AI adaptive tutoring system and method based on hierarchical learning state engine

Abstract

This invention discloses an AI adaptive tutoring system and method based on a hierarchical learning state engine, comprising an adaptive module; an L0 signal layer for collecting learner data in a learning scenario after learner authorization; an L1 state layer for analyzing learner real-time learning state through a multimodal fusion model; and an L2 profile layer for creating a personalized profile of the learner based on dimensions such as stress threshold, recovery speed, confusion tolerance, and motivation type. The beneficial effects of this invention are: real-time multimodal perception of learner's emotions, focus, and cognitive load; driving the selection of interactive intervention strategies through learner state to achieve differentiated interventions triggered for different states; realizing a complete signal → state → strategy → result chain, ensuring the interpretability and auditability of decisions; and continuously iterating and optimizing intervention strategies through learner profiles.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and education technology, and in particular to an AI adaptive tutoring system and method based on a hierarchical learning state engine. Background Technology

[0002] Existing AI tutoring systems in the education field mainly suffer from the following technical shortcomings: 1) Focusing only on the result layer and ignoring the process layer Traditional AI teaching systems primarily evaluate learning effectiveness based on outcome metrics such as answer accuracy and class duration, failing to perceive the learner's real-time psychological and cognitive state during the learning process. Research shows that when learners experience frustration or cognitive overload, even if they ultimately answer correctly, their learning efficiency and long-term memory consolidation significantly decline. Existing systems lack the ability to model this process layer.

[0003] 2) Fixed strategies, lacking state-aware personalized responses Existing AI tutors typically use fixed prompting templates or pre-defined error correction processes, failing to dynamically adjust interaction strategies based on the learner's current learning state (such as confusion, frustration, inattentiveness, or steady-state focus). This leads to: • Continuously subjecting learners to high cognitive load tasks when they are frustrated exacerbates their emotional deterioration; • Using the same teaching rhythm as a focused learner when they are inattentive is ineffective in rekindling their attention; • Failure to provide appropriate encouragement, follow-up questions, or scaffolding support at the right time.

[0004] 3) Lack of long-term learner profiles Existing systems typically process learning behavior at the level of a single class session, failing to accumulate learners' personalized traits (such as stress threshold, frustration recovery speed, and motivation type) across courses, thus hindering the achievement of true long-term personalization.

[0005] 4) Black-box decision-making, lacking explainability Most AI Tutors, when triggering a certain interaction strategy, cannot provide a traceable signal → state → strategy → result chain, making it difficult to meet the requirements of educational scenarios for interpretability, auditability, and security degradation. Summary of the Invention

[0006] The purpose of this invention is to provide an AI adaptive tutoring system and method based on a hierarchical learning state engine that solves or partially solves the above-mentioned technical problems.

[0007] To achieve the above objectives, the present invention provides the following technical solution: An AI adaptive tutoring system based on a hierarchical learning state engine includes an online learning platform for learners to learn online. The online learning platform includes a learner client and a server. The server includes an AI tutoring module and an adaptive module. The learner client allows learners to engage in interactive online learning through the AI ​​tutoring module. The adaptive module includes: The L0 signal layer is used to collect learner learning data in the learning scenario after the learner has authorized it. The L1 state layer is used to analyze the learner's real-time learning state through a multimodal fusion model, classify the real-time learning state, and select and trigger the corresponding interactive intervention strategy from the preset strategy library according to the category of the real-time learning state, so as to adjust the learning content and tutoring strategy of the AI ​​tutoring module. The L2 profiling layer is used to create personalized profiles of learners based on stress threshold, recovery speed, confusion tolerance, and motivation type, and the profile results are fed back to the L1 state layer to continuously optimize interactive intervention strategies.

[0008] Preferably, the learning data includes: video data, audio data, interactive behavior data, learning process data, and learning result data.

[0009] Preferably, the video data includes the learner's facial expressions, eye contact, and body movements during learning; the audio data includes the learner's speech rate, pauses, and intonation characteristics; the interactive behavior data includes events such as clicking, answering, dragging, saving, pausing, speeding up playback, triggering prompts, and page switching; the learning process data includes reaction time, number of error corrections, prompt requests, and knowledge point browsing paths; and the learning result data includes quiz scores, task completion rate, and knowledge point mastery assessment.

[0010] Preferably, the categories of the real-time learning states include confusion, frustration, inattentiveness, steady-state focus, and cognitive overload.

