Robot online teaching method and system based on artificial intelligence

Abstract

The invention relates to the technical field of online teaching, and discloses a robot online teaching method and system based on artificial intelligence, and the method comprises the steps: collecting the learning historical data of students, carrying out the processing and analysis of the learning historical data through employing a K-means clustering algorithm and a decision tree algorithm, building a course recommendation model based on the fusion of LSTM and an attention mechanism, and achieving the online teaching of a robot. Recommending personalized course contents to the students through a course recommendation model; performing real-time question answering on the student through a natural language processing technology, a knowledge graph and a BERT model in the process that the student learns the personalized course content; evaluating the real-time learning behavior data by adopting a multi-modal model; and adjusting the personalized course content based on an evaluation result output by the multi-modal model. The learning behavior mode and the knowledge mastering condition of students can be deeply known, and an educator can be helped to better know the characteristics and individual differences of student groups and make a more targeted teaching strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of online teaching, and in particular to an artificial intelligence-based robot online teaching method and system. Background Art

[0002] With the rapid development of Internet technology, the online education market has flourished, but traditional online teaching methods have significant defects. The teaching content is usually stereotyped and cannot be accurately adapted to the individual differences of students, and it is difficult to meet the diverse learning needs; the interactive links in the teaching process are inefficient, and it is difficult for teachers to timely understand the learning difficulties of students, and students' learning status and emotional problems cannot be discovered in time, and questions raised by students cannot be answered quickly and effectively; at the same time, the teaching effect evaluation method is single and cannot fully reflect the learning results and progress of students. Summary of the invention

[0003] The purpose of the present invention is to solve the above problems and to design a robot online teaching method and system based on artificial intelligence.

