An intelligent teaching management method based on virtual reality
By integrating the metadata and user characteristics of virtual teaching scenarios, a comprehensive level evaluation matrix for virtual teaching scenarios is constructed, which solves the problem of low personalization of virtual teaching scenarios and achieves more efficient personalized recommendations and learning experiences.
Patent Information
- Application Number
- CN202411335868.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-24
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-09-24
AI Technical Summary
The existing virtual teaching scenarios have low personalization and cannot fully utilize the advantages of virtual reality technology, resulting in poor learning results and low user satisfaction.
By obtaining historical teaching data and virtual teaching scene metadata, combining physiological characteristics and emotional characteristics, a fusion representation vector is generated using the feature fusion model, combining user individual attributes and scene difficulty coefficients, a comprehensive level evaluation matrix for virtual teaching scenes is constructed to achieve personalization recommend.
It improves the personalization of virtual teaching, improves learning effect and user experience, provides more accurate health status assessment and scenario difficulty prediction, and enhances the accuracy and targetedness of recommendations.
Smart Images

Figure CN118864194B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the technical field of virtual reality teaching, and in particular to an intelligent teaching management method based on virtual reality. Background Art
[0002] With the rapid development of information technology and continuous innovation in the field of education, virtual reality (VR) technology has been widely used in the field of teaching. With its immersive and interactive characteristics, VR technology provides learners with an immersive learning experience, stimulates learning interest, and improves learning efficiency. More and more educational institutions and enterprises are beginning to explore the combination of VR technology and teaching, and develop various virtual teaching scenarios and applications to meet the needs of different learners.
[0003] However, existing virtual teaching scenarios generally have the problem of low personalization. Although VR technology provides learners with an immersive learning environment, most virtual teaching scenarios adopt a "one-size-fits-all" approach, providing the same teaching content and experience to all learners without considering the individual differences and characteristics of learners. This lack of personalization in teaching cannot give full play to the advantages of VR technology, resulting in poor learning results and low user satisfaction.
[0004] In the related technology, for example, Chinese patent document CN117708571B provides a teaching management method and system based on virtual reality, which relates to the field of virtual reality teaching technology. The method includes: obtaining historical teaching data, extracting the user's physiological feature vector; constructing a physiological feature set of each virtual teaching scene, generating a user distribution feature vector of each physiological feature set; classifying multiple virtual teaching scenes; receiving a virtual teaching request of a target user, extracting the first target virtual teaching scene of the target user, and determining the physiological impact level of the first target virtual teaching scene; judging whether the physiological impact level of the first target virtual teaching scene is lower than the preset safety level, if so, constructing the first target virtual teaching scene, and performing virtual teaching on the target user based on the first target virtual teaching scene. Although the physiological characteristics of the user and the physiological impact level of the scene are taken into account, there are still some deficiencies in personalized recommendation. Summary of the invention
[0005] In response to the problem of low personalization in virtual teaching scenarios in the prior art, the present application provides an intelligent teaching management method based on virtual reality, which improves the user targeting in virtual teaching by integrating the user's physiological characteristics and emotional characteristics, as well as scene metadata.
[0006] The purpose of this application is achieved through the following technical solutions.
[0007] The present application provides an intelligent teaching management method based on virtual reality, including: S1, obtaining historical teaching data and virtual teaching scene metadata; wherein the historical teaching data includes physiological characteristic data and emotional characteristic data of multiple users in multiple virtual teaching scenes; the virtual teaching scene metadata includes the teaching scene knowledge point coverage and user completion time data; S2, preprocessing the acquired physiological characteristic data and emotional characteristic data; S3, inputting the preprocessed physiological characteristic data and emotional characteristic data into a feature fusion model, and obtaining a fusion representation vector containing physiological information and emotional information through the feature fusion model, and the feature fusion model adopts a multi-level structure design; S4, using the obtained fusion representation vector as input, through a multi-layer fully connected neural network, combined with the user's individual attribute data, to obtain the physiological health and emotional health of each user in each virtual scene as health state feature vector; wherein, the individual attributes of the user include age and gender; S5, taking the obtained virtual teaching scene metadata as input, using the pre-trained gradient boosting decision tree model to generate the scene difficulty coefficient; S6, according to the health state feature vectors of all users in each virtual scene and the corresponding scene difficulty coefficient, the average physiological health and average emotional health of each virtual teaching scene are calculated by the weighted average algorithm, and the grade evaluation matrix of the virtual teaching scene is constructed in combination with the scene difficulty coefficient, and the comprehensive grade score of the virtual teaching scene is generated by matrix transformation; S7, when receiving the teaching request of the target user, according to the historical teaching data of the target user, the matching degree between the target user and each virtual teaching scene is calculated by the collaborative filtering algorithm, and the recommendation priority ranking is generated in combination with the comprehensive grade score of the virtual teaching scene, and the top N virtual teaching scenes are selected and recommended to the target user.
[0008] Among them, virtual teaching scene metadata: refers to the metadata information of the virtual teaching scene, including the coverage rate of teaching scene knowledge points and the user completion time data. These data reflect the characteristics of the virtual teaching scene and the learning behavior of the user in the scene. Physiological characteristic data: refers to the user's physiological characteristic data collected in the virtual teaching scene, such as heart rate, electroencephalogram, skin conductance, etc. These data reflect the user's physiological state and reaction during the learning process. Emotional characteristic data: refers to the user's emotional characteristic data collected in the virtual teaching scene, such as facial expressions, voice intonation, body movements, etc. These data reflect the user's emotional changes and state during the learning process. Teaching scene knowledge point coverage rate: refers to the proportion of the number of knowledge points covered in the virtual teaching scene to the total number of knowledge points. It reflects the breadth and completeness of the scene's knowledge and is one of the important indicators for evaluating the quality of the scene. User completion time data: refers to the time it takes for the user to complete the learning task or achieve the learning goal in the virtual teaching scene. It reflects the difficulty of the scene and the user's learning efficiency and is one of the important indicators for evaluating the difficulty of the scene and the performance of the user. Physiological health: refers to evaluating the user's physiological health status by analyzing the user's physiological characteristic data in the virtual teaching scene. Physiological health reflects the impact of the scene on the user's physiological state, such as stress level, fatigue level, etc. Emotional health: refers to the evaluation of the user's emotional health state by analyzing the user's emotional characteristic data in the virtual teaching scene. Emotional health reflects the impact of the scene on the user's emotional state, such as pleasure, frustration, etc. Scene difficulty coefficient: refers to the scene difficulty evaluation value generated by the pre-trained gradient boosting decision tree model combined with the virtual teaching scene metadata. The scene difficulty coefficient reflects the difficulty of the scene and provides a reference for scene recommendation and learning path planning. Collaborative filtering algorithm: refers to a commonly used recommendation algorithm that predicts the user's interest in unknown items by analyzing the similarity between users and the similarity between items. In this solution, the collaborative filtering algorithm is used to calculate the matching degree between the target user and each virtual teaching scene to provide support for personalized recommendations.
[0009] Furthermore, S3, the feature fusion model adopts a multi-level structural design, including: the first layer uses a gated recurrent unit to extract the temporal dependency of physiological feature data and emotional feature data, adjusts the feature weights of different time steps by setting an attention mechanism, and performs weighted summation on the past hidden states according to the hidden state at the current moment and the attention weight to obtain the physiological feature representation and emotional feature representation in the time dimension; the second layer uses a multi-head attention mechanism to calculate the correlation between the physiological feature representation and the emotional feature representation output by the first layer in different feature subspaces, and performs multiple linear transformations on the physiological feature representation and the emotional feature representation through the multi-head attention mechanism to obtain feature representations in different subspaces, and then calculates the correlation between the physiological feature subspace and the emotional feature subspace through the attention mechanism to obtain a cross-modal attention matrix, and performs weighted fusion of the physiological feature representation and the emotional feature representation according to the attention matrix to obtain a semantically aligned cross-modal fusion feature representation; the third layer uses a multi-layer perceptron to perform semantic fusion on the output of the second layer The aligned feature representation is transformed nonlinearly, and high-level semantic information is extracted through multiple fully connected layers and nonlinear activation functions. The output of the third layer is then added element by element to the output of the second layer through residual connection to obtain a fused feature representation containing high-level semantic information and low-level feature information. Among them, the high-level semantic information represents the abstract feature pattern extracted after the cross-modal fusion feature representation is transformed by the multi-layer perceptron, which reflects the interactive relationship between physiological features and emotional features at a higher speech level. The low-level feature information represents the shallow feature representation retained in the cross-modal fusion feature representation, which reflects the local correlation between physiological features and emotional features at the original data level. The fourth layer uses a multi-scale self-attention mechanism to perform global information aggregation on the fusion features output by the third layer, and calculates the dependency between the fusion feature representation in different regions through multiple attention kernel functions of different sizes to obtain attention weights of different scales. The fusion feature representation is weighted pooled according to the attention weights to generate a hierarchical fusion representation vector output.
