AI-based Teaching Interaction Method and System

Abstract

The present invention proposes an AI-based teaching interaction method and system. The method includes: collecting multi-modal data of students; constructing an individual absorption model to calculate the knowledge absorption rate of students and predicting the trend of the knowledge absorption rate sequence of students; constructing a feature mapping to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, and constructing a dynamic adaptive content generator according to the feature mapping; obtaining the teaching content feature vector generated by the dynamic adaptive content generator and mapping it to the social behavior features of students to obtain a social feature vector. Dynamically adjusting the interactive recommendations and group assignments among students; constructing an adaptive learning curve model according to the second dynamic social weight update model and designing a feedback mechanism to update the interactive recommendations and group assignments among students; constructing a blockchain network to encrypt and upload the data output by the update model, and constructing a learning feedback smart contract. The present invention provides a more intelligent and efficient solution for modern education.

Description

Technical Field

[0001] The present invention belongs to the field of teaching interaction, and particularly relates to an AI-based teaching interaction method and system. Background Art

[0002] In modern education, personalized learning and teaching interaction have been increasingly emphasized. Due to the monotony and lack of pertinence of the traditional classroom teaching mode, it has been difficult to meet the diverse learning needs of students. To improve this situation, in recent years, teaching systems based on artificial intelligence (AI) have emerged. These systems provide personalized learning content and suggestions by analyzing students' learning data. However, there are still many limitations in the existing AI teaching systems in terms of realizing personalized teaching and enhancing interactivity.

[0003] Firstly, current AI teaching systems usually rely on students' historical learning data, such as exam scores and homework performance. These data can only reflect students' learning status at specific time points, but cannot comprehensively capture students' cognitive changes and emotional responses. The limitations of this static data result in the system being unable to adjust teaching content in real time, unable to respond promptly to the changing needs of students during the learning process, and lacking dynamic adaptability. In addition, the feedback mechanisms of existing systems are usually relatively single, mainly based on standardized scoring and simple prompts, lacking precise perception of students' in-depth understanding and knowledge absorption processes. Therefore, these systems often cannot effectively help students overcome learning difficulties or stimulate their learning interests.

[0004] Secondly, although some AI teaching systems have begun to introduce emotion computing and behavior analysis technologies, such as judging students' emotional states through facial expression recognition and speech emotion detection, the application of these technologies in the education field is not yet mature, facing challenges such as insufficient accuracy and inability to handle complex emotional states. This leads to easy misjudgment when the system judges students' learning status, thus affecting the effect of personalized teaching. In addition, existing systems pay more attention to the learning experience of individual students in design, but ignore the social interaction among students and the optimization of group learning. Modern learning research shows that social interaction and collaborative learning are crucial for students' knowledge construction and skill development, but the existing technology still has limited support in this regard. There is a lack of in-depth analysis of the interaction within the learning group, and it is unable to effectively promote knowledge exchange and collaborative cooperation among students.

[0005] Furthermore, the improvements in most AI teaching systems in terms of adaptability and personalization still remain at a superficial level. For example, simply recommending more similar learning content or adjusting the content difficulty without deeply analyzing the specific characteristics of students' learning curves and the knowledge absorption process. Such system designs result in the inability to provide the most optimized learning paths and strategy adjustments when facing students with different backgrounds and learning abilities. In addition, there are also certain challenges in data privacy and security in the existing technologies, especially when sharing and verifying learning data across platforms, where there is a risk of data leakage or tampering. These deficiencies significantly limit the potential and application effects of AI in the field of teaching interaction. Summary of the Invention

[0006] The object of the present invention is to propose an AI-based teaching interaction method and system. By combining multi-modal data fusion, adaptive learning algorithms, social network analysis, adaptive learning curve modeling, and blockchain technology, it significantly improves the personalization, interactivity, and security of the AI teaching interaction system, overcomes many defects of the existing technologies, and provides a more intelligent and efficient solution for modern education.

[0007] To achieve the above object, the present invention provides an AI-based teaching interaction method, and the method includes:

[0008] S1. Deploy sensor devices to collect students' multi-modal data, preprocess the multi-modal data, extract and fuse features from the preprocessed data to obtain a comprehensive feature vector, and store the comprehensive feature vector as a multi-modal data set; wherein, the comprehensive feature vector includes EEG spectrum features, heart rate variability features, eye movement features, and facial expression features;

[0009] S2. Receive the multi-modal data set sent by the sensor and construct an individual absorption model based on the multi-modal data set to calculate the knowledge absorption rate of the student at each time point and predict the trend of the student's knowledge absorption rate sequence;

[0010] S3. Construct a feature mapping to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, construct a dynamic first adaptive content generator according to the feature mapping, design a joint loss function, and jointly train and optimize the discriminator and the first adaptive content generator according to the feature mapping to obtain a second adaptive content generator;

[0011] S4. Obtain the teaching content feature vector generated by the second adaptive content generator, map it to the social behavior characteristics of students to obtain a social feature vector, construct a student social network using the social feature vector, build a first dynamic social weight update model based on the student social network, and design a group interaction optimization algorithm to adjust the weights of the first dynamic social weight update model to obtain a second dynamic social weight update model. Generate a new social interaction strategy according to the second dynamic social weight update model, and dynamically adjust the interaction recommendations and group assignments among students;

[0012] S5. Construct an adaptive learning curve model based on the second dynamic social weight update model, and design a feedback mechanism to update the interaction recommendations and group assignments among students;

[0013] S6. Construct a blockchain network to encrypt and upload the data output by the second dynamic social weight update model, and construct a learning feedback smart contract to automatically execute data verification, storage, and feedback optimization, and design a learning data consensus algorithm to judge the reception of data blocks; wherein, the learning data consensus algorithm is that all nodes use a voting mechanism to decide whether to accept a new data block, and only when all nodes vote through can the new data block be accepted;

[0014] Among them, in the student social network, each node represents a student, and the weight of the edge represents the strength of the social connection between students; the first dynamic social weight update model updates the weight of the edge in real time according to the knowledge absorption rate and social behavior of students; the calculation of the weight of the edge is as follows:

[0015] w i,j,t =α·sim(F social,i,t ,F social,j,t )+β·|KAR i,t -KAR j,t |

[0016] Among them, w i,j,t represents the edge weight between student i and student j at time point t; α and β represent weight coefficients, which are used to balance the influence of social behavior similarity and knowledge absorption rate difference; sim(F social,i,t ,F social,j,t ) represents the social feature similarity between student i and student j at time point t; |KAR i,t -KAR j,t | represents the difference in knowledge absorption rate between student i and student j at time point t;

[0017] Design a group interaction optimization algorithm to increase the interaction between the student with the highest knowledge absorption rate and the student with the lowest absorption rate. The group interaction optimization algorithm is expressed as follows:

[0018]

[0019] Among them, L opt represents the optimization objective function, which is used to minimize the difference in knowledge absorption rates among students in the social network; E represents the edge set in the social network; w i,j,t represents the weight of edges i and j; KAR i,t and KAR j,t represent the knowledge absorption rates of students i and j at time point t;

[0020] Update the edge weights according to the output of the group interaction optimization algorithm to obtain the second dynamic social weight update model. The edge weights of the second dynamic social weight update model are expressed as follows:

[0021] w i,j,t+1 = w i,j,t + γ·Δw i,j,t

[0022] where w i,j,t+1 represents the edge weight between student i and student j at time point t + 1; γ represents the learning rate, which is used to control the step size of weight update; Δw i,j,t represents the change amount of the edge weight.