[0004] A first aspect of the present invention provides an artificial intelligence-based robot online teaching method, the artificial intelligence-based robot online teaching method comprising the following steps: Sending a data collection instruction to the teaching robot, the teaching robot collects the student's learning history data after receiving the data collection instruction, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; A course recommendation model is constructed based on the fusion of LSTM and attention mechanism, and personalized course content is recommended to students through the course recommendation model according to the learning needs analysis results; While students are learning personalized course content, they can answer questions in real time through natural language processing technology, knowledge graphs, and BERT models; Acquire students' real-time learning behavior data, and evaluate the real-time learning behavior data using a multimodal model; Based on the evaluation results output by the multimodal model, the Kuhn-Munkres algorithm is used to adjust the personalized course content so that the teaching robot can complete the online teaching task for students.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the data collection instruction is sent to the teaching robot, and the teaching robot collects the student's learning history data after receiving the data collection instruction, and uses a K-means clustering algorithm and a decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results, including: Collect various learning history data of students on the online learning platform, wherein the learning history data at least includes course selection records, learning progress data, homework and test scores, and interactive behavior data; Performing data cleaning and data normalization processing on the learning history data to obtain preprocessed learning history data; Extracting candidate features from the preprocessed learning history data, and using the candidate features as data points input into the K-means clustering algorithm; Use K-means clustering algorithm to cluster the candidate features of students, assign data points to the group with the closest cluster center, then recalculate the center of each group, and iterate until the cluster center no longer changes or reaches the preset number of iterations, and output the results of K-means clustering algorithm, where the preset number of iterations is 100; The clustering results based on the K-means clustering algorithm use the decision tree algorithm to classify and predict the students' learning needs and obtain the results of the students' learning needs analysis.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the clustering results based on the K-means clustering algorithm use a decision tree algorithm to classify and predict the students' learning needs to obtain the students' learning needs analysis results, including: Build a decision tree: ; In the formula, represents a decision tree, Represents input data, Indicates that the decision tree parameters are determined by the leaf node area With number Calculated, Indicated by and The collection composed of Indicates the division area. Represents the decision value of the partition area, and the final decision tree is obtained after several iterations; Split by calculating information gain: ; In the formula, represents the improved information gain, Representation dataset The original entropy of Representation characteristics The set of all possible values ​​of Indicates when the feature The value is The data subset at time Representation Subset The number of samples, Representation dataset The total number of samples, Representation Subset The entropy of represents the number of features, Indicates The important weight coefficient of the feature, is an indicator function, if the feature With features In the value The following is relevant, =1, otherwise 0; The clustering results obtained by the K-means clustering algorithm are input into the constructed decision tree. Starting from the root node, it is traversed downward step by step according to the splitting conditions of each node until it reaches the leaf node to obtain the classification results of students' learning needs.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the course recommendation model is constructed based on the fusion of LSTM and attention mechanism, and personalized course content is recommended to students through the course recommendation model according to the learning needs analysis result, including: Obtain student demand data according to the learning demand analysis results, input the sequence information of the student demand data into the course recommendation model, and capture the time series characteristics of the changes in student demand at different learning stages through the forget gate, input gate and output gate of the LSTM layer; According to the hidden state of the LSTM layer output containing the characteristics of the student demand sequence , the score of each time step is calculated through the attention layer : ; In the formula, represents the weight matrix of the attention layer, represents the hyperbolic tangent function; Use the softmax function to score all time steps Normalize and get the attention weight: ; In the formula, represents the total number of time steps, Indicates The score of each task; The hidden state at each time step According to the attention weight Weighted to get the final context vector : ; The context vector obtained is fused with the course content feature vector through the fully connected layer, and the fused vector is input into the linear regression layer to output a recommendation score. The courses are ranked and recommended according to the scores to obtain personalized course content recommended to students.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, in the process of students learning personalized course content, providing real-time questions and answers for students through natural language processing technology, knowledge graph and BERT model, including: When students ask questions during their personalized course learning, natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic analysis of the questions to obtain the analysis results. The CRF conditional random field model is used for named entity recognition to extract the key concepts in the question. Let the text sequence of the question be X and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ; In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ; ; In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; Associating the extracted key concepts with a pre-built knowledge graph, and extracting information related to the question based on the associated knowledge graph nodes and edges, wherein the knowledge graph includes various knowledge points in the subject area and the relationships between them, and the key concepts include at least the subject name and knowledge points; The question analysis results obtained by natural language processing are integrated with the information in the knowledge graph, and the integrated text is input into the BERT model. The BERT model is used to encode and decode the input text to generate answer text for students' questions, so as to provide real-time answers to students' questions.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, obtaining the real-time learning behavior data of the students and evaluating the real-time learning behavior data using a multimodal model includes: Acquire text data of students during the learning process, and collect image data of students during the learning process through a camera system to obtain real-time learning behavior data of students, wherein the text data at least includes question content, discussion speeches, and homework answers, and the image data at least includes students' facial expressions, postures, learning environment, and pictures of textbooks and notes that students are viewing; Preprocessing the text data and image data in the real-time learning behavior data, and using the preprocessed real-time learning behavior data as input data of the multimodal model; The image features are detected by YOLOv5 in the multimodal model, and the attention mechanism is introduced into the feature extraction network of YOLOv5. The GIoU loss function is adopted, where the attention mechanism includes channel attention and spatial attention. The text features are extracted through the Transformer in the multimodal model. Sine position encoding is used in the Transformer to add position information to the text features. The number of heads in the multi-head attention mechanism is increased to capture more different text feature subspaces. The attention weights of different heads are dynamically adjusted according to the learning behavior data. The number of heads in the multi-head attention mechanism is 8. The image features extracted by YOLOv5 and the text features extracted by Transformer are fused, and the fusion weights are dynamically adjusted using the attention mechanism; The fully connected layer and softmax function of the multimodal model are used to classify and evaluate the fused features to determine the students’ learning behavior status and obtain the evaluation results.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, the evaluation result based on the multimodal model output uses a Kuhn-Munkres algorithm to adjust the personalized course content, including: Based on the evaluation results output by the multimodal model, the student's learning status set and the course content feature set of personalized course content are obtained to construct the student's learning status Course content characteristics The correlation matrix ; Use the Kuhn-Munkres algorithm to match so that the sum of the matched elements in the incidence matrix is ​​maximized and the optimal matching solution is obtained. The objective function is: ; In the formula, Indicates learning status Course content characteristics The degree of correlation, is a decision variable, representing the learning state Is it related to the course content characteristics? Match, if match then If it does not match, ; The constraints of the Kuhn-Munkres algorithm are: For each learning state , can only match one course content feature at most, i.e. ; For each course content feature , can only match one learning state at most, i.e. ; According to the optimal matching scheme, the course content characteristics corresponding to each student's learning status are determined to adjust the personalized course content, continuously collect students' learning behavior data, re-evaluate students' learning status, and use the Kuhn-Munkres algorithm to adjust the course content.

[0011] The second aspect of the present invention provides an artificial intelligence-based robot online teaching system, which includes a data analysis module, a course recommendation module, a real-time question-answering module, a learning evaluation module and a course adjustment module, wherein: The data analysis module is used to send data collection instructions to the teaching robot, and the teaching robot collects the student's learning history data after receiving the data collection instructions, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; A course recommendation module is used to build a course recommendation model based on the fusion of LSTM and attention mechanism, and recommend personalized course content to students through the course recommendation model according to the learning needs analysis results; The real-time question-answering module is used to answer students' questions in real time through natural language processing technology, knowledge graphs and BERT models while students are learning personalized course content; A learning evaluation module is used to obtain students' real-time learning behavior data and evaluate the real-time learning behavior data using a multimodal model; The course adjustment module is used to adjust the personalized course content based on the evaluation results output by the multimodal model using the Kuhn-Munkres algorithm, so as to complete the online teaching tasks for students through the teaching robot.