[0010] The parameters of the feature fusion model are optimized through a joint loss function, which includes a reconstruction loss function, a contrast loss function, and a classification loss function. The reconstruction loss function adopts the framework of a variational autoencoder, the contrast loss function adopts the paradigm of contrastive learning, and the classification loss function adopts the cross entropy function. During the model training process, the reconstruction loss function, the contrast loss function, and the classification loss function are weighted and summed to form a joint loss function through multi-task learning. During the model training process, the three loss functions are jointly optimized through multi-task learning: the reconstruction loss function is mainly aimed at the reconstruction of physiological feature data and emotional feature data, and the parameters of the encoder and decoder are optimized. The contrast loss function is mainly aimed at the contrastive learning of the fused feature representation vector, and the parameters of each layer in the feature fusion model are optimized, so that the fused feature representation can distinguish the physiological features and emotional features of different users. The classification loss function is mainly aimed at the classification task of the fused feature representation vector, and the parameters of each layer in the feature fusion model and the parameters of the classifier are optimized, so that the fused feature representation can accurately predict the user's label. In each training step, the gradients of the three loss functions are calculated and accumulated, and then used to update the parameters of the feature fusion model to achieve multi-task joint optimization. In this way, the feature fusion model can simultaneously learn the capabilities of effective feature reconstruction, feature comparison, and feature classification, thereby improving the generalization performance of the model.
[0011] Among them, Gated Recurrent Unit (GRU): a commonly used recurrent neural network structure for processing time series data. GRU controls the flow and update of information through a gating mechanism, and can effectively capture long-distance dependencies in time series data. In this scheme, GRU is used to extract the time series dependencies between physiological feature data and emotional feature data. Hidden State: In a recurrent neural network, the hidden state represents the network state at the current moment and contains information from the previous moment. In GRU, the hidden state is updated through a gating mechanism to capture the dynamic changes of time series data. Different feature subspaces: refers to mapping physiological feature representations and emotional feature representations to different feature subspaces to capture their correlation at different semantic levels. Through the multi-head attention mechanism, the correlation between physiological features and emotional features is calculated in different feature subspaces to achieve cross-modal feature fusion. Cross-modal attention weighting: refers to using the attention mechanism to calculate the correlation between physiological feature representations and emotional feature representations in different feature subspaces, and weightedly fuse them according to the attention weight. Cross-modal attention weighting can achieve semantic alignment of physiological features and emotional features, and strengthen the interaction and fusion between them. Semantically aligned feature representation: refers to the physiological feature representation and emotional feature representation obtained by cross-modal attention weighting, which are aligned and fused at the semantic level. Semantically aligned feature representation can capture the high-level semantic correlation between physiological features and emotional features.
[0012] Multi layer Perceptron (MLP): A feedforward neural network structure consisting of multiple fully connected layers and nonlinear activation functions. MLP is used to perform nonlinear transformation on input features to extract high-level semantic information. In this scheme, MLP is used to perform nonlinear transformation on semantically aligned feature representations. High-level semantic information: refers to the abstract feature pattern extracted after nonlinear transformation of semantically aligned feature representations by MLP. High-level semantic information reflects the interactive relationship between physiological features and emotional features at a higher semantic level and captures the complex correlation between them. Low-level feature information: refers to the shallow feature representation retained in the fused feature representation, reflecting the local correlation between physiological features and emotional features at the original data level. The low-level feature information is fused with the high-level semantic information through residual connections to retain feature information at different levels. Multi-scale self-attention mechanism: refers to the use of multiple attention kernel functions of different sizes to calculate the dependency between different regions of the fused feature representation and obtain attention weights of different scales. The multi-scale self-attention mechanism can capture the global information of the fused feature representation at different scales and realize hierarchical feature aggregation. Attention kernel function: refers to the function used in the self-attention mechanism to calculate the similarity or correlation between different regions. Attention kernel functions of different sizes can capture regional dependencies of different scales and realize multi-scale attention calculation. Multi-task learning method: refers to optimizing the loss functions of multiple related tasks at the same time during model training, improving the generalization ability of the model by sharing model parameters and joint optimization. In this scheme, the reconstruction loss function, contrast loss function and classification loss function are jointly optimized to realize multi-task learning.
[0013] Furthermore, the reconstruction loss function adopts the framework of variational autoencoder, including: mapping the original physiological feature data to the physiological latent space through the encoder to obtain physiological latent variables; wherein the physiological latent space is used to represent the compressed representation of the original physiological feature data; mapping the original emotional feature data to the emotional latent space through the encoder to obtain emotional latent variables; wherein the emotional latent space is used to represent the compressed representation of the original emotional feature data; reconstructing the physiological latent variables into physiological feature data through the decoder; reconstructing the emotional latent variables into emotional feature data through the decoder; calculating the physiological feature reconstruction loss based on the reconstructed physiological feature data and the original physiological feature data through the maximum average difference loss function; calculating the emotional feature reconstruction loss based on the reconstructed emotional feature data and the original emotional feature data through the maximum average difference loss function; adding the physiological feature reconstruction loss and the emotional feature reconstruction loss to obtain the reconstruction loss function.
[0014] Among them, Variational Auto encoder (VAE): a generative model consisting of an encoder and a decoder. The encoder maps the input data to the latent space to obtain the probability distribution of the latent variables; the decoder samples from the probability distribution of the latent variables and reconstructs the input data. VAE learns the compressed representation of data by minimizing the reconstruction loss and regularization terms (such as KL divergence). In this scheme, VAE is used to learn the compressed representation of physiological feature data and emotional feature data, and calculate the reconstruction loss. Physiological latent space: refers to the low-dimensional space to which the original physiological feature data is mapped by the encoder. The physiological latent space represents the compressed representation of the physiological feature data and captures the essential characteristics and patterns of the physiological feature data. By sampling and decoding in the physiological latent space, data similar to the original physiological feature data can be generated. Physiological latent variables: refers to the variables that represent the compressed representation of the original physiological feature data in the physiological latent space. The physiological latent variables are learned from the original physiological feature data by the encoder and contain the key information and features of the physiological feature data. The physiological latent variables can be regarded as the encoding of the physiological feature data in the low-dimensional space. Emotion latent variables: refers to variables that represent the compressed representation of the original emotion feature data in the emotion latent space. The emotion latent variables are learned from the original emotion feature data by the encoder and contain the key information and features of the emotion feature data. The emotion latent variables can be regarded as the encoding of the emotion feature data in the low-dimensional space.
[0015] In this scheme, the reconstruction loss function adopts the framework of variational autoencoder. The encoder maps the original physiological feature data and emotional feature data to the physiological latent space and emotional latent space respectively to obtain the corresponding physiological latent variables and emotional latent variables. Then, the decoder reconstructs the latent variables into the original physiological feature data and emotional feature data, and calculates the reconstruction loss. The reconstruction loss function measures the difference between the reconstructed data and the original data, reflecting the information loss of the model during the compression and reconstruction process. By minimizing the reconstruction loss function, the variational autoencoder learns the compressed representation of the physiological feature data and emotional feature data, namely the physiological latent variables and emotional latent variables. These latent variables capture the essential characteristics and patterns of the data and can be used for subsequent feature fusion and representation learning tasks. At the same time, by optimizing the reconstruction loss function, the variational autoencoder can generate data similar to the original data, realizing data compression and generation.
[0016] Furthermore, the maximum average difference loss function is calculated as follows ,in, represents the maximum average difference loss; and Respectively represent the i-th physiological characteristic data and emotional feature data Eigenmapping function mapped to the reproducing kernel Hilbert space RKHS; represents the RKHS space; and Respectively represent the original data of the ith physiological feature and the original data of the emotional feature; and Represents the number of samples of physiological characteristic data and emotional characteristic data respectively; and Respectively represent the i-th reconstructed physiological feature original data and emotional feature original data; represents the squared norm in the RKHS space.
[0017] The maximum average difference loss function measures the distribution difference between the reconstructed data and the original data by calculating the mean difference between the original data and the reconstructed data in the RKHS space, thereby guiding the optimization of the encoder and decoder to make the reconstructed data as close to the original data as possible.
[0018] Furthermore, the contrastive loss function adopts the paradigm of contrastive learning, including: taking the hierarchical fusion representation vector output by the fourth layer as the anchor vector; sampling the physiological feature data and emotional feature data of the same user based on the anchor vector to obtain a positive sample vector, and sampling the physiological feature data and emotional feature data of different users to obtain a negative sample vector; calculating the similarity between the anchor vector and the positive sample vector, and calculating the similarity between the anchor vector and the negative sample vector; based on the similarity between the anchor vector and the positive sample vector, calculating the similarity between the anchor vector and the negative sample vector, and calculating the contrastive loss function through Info NCE Loss.
[0019] Among them, the Contrastive Learning Paradigm is an unsupervised learning method that learns effective representations of data by learning to distinguish similar positive samples from dissimilar negative samples. The goal of contrastive learning is to maximize the similarity between pairs of positive samples while minimizing the similarity between pairs of negative samples. In this way, the model learns the discriminative representation of the data and is able to distinguish samples of different categories. In this scheme, contrastive learning is used to learn the discriminative representation of the fused representation vector to improve the effect of feature fusion. Anchor Vector: refers to a vector used as a reference in contrastive learning. The anchor vector is usually the output or intermediate representation of the model, which is used to compare with positive and negative samples.
[0020] In this scheme, the hierarchical fusion representation vector output by the fourth layer is used as an anchor vector as a reference vector for contrastive learning. Info NCE Loss (Info Noise-Contrastive Estimation Loss): A commonly used contrastive loss function that measures the difference between the similarity between the anchor vector and the positive sample vector and the similarity between the anchor vector and the negative sample vector. Info NCE Loss encourages the model to learn discriminative representations by comparing the similarity of positive sample pairs with the similarity of negative sample pairs.
[0021] In this scheme, the contrast loss function uses Info NCE Loss, which encourages the model to learn a discriminative fusion representation vector by calculating the similarity between the anchor vector and the positive sample vector and the similarity between the anchor vector and the negative sample vector. By minimizing Info NCE Loss, the model can map the physiological feature data and emotional feature data of the same user to similar representations, and map the data of different users to dissimilar representations, thereby achieving user-level feature alignment and discriminative representation learning.