[0011] Preferably, the interactive intervention strategies in the preset strategy library include: When the real-time learning state is one of confusion, strategies such as reducing task difficulty, providing knowledge scaffolding, or step-by-step questioning are triggered. When the real-time learning state is one of frustration, a low-load strategy is triggered: pause the current task, output emotional support words, or switch to a low-load interaction mode (such as reviewing and practicing knowledge points that have been mastered). When the real-time learning state is characterized by inattentiveness, a low-attention strategy is triggered: increase attention-awakening prompts, push short achievement tasks, or accelerate the teaching pace. When the real-time learning state is a steady state of focus, a high-attention strategy is triggered: maintain or increase the current challenge, increase teaching content, or reduce unnecessary interventions; When the real-time learning state is cognitive overload, a low-cognitive strategy is triggered: reduce information density, increase information processing time, or switch practice modes.

[0012] Preferably, the L2 profiling layer profiles learners based on the following dimensions: Stress threshold: the number of consecutive errors or the level of cognitive load that triggers sympathetic nerve activation signals in learners; Recovery speed: The type of strategy and time required for learners to recover from a state of frustration / distraction to a steady state of focus; Distress tolerance: The level of distress at which learners can maintain effective learning; Types of motivation: Learners' learning motivations can be categorized into three types: quick feedback-driven, social recognition-driven, and long-term achievement-driven. State response pattern classification: Learners are divided into frustration-sensitive, low-arousal distracted, and confusion-tolerant types based on clustering methods; The L2 profiling layer feeds back the learner's profile to the L1 state layer, and selects and triggers the corresponding interactive intervention strategy from the preset strategy library.

[0013] Preferably, the server also includes an interpretability recording module for recording and outputting the complete causal chain from the collected learning data to the output interactive intervention strategy.

[0014] This invention also provides an AI adaptive tutoring method based on a hierarchical learning state engine, which is executed using the aforementioned AI adaptive tutoring system based on a hierarchical learning state engine, and includes the following steps: S1. After obtaining the learner's authorization, collect the learner's learning data in the learning scenario; S2. Extract facial emotion features, speech prosody features, and behavioral rhythm features from the learning data; S3. Analyze the learner's real-time learning status and classify the real-time learning status; S4. Select and trigger the corresponding interactive intervention strategy from the preset strategy library; S5. Record and output the complete causal chain from the collected learning data to the output interactive intervention strategy.

[0015] Preferably, the features extracted in step S2 include: facial motion units, speech fundamental frequency and pause features, and behavioral response delay features.

[0016] Preferably, the following steps are also included: S6. Online strategy evaluation: The real-time learning status of learners after the interactive intervention strategy is triggered in step S4 is analyzed through the L1 state layer, and the interactive intervention strategy in step S4 is scored. When the score is less than the preset threshold, it is reverted to a conservative strategy or manual intervention is initiated. S7. Update the personalized parameters of the L2 profile layer based on the intervention results corresponding to the interactive intervention strategy in step S4, and continuously optimize the matching accuracy between the learner's profile and the interactive intervention strategy.

[0017] The beneficial effects of this invention are: it employs multimodal real-time perception of learners' emotions, concentration, and cognitive load, driving the selection of interactive intervention strategies based on learners' states, thus achieving differentiated interventions triggered for different states; it realizes a complete signal → state → strategy → result chain, ensuring the interpretability and auditability of decisions; and it continuously iterates and updates learner profiles. This invention can effectively reduce the incidence of learners dropping out; shorten the time for learners to recover from negative states (frustration / distraction) to a steady state of focus; increase the proportion of effective interaction rounds and the quality of error correction; improve knowledge mastery and learning efficiency per unit time; and provide an auditable, interpretable, and securely degradeable decision-making mechanism for AI education systems. Attached Figure Description

[0018] Figure 1 This is a flowchart of the AI ​​adaptive tutoring method based on a hierarchical learning state engine, as described in this invention. Detailed Implementation

[0019] The technical solution of the present invention will be further described in detail below with reference to specific embodiments.

[0020] This invention provides an AI adaptive tutoring system based on a hierarchical learning state engine, including an online learning platform for learners to learn online. The online learning platform includes a learner client and a server. The server includes an AI tutoring module and an adaptive module. The learner client allows learners to engage in interactive online learning through the AI ​​tutoring module. The adaptive module includes: The L0 signal layer is used to collect learner learning data in the learning scenario after the learner has authorized it. The L1 state layer is used to analyze the learner's real-time learning state through a multimodal fusion model, classify the real-time learning state, and select and trigger the corresponding interactive intervention strategy from the preset strategy library according to the category of the real-time learning state, so as to adjust the learning content and tutoring strategy of the AI ​​tutoring module. The L2 profiling layer is used to create personalized profiles of learners based on dimensions such as stress threshold, recovery speed, confusion tolerance, and motivation type, and the profile results are fed back to the L1 state layer to continuously optimize intervention strategies.