[0012] Optionally, in a second implementation of the second aspect of the present invention, the course recommendation module includes an information input submodule, a calculation submodule, a normalization processing submodule, a weighting submodule and a fusion submodule, wherein: An information input submodule, used to obtain student demand data according to the learning demand analysis results, input the sequence information of the student demand data into the course recommendation model, and capture the time series characteristics of the changes in student demand at different learning stages through the forget gate, input gate and output gate of the LSTM layer; The calculation submodule is used to calculate the hidden state containing the student demand sequence characteristics output by the LSTM layer , the score of each time step is calculated through the attention layer : ; In the formula, represents the weight matrix of the attention layer, represents the hyperbolic tangent function; Normalization submodule, used to use the softmax function to score all time steps Normalize and get the attention weight: ; In the formula, represents the total number of time steps, Indicates The score of each task; The weighted submodule is used to transform the hidden state of each time step According to the attention weight Weighted to get the final context vector : ; The fusion submodule is used to fuse the obtained context vector with the course content feature vector through the fully connected layer, input the fused vector into the linear regression layer, output a recommendation score, and rank and recommend courses according to the scores to obtain personalized course content recommended to students.

[0013] Optionally, in a third implementation of the second aspect of the present invention, the real-time question answering module includes a question analysis submodule, an extraction submodule, an association submodule, and an encoding and decoding submodule, wherein: The question analysis submodule is used to use natural language processing technology to perform lexical analysis, syntactic analysis and semantic analysis on questions raised by students in the process of learning personalized course content, and obtain the question analysis results; The extraction submodule is used to recognize named entities using the CRF conditional random field model to extract the key concepts in the question. Let the text sequence of the question be X and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ; In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ; ; In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; An association submodule is used to associate the extracted key concepts with a pre-built knowledge graph, and extract information related to the question based on the associated knowledge graph nodes and edges, wherein the knowledge graph includes various knowledge points in the subject area and the relationships between them, and the key concepts include at least the subject name and knowledge points; The encoding and decoding submodule is used to fuse the question analysis results obtained by natural language processing with the information in the knowledge graph. The fused text is input into the BERT model, and the BERT model is used to encode and decode the input text to generate answer text for questions raised by students, so as to provide real-time answers to students' questions.

[0014] In the technical solution provided by the present invention, by sending a data collection instruction to the teaching robot, the teaching robot collects the student's learning history data after receiving the data collection instruction, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; based on the fusion of LSTM and attention mechanism, a course recommendation model is constructed, and according to the learning needs analysis results, personalized course content is recommended to students through the course recommendation model; in the process of students learning personalized course content, real-time questions are answered for students through natural language processing technology, knowledge graph and BERT model; students' real-time learning behavior data are obtained, and the real-time learning behavior data is evaluated by a multimodal model; based on the evaluation results output by the multimodal model, the personalized course content is adjusted by the Kuhn-Munkres algorithm, so as to complete the online teaching task for students through the teaching robot. 1. It can deeply understand the students' learning behavior patterns and knowledge mastery, obtain accurate learning needs analysis results, provide data support for subsequent personalized education services, help educators better understand the characteristics and individual differences of student groups, and formulate more targeted teaching strategies. 2. Based on the results of the students' learning needs analysis, it can accurately recommend personalized course content that meets their learning progress, interests and abilities for each student, improve students' acceptance of the course and learning enthusiasm, avoid students getting lost in a large number of irrelevant courses, improve learning efficiency and learning effects, and promote students' personalized development. 3. It can promptly solve the problems encountered by students in learning, reduce learning obstacles, maintain the continuity and fluency of learning, enhance students' learning confidence and autonomous learning ability, and improve learning satisfaction. 4. It can comprehensively and objectively reflect the students' learning status, including information such as concentration and learning emotions, and provide educators with richer and more detailed feedback, so as to timely discover possible problems in the students' learning process and take corresponding intervention measures. 5. It can dynamically optimize the course content according to the students' real-time learning situation, ensure that the course is always closely matched with the students' learning needs, further improve the quality and effect of education and teaching, realize the efficient use of educational resources, and promote the intelligent and personalized development of education. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the following detailed description of the preferred embodiment.The drawings are only for the purpose of illustrating the preferred embodiments and are not to be construed as limiting the invention.