[0022] Furthermore, the contrast loss function expression is:
[0023] Where V represents the number of perspectives; N represents the total number of samples used for contrastive learning; represents the temperature coefficient; and Respectively represent the number of negative samples of physiological features and the number of negative samples of emotional features; and They represent the physiological feature representation and emotional feature representation of the i-th sample at the v-th perspective respectively; and They represent the positive sample physiological feature representation and positive sample emotional feature representation of the i-th sample at the v-th perspective respectively; and They respectively represent the j-th negative sample physiological feature representation and negative sample emotional feature representation of the ith sample at the v-th perspective.
[0024] is the temperature coefficient, where, in the contrast loss function, the temperature coefficient is an important hyperparameter that controls the degree of scaling of the similarity score, thereby affecting the gradient of the loss function and the optimization process. In the contrast loss function, we construct the loss function by calculating the similarity between the anchor vector and the positive sample vector, and the similarity between the anchor vector and the negative sample vector. The similarity is usually calculated using methods such as inner product or cosine similarity, and the original similarity scores may be distributed in a large range. Temperature coefficient Is a positive real number that scales the original similarity score to map it to a suitable scale. The temperature coefficient can control the degree of "softening" of the similarity score. When the temperature coefficient is large, the similarity score will be reduced, making the similarity difference between different samples smaller; when the temperature coefficient is small, the similarity score will be amplified, making the similarity difference between different samples larger. By adjusting the temperature coefficient, we can affect the gradient calculation and optimization process in the contrastive loss function. A larger temperature coefficient will make the gradient smaller, thereby slowing down the model optimization process, but may be more stable; a smaller temperature coefficient will make the gradient larger, thereby speeding up the model optimization process, but may cause training instability. A suitable temperature coefficient can help the model establish a suitable interval between positive and negative samples and improve the generalization ability of the model. If the temperature coefficient is set too large, the model may have difficulty distinguishing between positive and negative samples; if the temperature coefficient is set too small, the model may focus too much on the differences between individual samples and find it difficult to learn robust feature representations. The temperature coefficient is usually set to a positive real number, with a common value range of 0.1 to 1.
[0025] Furthermore, the classification loss function adopts a cross entropy function, including: inputting the hierarchical fusion representation vector output by the fourth layer into the classifier, performing feature transformation and nonlinear mapping on the fusion feature vector through the classifier to obtain a prediction probability vector; normalizing the prediction probability vector through a softmax function to obtain a normalized prediction probability vector; comparing the normalized prediction probability vector with the annotated true label element by element, and calculating the classification loss function based on the cross entropy function to measure the difference between the prediction probability and the true label.
[0026] Among them, classifier: a model used to map input features to predefined categories. The classifier predicts the category of new input samples by learning the mapping relationship between input features and category labels. In this scheme, the classifier receives the hierarchical fusion representation vector output by the fourth layer and maps it to the predicted category probability vector. The classifier is usually composed of a multi-layer fully connected neural network, and realizes the mapping of features to categories through nonlinear transformation and feature transformation.
[0027] Feature Transformation: refers to mapping input features to a new feature space through a transformation function. The purpose of feature transformation is to convert the original features into a more discriminative and informative representation to facilitate subsequent classification tasks. In this scheme, the classifier performs feature transformation on the fused feature vector and maps the fused features to the space where the prediction probability vector is located through linear or nonlinear transformation. Common feature transformation methods include linear transformation (such as fully connected layers) and nonlinear transformation (such as activation functions).
[0028] Nonlinear Mapping: refers to mapping input features to new feature spaces through nonlinear functions. Nonlinear mapping can capture complex nonlinear relationships between input features and enhance the representation capabilities of features. In neural networks, nonlinear mapping is usually implemented through activation functions, such as ReLU, sigmoid, tanh, etc. In this scheme, the classifier uses nonlinear mapping to convert the fused feature vector into a prediction probability vector, introduces nonlinear transformation, and improves the fitting ability of the classifier.
[0029] Predicted Probability Vector: refers to the vector output by the classifier that represents the predicted probability of each category. Each element of the predicted probability vector represents the estimated probability that the input sample belongs to the corresponding category. In this scheme, the classifier fuses the feature vector through feature transformation and nonlinear mapping to obtain the predicted probability vector, which represents the possibility that the input sample belongs to each category. The predicted probability vector is usually normalized by the softmax function so that the sum of its elements is 1, representing a probability distribution.
[0030] In this scheme, the classification loss function uses the cross entropy function to measure the difference between the predicted probability vector and the true label. Specifically, the hierarchical fusion representation vector output by the fourth layer is input into the classifier, and the predicted probability vector is obtained through feature transformation and nonlinear mapping. Then, the predicted probability vector is normalized by the softmax function to obtain a normalized vector representing the predicted probability of each category. Finally, the normalized predicted probability vector is compared element by element with the true label, and the classification loss function is calculated based on the cross entropy function. The cross entropy function measures the difference between the predicted probability and the true label, encouraging the model to learn the correct category prediction.
[0031] Furthermore, ,in, Indicates the predicted probability that the i-th sample belongs to the c-th class; The cth element of the one-hot encoding vector representing the true label of the i-th sample; represents the category measurement factor; The focusing parameter is the category-specific focusing parameter. Different categories have different focusing parameters. For categories that are difficult to distinguish, the larger the focusing parameter, the greater the penalty of the loss function for samples with a smaller prediction probability. This can be optimized through methods such as cross-validation. is the sample weight factor, which is used to adjust the contribution of different samples to the loss function. For samples that are difficult to classify correctly, the larger the weight factor is; ,in, represents the number of samples in the cth category, and N represents the total number of samples; ,in, represents the predicted probability of the i-th sample in the t-1th iteration.
[0032] Furthermore, S5 uses the obtained virtual teaching scene metadata as input, and uses the pre-trained gradient boosting decision tree model to generate the scene difficulty coefficient, including: extracting indicators such as knowledge point coverage and user completion time from the virtual teaching scene metadata as input features, and using the scene difficulty level annotated by experts as training labels to construct a scene difficulty classification data set; randomly dividing the scene difficulty classification data set into a training set and a test set, using the training set data to train the gradient boosting decision tree model, and optimizing the model hyperparameters through the grid search method, including the number of decision trees, the maximum tree depth, the learning rate, etc.; wherein the grid search method evaluates the model performance by traversing different combinations of hyperparameters using a cross-validation method, Select the hyperparameter combination with the best performance; use the test set data to evaluate the performance of the trained gradient boosting decision tree model, calculate the model's accuracy, precision, recall and other indicators, and perform model tuning; among them, the accuracy measures the proportion of samples predicted correctly by the model to the total number of samples, the precision measures the proportion of samples predicted to be positive that are truly positive, and the recall measures the proportion of samples predicted to be positive that are truly positive; deploy the trained gradient boosting decision tree model to the scene difficulty coefficient generation module, perform difficulty prediction on the new virtual teaching scene metadata obtained in step S1, and output the normalized scene difficulty coefficient. The value range of the difficulty coefficient is [0, 1], and the larger the value, the higher the scene difficulty.
[0033] Among them, the Grid Search Method is a commonly used hyperparameter tuning method, which is used to search for the optimal hyperparameter combination within the specified hyperparameter value range. Grid Search tries all possible combinations of hyperparameters through exhaustive search, and selects the optimal hyperparameter settings based on the performance evaluation results of the model on the validation set. Specifically, the value range of each hyperparameter is divided into several discrete values to form a hyperparameter grid. Then, for each hyperparameter combination in the grid, the corresponding model is trained and the performance is evaluated on the validation set. Finally, the hyperparameter combination with the best performance on the validation set is selected as the final hyperparameter setting.
[0034] In this solution, the grid search method is used to tune the hyperparameters of the gradient boosting decision tree model. A hyperparameter grid is formed by setting the number of decision trees, the maximum tree depth, and the range of learning rates. For each hyperparameter combination in the grid, the corresponding gradient boosting decision tree model is trained using the training set, and the model performance is evaluated on the validation set. Finally, the hyperparameter combination with the best performance on the validation set is selected as the final hyperparameter setting of the model. The grid search method can help find a hyperparameter combination that is more suitable for the task and improve the performance and generalization ability of the model.
[0035] Gradient Boosting Decision Tree (GBDT): A machine learning algorithm based on ensemble learning that makes predictions by iteratively building a series of decision tree models. The core idea of GBDT is to train the next decision tree based on the residual of the previous decision tree through gradient boosting to continuously reduce the prediction error. Specifically, the training process of GBDT is as follows: Initialize a constant prediction value, usually the average value of the training set labels. Calculate the residual (the difference between the true value and the predicted value) of the current model on the training set. Use the residual as the new target value and train a decision tree model to fit the residual. Add the prediction result of the newly trained decision tree model to the current prediction value to obtain the updated prediction value. Repeat until the predetermined number of iterations is reached or the stopping condition is met.