[0023] Preferably, the sensor device includes an EEG head-mounted device, a heart rate monitor, an eye tracker, and a facial expression recognition camera; the preprocessing includes time synchronization processing, applying an adaptive band-pass filter to the data of the EEG head-mounted device to remove noise in the EEG data, processing the data of the heart rate monitor using a window smoothing algorithm, using a peak detection algorithm to identify fixations and saccades in the eye movement data to obtain eye movement features, and using a convolutional neural network to extract features from the micro-expression data to obtain facial expression features; the feature extraction includes using the short-time Fourier transform on the data of the EEG head-mounted device to obtain the power spectrum features of the α band and the β band within different time windows to obtain EEG spectrum features, and calculating the heart rate variability features using the standard deviation of the R-R interval;

[0024] Use the joint feature space mapping method to map different modality data to a shared feature space.

[0025] Preferably, the individual absorption model is constructed as follows:

[0026] Standardize the student state at each moment on the comprehensive feature vector, and construct a multi-modal time series embedding model introducing a modal attention weight mechanism. Combine the output of the modal attention mechanism with the update formula of the multi-modal time series embedding model to obtain an updated representation of the student's learning state. Among them, the modal attention weight mechanism assigns a weight to the data of each modality, indicating its relative contribution to the knowledge absorption rate at the current time point. The multi-modal time series embedding model processes multi-modal data and captures temporal dynamic changes.

[0027] Calculate the knowledge absorption rate at each time point and predict the future knowledge absorption trend based on the hidden state output by the multi-modal time series embedding model and the modal attention weights.

[0028] Preferably, in calculating the knowledge absorption rate at each time point, introduce a knowledge absorption rate regularization term based on L2 regularization to avoid overfitting and encourage weight sparsity. The prediction of the future knowledge absorption trend is obtained by inputting the hidden state into a fully connected layer with an adaptive learning rate, and the output of the fully connected layer is the prediction of the knowledge absorption rate at several future time points.

[0029] Preferably, the first adaptive content generator generates a first generated teaching content feature vector according to the state vector and the parameters of the generator. The joint loss function L(G, D) is expressed as follows:

[0030]

[0031] where G(S t ) represents the content feature generated by the generator; D(G(S t )) represents the adaptability score of the discriminator to the generated content; S t represents the state vector; D represents the learning state discriminator; G represents the knowledge enhancement generator; represents the average content feature vector in the training set, which is used to calculate content diversity; λ represents the regularization parameter, which controls the influence of the diversity term;

[0032] where the knowledge enhancement generator generates content features and inputs them into the learning state discriminator for learning state discrimination. The learning state discriminator is used to discriminate whether the generated content is suitable for the current learning state.

[0033] Preferably, after the first adaptive content generator is trained using the joint loss function L(G, D), a second adaptive content generator is obtained. Use the second adaptive content generator to generate in real time the teaching content features adapted to the student's learning state, and convert these features into a second generated teaching content feature vector through a predefined content library and mapping rules.

[0034] Preferably, the S5 specifically includes:

[0035] S501. Extract the learning behavior characteristics of each student to generate a learning behavior feature vector, which is used to describe the learning behavior and state of student i at time point t;

[0036] S502. Construct an adaptive learning curve model based on the dynamic Bayesian network according to the learning behavior feature vector, capture the changes in the learning state of students and the potential dependencies in the time series according to the knowledge absorption rate and social behavior characteristics of students, and reflect the learning progress and knowledge mastery of students at different time points;

[0037] S503. Design a real-time feedback mechanism according to the adaptive learning curve model to provide students with personalized learning suggestions and strategy adjustments; wherein, the real-time feedback mechanism is set according to the learning efficiency function, and the learning efficiency function is used to evaluate the current learning state and potential improvement direction of each student;

[0038] S504. Generate personalized feedback suggestions for students according to the calculation results of the learning efficiency function and optimize the learning strategy.

[0039] Preferably, the learning efficiency function is expressed as follows:

[0040]

[0041] where E i,t represents the learning efficiency of student i at time point t; KAR i,t+1 and KAR i,t represent the knowledge absorption rates of student i at time points t + 1 and t; Δt represents the time interval; ω represents the weight parameter, which is used to control the influence of the behavior feature difference on the learning efficiency; represents the average learning behavior feature;

[0042] The knowledge absorption rate represents the state transition equation of the adaptive learning curve model and is calculated as follows:

[0043]

[0044] where KAR i,t+1 represents the knowledge absorption rate of student i at time point t + 1; γ and δ represent the model parameters, which are used to balance the current absorption rate and the new learning contribution; λ represents the learning rate parameter, which is used to adjust the influence of the learning behavior feature on the absorption rate; F learn,i,t represents the learning behavior feature of student i at time point t.

[0045] Preferably, the encryption uses the elliptic curve cryptography algorithm to encrypt the data; the verification formula of the learning feedback smart contract is expressed as follows:

[0046]

[0047] Among them, F' data,i,t represents the encrypted data of student i at time point t; F data,i,t represents the unencrypted learning data feature vector; V(F' data,i,t , key priv ) represents the verification function of the smart contract; key priv represents the private key of the blockchain network for data decryption; ECC_decrypt represents the elliptic curve cryptography algorithm.

[0048] In the second aspect of the present invention, an AI-based teaching interaction system is provided, and the system includes:

[0049] A teaching modality data collection module for deploying sensor devices to collect multi-modal data of students, preprocessing the multi-modal data, extracting and fusing features of the preprocessed data to obtain a comprehensive feature vector, and storing the comprehensive feature vector as a multi-modal data set; wherein, the comprehensive feature vector includes EEG spectrum features, heart rate variability features, eye movement features, and facial expression features;

[0050] A knowledge absorption analysis module for receiving the multi-modal data set sent by the sensor and constructing an individual absorption model according to the multi-modal data set to calculate the knowledge absorption rate of the student at each time point and predict the trend of the student's knowledge absorption rate sequence;

[0051] A teaching content output module for constructing a feature mapping to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, constructing a dynamic first adaptive content generator according to the feature mapping, designing a joint loss function to jointly train and optimize the discriminator and the first adaptive content generator according to the feature mapping to obtain a second adaptive content generator;

[0052] A teaching interaction module for obtaining the teaching content feature vector generated by the second adaptive content generator, mapping it to the social behavior features of the students to obtain a social feature vector, constructing a student social network using the social feature vector, constructing a first dynamic social weight update model according to the student social network, and designing a group interaction optimization algorithm to adjust the weights of the first dynamic social weight update model to obtain a second dynamic social weight update model, and generating a new social interaction strategy according to the second dynamic social weight update model to dynamically adjust the interaction recommendations and group assignments among students;

[0053] An interaction feedback module for constructing an adaptive learning curve model according to the second dynamic social weight update model and designing a feedback mechanism to update the interaction recommendations and group assignments among students;

[0054] A data encryption module is used to encrypt and upload the data output by the second dynamic social weight update model in the blockchain network, construct a learning feedback smart contract to automatically execute data verification, storage, and feedback optimization, and design a learning data consensus algorithm to judge the reception of data blocks; wherein, the learning data consensus algorithm is that all nodes use a voting mechanism to decide whether to accept a new data block, and only when all nodes vote through can the new data block be accepted;

[0055] Among them, in the student social network, each node represents a student, and the weight of the edge represents the intensity of social connection between students; the first dynamic social weight update model updates the weight of the edge in real time according to the knowledge absorption rate and social behavior of students; the calculation formula of the weight of the edge is as follows:

[0056] w i,j,t =α·sim(F social,i,t ,F social,j,t )+β·|KAR i,t -KAR j,t |

[0057] Among them, w i,j,t represents the edge weight between student i and student j at time point t; α and β represent weight coefficients, which are used to balance the influence of social behavior similarity and knowledge absorption rate difference; sim(F social,i,t ,F social,j,t ) represents the social feature similarity between student i and student j at time point t; |KAR i,t -KAR j,t | represents the difference in knowledge absorption rate between student i and student j at time point t;

[0058] Design a group interaction optimization algorithm to increase the interaction between the student with the highest knowledge absorption rate and the student with the lowest absorption rate. The group interaction optimization algorithm is expressed as follows:

[0059]

[0060] Among them, L opt represents the optimization objective function, which is used to minimize the difference in knowledge absorption rate among students in the social network; E represents the edge set in the social network; w i,j,t represents the weight of edge i and j; KAR i,t and KAR j,t represent the knowledge absorption rates of students i and j at time point t;

[0061] Update the edge weight according to the output of the group interaction optimization algorithm to obtain the second dynamic social weight update model. The edge weight of the second dynamic social weight update model is expressed as follows:

[0062] w i,j,t+1 =wi,j,t +γ·Δw i,j,t

[0063] where w i,j,t+1 represents the edge weight between student i and student j at time point t + 1; γ represents the learning rate, which is used to control the step size of weight update; Δw i,j,t represents the change in the edge weight.