[0016] Figure 1 A schematic diagram of a first embodiment of an artificial intelligence-based robot online teaching method provided by an embodiment of the present invention; Figure 2A schematic diagram of a second embodiment of an artificial intelligence-based robot online teaching method provided by an embodiment of the present invention; Figure 3 A schematic diagram of a third embodiment of an artificial intelligence-based robot online teaching method provided by an embodiment of the present invention; Figure 4 A schematic diagram of the structure of an artificial intelligence-based robot online teaching system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or system that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or systems.

[0018] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 A schematic diagram of a first embodiment of an artificial intelligence-based robot online teaching method provided by an embodiment of the present invention, the method specifically comprises the following steps: Step 101: Send a data collection instruction to the teaching robot. After receiving the data collection instruction, the teaching robot collects the student's learning history data, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; Specifically, in this embodiment, various learning history data of students on the online learning platform are collected, wherein the learning history data at least includes course selection records, learning progress data, homework and test scores, and interactive behavior data; the robots at least include head-mounted robots, humanoid robots, and hand-mounted robots, etc. These robots integrate multiple sensors such as image sensors, temperature and humidity sensors, and radar sensors; the robots can automatically obtain students' historical data from the network, and AI computing chips are implanted in the robots for analyzing and processing the data; Perform data cleaning and data normalization on the learning history data to obtain the preprocessed learning history data; Extract candidate features from the preprocessed learning history data and use the candidate features as data points for input into the K-means clustering algorithm; Use K-means clustering algorithm to cluster the candidate features of students, assign data points to the group with the closest cluster center, then recalculate the center of each group, and iterate until the cluster center no longer changes or reaches the preset number of iterations, and output the results of K-means clustering algorithm, where the preset number of iterations is 100; Specifically, determine the number of clusters: first determine how many groups to divide students into, which is the K value in the K-means algorithm. For example, based on teaching experience or educational goals, choose to divide students into 5 categories. Select the initial centroid: Before starting clustering, randomly select K data points from the student's feature data as the initial centroid. These centroids can be regarded as representative points of each cluster, and they will be continuously updated in subsequent calculations. Assign data points to clusters: Then, compare each student's feature data point with these K centroids. The comparison process is based on the similarity between the data point and the centroid, usually based on their similarity in various features. For example, if we mainly focus on students' grades and study time, then a student's grades and study time features are most similar to which centroid in these aspects, and the student is assigned to the corresponding cluster. Update the centroid: After assigning students to clusters, the features of all students in each cluster are recalculated to find the new centroid of each cluster. The new centroid reflects the average feature level of students in the cluster. For example, if there are several students in a cluster, their average grades, average study time, etc. are calculated to get the new centroid position. Iterative optimization: The above steps of allocating and updating the centroid are repeated until the student allocation in the cluster is basically stable, the centroid no longer changes significantly, or the predetermined number of iterations is reached. In this way, through multiple iterations, the grouping of students is continuously adjusted, and finally a stable clustering result is formed.

[0019] The clustering results based on the K-means clustering algorithm use the decision tree algorithm to classify and predict the students' learning needs and obtain the results of the students' learning needs analysis.

[0020] Specifically, in this embodiment, “using a decision tree algorithm to classify and predict the learning needs of students based on the clustering results of the K-means clustering algorithm to obtain the learning needs analysis results of students” also includes Build a decision tree: ; In the formula, represents a decision tree, Represents input data, Indicates that the decision tree parameters are determined by the leaf node area With number Calculated, Indicated by and The collection composed of Indicates the division area. Represents the decision value of the partition area, and the final decision tree is obtained after several iterations; Split by calculating information gain: ; In the formula, represents the improved information gain, Representation dataset The original entropy of Representation characteristics The set of all possible values ​​of Indicates when the feature The value is The data subset at time Representation Subset The number of samples, Representation dataset The total number of samples, Representation Subset The entropy of represents the number of features, Indicates The important weight coefficient of the feature, is an indicator function, if the feature With features In the value The following is relevant, =1, otherwise 0; The clustering results obtained by the K-means clustering algorithm are input into the constructed decision tree. Starting from the root node, it is traversed downward step by step according to the splitting conditions of each node until it reaches the leaf node to obtain the classification results of students' learning needs.

[0021] Step 102: Construct a course recommendation model based on the fusion of LSTM and attention mechanism, and recommend personalized course content to students through the course recommendation model according to the learning needs analysis results; Specifically, in this embodiment, student demand data is obtained according to the learning demand analysis results, and the sequence information of the student demand data is input into the course recommendation model, and the time series characteristics of the changes in student demand at different learning stages are captured through the forget gate, input gate and output gate of the LSTM layer; According to the hidden state of the LSTM layer output containing the characteristics of the student demand sequence , the score of each time step is calculated through the attention layer : ; In the formula, represents the weight matrix of the attention layer, represents the hyperbolic tangent function; Use the softmax function to score all time steps Normalize and get the attention weight: ; In the formula, represents the total number of time steps, Indicates The score of each task; The hidden state at each time step According to the attention weight Weighted to get the final context vector : ; The context vector obtained is fused with the course content feature vector through the fully connected layer, and the fused vector is input into the linear regression layer to output a recommendation score. The courses are ranked and recommended according to the scores to obtain personalized course content recommended to students.

