher acts as an agent The cumulative number of interactions exceeds the threshold Agent's tendency to participate Will The range is shifted 0.3 in the positive direction to simulate the educational phenomenon that sustained teacher attention gradually encourages introverted students to participate.

6. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The K module, each agent Maintaining the knowledge state vector ,in Indicates agent Knowledge units in the current course knowledge graph The degree of mastery, This is the set of knowledge graph nodes for this session. The knowledge graph is defined by the course configuration file during session initialization, and the node types must include at least: conceptual knowledge nodes, procedural knowledge nodes, and migration application nodes. The nodes are labeled with prerequisite relationships at each time step. The K module calls LLMAgent, which is based on the RAG+ role-based prompting framework, to process the current input: Input construction: The current agent Knowledge state vector Current teacher transcription fragments Blackboard writing semantics and agents Error type archive Concatenate the context and send it to the LLM Agent; LLM Agent Inference: The system prompts the agent to set its role as a learning and cognitive simulator, requiring it to update the mastery probability of knowledge units based on the teaching content and return the result in JSON format. This represents the increment in the degree of mastery; a negative value indicates forgetting or confusion. Status Update: ； The K module also outputs a real-time perplexity index. Defined as the set of knowledge units involved in the current teaching time step. Average level of lack of mastery: ； The D module is directly input as the core criterion for triggering behavior, and also serves as the raw data for evaluating the appropriateness of knowledge explanation in the subsystem.

7. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The behavior generation logic of module D is as follows: Module D employs a priority-based causal decision-making framework, judging decisions in the following order of priority: Decision Node 1: Teacher Attention Check If the teacher is in continuous Internal attention was not focused Pointing to a proxy This triggers a state of inattention, initiating a logic check for distraction: like Furthermore, more than two of the adjacent agents are already distracted / conversing: triggering conversation with the person at the table, duration... ； Otherwise: Triggers a daydreaming or spacing-out behavior, duration ； Decision Node 2: Perplexity Threshold Check If the teacher has sufficient attention ( In the near (Appears within), proceed to confusion level judgment: like Maintaining focused listening behavior led to a continuous improvement in knowledge of Module K. like : Shows mild confusion, the facial expression changes from calm to a slightly furrowed brow, without triggering any interaction; like : Entering the participation tendency judgment: like : Triggers the action of raising a hand to ask a question, and simultaneously generates a message based on the speaker. Ask virtual voice questions about knowledge points that you are currently confused about; like or : Demonstrate confused behavior, remain silent in a confused state, and do not actively interact; The behavior generation results of module D are presented to the teacher through VST's multi-channel output: Screen channel: Displays 2D facial animations and body status indicators corresponding to the behavior; Voice Channel: When a hand-raising action is triggered, a virtual student voice question generated based on TTS is played through the speaker. The content is determined by LLM based on the current knowledge point of confusion and... Archives are generated dynamically; System logs: Record the behavior type, trigger time, and trigger basis of each VST in real time, which is used for causal tracing analysis of the post-event evaluation subsystem.

8. The virtual student classroom simulation system based on the personality-knowledge-dynamic three-module framework as described in claim 1, characterized in that, The teacher behavior feedback and competency assessment subsystem generates an adaptive teaching competency assessment report with four dimensions: (1) Dimension 1: Classroom management ability Calculation metric: Total duration of distraction / conversation behavior Delayed intervention response after teachers detect inattentiveness The rate of resolution of distraction behavior after intervention Evaluation formula: ； (2) Dimension Two: Appropriateness of Knowledge Explanation Calculation metric: Average class confusion curve during the conversation The rate of decline (knowledge acquisition efficiency), the average knowledge gain of the class at the end of the course ; (3) Dimension Three: Emotional Connection Ability Calculation indicators: uniformity of teacher attention distribution, high The magnitude of the agent's emotional response to positive emotional signals from the teacher; (4) Dimension 4: Differentiated teaching response capability Calculation indicators: Judging whether the teacher adjusts the pace of explanation in a timely manner after the agent becomes confused at different levels of confusion by using changes in speech rate and blackboard writing frequency; and whether more spatial attention is given to the agent in areas of high confusion.

9. A virtual student classroom simulation method based on the personality-knowledge-dynamic three-module framework as described in claims 1-8, characterized in that, The method specifically includes: S1: Multimodal signal extraction. The multimodal data acquisition subsystem extracts three types of input signals in real time, aligns them with a unified timestamp, and forms a multimodal feature frame, which is then sent to the PDK engine. S2: PDK three-module virtual student state modeling, P module models the stable personality characteristics of each virtual student agent, K module dynamically tracks the knowledge mastery status of each virtual student agent on the current course content and outputs confusion index to drive subsequent behavior generation, D module combines the internal state output of P module and K module with real-time input signals and class environment, and generates the external behavior output of each virtual student agent at the current moment through causal decision logic. S3: The evaluation subsystem performs post-processing on the teacher behavior data and virtual student status response data recorded throughout the process, and automatically generates an adaptive teaching ability evaluation report in four dimensions.

Images