Dynamic teaching strategy optimization method for personalized learning

Abstract

This invention discloses a dynamic teaching strategy optimization method for personalized learning, relating to the field of intelligent teaching optimization. The method includes acquiring emotional, behavioral, and academic data; using a cognitive diagnostic unit to perform correlation and feature fusion processing on the behavioral and academic data to obtain a knowledge point mastery probability matrix and a knowledge gap association strength matrix; using an affective computing unit to extract and fuse features from the emotional data to obtain affective fusion features; classifying and scoring the affective fusion features to obtain affective intensity and affective stability scores; constructing a fusion reinforcement learning model; and generating and adjusting dynamic teaching strategies based on the knowledge point mastery probability matrix, the knowledge gap association strength matrix, the affective intensity score, and the affective stability score. This invention improves the timeliness and accuracy of cognitive diagnostic parameters using the above method, supporting precise identification of knowledge weaknesses.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching optimization, and in particular to a method for optimizing dynamic teaching strategies for personalized learning. Background Technology

[0002] With the deep integration of artificial intelligence and education, personalized learning has become a core direction for breaking through the bottlenecks of traditional large-scale teaching. However, current mainstream personalized learning systems still face many technical bottlenecks and application pain points, making it difficult to accurately adapt teaching strategies to the dynamic state of learners. The core problems are concentrated in data collection, analysis and adaptation, and strategy generation. First, the data collection dimensions are singular and one-sided. Most systems rely too much on structured data such as academic performance and homework accuracy, while generally ignoring unstructured biometric data such as facial expressions, voice tone, and heart rate variability, as well as process learning behavior data such as answer paths, resource interaction frequency, and discussion forum interaction details. This results in the construction of learner profiles being superficial and unable to fully depict their cognitive level, learning habits, and emotional state. Secondly, the teaching adaptation lacks real-time and flexibility. Existing analysis engines mostly adopt offline batch processing mode, which has high data transmission and analysis latency. It is difficult to quickly respond to sudden learning obstacles that learners encounter during the learning process (such as misunderstanding of knowledge points, emotional anxiety fluctuations, etc.). Often, adjustments are only made passively after learners show obvious learning lag, missing the best time for intervention. Third, the problem of homogenization in strategy generation is prominent. Most systems rely on preset fixed rules or simple knowledge point association logic to generate teaching strategies, without fully considering the cognitive differences, emotional tolerance and learning pace preferences of different learners. This results in insufficient strategy targeting and difficulty in adapting to the dynamic learning needs of learners.

[0003] These problems directly diminish the effectiveness of personalized learning systems. They not only fail to effectively improve learning efficiency and quality, but may also trigger learner resistance and anxiety due to inappropriate strategy adaptation, thus hindering the successful implementation of personalized education concepts. Against this backdrop, the education industry urgently needs a dynamic, end-to-end optimization solution that can overcome the shortcomings of existing technologies.

[0004] To address these issues, there is an urgent need for dynamic teaching strategy optimization methods tailored to personalized learning. Summary of the Invention

[0005] To address the aforementioned issues, this application proposes a dynamic teaching strategy optimization method for personalized learning. Based on multimodal data fusion and real-time intelligent analysis, it constructs a full-process dynamic teaching optimization system, upgrading the teaching model from passive adaptation to proactive prediction. This method accurately matches the dynamic needs of different learners, significantly improving the accuracy, real-time performance, and effectiveness of personalized learning, and providing reliable technical support for the large-scale implementation of personalized education.

[0006] Specifically, it includes the following: S1. Obtain multi-dimensional information collected in the classroom, and analyze and extract emotional data, behavioral data, and academic data from the multi-dimensional information collected in the classroom. S2. Construct a cognitive diagnostic unit based on the improved DINA algorithm. Use the cognitive diagnostic unit to perform correlation and feature fusion processing on behavioral data and academic data to obtain the knowledge point mastery probability matrix and the knowledge gap correlation strength matrix. S3. Construct an emotion computing unit based on an improved CNN algorithm. Use the emotion computing unit to extract and fuse emotion data to obtain emotion fusion features. Classify and score the emotion fusion features to obtain emotion intensity score and emotion stability score. S4. Construct a fusion reinforcement learning model to generate and adjust dynamic teaching strategies based on the knowledge point mastery probability matrix, the knowledge gap association strength matrix, the emotion intensity score, and the emotion stability score.