[0022] Step 103: While students are learning personalized course content, real-time questions are answered for students through natural language processing technology, knowledge graphs, and BERT models; Specifically, in this embodiment, when students ask questions during the process of learning personalized course content, natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic analysis of the questions to obtain question analysis results; The CRF conditional random field model is used for named entity recognition to extract the key concepts in the question. Let the text sequence of the question be X and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ; In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ; ; In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; The extracted key concepts are associated with the pre-built knowledge graph, and information related to the question is extracted based on the associated knowledge graph nodes and edges, where the knowledge graph includes various knowledge points in the subject area and the relationships between them. The key concepts include at least the subject name and knowledge points. The question analysis results obtained by natural language processing are integrated with the information in the knowledge graph, and the integrated text is input into the BERT model. The BERT model is used to encode and decode the input text to generate answer text for students' questions, so as to provide real-time answers to students' questions.

[0023] Specifically, in this embodiment, the text preprocessing module in the natural language processing technology is used to clean and normalize the questions raised by students. This includes removing stop words, performing stem extraction or word form restoration, correcting spelling errors, and segmenting the text, so that the question text can enter the subsequent analysis process in a clearer and more accurate form. For example, for a question input by a student such as "I want to know the detailed explanation of Newton's laws of physics", after preprocessing, it becomes a keyword form such as "Detailed explanation of Newton's laws of physics" to facilitate subsequent rapid matching and analysis.

[0024] Next, the preprocessed questions are input into the semantic understanding module built on the BERT model. The BERT model can deeply parse and encode the semantics of the questions and convert them into high-dimensional semantic vector representations. For example, for two seemingly different but semantically similar questions, "What is the scope of application of Newton's laws" and "Under what circumstances can Newton's laws be used", the BERT model can accurately capture their semantic similarities and generate similar semantic vectors, thus providing a basis for subsequent unified processing.

[0025] At the same time, the knowledge graph organizes and stores various subject knowledge in the form of a graph, which contains a large number of entities (such as concepts, people, events, etc.) and the relationships between them (such as causal relationships, belonging relationships, logical relationships, etc.). When the student's question enters the system, the system will quickly match entities and relationships in the knowledge graph. Taking the student's question "Why does the apple fall from the tree" as an example, the system will find entities such as "apple", "tree", and "fall" in the knowledge graph, and locate the relevant knowledge nodes and explanation paths through the relationships such as "gravity" between them, providing accurate knowledge support for answering questions.

[0026] Then, the semantic vector generated by the BERT model is fused with the relevant knowledge retrieved from the knowledge graph. Through the weighted sum fusion algorithm, the information of the two is organically integrated, so that the answer is not only based on the structured knowledge in the knowledge graph, but also fully considers the semantic details and contextual information of the question. For example, for a question about a historical event, the BERT model may understand the key semantic elements such as time and people in the question, while the knowledge graph provides detailed background, causal relationships and other knowledge of the historical event. The fusion of the two can generate a more comprehensive, accurate and consistent answer that meets the student's questioning intention. Finally, the fused result is converted into natural language text form and presented to students as a real-time answer to questions.

[0027] Step 104: obtaining students' real-time learning behavior data, and evaluating the real-time learning behavior data using a multimodal model; Specifically, in this embodiment, text data of students in the learning process is obtained, and image data of students in the learning process is collected through a camera system to obtain real-time learning behavior data of students, wherein the text data at least includes question content, discussion speeches, and homework answers, and the image data at least includes students' facial expressions, postures, learning environment, and pictures of textbooks and notes that students are viewing; Preprocessing the text data and image data in the real-time learning behavior data, and using the preprocessed real-time learning behavior data as input data of the multimodal model; The image features are detected by YOLOv5 in the multimodal model, and the attention mechanism is introduced into the feature extraction network of YOLOv5. The GIoU loss function is adopted, where the attention mechanism includes channel attention and spatial attention. The text features are extracted through the Transformer in the multimodal model. Sine position encoding is used in the Transformer to add position information to the text features. The number of heads in the multi-head attention mechanism is increased to capture more different text feature subspaces. The attention weights of different heads are dynamically adjusted according to the learning behavior data. The number of heads in the multi-head attention mechanism is 8. The image features extracted by YOLOv5 and the text features extracted by Transformer are fused, and the fusion weights are dynamically adjusted using the attention mechanism; The fully connected layer and softmax function of the multimodal model are used to classify and evaluate the fused features to determine the students’ learning behavior status and obtain the evaluation results.

