A multimodal learning state recognition and teaching intervention arrangement system

Abstract

The application provides a multimodal learning state recognition and teaching intervention arrangement system, which collects multimodal data in the learning process and adds timestamps, completes time alignment and quality evaluation of each mode, and converts into desensitization feature vectors and / or statistics and encrypted storage; based on the quality points weighted fusion of multimodal data, the learning state recognition result is output after consistency verification; the context state quantity is generated by synchronously analyzing the current learning task content; the intervention instruction is generated by comprehensively considering the above results; when the preset intervention condition is not met, the intervention is automatically inhibited or degraded; the learning response feedback after intervention is recorded continuously, and the intervention strategy parameters are optimized. The system realizes stable determination of the learning state through multi-source collection and quality evaluation in the learning process, and outputs controllable teaching intervention instructions through intervention arrangement and excessive intervention inhibition, thereby improving the reliability and safety in the case of noise, shielding, weak network or data loss.

Description

Technical Field

[0001] This invention relates to the field of online education technology, and in particular to a multimodal learning state recognition and teaching intervention scheduling system. Background Technology

[0002] Existing online / intelligent teaching systems typically consist of three parts: a learning terminal, a server, and data storage. The learning terminal displays teaching content and records interaction logs such as answer results, pause duration, and clicks. The server analyzes these interaction logs based on knowledge tracking or recommendation models and generates content recommendations or difficulty adjustment instructions. Data storage is used to save students' historical learning records and knowledge mastery levels.

[0003] The aforementioned systems are mostly one-way closed loops of "students explicitly submit data → server-side calculation → content distribution". The data sources are mainly the accuracy rate of answers, time spent, and click behavior. They usually do not include real-time perception of natural interaction signals such as facial expressions and voice during the learning process, and also lack real-time judgment and controllable intervention mechanisms for the learning status.

[0004] Furthermore, in real learning environments, there are situations such as occlusion, noise, weak networks, and lack of student cooperation. When relying on data from a single source (such as answer logs or a single sensor), the determination of learning status is prone to misjudgment, which may lead to incorrect or excessive intervention.

[0005] In summary, existing online / intelligent teaching systems have the following drawbacks: 1. The data source is too singular and it is difficult to reflect the real-time changes in the learning process, which causes the system to continue to push content at the established pace when students are confused, frustrated, bored or tired. 2. The lack of a real-time intervention scheduling mechanism oriented towards the learning process makes it impossible to control the "intervention type, intervention intensity, intervention rhythm and cooling window", which is prone to ineffective intervention or over-intervention. 3. Insufficient adaptability to complex environments, making it difficult to guarantee the reliability of judgments in the presence of noise, obstruction, weak network, or missing data; 4. The lack of a data processing mechanism that minimizes privacy poses a risk of privacy leakage during the raw audio and video capture and transmission. Summary of the Invention

[0006] In view of the above problems, the present invention is proposed to provide a multimodal learning state recognition and teaching intervention arrangement system that overcomes or at least partially solves the above problems.

[0007] This invention provides a multimodal learning state recognition and instructional intervention scheduling system, comprising: The multi-source acquisition unit is used to acquire facial image sequences, speech sequences, text input, and behavior logs during the learning process, and adds timestamps to each modality of data. The timing synchronization and modal quality assessment unit is used to time-align the data of each modality and output the quality score of each modality. The local characterization and privacy minimization processing unit is used to convert each modality data into desensitized feature vectors and / or statistics, and to store them encrypted. The credibility fusion and consistency verification unit is used to perform credibility fusion based on the quality scores of each modality data, and to perform consistency verification on the fusion results of different modalities, outputting learning state labels, confidence scores and interventionability scores. The learning task context parsing unit is used to obtain the current learning task information and generate context state variables based on the current learning task information. An intervention scheduling and intensity control unit is used to output intervention instructions based on learning state labels, confidence levels, interventionability scores, and contextual state quantities. The intervention instructions include at least intervention type, intervention intensity, intervention rhythm, and cool-down window parameters. The over-intervention suppression and safety control unit is used to suppress or downgrade intervention instructions when confidence is insufficient, the cooling window has not been reached, the maximum number of interventions has been reached, or the learning critical step is under the condition that the intervention instruction is suppressed or downgraded. The feedback recording and strategy update unit is used to record the learning response after the intervention is implemented and to update the intervention orchestration parameters of the intervention instructions.