[0064] The beneficial technical effects of the present invention are at least as follows:

[0065] (1) The present invention adopts a multi-modal data fusion technology, integrating the physiological data (such as brain waves, heart rate), behavioral data (such as eye movement, micro-expression), and traditional learning performance data of students. Through deep learning algorithms, the system constructs a dynamic individual knowledge absorption model, which can track the process of students' understanding and absorption of teaching content in real time. Compared with the existing technology that only relies on the static analysis of historical learning data, the multi-modal data fusion of the present invention can comprehensively reflect the real-time cognitive state and emotional response of students, thus realizing more accurate personalized teaching.

[0066] (2) The present invention introduces a reinforcement learning algorithm, allowing the system to dynamically adjust the teaching content and interaction strategies according to the students' knowledge absorption model and real-time learning performance. The system can not only adjust the content difficulty according to the current understanding level of students, but also adjust the teaching style and interaction method in real time through sentiment analysis. For example, when students feel frustrated, more encouragement or appropriate rest suggestions are given. This adaptive ability greatly enhances the flexibility and pertinence of teaching, overcoming the limitation that the existing system cannot respond to the changing needs of students in real time.

[0067] (3) Aiming at the defect of insufficient attention to student interaction in the existing technology, the present invention adopts social network analysis (SNA) technology to analyze and optimize the interaction structure and knowledge dissemination path within the learning group in real time. The system can identify the key nodes (such as knowledge disseminators) and potential knowledge islands (such as students with low participation) in the learning group, and dynamically adjust the group members and collaborative task configuration to promote the group learning efficiency and collaborative effect. This innovation significantly improves the interactivity and effectiveness of group teaching.

[0068] (4) The present invention proposes an adaptive learning curve modeling technology, which combines the multi-dimensional data and historical learning records of students to dynamically predict the learning progress and understanding depth of students. Based on the changes in the learning curve model, the system provides instant intelligent feedback and reinforcement suggestions to ensure that each student can receive personalized guidance suitable for their learning state. This intelligent feedback mechanism can correct learning deviations in time, stimulate students' learning interest, and solve the problems of single feedback and inability to deeply analyze the learning process in the existing system.

[0069] (5) The present invention also introduces blockchain technology to ensure the security and credibility of learning data. Through decentralized data storage and smart contract mechanisms, the system can securely and transparently share and verify learning data, preventing the risks of data tampering and leakage. This innovation not only solves the data privacy problems of the existing technologies but also promotes cross-platform data collaboration and trusted verification. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] The present invention will be further described with reference to the accompanying drawings. However, the embodiments in the drawings do not constitute any limitation to the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on the following drawings without creative efforts.

[0071] Figure 1 It is a flowchart of the AI-based teaching interaction method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0072] The embodiments of the present invention will be described in detail below. The examples of the embodiments are shown in the accompanying drawings, where the same or similar reference numerals denote the same or similar elements or elements with the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and should not be construed as a limitation to the present invention.

[0073] As Figure 1 shown, the AI-based teaching interaction method provided by the embodiment of the present invention includes the following steps S1-S6:

[0074] S1. Deploy sensor devices to collect multi-modal data of students, preprocess the multi-modal data, extract and fuse features from the preprocessed data to obtain a comprehensive feature vector, and store the comprehensive feature vector as a multi-modal data set; wherein, the comprehensive feature vector includes EEG spectrum features, heart rate variability features, eye movement features, and facial expression features.

[0075] Specifically, to ensure the time alignment of the multi-modal data, all sensor devices (EEG head-mounted device, heart rate monitor, eye tracker, facial expression recognition camera) need to be calibrated before data collection to ensure the collection accuracy. The EEG device is set to a sensitivity of 0.1 μV, the error of the heart rate monitor is controlled within ±0.5 bpm, the accuracy of the eye tracker is 0.05°, and the facial expression recognition camera adjusts the light to adapt to environmental changes.

[0076] The device time synchronization adopts the weighted minimum variance time synchronization algorithm (WMVTS) to calculate a global timestamp to ensure that the data of all devices is aligned at the same time point. This method assigns weights to the timestamps of each device and calculates the weighted average time as the synchronization time.

[0077]

[0078] Among them, T sync is the global timestamp, and T i is the timestamp of the i-th sensor, and w i is the weight of the sensor (assigned based on the time accuracy and latency of the device). For example, if there are three sensor devices in the present invention, their timestamps are T1 = 100ms, T2 = 105ms, T3 = 102ms, and the weights are w1 = 0.2, w2 = 0.5, w3 = 0.3 respectively, then the global timestamp

[0079] Furthermore, EEG data preprocessing is to apply an Adaptive Bandpass Filter (ABF) to remove noise (such as electromyogram noise and baseline drift) in the EEG data. The bandpass frequency range of the filter is from 0.5 Hz to 70 Hz.

[0080] X EEG = ABF(EEG, [0.5, 70])

[0081] Among them, X EEG is the filtered EEG signal data.

[0082] Heart rate data preprocessing is to use the Windowed Smoothing Algorithm (WSA) to process heart rate data, eliminate short-term fluctuations, and ensure data stability.

[0083]

[0084] Among them, X HR is the smoothed heart rate data, and N is the size of the smoothing window. Assuming N = 5, for HR = [70, 72, 71, 73, 75], then

[0085] Eye movement data preprocessing is to use the peak detection algorithm for eye movement data to identify fixations and saccades, and record the fixation duration D f and the saccade speed S s . Micro-expression data uses a Convolutional Neural Network (CNN) for feature extraction.

[0086] Furthermore, for EEG data, the Short-Time Fourier Transform (STFT) is used to obtain the power spectral features of the alpha band (8 - 13 Hz) and the beta band (13 - 30 Hz) within different time windows:

[0087] F EEG = STFT(X EEG )

[0088] where F EEG is a spectral feature matrix containing the power spectral density of each time window.

[0089] Calculate the heart rate variability (HRV) feature, and use the standard deviation of the R-R interval (SDNN) to quantify the heart rate variability:

[0090]

[0091] where F HR is the standard deviation of heart rate variability, RR i is the R-R interval of each heartbeat, is the mean of the R-R intervals.

[0092] Furthermore, fuse the extracted features, and use the Joint Feature Space Mapping (JFSM) method to map different modality data to a shared feature space. The present invention adopts a multi-modal feature fusion strategy to combine the EEG spectral feature F EEG , the heart rate variability feature F HR , the eye movement features (D f , S s ), and the facial expression feature F MF to form a comprehensive feature vector F Fusion :

[0093] F Fusion = JSFM(F EEG , F HR , D f , S s , F MF )

[0094] where F Fusion is the fused comprehensive feature vector.