[0021] Learning data includes: video data, audio data, interactive behavior data, learning process data, and learning outcome data.

[0022] Video data includes learners' facial expressions, eye contact, and body language during learning; audio data includes learners' speech rate, pauses, and intonation characteristics; interactive behavior data includes events such as clicking, answering, dragging, saving, pausing, speeding up playback, prompt triggering, and page switching; learning process data includes reaction latency, number of error corrections, prompt requests, and knowledge point browsing paths; and learning outcome data includes quiz scores, task completion rate, and knowledge point mastery assessment.

[0023] The categories of real-time learning states include confusion, frustration, inattentiveness, steady-state focus, and cognitive overload.

[0024] The interactive intervention strategies in the preset strategy library include: When the real-time learning status is confused, strategies such as reducing task difficulty, providing knowledge scaffolding, or step-by-step follow-up questions are triggered. When the real-time learning state is frustrating, trigger a low-load strategy: pause the current task, output emotional support words, or switch to a low-load interaction mode (such as reviewing and practicing knowledge points that have been mastered). When students are distracted during real-time learning, low-attention strategies are triggered, such as increasing attention-awakening prompts, pushing short achievement tasks, or accelerating the teaching pace. When the real-time learning state is a steady state of focus, high attention strategies are triggered: maintain or increase the current challenge, increase teaching content, or reduce unnecessary interventions; When the real-time learning state is characterized by cognitive overload, low-cognitive strategies are triggered: reducing information density, increasing information processing time, or switching practice modes.

[0025] The L2 profiling layer profiles learners based on the following dimensions: Stress threshold: the number of consecutive errors or the level of cognitive load that triggers sympathetic nerve activation signals in learners; Recovery speed: The type of strategy and time required for learners to recover from a state of frustration / distraction to a steady state of focus; Distress tolerance: The level of distress at which learners can maintain effective learning; Types of motivation: Learners' learning motivations can be categorized into three types: quick feedback-driven, social recognition-driven, and long-term achievement-driven. State response pattern classification: Learners are divided into frustration-sensitive, low-arousal distracted, and confusion-tolerant types based on clustering methods; The L2 profiling layer feeds back the learner's profile to the L1 state layer, and selects and triggers the corresponding interactive intervention strategy from the preset strategy library.

[0026] The server also includes an interpretability logging module, which records and outputs the complete causal chain from the collected learning data to the output interactive intervention strategy.

[0027] like Figure 1 As shown, the AI ​​adaptive tutoring system based on a hierarchical learning state engine of the present invention, and the AI ​​adaptive tutoring method based on a hierarchical learning state engine of the present invention, include the following steps: S1. After obtaining the learner's authorization, collect the learner's learning data in the learning scenario; S2. Extract facial emotion features, speech prosody features, and behavioral rhythm features from the learning data; S3. Analyze the learner's real-time learning status and classify the real-time learning status; S4. Select and trigger the corresponding interactive intervention strategy from the preset strategy library; S5. Record and output the complete causal chain from the collected learning data to the output interactive intervention strategy; S6. Online strategy evaluation: Analyze the learner's real-time learning status after the interactive intervention strategy is triggered in step S4 through the L1 state layer, score the interactive intervention strategy in step S4, and when the score is less than the preset threshold, revert to the conservative strategy or switch to manual intervention. S7. Based on the intervention results corresponding to the interactive intervention strategy in step S4, update the personalized parameters of the L2 profile layer to continuously optimize the matching accuracy between the learner's profile and the interactive intervention strategy.

[0028] The features extracted in step S2 include: facial action units, speech fundamental frequency and pause features, and behavioral response delay features.

[0029] Among them, the adaptive module forms a complete modeling chain from behavioral signals to real-time status and then to long-term profile:

[0030] Specifically: (I) The learning data collected by the L0 signal layer and the analysis of the learning data by the L1 state layer are typically collected and processed as follows: (1) Video data Real-time or offline frame sampling is performed on learner camera videos (collected with learner authorization) from online learning platforms to extract facial action unit (AU) features; Inferring instantaneous emotion labels (confusion, frustration, inattentiveness, steady-state focus, and cognitive overload) based on AU feature sequences. A temporal smoothing mechanism is introduced to suppress inference noise caused by changes in light or occlusion.