[0007] Preferably, the specific content of the knowledge point mastery probability matrix and knowledge gap association strength matrix obtained by using the cognitive diagnostic unit in S2 to perform correlation and feature fusion processing on behavioral data and academic data includes: Input academic data and behavioral data, and preprocess the academic data and behavioral data into standard cognitive initial data using min-max standardization, 3σ criterion and data smoothing; The standard initial cognitive data includes standard initial academic data X. a Standard answer path data X path Standard answer time series data X time ; An attention mechanism is introduced to dynamically assign weights to academic data, answer path data, and answer time series data. Features are then fused using a fully connected layer and ReLU activation. The expression is: ; ; in, , This is the attention weight matrix. For bias terms, The dimension of the input vector; Based on maximum likelihood estimation, a Bayesian prior distribution is introduced to dynamically update the parameters of the DINA inference layer; The parameters of the DINA inference layer include the knowledge point mastery state vector α, the slip probability s, and the guessing probability g; The DINA inference layer is used to perform inference on feature fusion to obtain dimension X. Knowledge points related to P, probability matrix M, dimension P. P's knowledge vulnerability association strength matrix C.

[0008] Preferably, the expression for introducing the Bayesian prior distribution based on maximum likelihood estimation is: ; Where P(α), P(s), and P(g) are the prior distributions of the parameters, using a Beta distribution. For input data, This is a state vector representing the mastery of knowledge points, where s is the probability of slipping and g is the probability of guessing. Parameter dynamic update expression: ; in, For learning rate, The gradient of the likelihood function. Add data at time t+1. Let t be the estimated value of the knowledge point mastery state vector. This is the updated estimate.

[0009] Preferably, the expression for the knowledge point mastery probability in the knowledge point mastery probability matrix M is: ; The expression for the knowledge vulnerability association strength in the knowledge vulnerability association strength matrix C is: ; in, Let be the probability that learner n has mastered knowledge point t. This contains all data from time 1 to time t. This represents the mastery status of the nth learner regarding the ith knowledge point, where 1 indicates mastery and 0 indicates non-mastery. Matrix elements This represents the strength of the impact of a gap in knowledge point i on the learning of knowledge point j. For covariance, Standard deviation, M is the probability vector for mastering the i-th knowledge point (an N-dimensional column vector representing the mastery probabilities of all N learners for that knowledge point); j Let j be the probability vector of mastering the j-th knowledge point; For vector Mi With M j covariance; , They are vectors and The standard deviation.

[0010] Preferably, the emotion data includes facial expression images, voice data, and physiological data; The emotion computing unit includes a multimodal input enhancement and preprocessing layer, a fusion feature extraction layer, a cross-modal attention fusion layer, a dynamic threshold optimization classification layer, and a dual output layer; The multimodal input enhancement and preprocessing layer is used to input facial expression images, speech data and physiological data, and removes background interference from the facial expression images and normalizes them to obtain the first data; The speech data is denoised and endpoint detected, then converted into a Mel spectrogram and augmented to obtain the second data. The physiological data is filtered and smoothed to obtain the third data; The fusion feature extraction layer includes image feature extraction, speech feature extraction, and physiological signal feature extraction; The image feature extraction process captures subtle facial features through 4 convolutional layers, 2 pooling layers, and 1 Transformer encoder, outputting an image feature vector F. img ; The speech feature extraction process involves extracting local features from the Mel spectrogram using a two-layer CNN, followed by capturing temporal variations in speech intonation using a Transformer encoder, and outputting a speech feature vector F. voice ; The physiological signal feature extraction employs a single fully connected layer to perform feature mapping on the preprocessed physiological signal, outputting a physiological feature vector F. phy ; The cross-modal attention fusion layer assigns weights to image feature vectors, speech feature vectors, and physiological feature vectors based on the Softmax function and fuses the features to obtain the fused feature F. fusion .

[0011] Preferably, the dynamic threshold optimization classification layer includes a classification layer and a dynamic threshold adjustment layer; The classification layer will fuse feature F fusion The input consists of two fully connected layers, and the output layer uses Softmax activation to output five emotions: focus, confusion, anxiety, fatigue, and irritability. The expression is as follows: ; in, This represents the original score for the k-th emotion category; The dynamic threshold adjustment layer dynamically adjusts the emotion classification threshold based on the learner's historical emotion data and the learning scenario. The expression is: ; in, Let be the classification threshold for the k-th emotion at time t. β is the initial threshold, and β is the adjustment coefficient. The historical average emotional intensity , which serves as the initial intensity benchmark for the k-th emotion category.