[0036] In the prediction stage, the prediction results of all decision trees are added to get the final prediction value. GBDT gradually reduces the prediction error and improves the model's fitting ability by continuously training new decision trees based on the residuals. At the same time, by integrating the prediction results of multiple decision trees, GBDT can effectively reduce the variance of a single decision tree and improve the generalization performance of the model. In this scheme, a pre-trained gradient boosting decision tree model is used to generate the scene difficulty coefficient. By extracting the knowledge point coverage and the user completion time as input features and using the annotated scene difficulty level as the training label, a scene difficulty classification dataset is constructed. Then, the gradient boosting decision tree model is trained using the training set, and the model's hyperparameters, such as the number of decision trees, the maximum tree depth, and the learning rate, are adjusted using the grid search method. After training, the model is evaluated using the test set to verify the generalization ability of the model. Finally, the trained gradient boosting decision tree model is used to predict the difficulty of the input virtual scene metadata and output the scene difficulty coefficient.
[0037] Further S6, generating a comprehensive grade score for the virtual teaching scene, including: obtaining the health status feature vector of all users in each virtual teaching scene output by step S4, wherein the health status feature vector includes two elements: physiological health and emotional health; for each virtual teaching scene, calculating the average value of the health status feature vectors of all users in the scene, and obtaining the average physiological health and average emotional health of the scene; constructing a triplet (scenario ID, health impact, difficulty coefficient) by combining the calculated average physiological health and average emotional health of each virtual teaching scene with the corresponding scene difficulty coefficient output by step S5; wherein the health impact includes two elements: average physiological health and average emotional health; sorting the constructed triples of all virtual teaching scenes according to the scene ID to form a matrix of N rows and 3 columns as a virtual teaching scene grade evaluation matrix; wherein N is the total number of virtual teaching scenes obtained in step S1, and the first column is the scene ID , the second column is the health impact, and the third column is the difficulty coefficient; perform data normalization on the generated virtual teaching scene grade evaluation matrix, perform Min-Max standardization on each element in the health impact in the scene dimension, scale the value range to [0, 1], and obtain a standardized health impact vector; perform weighted summation on the standardized health impact vector and the difficulty coefficient to generate a comprehensive grade score; wherein the weight can be adjusted according to actual needs, for example, a greater weight is given to the difficulty coefficient, and the comprehensive grade score will be more biased towards the impact of the scene difficulty; add the calculated comprehensive grade score to the last column of the generated virtual teaching scene grade evaluation matrix to obtain an extended evaluation matrix; according to the extended evaluation matrix, evaluate and rank the comprehensive impacts of different virtual teaching scenes, and output the comprehensive grade score of the virtual teaching scene; the scoring result can be used in step S7 to generate a recommended priority ranking of the virtual teaching scene in combination with the matching degree between the target user and each virtual teaching scene.
[0038] Triplet: In this solution, a triplet is a data structure used to represent the characteristics of a virtual teaching scene, which consists of three elements: scene ID, health impact, and difficulty coefficient. Specifically: Scene ID: a unique identifier for a virtual teaching scene, used to distinguish different scenes. Health impact: contains two elements, average physiological health and average emotional health, which indicate the degree of impact of the scene on the user's health. The average physiological health and average emotional health are calculated by averaging the physiological health and emotional health of all users in the scene. Difficulty coefficient: indicates the difficulty level of the scene, which is predicted by a pre-trained gradient boosting decision tree model. By constructing a triplet, the key features of each virtual teaching scene (scene ID, health impact, difficulty coefficient) are combined together to facilitate subsequent evaluation and sorting.
[0039] Min-Max Normalization on Scene Dimension: Min-Max normalization is a commonly used data normalization method that maps data to a specified range (usually [0, 1]). In this scheme, each element in the health impact quantity (average physical health and average emotional health) is Min-Max normalized on the scene dimension. The specific steps are as follows: Let the value of an element in the health impact quantity in different scenes be , where N is the total number of scenes. Calculate the minimum value of the element in all scenes and maximum value For each scene i, the corresponding element value Normalized to , the calculation formula is: , repeat until all elements in the health impact quantity are standardized. Through Min-Max standardization on the scenario dimension, the health impact quantities of different scenarios can be mapped to the same scale, eliminating the differences in dimensions and numerical ranges, which is convenient for subsequent weighted summation and comprehensive evaluation. The standardized health impact quantity is called a standardized health impact vector.
[0040] In this scheme, the key features of each virtual teaching scene are represented by constructing a triple (scenario ID, health impact, difficulty coefficient), and the triples of all scenes are sorted according to the scene ID to form a virtual teaching scene grade evaluation matrix. Then, each element in the health impact is normalized in the scene dimension by Min-Max to obtain a standardized health impact vector. Finally, the standardized health impact vector and the difficulty coefficient are weighted and summed to generate an initial comprehensive grade score, which is added to the last column of the evaluation matrix to obtain an extended evaluation matrix. According to the extended evaluation matrix, the comprehensive impact of different virtual teaching scenes is evaluated and sorted, and the comprehensive grade score of the virtual teaching scene is output.
[0041] Compared with the prior art, the advantages of this application are:
[0042] By acquiring historical teaching data and metadata of virtual teaching scenes, and preprocessing physiological characteristic data and emotional characteristic data, the data quality is improved, providing reliable input for subsequent steps and improving the stability and accuracy of the overall solution.
[0043] The reconstruction loss function adopts the framework of variational autoencoder, introduces physiological latent space and emotional latent space, maps the original feature data to the latent space through the encoder, obtains compressed representation, and then reconstructs the feature data through the decoder. This method can effectively extract low-dimensional representations of physiological and emotional features while retaining the essential attributes and key information of the features. The reconstruction loss is calculated using the maximum mean difference loss function. By measuring the mean difference between the original data and the reconstructed data in the reproducing kernel Hilbert space (RKHS), the difference in data distribution can be more accurately characterized, guiding the optimization of the encoder and decoder, making the reconstructed data closer to the original data, and improving the reconstruction quality.
[0044] The contrast loss function adopts the paradigm of contrastive learning, takes the hierarchical fusion representation vector output by the fourth layer as the anchor vector, samples positive samples in the feature data of the same user, and samples negative samples in the feature data of different users, and learns discriminative feature representations by maximizing the similarity between the anchor and the positive sample and minimizing the similarity between the anchor and the negative sample. The contrast loss function is calculated using Info NCE Loss, and the softmax normalization process can effectively measure the relative similarity between the anchor and the positive and negative samples, strengthen the attractiveness of the positive sample, weaken the influence of the negative sample, and make the learned feature representation have stronger discriminative ability.
[0045] The classification loss function uses the cross entropy function as the classification loss function. By inputting the hierarchical fusion representation vector output by the fourth layer into the classifier for feature transformation and nonlinear mapping, the predicted probability vector is obtained, and then compared with the true label. This can effectively measure the difference between the predicted result and the true label and guide the optimization of the classifier. The introduction of adaptive weighted FocalLoss can effectively alleviate the problem of category imbalance through the setting of category measurement factors and focus parameters, pay more attention to difficult-to-distinguish categories, improve the sensitivity of the classifier to minority class samples, and improve the overall classification performance.
[0046] By introducing user individual attribute data and combining it with the fusion representation vector, the physical health and emotional health of each user in each virtual scene are obtained through a multi-layer fully connected neural network to form a health status feature vector. This method takes into account individual differences among users, provides personalized health status assessment, and improves the accuracy and pertinence of health status assessment.
[0047] The scene difficulty coefficient is generated by the pre-trained gradient boosting decision tree model, and the scene difficulty classification data set is used for training and evaluation, which improves the reliability of scene difficulty prediction. The health status feature vectors of all users in each virtual scene and the corresponding scene difficulty coefficient are combined, and the comprehensive grade score of the virtual teaching scene is generated through weighted average algorithm, data normalization processing and matrix transformation, which fully considers the scene difficulty and user health status.
[0048] According to the historical teaching data of the target user, the matching degree between the target user and each virtual teaching scene is calculated through the collaborative filtering algorithm. Combined with the comprehensive grade score of the virtual teaching scene, a personalized recommendation priority ranking is generated, and the top N virtual teaching scenes are selected and recommended to the target user. This method makes full use of the user's historical data and the comprehensive evaluation results of the scene, provides highly personalized and accurate scene recommendations, and improves the user experience and teaching effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 is an exemplary flow chart of an intelligent teaching management method based on virtual reality according to some embodiments of the present application;
[0050] Figure 2 is an exemplary flow chart of obtaining a fused representation vector according to some embodiments of the present application;
[0051] Figure 3 is an exemplary flow chart of constructing a reconstruction loss function according to some embodiments of the present application;
[0052] Figure 4 is an exemplary flow chart of constructing a contrast loss function according to some embodiments of the present application;
[0053] Figure 5 is an exemplary flow chart of obtaining a scene difficulty coefficient according to some embodiments of the present application;
[0054] Figure 6 This is an exemplary flowchart for obtaining a comprehensive grade score of a virtual teaching scene according to some embodiments of the present application. DETAILED DESCRIPTION
[0055] The method and system provided in the embodiments of the present application are described in detail below with reference to the accompanying drawings.