[0008] Optionally, the timing synchronization and modal quality assessment unit is further configured to: align each modal data according to the timestamp of each modal data, and calculate at least one or more of the following as the quality score of each modal data: speech signal-to-noise ratio, face detection confidence, text completeness or behavior log missing rate.

[0009] Optionally, the credibility fusion and consistency verification unit is further configured to: determine the contribution of each modal data based on the quality score of each modal data; when the quality score of a modal data is lower than the quality score threshold, automatically reduce the contribution of the modal data or remove the modal data, and trigger a degradation path.

[0010] Optionally, the current learning task information includes at least one or more of the following: knowledge points, question types, error chain length, pause duration, and learning objectives.

[0011] Optionally, the intervention types include at least one or more of the following: encouragement prompts, rest prompts, difficulty adjustment, explanation method switching, question type switching, pace adjustment, and teacher intervention prompts.

[0012] Optionally, the system further includes a teacher strategy takeover interface unit. The teacher strategy takeover interface unit is used to receive adjustment information issued by the teacher or platform, and to take over or constrain the intervention program according to the adjustment information; the adjustment information includes at least one or more of the following: whitelist, blacklist, intensity limit, and cooling window parameters.

[0013] Optionally, the system further includes a weak network and offline degradation unit. The Weak Network and Offline Degradation Unit is used to reduce the sampling rate when the network quality is below the weak network judgment threshold, or to only upload desensitized feature vectors and / or statistics, and to set high priority for local basic prompt output.

[0014] Optionally, the system further includes a physiological signal assistance unit. A physiological signal auxiliary unit is used to collect skin conductance, skin temperature, or low-frequency statistics output by wearable devices, which serve as calibration factors for reliability fusion and consistency verification.

[0015] Optionally, the system further includes a class situation aggregation unit. The class status aggregation unit is used to aggregate the desensitized feature vectors and / or statistics of multiple learners and output the fatigue and / or engagement status trends at the class level.

[0016] Optionally, the degradation path is to maintain basic prompt output under weak network and / or occlusion conditions, and to synchronously update the desensitized feature vectors and / or statistics to update the intervention orchestration parameters after the network recovers.

[0017] This invention has the following advantages: The multimodal learning state recognition and instructional intervention scheduling system of this invention synchronously acquires multimodal data such as facial images, speech, text input, and behavior logs during the learning process through a multi-source acquisition unit, and adds timestamps. A temporal synchronization and modal quality assessment unit aligns the data and outputs quality scores for each modality. Combined with local featureization and privacy minimization processing, the data is converted into desensitized feature vectors and / or statistics and stored encrypted. A credibility fusion and consistency verification unit fuses and verifies the multimodal data based on the quality scores, generating learning state labels, confidence levels, and interventionizability scores. Simultaneously, a learning task context parsing unit obtains current task information and generates contextual state quantities. An intervention scheduling and intensity control unit outputs intervention instructions including type, intensity, pace, and cooling-off window based on the above information. An over-intervention suppression and safety control unit suppresses or downgrades instructions when confidence is insufficient, the cooling-off window has not been reached, the number of interventions exceeds the limit, or critical steps are taken, to avoid over-intervention. Finally, a feedback recording and strategy update unit dynamically adjusts the scheduling parameters based on the intervention response. During the learning process, the system achieves stable determination of the learning status through multi-source acquisition and quality assessment. It also improves the reliability and security in the presence of noise, occlusion, weak network, or missing data by controlling the output of teaching intervention instructions through intervention arrangement and over-intervention suppression. Attached Figure Description

[0018] Figure 1 This is a system structure diagram of a multimodal learning state recognition and teaching intervention arrangement system provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] This invention proposes a learning task context-driven multimodal learning state recognition and teaching intervention scheduling system, which is used to stably determine the learner's state during the learning process and output controllable intervention instructions to solve the problems of single data source, unstable determination, uncontrollable intervention and privacy risks in the prior art.