[0012] Preferably, the dual output layer is used to output emotion intensity score and emotion stability score; The expression for the emotional intensity rating is: ; in, Let k be the probability of the k-th emotion. The strength baseline score ranges from 1 to 5. The expression for the emotional stability score is: ; in, Let t be the emotional intensity score sequence from time 1 to time t, with values ​​ranging from 0 to 1.

[0013] Preferably, the construction process of the fusion reinforcement learning model includes: Establish the state space of the fusion reinforcement learning model. The expression for the state space is: ; The knowledge vulnerability association strength matrix, flattened into a vector, has dimension p. 2 , Assess emotional stability. Let p be the probability matrix of knowledge mastery at time t, flattened into a vector with dimension p. The emotional intensity score at time t. The learning progress at time t, with a value between 0 and 1; The total dimension of the state space is p+p 2 +2+1=p 2 +p+3; Define the action space: ; Among them, the selection of knowledge points is based on Select the knowledge points to learn; Dynamic matching of learning difficulty It is defined as: 1 = Basic consolidation, 2 = Ability enhancement, 3 = Expanding challenges; Learning duration matching The values ​​are defined as follows: 1 = 20 minutes, 2 = 30 minutes, 3 = 40 minutes; Teaching method matching It is defined as: 1 = video explanation, 2 = interactive practice, 3 = group discussion, 4 = experimental demonstration.

[0014] A dynamic weight adjustment mechanism is introduced to dynamically adjust the reward weights based on the learner's real-time status. Simultaneously, a constraint on the correlation strength of knowledge gaps is incorporated to construct a dynamic weight reward function. The expression for calculating the dynamic weights is as follows: ; ; ; in, The faster the improvement in knowledge points, the better. The larger the size, the higher the stability. The larger, ; The expression for the dynamic weighted reward function is: ; in, This is the associated reward coefficient, with a value of 0.1. The strength of the association between knowledge points i and j To improve the mastery of knowledge point i after learning j, students are encouraged to learn highly related knowledge points. Let be the reward value at time t. Rate the intensity of the target emotion. Let t be the learning time.

[0015] Preferably, the fusion reinforcement learning model uses DQN for basic path decision-making, PPO algorithm to optimize strategy updates, and introduces an experience replay pool for optimization. The expression for calculating the Q value is as follows: ; The expression for the PPO objective function is: ; The weight expression in the experience replay pool is: ; in, State at time t Next action The expected cumulative reward, As a discount factor, The strategy ratio, For the dominant function, Let E be the clipping factor and E be the expectation operator. For sample priority, Here, N is the importance sampling coefficient, and N is the total number of samples. For maximum weight, The instantaneous reward at time t. The objective function of PPO is used to optimize the policy parameters. Weights for the empirical replay samples.

[0016] Preferably, the teaching strategy is adjusted immediately when any of the following conditions are met: a) Severe mood fluctuations, i.e., emotional stability score <0.3; b. No improvement in learning, i.e., 3 consecutive moments. <0.01; c. Ineffective remedial learning of highly relevant knowledge points, i.e., knowledge gap correlation strength matrix There exists >0.8 and <0.02, ensuring the strategy adapts to real-time state changes.

[0017] In summary, the dynamic teaching strategy optimization method for personalized learning of this invention, compared with traditional technologies, integrates three types of multimodal data—biological characteristics, learning behavior, and academic performance—to construct a holographic learner profile, breaking through the limitations of single data. Relying on low-latency transmission through edge computing and combined with an improved algorithm, it quickly analyzes the learning status, enabling real-time response and adjustment of strategies. It accurately adapts to multiple dimensions, including hierarchical tasks, interdisciplinary connections, and tutoring techniques, taking into account knowledge acquisition, ability enhancement, and emotion regulation. Furthermore, it optimizes the algorithm model to improve the accuracy of diagnosis, emotion recognition, and path optimization.

[0018] The technical method of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the steps of the dynamic teaching strategy optimization method for personalized learning according to the present invention. Detailed Implementation

[0020] The technical method of the present invention will be further described below with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of this application.

[0021] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0022] Techniques, systems, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the instruction manual.

[0023] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] Example 1 This invention provides a method for optimizing dynamic teaching strategies for personalized learning, such as... Figure 1 As shown, the specific content is as follows: S1. Obtain multi-dimensional information collected in the classroom, and analyze and extract emotional data, behavioral data, and academic data from the multi-dimensional information collected in the classroom.