[0028] Step 105: Based on the evaluation results output by the multimodal model, the Kuhn-Munkres algorithm is used to adjust the personalized course content so as to complete the online teaching task for the students through the teaching robot.

[0029] Specifically, in this embodiment, based on the evaluation results output by the multimodal model, a set of student learning states and a set of course content features of personalized course content are obtained to construct a set of student learning states. Course content characteristics The correlation matrix ; Use the Kuhn-Munkres algorithm to match so that the sum of the matched elements in the incidence matrix is ​​maximized and the optimal matching solution is obtained. The objective function is: ; In the formula, Indicates learning status Course content characteristics The degree of correlation, is a decision variable, representing the learning state Is it related to the course content characteristics? Match, if match then If it does not match, ; The constraints of the Kuhn-Munkres algorithm are: For each learning state , can only match one course content feature at most, i.e. ; For each course content feature , can only match one learning state at most, i.e. ; According to the optimal matching scheme, the course content characteristics corresponding to each student's learning status are determined to adjust the personalized course content, continuously collect students' learning behavior data, re-evaluate students' learning status, and use the Kuhn-Munkres algorithm to adjust the course content.

[0030] Specifically, this implementation can also be implemented in the following ways: Integrate deep learning, machine vision, natural language processing, etc. into one robot platform. Adopt modular design to ensure that the various components of the robot platform can be modularly laid out for easy upgrading and maintenance.

[0031] Adopt a modular development approach, develop each component separately, and finally integrate it into a unified robot platform. Use deep learning framework and machine vision technology to develop visual recognition and tracking models. Develop voice recognition and natural language understanding models to achieve voice interaction. Integrate various sensors, actuators, controllers and other hardware into the mobile robot system. Develop SDK and API interfaces, support multiple programming languages, and facilitate secondary development and application expansion. Create teaching guides, PPTs, videos and other teaching resources, and provide online teaching platforms It uses one-to-one high-fidelity industrial scene reproduction technology and integrates multiple cutting-edge applications in the field of artificial intelligence to create an immersive learning experience for students, allowing them to efficiently learn the latest application examples of artificial intelligence technology and accelerate the integration of theoretical knowledge and practical skills. It fully supports the access of multi-brand depth cameras and integrates diverse hardware resources to achieve accurate and real-time depth perception and image processing. Users can deeply explore and master the complete artificial intelligence visual application system.

[0032] Equipped with a complete SDK development toolkit and open API interface, it enables developers to carry out secondary development flexibly and efficiently, meet the needs of diverse application scenarios, and promote the customization and scalability of products and services. Function-level algorithm source code facilitates developers to deeply understand and customize modifications, promotes efficient docking of technological innovation and projects, and realizes transparent controllability and deep optimization of algorithm applications.

[0033] In the technical solution provided by the present invention, the learning history data of students is collected by robots, and the learning history data is processed and analyzed by K-means clustering algorithm and decision tree algorithm to obtain the results of students' learning needs analysis; a course recommendation model is constructed based on the fusion of LSTM and attention mechanism, and personalized course content is recommended to students through the course recommendation model according to the results of learning needs analysis; in the process of students learning personalized course content, natural language processing technology, knowledge graph and BERT model are used to answer students' questions in real time; students' real-time learning behavior data is obtained, and real-time learning behavior data is evaluated by multimodal model; personalized course content is adjusted by Kuhn-Munkres algorithm based on the evaluation results output by multimodal model, so as to complete the online teaching tasks for students through teaching robots. 1. It can deeply understand students' learning behavior patterns and knowledge mastery, obtain accurate learning needs analysis results, provide data support for subsequent personalized education services, help educators better understand the characteristics and individual differences of student groups, and formulate more targeted teaching strategies. 2. Based on the results of the students' learning needs analysis, it can accurately recommend personalized course content that meets their learning progress, interests and abilities for each student, improve students' acceptance of the course and learning enthusiasm, avoid students getting lost in a large number of irrelevant courses, improve learning efficiency and learning effects, and promote students' personalized development. 3. It can promptly solve the problems encountered by students in learning, reduce learning obstacles, maintain the continuity and fluency of learning, enhance students' learning confidence and autonomous learning ability, and improve learning satisfaction. 4. It can comprehensively and objectively reflect the students' learning status, including information such as concentration and learning emotions, and provide educators with richer and more detailed feedback, so as to timely discover possible problems in the students' learning process and take corresponding intervention measures. 5. It can dynamically optimize the course content according to the students' real-time learning situation, ensure that the course is always closely matched with the students' learning needs, further improve the quality and effect of education and teaching, realize the efficient use of educational resources, and promote the intelligent and personalized development of education.