[0056] Example 1
[0057] Figure 1This is an exemplary flow chart of an intelligent teaching management method based on virtual reality shown in some embodiments of the present application, an intelligent teaching management method based on virtual reality, comprising: obtaining historical teaching data and virtual teaching scene metadata; wherein the historical teaching data includes physiological characteristic data and emotional characteristic data of multiple users in multiple virtual teaching scenes; the virtual teaching scene metadata includes the teaching scene knowledge point coverage and user completion time data; preprocessing the acquired physiological characteristic data and emotional characteristic data; inputting the preprocessed physiological characteristic data and emotional characteristic data into a feature fusion model, and obtaining a fusion representation vector containing physiological information and emotional information through the feature fusion model, wherein the feature fusion model adopts a multi-level structure design; using the obtained fusion representation vector as input, and obtaining the physiological information of each user in each virtual scene through a multi-layer fully connected neural network in combination with the user's individual attribute data. The physical health and emotional health are taken as health status feature vectors; the individual attributes of users include age and gender; the obtained virtual teaching scene metadata is taken as input, and the scene difficulty coefficient is generated by using the pre-trained gradient boosting decision tree model; according to the health status feature vectors of all users in each virtual scene and the corresponding scene difficulty coefficient, the average physiological health and average emotional health of each virtual teaching scene are calculated by the weighted average algorithm, and the grade evaluation matrix of the virtual teaching scene is constructed in combination with the scene difficulty coefficient, and the comprehensive grade score of the virtual teaching scene is generated by matrix transformation; when a teaching request of the target user is received, the matching degree between the target user and each virtual teaching scene is calculated by the collaborative filtering algorithm according to the historical teaching data of the target user, and the recommendation priority ranking is generated in combination with the comprehensive grade score of the virtual teaching scene, and the top N virtual teaching scenes are selected and recommended to the target user.
[0058] Data acquisition and preprocessing, acquisition of historical teaching data and virtual teaching scene metadata, acquisition of historical teaching data: extract historical teaching data from the database or log file of the virtual teaching system. Historical teaching data contains physiological characteristic data and emotional characteristic data of multiple users in multiple virtual teaching scenes. Physiological characteristic data can be collected through wearable devices or sensors, such as heart rate, electroencephalogram, skin conductance, etc. Emotional characteristic data can be obtained through facial expression recognition, voice emotion analysis and other technologies, such as smile degree, frown degree, tone of voice, etc. The acquired physiological characteristic data and emotional characteristic data are organized and stored according to user ID and timestamp to form a structured data set.
[0059] Obtaining virtual teaching scene metadata: Extracting virtual teaching scene metadata from the course design document or configuration file of the virtual teaching system. Virtual teaching scene metadata includes teaching scene knowledge point coverage and user completion time data. Teaching scene knowledge point coverage indicates the proportion of knowledge points covered by the scene to the total knowledge points, which can be calculated by analyzing the scene content and teaching objectives. User completion time data records the learning time of each user in the scene, which can be obtained through the user's operation log or interaction record. The obtained virtual teaching scene metadata is organized and stored according to the scene ID, and associated with the corresponding historical teaching data.
[0060] Preprocess the physiological characteristic data and emotional characteristic data: clean and filter the acquired physiological characteristic data and emotional characteristic data to remove outliers, missing values and noise data. Normalize the physiological characteristic data and scale different types of physiological characteristic data to the same numerical range, such as between 0 and 1. Standardize the emotional characteristic data and convert different types of emotional characteristic data into a standard normal distribution with a mean of 0 and a variance of 1. Time-align the physiological characteristic data and emotional characteristic data to ensure that the physiological characteristic and emotional characteristic data at each timestamp correspond to the same user and the same teaching scenario. Perform feature selection and dimensionality reduction on the preprocessed physiological characteristic data and emotional characteristic data to remove redundant and irrelevant features, reduce the data dimension, and improve the training efficiency of subsequent models.
[0061] Figure 2 This is an exemplary flow chart for obtaining a fused representation vector according to some embodiments of the present application, feature fusion, first layer: temporal feature extraction, using a gated recurrent unit (such as GRU or LSTM) to extract temporal features from physiological feature data and emotional feature data. For each time step, the physiological feature vector and the emotional feature vector are used as input, and the hidden state is calculated through the gated recurrent unit. An attention mechanism is set to calculate the attention weight of the hidden state of the past time step based on the hidden state at the current moment. The past hidden states are weighted and summed according to the attention weights to obtain the physiological feature representation and emotional feature representation at the current moment. The feature weights of different time steps are adjusted through the attention mechanism to highlight the feature representation of key moments.
[0062] The second layer: cross-modal feature alignment, using a multi-head attention mechanism to perform cross-modal feature alignment on the physiological feature representation and emotional feature representation output by the first layer. Map the physiological feature representation and emotional feature representation to multiple feature subspaces respectively, and calculate their correlation in different subspaces. Through the cross-modal attention mechanism, with the physiological feature representation as the query and the emotional feature representation as the key value, calculate the attention weights of the physiological feature representation and the emotional feature representation. Perform a weighted summation of the emotional feature representation according to the attention weights to obtain the emotional feature representation that is semantically aligned with the physiological feature representation. Similarly, with the emotional feature representation as the query and the physiological feature representation as the key value, calculate the attention weights of the emotional feature representation and the physiological feature representation to obtain the physiological feature representation that is semantically aligned with the emotional feature representation.
[0063] The third layer: extracts high-level semantic information. It uses a multi-layer perceptron to perform nonlinear transformation on the semantically aligned feature representation output by the second layer. The feature representation is transformed through multiple fully connected layers and nonlinear activation functions (such as ReLU) to extract high-level semantic information. Through the residual connection method, the output of the third layer is added element by element to the output of the second layer to obtain a fused feature representation containing high-level semantic information and low-level feature information. The high-level semantic information represents the abstract feature pattern extracted after the cross-modal fusion feature representation is transformed by the multi-layer perceptron, reflecting the interactive relationship between physiological features and emotional features at a higher semantic level. The low-level feature information represents the shallow feature representation retained in the cross-modal fusion feature representation, reflecting the local correlation between physiological features and emotional features at the original data level.
[0064] The fourth layer: multi-scale feature aggregation, using a multi-scale self-attention mechanism to perform global information aggregation on the fused features output by the third layer. Use multiple attention kernel functions of different sizes (such as convolution kernels of different sizes) to extract local area features from the fused feature representation. For the output of each attention kernel function, the self-attention mechanism is used to calculate the dependency between different regions of the fused feature representation to obtain attention weights of different scales. The fused feature representation is weighted pooled according to the attention weights of different scales to obtain fused representation vectors of different scales. The fused representation vectors of different scales are concatenated to form the final hierarchical fused feature representation output.
[0065] The parameters of the feature fusion model are optimized through a joint loss function, which includes a reconstruction loss function, a contrast loss function, and a classification loss function. The reconstruction loss function uses the framework of a variational autoencoder, the contrast loss function uses the paradigm of contrastive learning, and the classification loss function uses a cross entropy function. During the model training process, the reconstruction loss function, the contrast loss function, and the classification loss function are weighted and summed to form a joint loss function through multi-task learning. Multiply the reconstruction loss function by the weight coefficient , the contrast loss function is multiplied by the weight coefficient , the classification loss function is multiplied by the weight coefficient . Add the weighted reconstruction loss function, contrast loss function and classification loss function to get the joint loss function: , the weight coefficients can be optimized by empirical settings or by hyperparameter search methods to balance the impact of different tasks on model training. Use optimization algorithms such as Adam and SGD to optimize the joint loss function. Calculate the gradient of the joint loss function to the model parameters, and use the optimization algorithm to update the model parameters. Through multiple training iterations, minimize the joint loss function so that the model can achieve good performance on reconstruction tasks, comparison tasks, and classification tasks at the same time.
[0066] Figure 3 This is an exemplary flow chart for constructing a reconstruction loss function according to some embodiments of the present application. The reconstruction loss function adopts the framework of a variational autoencoder. Physiological feature encoding: the original physiological feature data is mapped to a physiological latent space through an encoder. The encoder can use structures such as a multi-layer perceptron (MLP) or a convolutional neural network (CNN). The original physiological feature data is compressed into a low-dimensional latent space by the encoder to obtain physiological latent variables. The physiological latent variables represent the compressed representation of the original physiological feature data and capture the key information of the physiological features. Emotional feature encoding: the original emotional feature data is mapped to an emotional latent space through an encoder. The encoder can use structures such as a multi-layer perceptron (MLP) or a convolutional neural network (CNN). The original emotional feature data is compressed into a low-dimensional latent space by the encoder to obtain emotional latent variables. The emotional latent variables represent the compressed representation of the original emotional feature data and capture the key information of the emotional features.
[0067] Physiological feature reconstruction: The physiological latent variables are reconstructed into physiological feature data through the decoder. The decoder can use structures such as multi-layer perceptron (MLP) or deconvolutional neural network. The physiological latent variables are mapped back to the dimensions of the original physiological feature data through the decoder to obtain reconstructed physiological feature data. The reconstructed physiological feature data should be as similar as possible to the original physiological feature data to ensure that the encoder and decoder can effectively capture the key information of physiological features. Emotional feature reconstruction: The emotional latent variables are reconstructed into emotional feature data through the decoder. The decoder can use structures such as multi-layer perceptron (MLP) or deconvolutional neural network. The emotional latent variables are mapped back to the dimensions of the original emotional feature data through the decoder to obtain reconstructed emotional feature data. The reconstructed emotional feature data should be as similar as possible to the original emotional feature data to ensure that the encoder and decoder can effectively capture the key information of emotional features.
[0068] Physiological feature reconstruction loss calculation: Based on the reconstructed physiological feature data and the original physiological feature data, the physiological feature reconstruction loss is calculated by the Maximum Mean Discrepancy (MMD) loss function. The MMD loss function measures the difference between the distribution of the two sets of data, maps the data to a high-dimensional feature space through a kernel function, and calculates the mean difference between the two sets of data in the feature space. The physiological feature reconstruction loss represents the difference between the reconstructed physiological feature data and the original physiological feature data, and is used to optimize the parameters of the encoder and decoder so that the reconstructed physiological feature data is as close to the original data as possible. Emotional feature reconstruction loss calculation: Based on the reconstructed emotional feature data and the original emotional feature data, the emotional feature reconstruction loss is calculated by the Maximum Mean Discrepancy (MMD) loss function. The MMD loss function measures the difference between the distribution of the two sets of data, maps the data to a high-dimensional feature space through a kernel function, and calculates the mean difference between the two sets of data in the feature space. The emotional feature reconstruction loss represents the difference between the reconstructed emotional feature data and the original emotional feature data, and is used to optimize the parameters of the encoder and decoder so that the reconstructed emotional feature data is as close to the original data as possible.