[0026] Specifically, a fusion approach combining sensor acquisition, platform recording, and API calls is used to collect three types of core data: Biometric data: Facial expressions (micro-expressions such as focus and confusion) are collected through classroom cameras, and physiological data such as heart rate, blood oxygen, and attention curves are collected by wearable devices.

[0027] Learning behavior data: Based on the interactive learning platform, unstructured data such as answer paths (problem-solving steps, error nodes), resource click frequency (number of video pauses, repeated access records of courseware), and interactive text in the discussion area are recorded.

[0028] Academic data: By calling the API interface of the academic affairs and learning platform, structured data such as homework accuracy rate, knowledge point mastery score, and historical assessment results can be obtained.

[0029] Edge computing technology is used to perform local preprocessing of multimodal data, including denoising, feature extraction, and min-max normalization (mapping the data to the [0,1] interval); data anonymization is used to ensure privacy and security, while low-latency data upload is achieved to ensure the real-time nature of subsequent analysis.

[0030] Based on preprocessed data collected by S1, a dynamic adaptive analysis engine is built through three core algorithm models to accurately identify learners' knowledge gaps, emotional states, and learning needs. Specifically, this is achieved through: S2. Construct a cognitive diagnostic unit based on the improved DINA algorithm. Use the cognitive diagnostic unit to perform correlation and feature fusion processing on behavioral data and academic data to obtain the knowledge point mastery probability matrix and the knowledge gap correlation strength matrix. Furthermore, in S2, the cognitive diagnostic unit performs correlation and feature fusion processing on behavioral and academic data to obtain the knowledge point mastery probability matrix M and the knowledge gap association strength matrix C. The specific contents of these matrices include: Input academic data and behavioral data, and preprocess the academic data and behavioral data into standard cognitive initial data using min-max standardization, 3σ criterion and data smoothing; The standard initial cognitive data includes standard initial academic data X. a Standard answer path data X path Standard answer time series data X time ; An attention mechanism is introduced to dynamically assign weights to academic data, answer path data, and answer time series data. Features are then fused using a fully connected layer and ReLU activation. The expression is: ; ; in, , This is the attention weight matrix. For bias terms, The dimension of the input vector; Based on maximum likelihood estimation, a Bayesian prior distribution is introduced to dynamically update the parameters of the DINA inference layer; The parameters of the DINA inference layer include the knowledge point mastery state vector. The probability of slipping is s, and the probability of guessing is g; The DINA inference layer is used to perform inference on feature fusion to obtain dimension X. Knowledge points related to P, probability matrix M, dimension P. P's knowledge vulnerability association strength matrix C.

[0031] Furthermore, the expression for the Bayesian prior distribution introduced based on maximum likelihood estimation is as follows: ; Where P(α), P(s), and P(g) are the prior distributions of the parameters, using a Beta distribution. For input data, This is a state vector representing the mastery of knowledge points, where s is the probability of slipping and g is the probability of guessing. Parameter dynamic update expression: ; in, For learning rate, The gradient of the likelihood function. Add data at time t+1. Let t be the estimated value of the knowledge point mastery state vector. This is the updated estimate.

[0032] Furthermore, the expression for the probability of mastering a knowledge point in the knowledge point mastery probability matrix M is as follows: ; The expression for the knowledge vulnerability association strength in the knowledge vulnerability association strength matrix C is: ; in, Let be the probability that learner n has mastered knowledge point t. This contains all data from time 1 to time t. This represents the mastery status of the nth learner regarding the ith knowledge point, where 1 indicates mastery and 0 indicates non-mastery. Matrix elements This represents the strength of the impact of a gap in knowledge point i on the learning of knowledge point j. For covariance, Standard deviation, M is the probability vector for mastering the i-th knowledge point (an N-dimensional column vector representing the mastery probabilities of all N learners for that knowledge point); j Let j be the probability vector of mastering the j-th knowledge point; For vector M i With M j covariance; , They are vectors and The standard deviation.