[0095] Furthermore, finally store the fused comprehensive feature vector as a multi-modal dataset D multi , for the construction of the subsequent individual knowledge absorption model:

[0096]

[0097] S2. Receive the multi-modal dataset sent by the sensor and construct an individual absorption model based on the multi-modal dataset to calculate the knowledge absorption rate of the student at each time point and predict the trend of the student's knowledge absorption rate sequence.

[0098] Specifically, using the multi-modal dataset generated in step S1 as input, an innovative individual knowledge absorption model (KnowledgeAbsorptionModel, KAM) is constructed. This model captures the student's real-time understanding of the learning content and the knowledge absorption rate, and predicts their future learning state. This model needs to be able to handle the heterogeneity and temporal dynamic changes of multi-modal data, and provide in-depth insights into the student's learning state.

[0099] Specifically, the input data is the fused feature vector F multi in the multi-modal dataset D Fusion,t , and these vectors represent the student state at each time point t, including EEG spectrum features, heart rate variability features, eye movement features, and facial expression features. To capture the temporal dynamic characteristics of the data, the present invention first normalizes the data so that the mean of each feature is 0 and the variance is 1:

[0100]

[0101] where is the normalized feature vector, μ F and σ F are the mean and standard deviation of the feature vector, respectively.

[0102] To extract the time series features of the student's learning state, the present invention uses an extended gated recurrent unit (GRU) model, called the multi-modal time series embedding model (MultimodalTimeSeriesEmbeddingModel, MTSEM). This model can handle multi-modal data and capture temporal dynamic changes, especially in terms of the impact of the student's emotional and physiological responses on learning.

[0103] To make full use of the relationships between different modal data, the present invention adds a new modality attention mechanism (ModalityAttentionMechanism, MAM) on the basis of GRU. This mechanism assigns a weight to the data of each modality, indicating its relative contribution to the knowledge absorption rate at the current time point. Let the feature vector of each modality be M i,t , then the calculation of the modality attention weight α i,t is:

[0104]

[0105] where α i,t is the attention weight of modality i at time point t; W m is the modality weight matrix, used to map the modality features to the attention weight space; M i,tis the feature vector of modality i at time point t.

[0106] For example, assume there are feature vectors M of three modalities 1,t = [0.2, 0.3], M 2,t = [-0.1, 0.5], M 3,t = [0.4, -0.2], and weight matrix W m = [0.5, 0.1]. The calculated attention weights are α 1,t = 0.3, α 2,t = 0.4, α 3,t = 0.3.

[0107] The output of the modality attention mechanism will be combined with the update formula of GRU to update the learning state representation of the student:

[0108]

[0109] where h t is the hidden state of GRU, representing the learning state representation at time point t; z t is the update gate, calculated by combining the attention weights, used to control the combination of the previous state h t-1 and the current candidate state ; is the candidate state, calculated using the current modality weights and input features.

[0110] Assume the input feature vector is F Fusion,t = [0.5, -0.1, 0.3,...], the previous state h t-1 = [0.2, 0.4, 0.6,...], the update gate z t = [0.7, 0.5, 0.3,...] and the candidate state can obtain the current hidden state h t .

[0111] Furthermore, use the hidden state h t output by GRU and the modality attention weights to calculate the Knowledge Absorption Rate (KAR) at each time point. For this purpose, the present invention introduces a new Knowledge Absorption Rate Regularization Term (KARRT), introducing L2 regularization into the calculation formula to avoid overfitting and encourage weight sparsity:

[0112]

[0113] where KAR tis the knowledge absorption rate at time point t; W k and b k are trainable weight matrices and biases; σ is the sigmoid activation function that normalizes the output between 0 and 1; is the L2 regularization term that controls the magnitude of the weights and avoids overfitting.

[0114] For example, assume the weight matrix W k = [0.2, 0.5, -0.3], the bias b k = 0.1, the hidden state h t = [0.6, 0.3, 0.9], the regularization parameter λ = 0.01, and the calculated

[0115] KAR t = σ(0.2·0.6 + 0.5·0.3 - 0.3·0.9 + 0.1) - 0.01(0.2 2 + 0.5 2 + (-0.3) 2 ) ≈ 0.515 - 0.011 = 0.504

[0116] Finally, the present invention uses the hidden state of the GRU and the attention mechanism to predict the future knowledge absorption trend. This is accomplished by inputting the hidden state into a fully connected layer with an adaptive learning rate (AdaptiveLearningRateFullyConnectedLayer, ALR-FC). The output of the fully connected layer is the prediction of the knowledge absorption rate at several future time points, which is used to guide the adjustment of teaching strategies:

[0117] KAR t+1:t+k = ALR-FC(h t , α t ; θ ALR-FC )

[0118] where, KAR t+1:t+k is the prediction of the knowledge absorption rate at the next k time points; ALR-FC is a fully connected layer with an adaptive learning rate, and the parameters are θ ALR-FC ; α t is the modal attention weight used to adjust the learning rate.

[0119] Assume the hidden state h t = [0.6, 0.2, 0.5,...] at the current time point t, the modal attention weight is α t = [0.3, 0.4, 0.3], the fully connected layer calculates the prediction for the next k = 2 time points, and the output is KAR t+1 = 0.65, KAR t+2 = 0.68.

[0120] Through the above steps, the present invention constructs an innovative individual knowledge absorption model, which can reflect and predict the learning status of students in real time. Each step and formula takes into account the special requirements in the patent scenario, such as the complexity of multi-modal data, the time-dynamic characteristics, and the requirements for prediction accuracy and robustness. The knowledge absorption rate and trend prediction output by the model will be used for the generation of personalized teaching content in the next step to enhance the teaching interaction effect.

[0121] S3. Construct a feature mapping to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, construct a dynamic first adaptive content generator according to the feature mapping, and design a joint loss function to jointly train and optimize the discriminator and the first adaptive content generator according to the feature mapping to obtain a second adaptive content generator.

[0122] Specifically, the input data is the knowledge absorption rate sequence KAR 1:t ={KAR1, KAR2, …, KAR t} and the future predicted absorption rate trend KAR t+1:t+k . These data reflect the learning status of students at each time point and the possible trends of future learning.

[0123] The present invention first performs a feature mapping on these input data to generate a state vector S t , which contains the current and predicted learning status information. This state vector is used to characterize the current learning status and trend of students.

[0124] S t =φ(KAR 1:t , KAR t+1:t+k )

[0125] Where φ is a feature mapping function used to transform the knowledge absorption rate and trend into a comprehensive state representation.

[0126] Furthermore, in order to generate personalized teaching content, the present invention designs a dynamic adaptive content generator (Dynamic Adaptive Content Generator, DACG). This generator generates a feature vector C t of teaching content based on the state vector S t , and this feature vector describes the attributes of the teaching content suitable for the current time point t (such as difficulty, type, media format, etc.).

[0127] The present invention uses a variant of the Generative Adversarial Network (GAN) to generate teaching content feature vectors. Specifically, a Knowledge-Enhanced Generator (KEG) is employed to generate content features, and a Learning State Discriminator (LSD) is used to determine whether the generated content is suitable for the current learning state. The output of the generator is represented by the following formula:

[0128] C t =KEG(S t ;θ G )

[0129] where C t is the generated teaching content feature vector, S t is the state vector, and θ G is the parameter of the generator. The generator generates a content feature vector that meets the current learning needs of the student based on the input state vector S t .

[0130] Furthermore, to train the generator and the discriminator, the present invention introduces a joint loss function L(G, D). This loss function combines the adaptability score of the content (the degree of match with the student's knowledge absorption state) and the content diversity regularization term (ensuring the diversity of the generated content and avoiding mode collapse). The output of the discriminator is used to update the parameters of the generator, enabling the generator to produce content features that better meet the needs of the student.