[0034] See also Figure 2 , a schematic diagram of a second embodiment of an artificial intelligence-based robot online teaching method provided by an embodiment of the present invention, in which natural language processing technology, knowledge graph and BERT model are used to answer students' questions in real time during the process of students learning personalized course content, including: Step 201: When a student asks a question during the process of learning personalized course content, natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic analysis of the question to obtain a question analysis result; Step 202: Named entity recognition is performed using the CRF conditional random field model to extract key concepts in the question. Let the text sequence of the question be X, and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ; In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ; ; In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; Step 203: Associating the extracted key concepts with the pre-built knowledge graph, and extracting information related to the question based on the associated knowledge graph nodes and edges, wherein the knowledge graph includes various knowledge points in the subject area and the relationships between them, and the key concepts include at least the subject name and knowledge points; Step 204: Fuse the question analysis results obtained by natural language processing with the information in the knowledge graph, input the fused text into the BERT model, use the BERT model to encode and decode the input text, and generate answer text for the questions raised by students, so as to provide real-time answers to students' questions.

[0035] Specifically, if students ask questions by voice, the system uses speech recognition technology to convert speech into text. A deep learning speech recognition model is used, such as a model based on a recurrent neural network (RNN) or a convolutional neural network (CNN) combined with RNN. The converted text is preliminarily cleaned and standardized to remove noise, irregular punctuation, and other issues for subsequent natural language processing and knowledge graph association.

[0036] Specifically, this embodiment also includes using Bi-LSTM (bidirectional long short-term memory network) to automatically learn the context representation of the sequence, and then using its output as the input of the CRF conditional random field model. After improving the CRF conditional random field model, feature named entity recognition is performed to extract the key concepts in the question. The improvement process is as follows: Assume the text sequence of the question is , the corresponding label sequence is , The first step is to build Bi-LSTM: Input layer: Assume the input sequence data is a text sequence , where T is the sequence length, is a character vector with dimension d, ; Bi-LSTM layer: The input sequence It is fed into the Bi-LSTM layer, which consists of a forward LSTM and a backward LSTM. For each time t, the hidden state of the forward LSTM is expressed as , the hidden state of the backward LSTM is expressed as ; The calculation process is as follows: ; in, , , They are input gate, forget gate, and output gate. is a candidate memory cell, is a memory unit, is the sigmoid function, represents element-wise multiplication, , , , , , , , is the corresponding weight matrix; The backward LSTM calculation formula is similar, but the input order is reversed, and the final result is the backward hidden state , concatenate the forward and backward hidden states to get the output representation of Bi-LSTM at time t: ; in, represents the number of LSTM hidden units. After the Bi-LSTM layer, the output of the entire sequence is , the dimension is .

[0037] Step 2: Use Bi-LSTM output as input to CRF; Construct CRF layer: Let the label sequence be ,in Define a transfer matrix for the number of label types ,in Indicates from the label Transfer to label The probability score (in the unnormalized case) for each position in the input sequence , Bi - feature vector output by LSTM Map it to the label space through a linear layer (fully connected layer) to get the score of each label. Let the weight matrix of the linear layer be , the bias vector is , then at time ,Label The score (unnormalized) is: + ; The entire sequence is for the label sequence The score (unnormalized) is defined as: ; Step 3: CRF decoding and training: In the prediction stage, the Viterbi algorithm is used to solve the score Maximum tag sequence , the Viterbi algorithm is based on the idea of ​​dynamic programming and defines For at the moment Arrival Tags The score of the maximum score path is calculated by the following recursive formula: ; Recursion (for ): ; Backtrack to find the optimal path: record the optimal previous label corresponding to each label at each moment , and finally from Find the label corresponding to the maximum score , and then by backtracking Get the entire optimal label sequence .

[0038] Training phase: The maximum likelihood estimation is used to train the model, with the goal of maximizing the log-likelihood probability of the label sequence on the training data.

[0039] Assume the training data set is ,in is the number of training samples; The log-likelihood function is defined as: ; in, Including the weight parameters of Bi-LSTM and the transfer matrix of CRF, Given an input sequence Next label sequence The probability is calculated based on the normalized score, namely: ; in, is the input sequence All possible label sequences are set. During the training process, the model parameters are updated by the stochastic gradient descent algorithm. , so that the log-likelihood function maximize.

[0040] When the maximized log-likelihood function is obtained Finally, the extracted key concepts are associated with the pre-built knowledge graph, and information related to the problem is extracted based on the associated knowledge graph nodes and edges. The knowledge graph includes various knowledge points in the subject area and the relationships between them. The key concepts include at least the subject name and knowledge points.