[0021] The main process flow of this system is as follows: Step 1: Multi-source data acquisition. Collect facial image sequences, speech sequences, text input, and behavior logs, and timestamp them.

[0022] Step 2: Timing synchronization and modal quality assessment. Time-align each modality and output quality scores for subsequent fusion and degradation.

[0023] Step 3: Local Feature Encoding and Privacy Minimization. The raw data is converted into de-identified feature vectors and / or statistics, and stored encrypted. Upload is controlled according to a policy.

[0024] Step 4: Credibility fusion and consistency verification, and output learning status. Based on the quality score, perform credibility fusion and consistency verification, and output learning status label, confidence level, and interventionability score; when the modality quality is insufficient, enter the degradation path.

[0025] Step 5: Intervention Orchestration and Safety Control. Based on the learning task context state, generate intervention instructions including intervention type, intensity, pace, and cooldown window. Suppress / degrade these instructions using over-intervention suppression logic before outputting them. Record intervention feedback and update policy parameters.

[0026] Reference Figure 1 The system of this invention adopts a modular structure and can be deployed in a hybrid architecture of learning terminals and servers. The learning terminal includes at least a camera, a microphone, and an interactive input interface; the server is used to store strategy parameters and output teaching content. The modules are connected to each other through data interfaces to form a closed loop.

[0027] (1) Multi-source acquisition unit ①: includes camera, microphone, text input interface and behavior log recorder, used to acquire facial image sequence, voice sequence, text input and behavior log, and timestamp each data.

[0028] (2) Timing synchronization and modal quality assessment unit ②: used to perform time alignment of each modality and output modal quality score; the quality score includes at least one or more of the following: speech signal-to-noise ratio, face detection confidence, text integrity or behavior log missing rate.

[0029] (3) Local characterization and privacy minimization processing unit ③: used to convert the original audio and video, text data and behavioral data into desensitized feature vectors and / or statistics, and store them in encryption; preferably, only the desensitized feature vectors and / or statistics are uploaded and the original audio and video are not uploaded.

[0030] (4) Credibility Fusion and Consistency Verification Unit ④: Based on the modality quality score, credibility fusion is performed, and consistency verification is performed on the output results of different modalities. The learning state label, confidence level, and interventionizability score are output. When the quality of a modality is lower than the threshold, its contribution is automatically reduced or it is removed, and a degradation path is triggered (e.g., judgment is based solely on behavioral logs and text). The learning state label is preferably a label related to the teaching scenario, including but not limited to: confusion, frustration, boredom, engagement, and fatigue. At the same time, the confidence level and interventionizability score are output to avoid triggering strong intervention under low confidence conditions.

[0031] (5) Learning task context parsing unit ⑤: used to obtain current learning task information and generate context state quantity. The learning task information includes at least one or more of the following: knowledge points, question type, error chain length, pause duration, and learning objectives.

[0032] (6) Intervention arrangement and intensity control unit ⑥: used to output intervention instructions according to the learning status. The intervention instructions include at least intervention type, intervention intensity, intervention rhythm and cooling window parameters; wherein the intervention type includes at least one or more of the following: encouragement prompt, rest prompt, difficulty adjustment, explanation method switching, question type switching, rhythm adjustment and teacher intervention prompt.

[0033] (7) Over-intervention suppression and safety control unit: used to suppress or downgrade intervention instructions under conditions such as insufficient confidence, failure to reach the cooling window, reaching the upper limit of the number of interventions, or being in a critical learning step; preferably, only output prompts without changing the teaching content when the confidence is below the threshold.

[0034] (8) Feedback Recording and Strategy Update Unit ⑧: Used to record learning responses after intervention (such as continued learning, pause, changes in answering time, etc.) and to update intervention programming parameters.

[0035] (9) Teacher Strategy Takeover Interface Unit (⑨, optional): Used to receive one or more of the parameters such as whitelist, blacklist, intensity limit, and cooling window issued by the teacher / platform end, and to take over or constrain the intervention arrangement.

[0036] (10) Weak network and offline degradation unit (⑩, optional): used to reduce the sampling rate or only upload desensitized feature vectors and / or statistics when the network quality is below the threshold, and to complete the status determination locally, and ensure that necessary prompts can still be output locally.