[0069] Reconstruction loss function calculation: Add the physiological feature reconstruction loss and the emotional feature reconstruction loss to get the total reconstruction loss function. The reconstruction loss function represents the performance of the encoder and decoder in reconstructing physiological feature data and emotional feature data, and is used to optimize the parameters of the variational autoencoder. By minimizing the reconstruction loss function, the encoder can map the original feature data to a meaningful latent space, and the decoder can reconstruct feature data similar to the original data from the latent space. The calculation formula of the maximum average difference loss function is as follows: .
[0070] Figure 4 This is an exemplary flow chart for constructing a contrast loss function according to some embodiments of the present application. The contrast loss function adopts the paradigm of contrast learning to obtain an anchor vector: the hierarchical fusion representation vector output by the fourth layer is used as the anchor vector. The anchor vector represents a fusion representation of the physiological characteristics and emotional characteristics of the current user at a specific moment. Sampling positive sample vectors: The positive sample vector is obtained by sampling the physiological characteristic data and emotional characteristic data of the same user based on the anchor vector. The positive sample vector represents a fusion representation of the physiological characteristics and emotional characteristics of the same user at other moments as the anchor vector. The positive sample vector can be obtained by random sampling or time window sampling to ensure that the positive sample vector has a certain correlation with the anchor vector.
[0071] Sampling negative sample vectors: Sampling negative sample vectors from physiological feature data and emotional feature data of different users. The negative sample vector represents the fusion representation of physiological features and emotional features of different users with the anchor vector. Obtain negative sample vectors by random sampling or batch sampling to ensure that the negative sample vectors and the anchor vectors have certain differences. Calculate the similarity between the anchor vector and the positive sample vector: Use cosine similarity or dot product to calculate the similarity between the anchor vector and the positive sample vector. The higher the similarity, the closer the distance between the anchor vector and the positive sample vector in the feature space, and the greater the possibility that they belong to the same user.
[0072] Calculate the similarity between the anchor vector and the negative sample vector: Use cosine similarity or dot product to calculate the similarity between the anchor vector and the negative sample vector. The lower the similarity, the farther the anchor vector and the negative sample vector are in the feature space, and the greater the possibility that they belong to different users. Calculate the contrast loss function: Based on the similarity between the anchor vector and the positive sample vector and the similarity between the anchor vector and the negative sample vector, the contrast loss function is calculated using Info NCE Loss. The contrast loss function expression is:
[0073]
[0074] The goal of the contrastive loss function is to maximize the similarity between the anchor vector and the positive sample vector, while minimizing the similarity between the anchor vector and the negative sample vector. By minimizing the contrastive loss function, the model can learn to distinguish the feature representation of the same user from different users, improving the discriminability and generalization ability of the feature representation.
[0075] Optimize the contrastive loss function: Use an optimization algorithm, such as Adam or SGD, to optimize the contrastive loss function. Calculate the gradient of the contrastive loss function with respect to the model parameters, and use the optimization algorithm to update the model parameters. Through multiple training iterations, minimize the contrastive loss function so that the model learns feature representations that are discriminative and generalizable.
[0076] The classification loss function uses the cross entropy function, including: Input of the fused feature vector: The hierarchical fused representation vector output by the fourth layer is used as the input of the classifier. The fused feature vector contains the representation of physiological features and emotional features at different abstract levels, and contains rich user status information. Design of the classifier: The classifier can use a multi-layer perceptron (MLP) or other neural network structure. The classifier performs feature transformation and nonlinear mapping on the fused feature vector to extract high-level features related to the classification task. The output of the classifier is a prediction probability vector, which represents the probability that the fused feature vector belongs to each category. Normalization of the prediction probability vector: The prediction probability vector output by the classifier is normalized by the softmax function. The softmax function maps each element in the prediction probability vector to the range of (0, 1) and ensures that the sum of all elements is 1. The normalized prediction probability vector represents the normalized probability distribution of the fused feature vector in each category.
[0077] Representation of true labels: The annotated true labels are represented as one-hot encoding. For each sample, the true label is a vector with the same length as the number of categories, where the position corresponding to the true category is 1 and the other positions are 0. Calculation of cross entropy loss function: Compare the normalized predicted probability vector with the true label element by element and calculate the cross entropy loss function. The expression of the cross entropy loss function is as follows: ,in, Indicates the predicted probability that the i-th sample belongs to the c-th class; The cth element of the one-hot encoding vector representing the true label of the i-th sample; represents the category measurement factor; Represents category-specific focusing parameters, and different categories set different focusing parameters; Represents the sample weight factor, which is used to adjust the contribution of different samples to the loss function; ,in, represents the number of samples in the cth category, and N represents the total number of samples; ,in, represents the predicted probability of the i-th sample in the t-1th iteration.
[0078] The fused representation vector is input into a multi-layer fully connected neural network, and combined with the user's individual attribute data, the physiological health and emotional health of each user in each virtual scene are obtained as the health state feature vector. The acquisition of the fused representation vector: Through the previous feature extraction and fusion process, the fused representation vector of each user in different virtual scenes is obtained. The fused representation vector contains a comprehensive representation of the physiological characteristics and emotional characteristics of the user in a specific virtual scene. The acquisition of user individual attribute data: Collect the user's individual attribute data, such as age, gender, height, weight, previous medical history, etc. The user's individual attribute data is preprocessed, such as numericalization and normalization, so as to be combined with the fused representation vector. The design of a multi-layer fully connected neural network: Construct a multi-layer fully connected neural network to map the fused representation vector and the user's individual attribute data to the health state feature vector. The input of the neural network is the concatenated vector of the fused representation vector and the user's individual attribute data. The output of the neural network is a health state feature vector, which contains two dimensions: physiological health and emotional health. The number of hidden layers of the neural network and the number of neurons in each layer can be adjusted according to the data scale and complexity. Data input and forward propagation: The concatenated vector of the fusion representation vector and the user's individual attribute data is input into the multi-layer fully connected neural network. Through forward propagation calculation, the input data is passed to each hidden layer of the neural network layer by layer, and the activation function is applied for nonlinear transformation. In the output layer of the neural network, the predicted values of physical health and emotional health are obtained. Acquisition of health status feature vector: The predicted values of physical health and emotional health output by the neural network are combined into a health status feature vector. The health status feature vector represents the overall health status of each user in each virtual scene, including both physical health and emotional health.
[0079] Figure 5According to the exemplary flow chart of obtaining the scene difficulty coefficient shown in some embodiments of the present application, the virtual teaching scene metadata is input into the pre-trained gradient boosting decision tree model to generate the scene difficulty coefficient, and the scene difficulty classification data set is constructed: the knowledge point coverage and user completion time are extracted from the virtual teaching scene metadata as input features. The knowledge point coverage rate indicates the proportion of the number of knowledge points covered in the scene to the total number of knowledge points, reflecting the knowledge breadth of the scene. The user completion time indicates the average time for the user to complete the task or learning content in the scene, reflecting the difficulty and complexity of the scene. The scene difficulty is manually annotated, and the difficulty is divided into multiple levels, such as simple, medium, difficult, etc., as training labels. The extracted input features and the annotated difficulty levels are combined into a scene difficulty classification data set. Data set division: The scene difficulty classification data set is randomly divided into a training set and a test set. The training set is used to train the gradient boosting decision tree model, and the test set is used to evaluate the performance of the model. The division ratio can be adjusted according to the amount of data and the complexity of the model. Usually, the training set accounts for a larger proportion, such as 70% or 80%. Training of the gradient boosting decision tree model: The gradient boosting decision tree model is trained using the training set. Gradient boosted decision tree is an ensemble learning algorithm that improves prediction performance by iteratively building multiple decision trees and weightedly combining their predictions. During the training process, the model parameters are optimized by minimizing the loss function, such as logarithmic loss or exponential loss. The hyperparameters of the model are tuned using the grid search method, such as the number of decision trees, maximum tree depth, and learning rate. The optimal hyperparameter combination is selected through cross-validation or validation set to obtain the best performing model. Performance evaluation of the model: The performance of the trained gradient boosted decision tree model is evaluated using the test set. Evaluation indicators such as accuracy, precision, recall, F1 score, etc. are used to measure the prediction performance of the model on the test set. The confusion matrix of the model is analyzed to understand the classification of the model at different difficulty levels.
[0080] Based on the evaluation results, the model can be further optimized or the data set can be adjusted to improve the generalization ability of the model. Generation of scene difficulty coefficient: The metadata of the virtual teaching scene is input into the trained gradient boosting decision tree model. The model predicts the difficulty level of the scene based on the input knowledge point coverage and the user completion time. The predicted difficulty level is converted into a scene difficulty coefficient, for example, simple corresponds to 0.2, medium corresponds to 0.5, difficult corresponds to 0.8, etc. Output the scene difficulty coefficient to indicate the difficulty level of the virtual teaching scene. Application of scene difficulty coefficient: The generated scene difficulty coefficient is combined with other features for subsequent tasks such as recommendation systems, learning path planning, etc. According to the scene difficulty coefficient, virtual teaching scenes of appropriate difficulty can be recommended to users to achieve personalized learning. In learning path planning, the learning order of the scenes can be reasonably arranged according to the scene difficulty coefficient, and the learning difficulty can be gradually increased.