[0033] S3. Construct an emotion computing unit based on an improved CNN algorithm. Use the emotion computing unit to extract and fuse emotion data to obtain emotion fusion features. Classify and score the emotion fusion features to obtain emotion intensity score and emotion stability score. Furthermore, the emotion data includes facial expression images, voice data, and physiological data; The emotion computing unit includes a multimodal input enhancement and preprocessing layer, a fusion feature extraction layer, a cross-modal attention fusion layer, a dynamic threshold optimization classification layer, and a dual output layer; The multimodal input enhancement and preprocessing layer is used to input facial expression images, speech data and physiological data, and removes background interference from the facial expression images and normalizes them to obtain the first data; The speech data is denoised and endpoint detected, then converted into a Mel spectrogram and augmented to obtain the second data. The physiological data is filtered and smoothed to obtain the third data; The fusion feature extraction layer includes image feature extraction, speech feature extraction, and physiological signal feature extraction; The image feature extraction process captures subtle facial features through 4 convolutional layers, 2 pooling layers, and 1 Transformer encoder, outputting an image feature vector F. img ; The speech feature extraction process involves extracting local features from the Mel spectrogram using a two-layer CNN, followed by capturing temporal variations in speech intonation using a Transformer encoder, and outputting a speech feature vector F. voice ; The physiological signal feature extraction employs a single fully connected layer to perform feature mapping on the preprocessed physiological signal, outputting a physiological feature vector F. phy ; The cross-modal attention fusion layer assigns weights to image feature vectors, speech feature vectors, and physiological feature vectors based on the Softmax function and fuses the features to obtain the fused feature F. fusion .

[0034] Furthermore, the dynamic threshold optimization classification layer includes a classification layer and a dynamic threshold adjustment layer; The classification layer will fuse feature F fusion The input consists of two fully connected layers, and the output layer uses Softmax activation to output five emotions: focus, confusion, anxiety, fatigue, and irritability. The expression is as follows: ; in, This represents the original score for the k-th emotion category; The dynamic threshold adjustment layer dynamically adjusts the emotion classification threshold based on the learner's historical emotion data and the learning scenario. The expression is: ; in, Let be the classification threshold for the k-th emotion at time t. β is the initial threshold, and β is the adjustment coefficient. The historical average emotional intensity , which serves as the initial intensity benchmark for the k-th emotion category.

[0035] Furthermore, the dual output layer is used to output emotion intensity scores and emotion stability scores; The expression for the emotional intensity rating is: ; in, Let k be the probability of the k-th emotion. The strength baseline score ranges from 1 to 5. The expression for the emotional stability score is: ; in, Let t be the emotional intensity score sequence from time 1 to time t, with values ​​ranging from 0 to 1.

[0036] S4. Construct a fusion reinforcement learning model to generate and adjust dynamic teaching strategies based on the knowledge point mastery probability matrix, the knowledge gap association strength matrix, the emotion intensity score, and the emotion stability score.

[0037] Furthermore, the construction process of the fusion reinforcement learning model includes: Establish the state space of the fusion reinforcement learning model. The expression for the state space is: ; The knowledge vulnerability association strength matrix, flattened into a vector, has dimension p. 2 , Assess emotional stability. Let p be the probability matrix of knowledge mastery at time t, flattened into a vector with dimension p. The emotional intensity score at time t. The learning progress at time t, with a value between 0 and 1; The total dimension of the state space is p+p 2 +2+1=p 2 +p+3; Define the action space: ; Among them, the selection of knowledge points is based on Select the knowledge points to learn; Dynamic matching of learning difficulty It is defined as: 1 = Basic consolidation, 2 = Ability enhancement, 3 = Expanding challenges; Learning duration matching The values ​​are defined as follows: 1 = 20 minutes, 2 = 30 minutes, 3 = 40 minutes; Teaching method matching It is defined as: 1 = video explanation, 2 = interactive practice, 3 = group discussion, 4 = experimental demonstration.

[0038] A dynamic weight adjustment mechanism is introduced to dynamically adjust the reward weights based on the learner's real-time status. Simultaneously, a constraint on the correlation strength of knowledge gaps is incorporated to construct a dynamic weight reward function. The expression for calculating the dynamic weights is as follows: ; ; ; in, The faster the improvement in knowledge points, the better. The larger the size, the higher the stability. The larger, ; The expression for the dynamic weighted reward function is: ; in, This is the associated reward coefficient, with a value of 0.1. The strength of the association between knowledge points i and j To improve the mastery of knowledge point i after learning j, students are encouraged to learn highly related knowledge points. Let be the reward value at time t. Rate the intensity of the target emotion. Let t be the learning time.