[0131]

[0132] where G(S t ) is the content feature generated by the generator; D(G(S t )) is the adaptability score of the discriminator for the generated content; is the average content feature vector in the training set, used to calculate content diversity; λ is the regularization parameter, controlling the influence of the diversity term.

[0133] For example, assume that the output of the generator is C t =[0.6, 0.8, 0.1], the output of the discriminator for the true state vector S t is 0.9, the output for the generated content is 0.1, and the content diversity regularization term is The joint loss function can be calculated as:

[0134] L(G, D)=log(0.9)+log(1 - 0.1)+0.1×0.05= - 0.1053

[0135] Furthermore, once the training is completed, the present invention uses the generator to generate in real time the teaching content feature C that adapts to the learning state of the students t , and converts these features into actual teaching content through a predefined content library and mapping rules. For example, if C t represents a content with a medium difficulty level and a video type, the system will select a teaching video that meets these features from the content library.

[0136] The system dynamically adjusts the content according to the feedback of the new knowledge absorption rate KAR t+1 to ensure that the generated teaching content can continue to adapt to the learning state of the students.

[0137] C t+1 = KEG(S t+1 ; θ G ), where S t+1 = φ(KAR 1:t+1 , KAR t+2:t+k+1 )

[0138] Assume that at time point t+1, the new state vector S t+1 is generated by the updated knowledge absorption rate sequence KAR 1:t+1 and the predicted future absorption rate trend KAR t+2:t+k+1 . The generator uses the updated state to generate a new content feature C t+1 .

[0139] Through this solution, the dynamic adaptive teaching content generation system can generate and adjust teaching content in real time according to the learning state of the students, ensuring that each student's learning experience is personalized and can maximize their knowledge absorption rate.

[0140] S4. Obtain the teaching content feature vector generated by the second adaptive content generator, map it to the social behavior features of the students to obtain a social feature vector, construct a student social network using the social feature vector, construct a first dynamic social weight update model according to the student social network, and design a group interaction optimization algorithm to adjust the weights of the first dynamic social weight update model to obtain a second dynamic social weight update model, and generate a new social interaction strategy according to the second dynamic social weight update model, and dynamically adjust the interaction recommendation and group allocation among the students.

[0141] Specifically, the input data is the teaching content feature vector C t generated in step 3, and the knowledge absorption rate sequence KAR 1:t of the students. First, the present invention maps these data to the social behavior features of the students to generate a social feature vector F social,t , representing the state of the students in the social network.

[0142] F social,t = ψ(C t , KAR 1:t )

[0143] where ψ is a feature mapping function that converts teaching content features and knowledge absorption rates into social behavior features.

[0144] Furthermore, using the generated social feature vector F social,t , the present invention constructs a social network among students. Each node represents a student, and the weight of the edge represents the strength of the social connection between students. The present invention designs a Dynamic Social Weight Update Model (DSWUM) to update the weight of the edge in real time according to the knowledge absorption rate and social behavior of students. The edge weight w i,j,t between nodes in the social network is calculated by the following formula:

[0145] w i,j,t = α·sim(F social,i,t , F social,j,t ) + β·|KAR i,t - KAR j,t |

[0146] where w i,j,t is the edge weight between student i and student j at time point t; α and β are weight coefficients used to balance the influence of social behavior similarity and knowledge absorption rate difference; sim(F social,i,t , F social,j,t ) is the social feature similarity between student i and student j at time point t; |KAR i,t - KAR j,t | is the knowledge absorption rate difference between student i and student j at time point t.

[0147] For example, assume that the social feature vectors of student i and student j are F social,i,t = [0.6, 0.8] and F social,j,t = [0.5, 0.7], the knowledge absorption rates are KAR i,t = 0.7 and KAR j,t = 0.6 respectively, and the weight coefficients are α = 0.5 and β = 0.3. Then the edge weight w i,j,t = 0.5·sim([0.6, 0.8], [0.5, 0.7]) + 0.3·|0.7 - 0.6|. Assume that the similarity sim([0.6, 0.8], [0.5, 0.7]) = 0.95, then w i,j,t = 0.5·0.95 + 0.3·0.1 = 0.505.

[0148] Furthermore, based on the constructed social network, the present invention uses the Group Interaction Optimization Algorithm (GIOA) to optimize the interaction among students. The core of this algorithm is to promote effective interaction among students by adjusting the weights of the edges, increasing the interaction between students with higher knowledge absorption rates and those with lower absorption rates, so as to achieve the effect of complementary learning. The optimization process is achieved by minimizing the following objective function L opt implemented as:

[0149]

[0150] where L opt is the optimization objective function, used to minimize the difference in knowledge absorption rates among students in the social network; E is the set of edges in the social network; w i,j,t is the weight of edge i and j; KAR i,t and KAR j,t are the knowledge absorption rates of students i and j at time point t.

[0151] For example, if the edge set E contains two edges (i, j) and (k, l), with weights w i,j,t = 0.5 and w k,l,t = 0.8 respectively, and the knowledge absorption rates are KAR i,t = 0.7, KAR j,t = 0.6, KAR k,t = 0.8 and KAR l,t = 0.5, then the objective function is:

[0152]

[0153] Furthermore, based on the optimization results, the present invention generates new social interaction strategies to improve the learning effect by adjusting the interaction recommendations and group assignments among students. The system continuously monitors the knowledge absorption rate KAR t+1 of students and updates the social network model to ensure that the interaction among students remains at the optimal level.

[0154] The formula for updating the edge weight is:

[0155] w i,j,t+1 = w i,j,t + γ·Δw i,j,t

[0156] where w i,j,t+1 is the edge weight between student i and student j at time point t + 1; γ is the learning rate, used to control the step size of weight update; Δw i,j,t is the change in the edge weight, output by the optimization algorithm.

[0157] S5. Construct an adaptive learning curve model according to the second dynamic social weight update model, and design a feedback mechanism to update the interactive recommendations and group assignments among students.

[0158] Specifically, the input data includes the edge weight w output by the social network model i,j,t and the updated knowledge absorption rate sequence KAR i,1:t . First, the present invention extracts the learning behavior characteristics of each student to generate a feature vector F learn,i,t , which is used to describe the learning behavior and state of student i at time point t.

[0159] F learn,i,t = η(w i,j,t , KAR i,1:t )

[0160] where η is a feature extraction function for converting the edge weight of the social network and the knowledge absorption rate into learning behavior characteristics.

[0161] Furthermore, using the extracted learning behavior characteristics F learn,i,t , the present invention constructs an adaptive learning curve model (Adaptive Learning Curve Model, ALCM) to reflect the learning progress and knowledge mastery of students at different time points. This model uses a dynamic Bayesian network (Dynamic Bayesian Network, DBN) that can capture the changes in students' learning states and the potential dependencies in the time series. The core of the learning curve model is to predict future learning performance based on the knowledge absorption rate and social behavior characteristics of students. The state transition equation of the model is:

[0162]

[0163] where KAR i,t+1 is the knowledge absorption rate of student i at time point t + 1; γ and δ are model parameters used to balance the current absorption rate and the new learning contribution; λ is a learning rate parameter used to adjust the influence of learning behavior characteristics on the absorption rate; F learn,i,t is the learning behavior characteristic of student i at time point t.