[0031] (2) Audio data Extract prosodic features of learners’ speech, such as fundamental frequency (F0), speech rate, pause duration, and volume variation; Detect abnormally long pauses (≥3 seconds), repetitive repetition, and a drop in tone, among other negative emotional signals; Combine speech content (ASR to text) to perform semantic-level emotion-assisted judgment.

[0032] (3) Interactive behavior data Analyze learners' interactive behavior events across all scenarios on online learning platforms: clicks, responses, drags, favorites, pauses, playback speed adjustments, prompts, error correction rounds, page switching, etc. Extract behavioral features that reflect cognitive load and attention (such as the standard deviation of reaction time from baseline). Timestamp-aligned fusion with video / audio features.

[0033] Learning process data and learning outcome data Analyze learners’ learning process data in various scenarios on online learning platforms: answer sequence, knowledge point coverage path, browsing / stay time, prompt request behavior, number of error correction rounds and rhythm characteristics; Integrate learning outcome data: quiz scores, completion rate, progress assessment, in-class assessment and unit test results; The learning process and result data are time-stamped with video / audio / interactive behavior to construct a complete multi-dimensional learning behavior vector, which is used for state inference in the L1 state layer and long-term profile modeling in the L2 profile layer.

[0034] (5) Multimodal fusion model analysis of learners’ real-time learning status Employing a multimodal fusion model (such as an attention mechanism or ensemble learning architecture), the learning data from the L0 signal layer is input, and the probability distribution of the real-time learning state is output. Introduce state transition constraints (such as a frustrated state cannot instantly transform into a focused state without intervention) to improve temporal consistency; It supports generalization across teachers and course lines, and reduces the interference of differences in teacher styles through domain adaptation.

[0035] (II) Refer to the table below for the selection of interactive intervention strategies at the L1 state level:

[0036] For a practical example: In a one-on-one online English class, learners authorize the platform to collect camera video, microphone audio, and interaction logs before registration or class. This invention integrates and analyzes these multiple data types, collecting learning data and inferring the learner's state every 10 seconds. When a learner is detected to be in a frustrated state (three consecutive error corrections + facial AU activation + continuously declining tone), the strategy engine triggers an emotional support + reduced-difficulty follow-up questioning strategy: pausing the current grammar practice, switching to reviewing relevant vocabulary already mastered by the learner, accompanied by encouraging verbal remarks, and gradually advancing the lesson after the learner's state returns to a stable state. The entire chain of this process is recorded for L2 profile updates and effect evaluation.

[0037] Definitions of terms used in this invention: The autonomic nervous system (ANS) and related neurochemicals serve as a mechanistic explanatory layer for human systemic behavior, providing a theoretical basis for interpretable outputs. Sympathetic activation (stress state / frustration): corresponds to attention deficit, decreased comprehension, and increased escape tendency; stress reduction strategies are triggered after system detection. Parasympathetic activation (safe state): corresponds to working memory recovery and information is more easily transferred to long-term memory; the system goal is to activate this state. Dopamine pathway (motivation / expectation): design with fast feedback, sense of progress and challenge; the system reinforces sustained engagement through immediate positive feedback; Serotonin (stable baseline mood): provides stable mood support, maintained through consistent rhythm and positive encouragement; Endorphins (persistence): The system promotes learners' steady-state focus during more challenging tasks through a moderate difficulty gradient design.

[0038] The above are merely embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. An AI adaptive tutoring system based on a hierarchical learning state engine, comprising an online learning platform for learners to learn online, characterized in that, The online learning platform includes a learner's end and a server end. The server end includes an AI tutoring module and an adaptive module. The learner's end allows learners to conduct online interactive learning through the AI ​​tutoring module. The adaptive module includes: The L0 signal layer is used to collect learner learning data in the learning scenario after the learner has authorized it. The L1 state layer is used to analyze the learner's real-time learning state through a multimodal fusion model, classify the real-time learning state, select and trigger the corresponding interactive intervention strategy from the preset strategy library according to the category of the real-time learning state, and dynamically adjust the interactive intervention strategy in combination with the personalized profile results of the L2 profile layer, so as to adjust the learning content and tutoring strategy of the AI ​​tutoring module. The L2 profiling layer is used to create personalized profiles of learners based on stress threshold, recovery speed, confusion tolerance, and motivation type, and the profile results are fed back to the L1 state layer to continuously optimize interactive intervention strategies.