[0039] Furthermore, the fusion reinforcement learning model uses DQN for basic path decision-making, PPO algorithm to optimize strategy updates, and introduces an experience replay pool for optimization. The expression for calculating the Q value is as follows: ; The expression for the PPO objective function is: ; The weight expression in the experience replay pool is: ; in, State at time t Next action The expected cumulative reward, As a discount factor, The strategy ratio, For the dominant function, Let E be the clipping factor and E be the expectation operator. For sample priority, Here, N is the importance sampling coefficient, and N is the total number of samples. For maximum weight, The instantaneous reward at time t. The objective function of PPO is used to optimize the policy parameters. Weights for the empirical replay samples.

[0040] Furthermore, the teaching strategy will be adjusted immediately when any of the following conditions are met: a) Severe mood fluctuations, i.e., emotional stability score <0.3; b. No improvement in learning, i.e., 3 consecutive moments. <0.01; c. Ineffective remedial learning of highly relevant knowledge points, i.e., knowledge gap correlation strength matrix There exists >0.8 and <0.02, ensuring the strategy adapts to real-time state changes.

[0041] Example 2 Tiered task generation: Based on the knowledge mastery probability matrix output by the cognitive diagnostic model, a priority scoring mechanism is constructed to generate differentiated tasks for learners with different levels of foundation: scaffold-style practice questions (from basic to comprehensive, with accompanying explanations and hints) are pushed to learners with weak foundations; project-based tasks (cross-chapter integrated applications, subject-specific practice projects) are recommended to advanced learners.

[0042] Interdisciplinary content association: By calling the domain knowledge graph API, interdisciplinary content is automatically associated based on the current learning knowledge points (such as associating the mathematical concept of "proportion" with the physical concept of "density calculation"), generating interdisciplinary resource packages and enhancing knowledge transfer capabilities.

[0043] Personalized tutoring script generation: Integrating a large language model (LLM), learners are inputted with their emotional state, knowledge gaps, and task completion status to generate appropriate scripts: encouraging scripts for anxious learners, guiding scripts for confused learners, and extended suggestions for high-achieving learners.

[0044] By leveraging a priority scoring mechanism, knowledge graph association algorithm, LLM Prompt optimization algorithm, and collaborative efforts with front-end data acquisition hardware and learning platform, the technology is successfully implemented, exhibiting clear technical characteristics and repeatability.

[0045] Learners engage in learning according to the generated personalized teaching strategies, and the system synchronously monitors the effectiveness of the strategies in real time. Specifically, this is achieved through: We collect two core performance data points through the learning platform and classroom monitoring equipment: Classroom performance matrix: frequency of raising hands, quality of interactive participation, and duration of attention span; Task completion effectiveness score: accuracy of answers, timeliness of task completion, and depth of exploration of extended content.

[0046] Set thresholds for abnormal behaviors (such as excessive consecutive error rates, persistent distraction, and task delays), and add deviation labels to behaviors that trigger these thresholds to provide a basis for subsequent corrections.

[0047] For detected abnormal behavior, the system automatically matches and corrects the policy library to achieve real-time policy adjustments. Specifically: Based on the type of deviation label, the corresponding solution is matched from the correction strategy library: for example, for knowledge gap anomalies, targeted supplementary learning resources are pushed; for emotional fluctuation anomalies, the learning pace is adjusted; and for inefficiency anomalies, the teaching method is switched (such as changing video lectures to interactive exercises).

[0048] The adjusted strategy is pushed to learners in real time to ensure that the learning process is not interrupted. At the same time, the state change data before and after the adjustment is recorded to provide samples for model iteration.

[0049] The monitoring and adjustment data are fed back to the core model to achieve a closed loop of parameter updates and strategy optimization. Specifically: A Bayesian network is used to construct a dynamically updated model. Input policy execution effect data and update learner profile parameters, cognitive diagnostic model weights, and sentiment computing model thresholds.

[0050] Regularly summarize all data and iteratively train the reinforcement learning path optimization model (Q network parameters) and LLM coaching script generation logic to improve the adaptation accuracy of subsequent strategies.

[0051] The iteratively updated model is then reapplied to the dynamic teaching strategy optimization method.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical methods of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical methods of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical methods to deviate from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for optimizing dynamic teaching strategies for personalized learning, characterized in that, Includes the following steps: S1. Obtain multi-dimensional information collected in the classroom, and analyze and extract emotional data, behavioral data, and academic data from the multi-dimensional information collected in the classroom. S2. Construct a cognitive diagnostic unit based on the improved DINA algorithm. Use the cognitive diagnostic unit to perform correlation and feature fusion processing on behavioral data and academic data to obtain the knowledge point mastery probability matrix and the knowledge gap correlation strength matrix. S3. Construct an emotion computing unit based on an improved CNN algorithm. Use the emotion computing unit to extract and fuse emotion data to obtain emotion fusion features. Classify and score the emotion fusion features to obtain emotion intensity score and emotion stability score. S4. Construct a fusion reinforcement learning model to generate and adjust dynamic teaching strategies based on the knowledge point mastery probability matrix, the knowledge gap association strength matrix, the emotion intensity score, and the emotion stability score.