[0081] Figure 6 This is an exemplary flowchart for obtaining a comprehensive grade score of a virtual teaching scene according to some embodiments of the present application, obtaining the physiological health and emotional health of each virtual teaching scene, and constructing a virtual teaching scene grade evaluation matrix to generate a comprehensive grade score, and obtaining the health impact of the virtual teaching scene: for each virtual teaching scene, obtain the physiological health and emotional health data of all users. Calculate the average physiological health and average emotional health of each scene as the health impact of the scene. The health impact reflects the overall impact of the scene on the physiological and emotional health of the user. Construct a virtual teaching scene triple: For each virtual teaching scene, construct a triple (scene ID, health impact, difficulty coefficient). The scene ID is the unique identifier of the scene, which is used to distinguish different scenes. The health impact contains two elements, average physiological health and average emotional health, which represent the impact of the scene on the health of the user. The difficulty coefficient is generated by a pre-trained gradient boosting decision tree model, which represents the difficulty level of the scene.
[0082] Generate a virtual teaching scene grade evaluation matrix: Sort all the triplets of virtual teaching scenes according to the scene ID. Form a matrix with N rows and 3 columns as the virtual teaching scene grade evaluation matrix. The first column of the matrix is the scene ID, the second column is the health impact, and the third column is the difficulty coefficient. N is the total number of virtual teaching scenes obtained. Data normalization: Perform data normalization on the virtual teaching scene grade evaluation matrix. Perform Min-Max normalization on the scene dimension for each element in the health impact (average physiological health and average emotional health). Min-Max normalization maps the data to the [0, 1] interval to make the health impact of different scenes comparable. The normalization formula is: (x-min) / (max-min), where x is the original value, min and max are the minimum and maximum values of the element in all scenes, respectively. After normalization, a standardized health impact vector is obtained. Calculate the initial comprehensive grade score: Perform weighted summation on the standardized health impact vector and the difficulty coefficient to generate the initial comprehensive grade score. The weighted summation formula is: Comprehensive level score = w1*standardized physical health + w2*standardized emotional health + w3*difficulty coefficient. w1, w2, and w3 are weight coefficients, which are set according to actual needs and expert experience to reflect the importance of different factors.
[0083] Construct an extended evaluation matrix: Add the initial comprehensive grade score to the last column of the virtual teaching scenario grade evaluation matrix. Get an extended evaluation matrix, which contains the scenario ID, health impact, difficulty coefficient, and comprehensive grade score. Evaluate and sort virtual teaching scenarios: According to the extended evaluation matrix, evaluate and sort the comprehensive impact of different virtual teaching scenarios. The higher the comprehensive grade score, the better the comprehensive impact of the scenario, that is, it performs better in terms of health impact and difficulty. Sort the virtual teaching scenarios in descending order according to the comprehensive grade score to get the comprehensive grade score of the scenario. Output the comprehensive grade score results of the sorted virtual teaching scenarios.
[0084] When receiving a teaching request from a target user, the matching degree between the target user and each virtual teaching scenario is calculated through a collaborative filtering algorithm based on the target user's historical teaching data. Combined with the comprehensive grade score of the virtual teaching scenario, a personalized recommendation priority ranking is generated, and the top N virtual teaching scenarios are selected and recommended to the target user. Receive a teaching request from a target user: When the system receives a teaching request initiated by a target user, the personalized recommendation process is triggered. The teaching request contains the user's identification information, such as the user ID, to facilitate the acquisition of the user's historical teaching data. Obtain the target user's historical teaching data: According to the target user's identification information, the user's historical teaching data is obtained from the teaching database. The historical teaching data includes information such as the virtual teaching scenarios that the user has learned in the past, the completion status, and the length of study. These data reflect the user's learning preferences, ability level, and interests.
[0085] Calculate the matching degree between the target user and the virtual teaching scene: Use the collaborative filtering algorithm to calculate the matching degree between the target user and each virtual teaching scene. The collaborative filtering algorithm is based on the similarity between users. By analyzing the user's historical behavior data, it predicts the user's interest in scenes that have not been learned. Common collaborative filtering algorithms include user-based collaborative filtering and item-based collaborative filtering. User-based collaborative filtering calculates the similarity between users to find users similar to the target user and recommends scenes based on the behavior of similar users. Item-based collaborative filtering calculates the similarity between scenes to find scenes similar to the scenes that the target user has learned and recommends them. Calculate the matching score between the target user and each virtual teaching scene.
[0086] Combined with the comprehensive grade score of the virtual teaching scene: Combine the match score between the target user and the virtual teaching scene with the comprehensive grade score of the virtual teaching scene. The comprehensive grade score of the virtual teaching scene is a comprehensive score obtained by comprehensively considering factors such as the health impact and difficulty of the scene. The match score and the comprehensive grade score are weighted and summed to obtain the final recommendation score. The weighted summation formula is: Recommendation score = w1*match score + w2*comprehensive grade score. w1 and w2 are weight coefficients, which are set according to actual needs and expert experience to reflect the importance of different factors.
[0087] Generate personalized recommendation priority ranking: Sort all virtual teaching scenarios in descending order according to the recommendation score. The higher the recommendation score, the higher the match between the scenario and the target user, and the better the overall impact. After sorting, a personalized recommendation priority list is obtained. Select the top N virtual teaching scenarios for recommendation: According to the recommendation priority list, select the top N virtual teaching scenarios as the recommendation results. The value of N can be set according to actual needs and user acceptance, and 3-5 scenarios are usually selected. Recommend the selected N virtual teaching scenarios to the target user to provide personalized learning suggestions.
[0088] Presentation and feedback of recommendation results: present the recommended virtual teaching scenarios to target users in a suitable form, such as lists, cards, etc. Provide brief information about the scenarios, such as scenario name, difficulty level, learning objectives, etc., to help users quickly understand the characteristics of the scenarios. Collect user feedback on recommendation results, such as clicks, learning completion status, etc., to optimize recommendation algorithms and models.
[0089] Example 2
[0090] There is a virtual teaching system that contains the following data: Historical teaching data: There are 10,000 teaching records, each of which contains a user's physiological characteristic data (such as heart rate, EEG, etc.) and emotional characteristic data (such as emotional state, stress level, etc.) in a virtual teaching scene, as well as the user's individual attributes (age, gender). Virtual teaching scene metadata: There are 100 virtual teaching scenes, and the metadata of each scene includes the knowledge point coverage and the average user completion time, as well as the pre-annotated scene difficulty level.
[0091] The physiological characteristics and emotional characteristics in the historical teaching data are preprocessed, including data cleaning, outlier processing, missing value filling and normalization. The preprocessed physiological characteristic data and emotional characteristic data are input into the feature fusion model: First layer: GRU is used to extract time series features, and the attention mechanism is set to adjust the weights of different time steps to obtain physiological feature representation (dimension is 128) and emotional feature representation (dimension is 128). Second layer: A 4-head attention mechanism is used to calculate the correlation of feature representations in different subspaces to obtain semantically aligned feature representations (dimension is 128). Third layer: A 3-layer MLP (hidden layer dimension is 256) is used to extract high-level semantics, and a residual connection is made with the output of the second layer to obtain a fused feature representation (dimension is 128). Fourth layer: Self-attention kernels of 3 scales (2, 4, 8) are used to perform global aggregation of fused features at different scales to obtain a hierarchical fused representation vector (dimension is 128).
[0092] Loss function: weighted sum of reconstruction loss (maximum average difference), contrast loss (InfoNCE) and classification loss (weighted cross entropy), with a weight ratio of 1:1:1, and trained for 50 epochs using the Adam optimizer. The fusion representation vector is concatenated with the user's individual attribute data and input into a 3-layer fully connected network (hidden layer dimension is 64, using ReLU activation) to predict physiological health and emotional health (values between 0-1) as health status feature vectors.
[0093] Input the metadata of the virtual teaching scene into the pre-trained GBDT model to generate the scene difficulty coefficient (a value between 0 and 1): extract the knowledge point coverage and average completion time of the scene as features, and the difficulty level as a label to construct a training set (80%) and a test set (20%). Through grid search tuning, the accuracy of the GBDT model with 100 trees, a maximum depth of 6, and a learning rate of 0.1 on the test set reached 95%. Use the trained GBDT model to predict the difficulty of each scene's metadata and obtain the scene difficulty coefficient.
[0094] Generate a comprehensive grade score for virtual teaching scenarios: Calculate the average physical health and average emotional health of all users in each scenario. Construct a triple (scenario ID, health impact, difficulty coefficient) to form a 100*3 scenario grade evaluation matrix. Perform Min-Max normalization on the health impact, and use the weighted sum of the health impact and the difficulty coefficient as the initial comprehensive score (health impact weight 0.7, difficulty coefficient weight 0.3). Add the comprehensive score to the evaluation matrix and sort it to get the comprehensive grade score of the scenario.
[0095] Personalized scenario recommendation: Receive the target user's teaching request and extract the most recent 10 historical teaching records. Use the collaborative filtering algorithm to calculate the similarity between the target user and other users, and predict the target user's matching degree for each scenario based on the scenario learning records of similar users. Add the weighted sum of the scenario matching degree and the scenario comprehensive grade score (weight ratio 1:1) to obtain the final recommendation priority. Select the top-5 scenarios with the highest recommendation priority as the teaching scenarios recommended to the target user.