[0037] (11) Physiological signal auxiliary unit (optional): used to collect skin conductance, skin temperature or low-frequency statistics output by wearable devices. Its output is only used as a calibration factor for credibility fusion and consistency verification, and is used to reduce the probability of false triggering when there is occlusion or high noise. The physiological signal is not a necessary input for learning state recognition, nor is it the only basis for intervention triggering.

[0038] (12) Class Status Aggregation Unit (optional): Aggregate the desensitized feature vectors and / or statistics of multiple learners to output the fatigue and / or engagement status trends at the class level for teacher reference.

[0039] Note: Each of the above units can be implemented by hardware circuits, software modules executed by a processor, or a combination thereof.

[0040] The aforementioned threshold parameters include modal quality threshold, confidence threshold, cooling window parameter, upper limit of intervention times, and weak network judgment threshold, etc., all of which are configurable strategy parameters. They are preferably issued by the teacher strategy takeover interface unit or platform strategy and stored in the system's strategy parameter storage area, and can be updated during operation.

[0041] Optional units do not upload the original audio and video, but only upload the desensitized feature vector and / or statistics.

[0042] Through the aforementioned system modules, this system can make a degradation judgment based on quality assessment when there is occlusion, noise, weak network, or missing data, avoiding false triggering of a single modality; and achieves a "controllable and safe" closed loop of teaching adjustment through intervention in program arrangement and suppression of over-intervention. Its technical effects include, but are not limited to: (1) Stability improvement: Through modal quality assessment and consistency verification, the system automatically degrades when there is single modal distortion (such as speech noise or face occlusion), reducing the probability of false triggering of intervention; (2) Improved controllability of intervention: By scheduling interventions (type, intensity, rhythm, cooling-off window) and suppressing over-intervention, the frequency and intensity of interventions can be controlled, reducing ineffective or excessive interventions; (3) Enhanced privacy and compliance: By local characterization and privacy minimization, it is preferred to upload only de-identified features / statistics to reduce the risk of leakage of original audio and video; (4) Improved system adaptability: Through weak network and offline degradation strategies, basic status judgment and necessary prompt output can still be maintained when network conditions are poor.

[0043] The following are three specific examples of the present invention: 1. K-12 homework scenarios (math / physics problems) (1) Scene description: Learners can practice math / physics at home using tablets or mobile phones. The learning device is equipped with a front-facing camera, microphone, and touchscreen input, and records answer logs (answer time, number of errors, number of replays, pause duration, etc.). The system can connect to a question bank / course server for distributing questions and explanations.

[0044] (2) Technical application solution: After the learning terminal collects facial, voice, text, and behavior logs, the temporal synchronization and modal quality assessment unit first outputs a quality score; in the case of occlusion or noise, the quality score is automatically downgraded, preferably based solely on behavior logs and text to determine the learning state; then the learning task context parsing unit generates a context state quantity by combining the current knowledge points and error chains; the intervention orchestration and intensity control unit outputs intervention instructions including intervention type, intensity, and cooling window, and the over-intervention suppression and safety control unit suppresses / downgrades the intervention when the confidence level is insufficient or the cooling window has not been reached, to avoid excessive interruption of learning; finally, the intervention feedback is recorded to update the orchestration parameters.

[0045] (3) Example output and intervention: When the error is identified as "confused / frustrated" and the error chain length continues to increase, and similar question types are answered incorrectly consecutively: an intervention of "switching the explanation method (step-by-step example) + intensity 2" is generated; if it is in a critical calculation step, it is downgraded to "hint card (guide the next step) + intensity 1".

[0046] When "fatigue" is identified and the pause duration increases and attention decreases: generate "short rest prompt + intensity 2", and at the same time slow down the pace of the questions or switch to shorter question stems.

[0047] When the confidence level is insufficient: only output "hint_only", without changing the difficulty of the question or the learning path.

[0048] Even under weak network / obstruction conditions, basic prompts can still be maintained through a downgraded path (e.g., providing mild prompts based only on behavior logs and text), and de-identified statistics can be synchronized for policy updates after the network is restored.