2. The method for optimizing dynamic teaching strategies for personalized learning according to claim 1, characterized in that, S2 uses a cognitive diagnostic unit to perform correlation and feature fusion processing on behavioral and academic data to obtain the knowledge point mastery probability matrix and the knowledge gap correlation strength matrix. The specific contents of these matrices include: Input academic data and behavioral data, and preprocess the academic data and behavioral data into standard cognitive initial data using min-max standardization, 3σ criterion and data smoothing; The standard initial cognitive data includes standard initial academic data X. a Standard answer path data X path Standard answer time series data X time ; An attention mechanism is introduced to dynamically assign weights to academic data, answer path data, and answer time series data. Features are then fused using a fully connected layer and ReLU activation. The expression is: ； ； in, , This is the attention weight matrix. For bias terms, The dimension of the input vector; Based on maximum likelihood estimation, a Bayesian prior distribution is introduced to dynamically update the parameters of the DINA inference layer; The parameters of the DINA inference layer include the knowledge point mastery state vector. The probability of slipping is s, and the probability of guessing is g; The DINA inference layer is used to perform inference on feature fusion to obtain dimension X. The probability matrix for mastering knowledge points of P, with dimension P. P's knowledge vulnerability association strength matrix.

3. The method for optimizing dynamic teaching strategies for personalized learning according to claim 1, characterized in that, The expression for the Bayesian prior distribution introduced based on maximum likelihood estimation is: ； Where P(α), P(s), and P(g) are the prior distributions of the parameters, using a Beta distribution. For input data, This is a state vector representing the mastery of knowledge points, where s is the probability of slipping and g is the probability of guessing. Parameter dynamic update expression: ； in, For learning rate, The gradient of the likelihood function. Add data at time t+1. Let t be the estimated value of the knowledge point mastery state vector. This is the updated estimate.

4. The method for optimizing dynamic teaching strategies for personalized learning according to claim 3, characterized in that, The expression for the probability of mastering a knowledge point in the knowledge point mastery probability matrix M is: ； The expression for the knowledge vulnerability association strength in the knowledge vulnerability association strength matrix C is: ； in, Let be the probability that learner n has mastered knowledge point t. This contains all data from time 1 to time t. This represents the mastery status of the nth learner regarding the ith knowledge point, where 1 indicates mastery and 0 indicates non-mastery. Matrix elements This represents the strength of the impact of a gap in knowledge point i on the learning of knowledge point j. For covariance, Standard deviation, M represents the probability vector of mastering the i-th knowledge point; j Let j be the probability vector of mastering the j-th knowledge point; For vector M i With M j covariance; , They are vectors and The standard deviation.

5. The method for optimizing dynamic teaching strategies for personalized learning according to claim 4, characterized in that, The emotional data includes facial expression images, voice data, and physiological data; The emotion computing unit includes a multimodal input enhancement and preprocessing layer, a fusion feature extraction layer, a cross-modal attention fusion layer, a dynamic threshold optimization classification layer, and a dual output layer; The multimodal input enhancement and preprocessing layer is used to input facial expression images, speech data and physiological data, and removes background interference from the facial expression images and normalizes them to obtain the first data; The speech data is denoised and endpoint detected, then converted into a Mel spectrogram and augmented to obtain the second data. The physiological data is filtered and smoothed to obtain the third data; The fusion feature extraction layer includes image feature extraction, speech feature extraction, and physiological signal feature extraction; The image feature extraction process captures subtle facial features through 4 convolutional layers, 2 pooling layers, and 1 Transformer encoder, outputting an image feature vector F. img ; The speech feature extraction process involves extracting local features from the Mel spectrogram using a two-layer CNN, followed by capturing temporal variations in speech intonation using a Transformer encoder, and outputting a speech feature vector F. voice ; The physiological signal feature extraction employs a single fully connected layer to perform feature mapping on the preprocessed physiological signal, outputting a physiological feature vector F. phy ; The cross-modal attention fusion layer assigns weights to image feature vectors, speech feature vectors, and physiological feature vectors based on the Softmax function and fuses the features to obtain the fused feature F. fusion .