Claims
1. An intelligent teaching management method based on virtual reality, characterized in that: S1, obtaining historical teaching data and virtual teaching scene metadata; wherein the historical teaching data includes physiological characteristic data and emotional characteristic data of multiple users in multiple virtual teaching scenes; the virtual teaching scene metadata includes the teaching scene knowledge point coverage rate and user completion time data; S2, preprocessing the acquired physiological characteristic data and emotional characteristic data; S3, inputting the preprocessed physiological feature data and emotional feature data into a feature fusion model, and obtaining a fusion representation vector containing physiological information and emotional information through the feature fusion model, wherein the feature fusion model adopts a multi-level structure design; S4, taking the obtained fusion representation vector as input, through a multi-layer fully connected neural network, combined with the user's individual attribute data, to obtain the physiological health and emotional health of each user in each virtual scene as a health status feature vector; wherein the user's individual attributes include age and gender; S5, taking the obtained virtual teaching scene metadata as input, and using the pre-trained gradient boosting decision tree model to generate the scene difficulty coefficient; S6, according to the health status feature vectors of all users in each virtual scene and the corresponding scene difficulty coefficient, the average physiological health and average emotional health of each virtual teaching scene are calculated by weighted average algorithm, and the grade evaluation matrix of the virtual teaching scene is constructed in combination with the scene difficulty coefficient, and the comprehensive grade score of the virtual teaching scene is generated by matrix transformation; S7, when receiving a teaching request from a target user, the matching degree between the target user and each virtual teaching scene is calculated by a collaborative filtering algorithm according to the historical teaching data of the target user, and a recommendation priority ranking is generated by combining the comprehensive grade scores of the virtual teaching scenes, and the top N virtual teaching scenes are selected and recommended to the target user; S3, the feature fusion model adopts a multi-level structure design, including: The first layer uses a gated recurrent unit to extract the temporal dependency between physiological feature data and emotional feature data. It adjusts the feature weights of different time steps by setting an attention mechanism, and performs a weighted summation of past hidden states based on the current hidden state and attention weight to obtain physiological feature representation and emotional feature representation in the time dimension. The second layer uses a multi-head attention mechanism to calculate the correlation between physiological feature representation and emotional feature representation in different feature subspaces, and obtains semantically aligned feature representation through cross-modal attention weighting; The third layer uses a multi-layer perceptron to perform nonlinear transformation on the semantically aligned feature representation output by the second layer, extracts high-level semantic information through multiple fully connected layers and nonlinear activation functions, and then adds the output of the third layer to the output of the second layer element by element through residual connection to obtain a fused feature representation containing high-level semantic information and low-level feature information; among them, the high-level semantic information represents the abstract feature pattern extracted after the cross-modal fusion feature representation is transformed by the multi-layer perceptron, reflecting the interactive relationship between physiological features and emotional features at a higher speech level; the low-level feature information represents the shallow feature representation retained in the cross-modal fusion feature representation, reflecting the local correlation between physiological features and emotional features at the original data level; The fourth layer uses a multi-scale self-attention mechanism to perform global information aggregation on the fused features output by the third layer. The dependency between different regions of the fused feature representation is calculated through multiple attention kernel functions of different sizes to obtain attention weights of different scales. The fused feature representation is weighted pooled according to the attention weights to generate a hierarchical fused representation vector output. The parameters of the feature fusion model are optimized through a joint loss function, which includes a reconstruction loss function, a contrast loss function, and a classification loss function. The reconstruction loss function adopts the framework of a variational autoencoder, the contrast loss function adopts the paradigm of contrastive learning, and the classification loss function adopts a cross entropy function. During the model training process, the reconstruction loss function, contrast loss function and classification loss function are weighted and summed to form a joint loss function through multi-task learning. The reconstruction loss function adopts the framework of variational autoencoder, including: The original physiological feature data is mapped to the physiological latent space through the encoder to obtain the physiological latent variables; wherein the physiological latent space is used to represent the compressed representation of the original physiological feature data; The original emotion feature data is mapped to the emotion latent space through the encoder to obtain the emotion latent variable; wherein the emotion latent space is used to represent the compressed representation of the original emotion feature data; Reconstruct the physiological latent variables into physiological feature data through the decoder; Reconstruct the emotional latent variables into emotional feature data through the decoder; Based on the reconstructed physiological feature data and the original physiological feature data, the physiological feature reconstruction loss is calculated by using the maximum average difference loss function; Based on the reconstructed emotion feature data and the original emotion feature data, the emotion feature reconstruction loss is calculated by the maximum average difference loss function; Add the physiological feature reconstruction loss and the emotional feature reconstruction loss to obtain the reconstruction loss function; The calculation formula of the maximum average difference loss function is as follows: ; in, represents the maximum average difference loss; and Respectively represent the i-th physiological characteristic data and emotional feature data Eigenmapping function mapped to the reproducing kernel Hilbert space RKHS; represents the RKHS space; and Respectively represent the original data of the ith physiological feature and the original data of the emotional feature; and Represents the number of samples of physiological characteristic data and emotional characteristic data respectively; and Respectively represent the i-th reconstructed physiological feature original data and emotional feature original data; represents the squared norm in the RKHS space; The contrastive loss function adopts the paradigm of contrastive learning, including: The hierarchical fusion representation vector output by the fourth layer is used as the anchor vector; Based on the anchor vector, positive sample vectors are obtained by sampling the physiological feature data and emotional feature data of the same user, and negative sample vectors are obtained by sampling the physiological feature data and emotional feature data of different users; Calculate the similarity between the anchor vector and the positive sample vector, and calculate the similarity between the anchor vector and the negative sample vector; Based on the similarity between the anchor vector and the positive sample vector, the similarity between the anchor vector and the negative sample vector is calculated, and the contrast loss function is calculated through Info NCE Loss; The contrast loss function expression is: ; Where V represents the number of perspectives; N represents the total number of samples used for contrastive learning; represents the temperature coefficient; and Respectively represent the number of negative samples of physiological features and the number of negative samples of emotional features; and They represent the physiological feature representation and emotional feature representation of the i-th sample at the v-th perspective respectively; and They represent the positive sample physiological feature representation and positive sample emotional feature representation of the i-th sample at the v-th perspective respectively; and They represent the jth negative sample physiological feature representation and negative sample emotional feature representation of the i-th sample at the v-th perspective respectively; The classification loss function uses the cross entropy function, including: The hierarchical fusion representation vector output by the fourth layer is input into the classifier, and the fusion feature vector is subjected to feature transformation and nonlinear mapping by the classifier to obtain a prediction probability vector; The predicted probability vector is normalized by the softmax function to obtain a normalized predicted probability vector; Compare the normalized predicted probability vector with the annotated true label element by element, and calculate the classification loss function based on the cross entropy function to measure the difference between the predicted probability and the true label; The cross entropy function calculation formula is as follows: ; in, Indicates the predicted probability that the i-th sample belongs to the c-th class; The cth element of the one-hot encoding vector representing the true label of the i-th sample; represents the category measurement factor; Represents category-specific focusing parameters, and different categories set different focusing parameters; Represents the sample weight factor, which is used to adjust the contribution of different samples to the loss function; ; in, represents the number of samples in the cth category, and N represents the total number of samples; ; in, represents the predicted probability of the i-th sample in the t-1th iteration.
2. The intelligent teaching management method based on virtual reality according to claim 1, characterized in that: S5, taking the obtained virtual teaching scene metadata as input, using the pre-trained gradient boosting decision tree model to generate the scene difficulty coefficient, including: Extract knowledge point coverage and user completion time from the metadata of virtual teaching scenes as input features, annotate scene difficulty levels as training labels, and construct a scene difficulty classification dataset; The scene difficulty classification dataset is randomly divided into a training set and a test set. The gradient boosting decision tree model is trained using the training set, and the model hyperparameters are adjusted using the grid search method. The hyperparameters include the number of decision trees, the maximum tree depth, and the learning rate. Use the test set to evaluate the performance of the trained gradient boosting decision tree model; The trained gradient boosting decision tree model is used to predict the difficulty of the input virtual scene metadata and output the scene difficulty coefficient.
3. The intelligent teaching management method based on virtual reality according to claim 2 is characterized in that: S6, generating a comprehensive grade score of the virtual teaching scene, including: Obtaining physical and emotional health for each virtual teaching scenario; For each virtual teaching scenario, the average physical health and average emotional health of all users were calculated; According to the average physical health and average emotional health of each virtual teaching scene, as well as the corresponding scene difficulty coefficient, a triplet (scene ID, health impact, difficulty coefficient) is constructed; among them, the health impact includes two elements: average physical health and average emotional health; Sort all the triplets of virtual teaching scenes according to the scene ID to form a matrix of N rows and 3 columns as the virtual teaching scene level evaluation matrix; where N is the total number of virtual teaching scenes obtained, the first column is the scene ID, the second column is the health impact, and the third column is the difficulty coefficient; The virtual teaching scene level evaluation matrix is normalized, and each element in the health impact quantity is normalized in the scene dimension to obtain a standardized health impact vector. The standardized health impact vector and difficulty coefficient are weighted and summed to generate an initial comprehensive grade score; Add the initial comprehensive grade score to the last column of the virtual teaching scenario grade evaluation matrix to obtain an extended evaluation matrix; According to the extended evaluation matrix, the comprehensive impact of different virtual teaching scenarios is evaluated and ranked, and the comprehensive grade score of the virtual teaching scenarios is output.