[0049] 2. Classroom tablet learning scenario (teacher-managed strategy) (1) Scene description: In school classrooms, students use standardized tablets for in-class exercises and quizzes. Teachers can view class trends on their own (lectern computer / teacher tablet) and take over or restrict intervention strategies (e.g., limiting intervention intensity, setting cool-down periods, setting whitelists / blacklists).

[0050] (2) Technical application solution: After the learning terminal collects facial, voice, text, and behavior logs, the temporal synchronization and modal quality assessment unit first outputs a quality score; in the case of occlusion or noise, the quality score is automatically downgraded, preferably based solely on behavior logs and text to determine the learning state; then the learning task context parsing unit generates a context state quantity by combining the current knowledge points and error chains; the intervention orchestration and intensity control unit outputs intervention instructions including intervention type, intensity, and cooling window, and the over-intervention suppression and safety control unit suppresses / downgrades the intervention when the confidence level is insufficient or the cooling window has not been reached, to avoid excessive interruption of learning; finally, the intervention feedback is recorded to update the orchestration parameters.

[0051] (3) Example of teacher strategy takeover: Teachers should set the following policies in the "Exam / Test Mode": disable "automatic difficulty reduction / automatic question type switching", allow only "hint / rest type" interventions, and set the intensity limit to 2.

[0052] When the class as a whole shows an "increasing fatigue trend", the teacher can allow the system to output a "short rest reminder (intensity 1)" to the whole class, but set a global cooldown window (e.g., not to repeat within 15 minutes).

[0053] When individual students experience repeated "frustration" with high confidence, the system first provides a "prompt card (intensity 1)". If there is no improvement after multiple consecutive attempts and the cooldown window is not triggered, it can be upgraded to "teaching method switch (intensity 2)", but it is still subject to the teacher's intensity limit.

[0054] Even under weak network / occlusion conditions, basic prompts can still be maintained through a downgraded path (e.g., providing conservative prompts based solely on answer time, error chains, and pause duration), and desensitized statistics can be uploaded to update intervention programming parameters after the classroom network is restored.

[0055] 3. Language speaking practice scenarios (primarily speech-based, automatically downgrading in noisy environments) (1) Scene description: Learners use their mobile phones to practice foreign language speaking in commuting / public environments. The learning terminal collects voice, facial (optional), text input (reading content / error correction prompts), and behavior logs (number of repetitions, pauses, playback times, etc.). Due to the greater prevalence of environmental noise and occlusion, the system needs to have stronger quality assessment and degradation capabilities.

[0056] (2) Technical application solution: After the learning terminal collects facial, voice, text, and behavior logs, the temporal synchronization and modal quality assessment unit first outputs a quality score; in the case of occlusion or noise, the quality score is automatically downgraded, preferably based solely on behavior logs and text to determine the learning state; then the learning task context parsing unit generates a context state quantity by combining the current knowledge points and error chains; the intervention orchestration and intensity control unit outputs intervention instructions including intervention type, intensity, and cooling window, and the over-intervention suppression and safety control unit suppresses / downgrades the intervention when the confidence level is insufficient or the cooling window has not been reached, to avoid excessive interruption of learning; finally, the intervention feedback is recorded to update the orchestration parameters.

[0057] (3) Typical behaviors and interventions under noise / obstruction: When the speech signal-to-noise ratio is low (quality score is insufficient): the contribution of the speech modality is automatically reduced, and the "frustration / confusion" is determined based on behavioral logs such as "increased number of repeated readings, longer pauses, and increased number of playbacks" and text correction results.

[0058] When the question is judged as "confused" and the knowledge point is pronunciation rules / connected speech: "Explanation method switch (pronunciation mouth shape prompt + short example sentence) + intensity 2" is generated; if it is in a key practice step or the cooldown window has not been reached, it is downgraded to "Prompt card (give a key pronunciation point) + intensity 1".

[0059] When the exercise is judged to be "bored / fatigued" and the practice time is too long: generate "rhythm adjustment (shorten sentences, reduce practice density) + rest prompt (intensity 1~2)" and set a cooldown window to avoid frequent interruptions.

[0060] Even under weak network / obstruction conditions, basic prompts can still be maintained through a degradation path (e.g., providing gentle prompts based on behavior logs and text correction), and after the network is restored, de-identified statistics will be synchronized for subsequent policy updates and personalized orchestration.