[0164] For example, assume that the current knowledge absorption rate of student i is KAR i,t = 0.7, the learning behavior characteristic is F learn,i,t = 1.2, the model parameters are γ = 0.6, δ = 0.4, and λ = 0.5. Then the predicted knowledge absorption rate for the future is:

[0165] KAR i,t+1 = 0.6·0.7 + 0.4·(1 - e -0.5·1.2)≈0.42 + 0.4·(1 - 0.5488) = 0.42 + 0.1805 = 0.6005

[0166] Furthermore, according to the adaptive learning curve model, the present invention designs a real-time feedback mechanism to provide students with personalized learning suggestions and strategy adjustments. The feedback mechanism is based on the "Learning Efficiency Function" (LEF), which is used to evaluate the current learning status and potential improvement directions of each student. The learning efficiency function is defined as:

[0167]

[0168] where E i,t is the learning efficiency of student i at time point t; KAR i,t+1 and KAR i,t are the knowledge absorption rates of student i at time points t + 1 and t; Δt is the time interval; ω is the weight parameter used to control the impact of behavioral feature differences on learning efficiency; is the average learning behavioral feature.

[0169] For example, assume that the knowledge absorption rate of student i increases from KAR i,t = 0.6 to KAR i,t+1 = 0.7, the learning behavioral feature is F learn,i,t = 1.2, the average learning behavioral feature is the time interval Δt = 1, and the weight parameter ω = 0.3. Then the learning efficiency is:

[0170]

[0171] Furthermore, based on the calculation results of the learning efficiency function, the system generates personalized feedback suggestions for students to help them optimize their learning strategies. The system continuously monitors the students' learning progress and knowledge absorption rates, and dynamically adjusts the learning curve model parameters to ensure that the model can accurately reflect the students' learning status and improvement directions.

[0172] The dynamic update formula for the model parameters is:

[0173]

[0174] where θ t+1 and θ t are the model parameters at time points t + 1 and t; η is the learning rate; is the gradient of the loss function with respect to the model parameters, which is used to guide the parameter update.

[0175] S6. Construct a blockchain network to encrypt and upload the data output by the second dynamic social weight update model, construct a learning feedback smart contract to automatically execute data verification, storage, and feedback optimization, and design a learning data consensus algorithm to judge the reception of data blocks; wherein, the learning data consensus algorithm is that all nodes use a voting mechanism to decide whether to accept a new data block, and only when all nodes vote to pass can the new data block be accepted.

[0176] Specifically, the input is the output data of the adaptive learning curve model and the feedback mechanism in step 5, and these data include the knowledge absorption rate sequence KAR of students i,1:t and the learning efficiency E i,t . The present invention uploads these data to the blockchain network to ensure the security and transparency of the data, and at the same time uses smart contracts to optimize the feedback mechanism. The following are the specific scheme steps:

[0177] The input data includes the learning data feature vector F data,i,t =(KAR i,1:t , E i,t ). First, the present invention encrypts these data to generate a ciphertext F' data,i,t to ensure the security of the data before uploading it to the blockchain. The present invention uses the method based on the elliptic curve cryptography (ECC) to encrypt the data.

[0178] F' data,i,t =ECC_encrypt(F data,i,t , key pub )

[0179] wherein, F' data,i,t is the encrypted data of student i at time point t; F data,i,t is the unencrypted learning data feature vector; key pub is the public key of the blockchain network for data encryption.

[0180] For example, assume F data,i,t =(0.7, 0.5), key pub =12345, then the encrypted data F' data,i,t =ECC_encrypt((0.7, 0.5), 12345)=(0.9, 0.8).

[0181] Furthermore, to manage the secure sharing and feedback optimization of learning data on the blockchain, the present invention defines a Learning Feedback Smart Contract (LFSC). This smart contract automatically performs data verification, storage, and feedback optimization. The core of the smart contract is to verify the authenticity and integrity of the uploaded data and execute feedback optimization according to predefined rules. The core verification formula of the contract is:

[0182]

[0183] Where V(F' data,i,t , key priv ) is the verification function of the smart contract; key priv is the private key of the blockchain network, used for data decryption;? means the first item is equal to the second item, and then it is judged whether the third item is equal to the first two items.

[0184] The contract verifies the integrity of the data by decrypting the encrypted data and comparing it with the original data. For example, assume F' data,i,t =(0.9, 0.8), key priv =54321, the decryption result is ECC_decrypt((0.9, 0.8), 54321)=(0.7, 0.5), which matches the original data F data,i,t =(0.7, 0.5), then the verification passes.

[0185] Furthermore, once the data passes the verification, the smart contract stores the data on the blockchain and performs dynamic feedback optimization according to the learning efficiency E i,t . The present invention designs a Feedback Optimization Function (FOF) to adjust the learning curve model parameters and optimize the learning strategy. The feedback optimization function is defined as:

[0186]

[0187] Where θ t+1 is the model parameter at time point t + 1; θ t is the model parameter at time point t; μ is the optimization step size; is the gradient of the learning efficiency with respect to the model parameters, used to guide parameter update.

[0188] For example, assume E i,t =0.04, the model parameter θ t =0.5, the optimization step size μ = 0.1, calculate the gradient Then the updated model parameter is:

[0189] θt+1 = 0.5 + 0.1·0.3 = 0.53

[0190] Furthermore, the blockchain network adopts the Learning Data Consensus Algorithm (LDCA) to ensure the secure sharing and transparency of data. All nodes use a voting mechanism to decide whether to accept a new data block. The learning data contained in each data block is verified by multiple signatures to ensure the immutability of the data. The data update process uses the following formula:

[0191] D t+1 = D t + σ·ΔD t

[0192] where D t+1 is the learning data state at time point t + 1; D t is the learning data state at time point t; σ is the consensus parameter used to control the data update step size; ΔD t is the change in the data state, output by the consensus algorithm.

[0193] Through this solution, the secure sharing and feedback optimization of learning data based on blockchain can dynamically adjust learning strategies according to students' learning status and feedback, ensuring the security, transparency, and immutability of learning data.

[0194] In another embodiment of the present invention, an AI-based teaching interaction system is provided. The system includes:

[0195] A teaching modality data collection module 101, which is used to deploy sensor devices to collect students' multimodal data, preprocess the multimodal data, extract and fuse features from the preprocessed data to obtain a comprehensive feature vector, and store the comprehensive feature vector as a multimodal data set; wherein, the comprehensive feature vector includes EEG spectrum features, heart rate variability features, eye movement features, and facial expression features;

[0196] A knowledge absorption analysis module 102, which is used to receive the multimodal data set sent by the sensor and construct an individual absorption model based on the multimodal data set to calculate the knowledge absorption rate of the student at each time point and predict the trend of the student's knowledge absorption rate sequence;

[0197] A teaching content output module 103, which is used to construct a feature mapping to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, construct a dynamic first adaptive content generator according to the feature mapping, design a joint loss function, and jointly train and optimize the discriminator and the first adaptive content generator according to the feature mapping to obtain a second adaptive content generator;

[0198] The teaching interaction module 104 is used to obtain the teaching content feature vector generated by the second adaptive content generator, map it to the social behavior features of students to obtain the social feature vector, construct the student social network using the social feature vector, build the first dynamic social weight update model according to the student social network, and design a group interaction optimization algorithm to adjust the weights of the first dynamic social weight update model to obtain the second dynamic social weight update model. Generate new social interaction strategies according to the second dynamic social weight update model, and dynamically adjust the interaction recommendations and group assignments among students;

[0199] The interaction feedback module 105 is used to construct an adaptive learning curve model according to the second dynamic social weight update model, and design a feedback mechanism to update the interaction recommendations and group assignments among students;

[0200] The data encryption module 106 is used to construct a blockchain network to encrypt and upload the data output by the second dynamic social weight update model, and construct a learning feedback smart contract to automatically execute data verification, storage, and feedback optimization, and design a learning data consensus algorithm to judge the reception of data blocks; among them, the learning data consensus algorithm is that all nodes use a voting mechanism to decide whether to accept a new data block, and only when all nodes vote through can the new data block be accepted;