6. The method for optimizing dynamic teaching strategies for personalized learning according to claim 5, characterized in that, The dynamic threshold optimization classification layer includes a classification layer and a dynamic threshold adjustment layer; The classification layer will fuse feature F fusion The input consists of two fully connected layers, and the output layer uses Softmax activation to output five emotions: focus, confusion, anxiety, fatigue, and irritability. The expression is as follows: ； in, This represents the original score for the k-th emotion category; The dynamic threshold adjustment layer dynamically adjusts the emotion classification threshold based on the learner's historical emotion data and the learning scenario. The expression is: ； in, Let be the classification threshold for the k-th emotion at time t. β is the initial threshold, and β is the adjustment coefficient. The historical average emotional intensity , which serves as the initial intensity benchmark for the k-th emotion category.

7. The method for optimizing dynamic teaching strategies for personalized learning according to claim 6, characterized in that, The dual-output layer is used to output emotion intensity scores and emotion stability scores; The expression for the emotional intensity rating is: ； in, Let k be the probability of the k-th emotion. The strength baseline score ranges from 1 to 5. The expression for the emotional stability score is: ； in, Let t be the emotional intensity score sequence from time 1 to time t, with values ​​ranging from 0 to 1.

8. The method for optimizing dynamic teaching strategies for personalized learning according to claim 6, characterized in that, The construction process of the fusion reinforcement learning model includes: Establish the state space of the fusion reinforcement learning model. The expression for the state space is: ； The knowledge vulnerability association strength matrix, flattened into a vector, has dimension p. 2 , Assess emotional stability. Let p be the probability matrix of knowledge mastery at time t, flattened into a vector with dimension p. The emotional intensity score at time t. The learning progress at time t, with a value between 0 and 1; The total dimension of the state space is p+p 2 +2+1=p 2 +p+3; Define the action space: ; Among them, the selection of knowledge points is based on Select the knowledge points to learn; Dynamic matching of learning difficulty It is defined as: 1 = Basic consolidation, 2 = Ability enhancement, 3 = Expanding challenges; Learning duration matching The values ​​are defined as follows: 1 = 20 minutes, 2 = 30 minutes, 3 = 40 minutes; Teaching method matching It is defined as: 1 = video explanation, 2 = interactive practice, 3 = group discussion, 4 = experimental demonstration; A dynamic weight adjustment mechanism is introduced to dynamically adjust the reward weights based on the learner's real-time status. Simultaneously, a constraint on the correlation strength of knowledge gaps is incorporated to construct a dynamic weight reward function. The expression for calculating the dynamic weights is as follows: ； ； ； in, The faster the improvement in knowledge points, the better. The larger the size, the higher the stability. The larger, ; The expression for the dynamic weighted reward function is: ； in, This is the associated reward coefficient, with a value of 0.

1. The strength of the association between knowledge points i and j To improve the mastery of knowledge point i after learning j, students are encouraged to learn highly related knowledge points. Let be the reward value at time t. Rate the intensity of the target emotion. Let t be the learning time.

9. The method for optimizing dynamic teaching strategies for personalized learning according to claim 6, characterized in that, The fusion reinforcement learning model uses DQN for basic path decision-making, PPO algorithm to optimize strategy updates, and introduces an experience replay pool for optimization. The expression for calculating the Q value is as follows: ； The expression for the PPO objective function is: ； The weight expression in the experience replay pool is: ； in, State at time t Next action The expected cumulative reward, As a discount factor, The strategy ratio, For the dominant function, Let E be the clipping factor and E be the expectation operator. For sample priority, Here, N is the importance sampling coefficient, and N is the total number of samples. For maximum weight, The instantaneous reward at time t. The objective function of PPO is used to optimize the policy parameters. Weights for the empirical replay samples.

10. The method for optimizing dynamic teaching strategies for personalized learning according to claim 9, characterized in that, The teaching strategy will be adjusted immediately when any of the following conditions are met: a) Severe mood fluctuations, i.e., emotional stability score <0.3; b. No improvement in learning, i.e., 3 consecutive moments. <0.01; c. Ineffective remedial learning of highly relevant knowledge points, i.e., knowledge gap correlation strength matrix There exists >0.8 and <0.02, ensuring the strategy adapts to real-time state changes.

Images