[0061] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0062] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, electronic devices, storage media, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0063] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0064] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0065] The above provides a detailed description of the multimodal learning state recognition and teaching intervention scheduling system provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application. The above embodiments are merely preferred embodiments given to fully illustrate this invention, and the protection scope of this invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on this invention are all within the protection scope of this invention.

Claims

1. A multimodal learning state recognition and instructional intervention scheduling system, characterized in that, The system includes: The multi-source acquisition unit is used to acquire facial image sequences, speech sequences, text input, and behavior logs during the learning process, and adds timestamps to each modality of data. The timing synchronization and modal quality assessment unit is used to time-align the data of each modality and output the quality score of each modality. The local characterization and privacy minimization processing unit is used to convert each modality data into desensitized feature vectors and / or statistics, and to store them encrypted. The credibility fusion and consistency verification unit is used to perform credibility fusion based on the quality scores of each modality data, and to perform consistency verification on the fusion results of different modalities, outputting learning state labels, confidence scores and interventionability scores. The learning task context parsing unit is used to obtain the current learning task information and generate context state variables based on the current learning task information. An intervention scheduling and intensity control unit is used to output intervention instructions based on learning state labels, confidence levels, interventionability scores, and contextual state quantities. The intervention instructions include at least intervention type, intervention intensity, intervention rhythm, and cool-down window parameters. The over-intervention suppression and safety control unit is used to suppress or downgrade intervention instructions when confidence is insufficient, the cooling window has not been reached, the maximum number of interventions has been reached, or the learning critical step is under the condition that the intervention instruction is suppressed or downgraded. The feedback recording and strategy update unit is used to record the learning response after the intervention is implemented and to update the intervention orchestration parameters of the intervention instructions.

2. The system according to claim 1, characterized in that, The timing synchronization and modal quality assessment unit is also used to: align each modal data according to the timestamp of each modal data, and calculate at least one or more of the following as the quality score of each modal data: speech signal-to-noise ratio, face detection confidence, text completeness or behavior log missing rate.

3. The system according to claim 1, characterized in that, The credibility fusion and consistency verification unit is also used to: determine the contribution of each modality data based on the quality score of each modality data; when the quality score of a modality data is lower than the quality score threshold, automatically reduce the contribution of the modality data or remove the modality data, and trigger a degradation path.

4. The system according to claim 1, characterized in that, The current learning task information includes at least one or more of the following: knowledge points, question types, error chain length, pause duration, and learning objectives.

5. The system according to claim 1, characterized in that, The intervention types include at least one or more of the following: encouragement prompts, rest prompts, difficulty adjustment, switching of explanation methods, switching of question types, adjustment of pace, and teacher intervention prompts.

6. The system according to claim 1, characterized in that, The system also includes a teacher strategy takeover interface unit. The teacher strategy takeover interface unit is used to receive adjustment information issued by the teacher or platform, and to take over or constrain the intervention program according to the adjustment information; the adjustment information includes at least one or more of the following: whitelist, blacklist, intensity limit, and cooling window parameters.

7. The system according to claim 1, characterized in that, The system also includes a weak network and offline degradation unit. The Weak Network and Offline Degradation Unit is used to reduce the sampling rate when the network quality is below the weak network judgment threshold, or to only upload desensitized feature vectors and / or statistics, and to set high priority for local basic prompt output.

8. The system according to claim 1, characterized in that, The system also includes a physiological signal auxiliary unit. A physiological signal auxiliary unit is used to collect skin conductance, skin temperature, or low-frequency statistics output by wearable devices, which serve as calibration factors for reliability fusion and consistency verification.

9. The system according to claim 1, characterized in that, The system also includes a class status aggregation unit. The class status aggregation unit is used to aggregate the desensitized feature vectors and / or statistics of multiple learners and output the fatigue and / or engagement status trends at the class level.

10. The system according to claim 3, characterized in that, The degradation path is to maintain basic prompt output under weak network and / or occlusion conditions, and to synchronously update the desensitized feature vectors and / or statistics to update the intervention orchestration parameters after the network recovers.

Images