[0201] Among them, in the student social network, each node represents a student, and the weight of the edge represents the strength of the social connection between students; the first dynamic social weight update model updates the weight of the edge in real time according to the knowledge absorption rate and social behavior of students; the calculation of the weight of the edge is as follows:

[0202] w i,j,t =α·sim(F social,i,t ,F social,j,t )+β·|KAR i,t -KAR j,t |

[0203] Among them, w i,j,t represents the edge weight between student i and student j at time point t; α and β represent weight coefficients, which are used to balance the influence of social behavior similarity and knowledge absorption rate difference; sim(F social,i,t ,F social,j,t ) represents the social feature similarity between student i and student j at time point t; |KAR i,t -KAR j,t | represents the difference in knowledge absorption rate between student i and student j at time point t;

[0204] Design a group interaction optimization algorithm to increase the interaction between the student with the highest knowledge absorption rate and the student with the lowest absorption rate. The group interaction optimization algorithm is expressed as follows:

[0205]

[0206] Among them, L opt represents the optimization objective function, which is used to minimize the difference in knowledge absorption rates among students in the social network; E represents the set of edges in the social network; w i,j,t represents the weight of edges i and j; KAR i,t and KAR j,t represent the knowledge absorption rates of students i and j at time point t;

[0207] Update the edge weights according to the output of the group interaction optimization algorithm to obtain the second dynamic social weight update model. The edge weights of the second dynamic social weight update model are expressed as follows:

[0208] w i,j,t+1 = w i,j,t + γ·Δw i,j,t

[0209] where w i,j,t+1 represents the edge weight between student i and student j at time point t + 1; γ represents the learning rate, which is used to control the step size of weight update; Δw i,j,t represents the change in edge weight.

[0210] It should be noted that the above-described work process is only illustrative and does not limit the protection scope of the present invention. In actual applications, those skilled in the art can select some or all of them according to actual needs to achieve the purpose of the solution of this embodiment, and no limitation is imposed here.

[0211] In addition, for the technical details not described in detail in this embodiment, reference can be made to the parameter operation method provided in any embodiment of the present invention, which will not be elaborated here.

[0212] It should be noted that in this article, the term "including", "comprising" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or system including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or system. Without further limitation, an element defined by the statement "including one..." does not exclude the existence of another identical element in the process, method, article or system including that element.

[0213] The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments.

[0214] Through the description of the above embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus a necessary general hardware platform. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as a read-only memory / random access memory, magnetic disk, optical disk), and includes several instructions for causing a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in various embodiments of the present invention.

[0215] The above are only the preferred embodiments of the present invention, and do not limit the patent scope of the present invention accordingly. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. The AI-based teaching interactive method is characterized by: The method comprises: S1. Deploy sensor equipment to collect multimodal data of students and preprocess the multimodal data, extract and fuse features of the preprocessed data to obtain a comprehensive feature vector, and store the comprehensive feature vector as a multimodal data set; wherein the comprehensive feature vector includes EEG spectrum features, heart rate variability features, eye movement features and facial expression features; S2, receiving a multimodal data set sent by a sensor and constructing an individual absorption model based on the multimodal data set to calculate the student's knowledge absorption rate at each time point and predict the student's knowledge absorption rate sequence trend; wherein the knowledge absorption rate is the student's future knowledge absorption trend predicted by using the hidden state and attention mechanism of GRU; S3, constructing a feature map to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, constructing a dynamic first adaptive content generator according to the feature map, and designing a joint loss function to jointly train and optimize the discriminator and the first adaptive content generator according to the feature map to obtain a second adaptive content generator; S4. Obtain the teaching content feature vector generated by the second adaptive content generator, and convert it into a social feature vector through a feature mapping function in combination with the knowledge absorption rate, use the social feature vector to build a student social network, build a first dynamic social weight update model based on the student social network, and design a group interaction optimization algorithm to adjust the weight of the first dynamic social weight update model to obtain a second dynamic social weight update model, generate a new social interaction strategy based on the second dynamic social weight update model, and dynamically adjust the interaction recommendation and group allocation between students; S5. Construct an adaptive learning curve model based on the second dynamic social weight update model, and design a feedback mechanism to update the interactive recommendation and group allocation between students; the adaptive learning curve model uses a dynamic Bayesian network to predict future learning performance based on students' knowledge absorption rate and social behavior characteristics; S6. Construct a blockchain network to encrypt and upload the data output by the second dynamic social weight update model, and construct a learning feedback smart contract to automatically perform data verification, storage and feedback optimization, and design a learning data consensus algorithm to judge the reception of data blocks; wherein, the learning data consensus algorithm is that all nodes decide whether to accept new data blocks through a voting mechanism, and only if all nodes vote in favor of accepting new data blocks will they accept new data blocks; In the student social network, each node represents a student, and the weight of the edge represents the strength of the social connection between students; the first dynamic social weight update model updates the weight of the edge in real time according to the students' knowledge absorption rate and social behavior; the weight of the edge is calculated as follows: In i,j,t =α sim(F social,i,t ,F social,j,t )+β |PENALTY i,t -PENALTY j,t | Among them, w i,j,t represents the edge weight between student i and student j at time point t; α and β represent weight coefficients, which are used to balance the influence of social behavior similarity and knowledge absorption rate difference; sim(F social,i,t ,F social,j,t ) represents the similarity of social characteristics between student i and student j at time point t; |KAR i,t -KAR j,t | represents the difference in knowledge absorption rate between student i and student j at time point t; The group interaction optimization algorithm is used to increase the interaction between students with the highest knowledge absorption rate and students with the lowest knowledge absorption rate. The group interaction optimization algorithm is expressed as follows: Among them, L opt represents the optimization objective function, which is used to minimize the difference in knowledge absorption rate among students in the social network; E represents the edge set in the social network; KAR i,t and KAR j,t represents the knowledge absorption rate of students i and j at time point t; The edge weights are updated according to the output of the group interaction optimization algorithm to obtain a second dynamic social weight update model. The edge weights of the second dynamic social weight update model are expressed as follows: In i,j,t+1 =in i,j,t +γ Δw i,j,t Among them, w i,j,t+1 represents the edge weight between student i and student j at time point t+1; γ represents the learning rate, which is used to control the step size of weight update; Δw i,j,t Indicates the change in edge weight.

2. The AI-based teaching interactive method according to claim 1, characterized in that: The sensor device includes an EEG headset, a heart rate monitor, an eye tracker and a facial expression recognition camera; the preprocessing includes time synchronization processing, applying an adaptive bandpass filter to the data of the EEG headset to remove noise in the EEG data, using a window smoothing algorithm to process the data of the heart rate monitor, using a peak detection algorithm to identify gaze and saccade processing on the eye movement data to obtain eye movement features, and using a convolutional neural network to extract features from micro-expression data to obtain facial expression features; the feature extraction includes using short-time Fourier transform on the data of the EEG headset to obtain power spectrum features of the α band and the β band in different time windows to obtain EEG spectrum features, and using the standard deviation of the RR interval to calculate the heart rate variability features; A joint feature space mapping method is used to map different modality data into a shared feature space.

3. The AI-based teaching interactive method according to claim 1 is characterized in that: The individual absorption model is constructed as follows: The student status at each moment on the comprehensive feature vector is standardized, and a multimodal time series embedding model that introduces a modal attention weight mechanism is constructed. The output of the modal attention mechanism is combined with the update formula of the multimodal time series embedding model to obtain an updated representation of the student's learning status; wherein the modal attention weight mechanism assigns a weight to each modal data to indicate its relative contribution to the knowledge absorption rate at the current time point, and the multimodal time series embedding model processes the multimodal data and captures temporal dynamic changes; The knowledge absorption rate at each time point is calculated and the future knowledge absorption trend is predicted based on the hidden states and modal attention weights output by the multimodal time series embedding model.