2. The AI ​​adaptive tutoring system based on a hierarchical learning state engine according to claim 1, characterized in that: The learning data includes: video data, audio data, interactive behavior data, learning process data, and learning result data.

3. The AI ​​adaptive tutoring system based on a hierarchical learning state engine according to claim 2, characterized in that: The video data includes learners' facial expressions, eye contact, and body movements during learning; the audio data includes learners' speech rate, pauses, and intonation characteristics; the interactive behavior data includes events such as clicking, answering, dragging, saving, pausing, playing at double speed, triggering prompts, and page switching; the learning process data includes reaction time, number of error corrections, prompt requests, and knowledge point browsing paths; and the learning result data includes quiz scores, task completion rate, and knowledge point mastery assessment.

4. The AI ​​adaptive tutoring system based on a hierarchical learning state engine according to claim 1, characterized in that: The categories of real-time learning states include confusion, frustration, inattentiveness, steady-state focus, and cognitive overload.

5. The AI ​​adaptive tutoring system based on a hierarchical learning state engine according to claim 4, characterized in that: The interactive intervention strategies in the preset strategy library include: When the real-time learning state is one of confusion, strategies such as reducing task difficulty, providing knowledge scaffolding, or step-by-step questioning are triggered. When the real-time learning state is one of frustration, a low-load strategy is triggered: pause the current task, output emotional support words, or switch to a low-load interaction mode (such as reviewing and practicing knowledge points that have been mastered). When the real-time learning state is characterized by inattentiveness, a low-attention strategy is triggered: increase attention-awakening prompts, push short achievement tasks, or accelerate the teaching pace. When the real-time learning state is a steady-state focus, a high-attention strategy is triggered: maintain or increase the current challenge, increase teaching content, or reduce unnecessary interventions; When the real-time learning state is cognitive overload, a low-cognitive strategy is triggered: reduce information density, increase information processing time, or switch practice modes.

6. The AI ​​adaptive tutoring system based on a hierarchical learning state engine according to claim 1, characterized in that: The L2 profiling layer profiles learners based on the following dimensions: Stress threshold: the number of consecutive errors or the level of cognitive load that triggers sympathetic nerve activation signals in learners; Recovery speed: The type of strategy and time required for learners to recover from a state of frustration / distraction to a steady state of focus; Distress tolerance: The level of distress at which learners can maintain effective learning; Types of motivation: Learners' learning motivations can be categorized into three types: quick feedback-driven, social recognition-driven, and long-term achievement-driven. State response pattern classification: Learners are divided into frustration-sensitive, low-arousal distracted, and confusion-tolerant types based on clustering methods; The L2 profiling layer feeds back the learner's profile to the L1 state layer, and selects and triggers the corresponding interactive intervention strategy from the preset strategy library.

7. The AI ​​adaptive tutoring system based on a hierarchical learning state engine according to claim 1, characterized in that: The server also includes an interpretability recording module, which records and outputs the complete causal chain from the collected learning data to the output interactive intervention strategy.

8. An AI adaptive tutoring method based on a hierarchical learning state engine, using the AI ​​adaptive tutoring system based on a hierarchical learning state engine as described in claim 1, characterized in that, Includes the following steps: S1. After obtaining the learner's authorization, collect the learner's learning data in the learning scenario; S2. Extract facial emotion features, speech prosody features, and behavioral rhythm features from the learning data; S3. Analyze the learner's real-time learning status and classify the real-time learning status; S4. Select and trigger the corresponding interactive intervention strategy from the preset strategy library; S5. Record and output the complete causal chain from the collected learning data to the output interactive intervention strategy.

9. The AI ​​adaptive tutoring method based on a hierarchical learning state engine according to claim 8, characterized in that: The features extracted in step S2 include: facial motion units, speech fundamental frequency and pause features, and behavioral response delay features.

10. The AI ​​adaptive tutoring method based on a hierarchical learning state engine according to claim 8, characterized in that: It also includes the following steps: S6. Online strategy evaluation: The real-time learning status of learners after the interactive intervention strategy is triggered in step S4 is analyzed through the L1 state layer, and the interactive intervention strategy in step S4 is scored. When the score is less than the preset threshold, it is reverted to a conservative strategy or manual intervention is initiated. S7. Update the personalized parameters of the L2 profile layer based on the intervention results corresponding to the interactive intervention strategy in step S4, and continuously optimize the matching accuracy between the learner's profile and the interactive intervention strategy.

Images