4. The AI-based teaching interactive method according to claim 3 is characterized in that: In the calculation of the knowledge absorption rate at each time point, a knowledge absorption rate regularization term based on L2 regularization is introduced to avoid overfitting and encourage weight sparsity; the prediction of the future knowledge absorption trend is obtained by inputting the hidden state into the fully connected layer output of the adaptive learning rate, and the prediction of the future knowledge absorption trend is the prediction of the knowledge absorption rate at several future time points.

5. The AI-based teaching interactive method according to claim 1, characterized in that: The first adaptive content generator generates a first teaching content feature vector according to the state vector and the parameters of the generator; the joint loss function L(G, D) is expressed as follows: Among them, G(S t ) represents the content features generated by the generator; D(G(S t )) represents the adaptability score of the discriminator to the generated content; S t represents the state vector; D represents the learning state discriminator; G represents the knowledge enhancement generator; represents the average content feature vector in the training set, which is used to calculate content diversity; λ represents the regularization parameter, which controls the influence of the diversity term; The knowledge enhancement generator generates content features and inputs them into a learning state discriminator for learning state discrimination. The learning state discriminator is used to discriminate whether the generated content is suitable for the current learning state.

6. The AI-based teaching interactive method according to claim 5 is characterized in that: The first adaptive content generator is trained using the joint loss function L(G,D) to obtain the second adaptive content generator, which is used to generate teaching content features adapted to the student's learning status in real time, and converts these features into a second teaching content feature vector through a predefined content library and mapping rules.

7. The AI-based teaching interactive method according to claim 1, characterized in that: The S5 specifically includes: S501, extracting the learning behavior characteristics of each student, generating a learning behavior characteristic vector for describing the learning behavior and state of student i at time point t; S502. Construct an adaptive learning curve model based on a dynamic Bayesian network according to the learning behavior feature vector, capture the changes in students' learning status and the potential dependencies of time series according to students' knowledge absorption rate and social behavior characteristics, and reflect students' learning progress and knowledge mastery at different time points; S503. Designing a real-time feedback mechanism based on the adaptive learning curve model to provide students with personalized learning suggestions and strategy adjustments; wherein the real-time feedback mechanism is set according to a learning efficiency function, and the learning efficiency function is used to evaluate each student's current learning status and potential improvement direction; S504. Generate personalized feedback suggestions for students based on the calculation results of the learning efficiency function, and optimize learning strategies.

8. The AI-based teaching interactive method according to claim 7, characterized in that: The learning efficiency function is expressed as follows: Among them, E i,t represents the learning efficiency of student i at time point t; KAR i,t+1 and KAR i,t represents the knowledge absorption rate of student i at time points t+1 and t; Δt represents the time interval; ω represents the weight parameter, which is used to control the impact of behavioral characteristic differences on learning efficiency; represents the average learning behavior characteristics; The knowledge absorption rate represents the state transition equation of the adaptive learning curve model and is calculated as follows: Among them, KAR i,t+1 represents the knowledge absorption rate of student i at time point t+1; γ and δ represent model parameters, which are used to balance the current absorption rate and new learning contribution; λ represents the learning rate parameter, which is used to adjust the impact of learning behavior characteristics on the absorption rate; F learn,i,t Represents the learning behavior characteristics of student i at time point t.

9. The AI-based teaching interactive method according to claim 1, characterized in that: The encryption adopts elliptic curve cryptography algorithm to encrypt data; the verification formula of the learning feedback smart contract is expressed as follows: Among them, F' data,i,t represents the encrypted data of student i at time point t; F data,i,t represents the unencrypted learning data feature vector; V(F' data,i,t ,key priv ) represents the verification function of the smart contract; key priv Indicates the private key of the blockchain network, which is used for data decryption; ECC_decrypt indicates an elliptic curve cryptographic algorithm; ? indicates that the first item is equal to the second item, and then determines whether the third item is equal to the first two items.

10. The AI-based teaching interactive system is characterized by: The system comprises: A teaching modality data collection module is used to deploy sensor equipment to collect students' multimodal data and preprocess the multimodal data, and perform feature extraction and fusion on the preprocessed data to obtain a comprehensive feature vector, and store the comprehensive feature vector as a multimodal data set; wherein the comprehensive feature vector includes EEG spectrum features, heart rate variability features, eye movement features and facial expression features; A knowledge absorption analysis module is used to receive a multimodal data set sent by a sensor and construct an individual absorption model based on the multimodal data set to calculate the student's knowledge absorption rate at each time point and predict the student's knowledge absorption rate sequence trend; wherein the knowledge absorption rate is the student's future knowledge absorption trend predicted by using the hidden state and attention mechanism of GRU; The teaching content output module is used to construct a feature map to embed the knowledge absorption rate and the predicted knowledge absorption rate sequence, construct a dynamic first adaptive content generator according to the feature map, and design a joint loss function to jointly train and optimize the discriminator and the first adaptive content generator according to the feature map to obtain a second adaptive content generator; A teaching interaction module is used to obtain the teaching content feature vector generated by the second adaptive content generator, and convert it into a social feature vector through a feature mapping function in combination with the knowledge absorption rate, use the social feature vector to build a student social network, build a first dynamic social weight update model based on the student social network, and design a group interaction optimization algorithm to adjust the weight of the first dynamic social weight update model to obtain a second dynamic social weight update model, generate a new social interaction strategy based on the second dynamic social weight update model, and dynamically adjust the interaction recommendation and group allocation between students; An interactive feedback module is used to construct an adaptive learning curve model based on the second dynamic social weight update model, and to design a feedback mechanism to update interactive recommendations and group assignments between students; the adaptive learning curve model uses a dynamic Bayesian network to predict future learning performance based on students' knowledge absorption rate and social behavior characteristics; A data encryption module is used to build a blockchain network to encrypt and upload the data output by the second dynamic social weight update model, and to build a learning feedback smart contract to automatically perform data verification, storage and feedback optimization, and to design a learning data consensus algorithm to judge the reception of data blocks; wherein, the learning data consensus algorithm is that all nodes decide whether to accept new data blocks through a voting mechanism, and only when all nodes vote in favor will they accept new data blocks; In the student social network, each node represents a student, and the weight of the edge represents the strength of the social connection between students; the first dynamic social weight update model updates the weight of the edge in real time according to the students' knowledge absorption rate and social behavior; the weight of the edge is calculated as follows: In i,j,t =α sim(F social,i,t ,F social,j,t )+β |PENALTY i,t -PENALTY j,t | Among them, w i,j,t represents the edge weight between student i and student j at time point t; α and β represent weight coefficients, which are used to balance the influence of social behavior similarity and knowledge absorption rate difference; sim(F social,i,t ,F social,j,t ) represents the similarity of social characteristics between student i and student j at time point t; |KAR i,t -KAR j,t | represents the difference in knowledge absorption rate between student i and student j at time point t; The group interaction optimization algorithm is used to increase the interaction between students with the highest knowledge absorption rate and students with the lowest knowledge absorption rate. The group interaction optimization algorithm is expressed as follows: Among them, L opt represents the optimization objective function, which is used to minimize the difference in knowledge absorption rate among students in the social network; E represents the edge set in the social network; KAR i,t and KAR j,t represents the knowledge absorption rate of students i and j at time point t; The edge weights are updated according to the output of the group interaction optimization algorithm to obtain a second dynamic social weight update model. The edge weights of the second dynamic social weight update model are expressed as follows: In i,j,t+1 =in i,j,t +γ Δw i,j,t Among them, w i,j,t+1 represents the edge weight between student i and student j at time point t+1; γ represents the learning rate, which is used to control the step size of weight update; Δw i,j,t Indicates the change in edge weight.

Images