[0041] The beneficial effect lies in that real-time question answering has significant beneficial effects by using natural language processing technology, knowledge graph and BERT model. First, natural language processing technology can accurately understand the diverse ways of students asking questions and improve the accuracy and efficiency of problem handling. Secondly, knowledge graph integrates rich structured knowledge, provides comprehensive and systematic subject content for question answering, and broadens students' knowledge horizons. Finally, the powerful semantic understanding ability of the BERT model ensures in-depth analysis of complex problems, making the answers more in line with the core of students' questions, thereby effectively improving students' learning experience and enhancing their mastery of knowledge. By adopting the CRF conditional random field model improved by the Bi-LSTM network for named entity recognition, it can bring many benefits to students learning personalized courses. With the powerful context capture ability of Bi-LSTM and the effective constraint of CRF on the label sequence, the model can accurately extract the key concepts in the question. This helps to quickly locate students' knowledge blind spots and core doubts, and improve the pertinence and efficiency of question answering. For example, in complex subject questions, key entities such as professional terms, character names, and events are accurately identified, making subsequent learning resource recommendations and question answers more accurate, thereby optimizing students' learning paths.

[0042] See also Figure 3, a schematic diagram of a third embodiment of an artificial intelligence-based robot online teaching method provided by an embodiment of the present invention, obtaining students' real-time learning behavior data, and using a multimodal model to evaluate the real-time learning behavior data, including: Step 301: Acquire text data of students during the learning process, and collect image data of students during the learning process through a camera system to obtain real-time learning behavior data of students, wherein the text data at least includes question content, discussion speeches, and homework answers, and the image data at least includes students' facial expressions, postures, learning environment, and pictures of textbooks and notes that students are viewing; Step 302: preprocessing the text data and image data in the real-time learning behavior data, and using the preprocessed real-time learning behavior data as input data of the multimodal model; Step 303: Detect image features through YOLOv5 in the multimodal model, introduce an attention mechanism into the feature extraction network of YOLOv5, and adopt a GIoU loss function, wherein the attention mechanism includes channel attention and spatial attention; Step 304: extract text features through the Transformer in the multimodal model, use sinusoidal position encoding in the Transformer to add position information to the text features, increase the number of heads of the multi-head attention mechanism to capture more different text feature subspaces, and dynamically adjust the attention weights of different heads according to the learning behavior data, where the number of heads of the multi-head attention mechanism is 8; Step 305: Fuse the image features extracted by YOLOv5 and the text features extracted by Transformer, and dynamically adjust the fusion weights using the attention mechanism; Step 306: Use the fully connected layer and softmax function of the multimodal model to classify and evaluate the fused features to determine the student's learning behavior status and obtain an evaluation result.

[0043] Its beneficial effect lies in that there are many significant beneficial effects through the way of multimodal data collection and processing. First, the comprehensive acquisition of text and image data can capture the details of the students' learning process in all aspects, whether it is questions at the knowledge level, communication and interaction, or the intuitive presentation of the learning status, it can be accurately grasped. Secondly, the preprocessing of data and the introduction of advanced mechanisms and technologies in the model, such as the attention mechanism in YOLOv5, the GIoU loss function, and the optimization measures in Transformer, have greatly improved the accuracy and comprehensiveness of feature extraction, and can more keenly discover subtle changes and potential problems in students' learning behavior. Furthermore, by dynamically adjusting the attention weight for feature fusion, multimodal information can be efficiently integrated, and ultimately the evaluation results of students' learning behavior status are more accurate and reliable, which helps teachers to understand students' situations in a timely manner, thereby providing a strong basis for personalized teaching intervention and learning strategy adjustment, and effectively improving teaching quality and students' learning outcomes.

[0044] See also Figure 4 , a schematic diagram of a structure of an artificial intelligence-based robot online teaching system provided by an embodiment of the present invention, the system includes The data analysis module 401 is used to send a data collection instruction to the teaching robot. After receiving the data collection instruction, the teaching robot collects the student's learning history data, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; Specifically, the data analysis module in this embodiment includes a data collection submodule, a data processing submodule, a feature extraction submodule, a cluster analysis submodule and a classification prediction submodule, wherein: The data collection submodule is used to collect various learning history data of students on the online learning platform, where the learning history data at least includes course selection records, learning progress data, homework and test scores, and interactive behavior data; The data processing submodule is used to perform data cleaning and data normalization on the learning history data to obtain the pre-processed learning history data; The feature extraction submodule is used to extract candidate features from the preprocessed learning history data and use the candidate features as data points for input into the K-means clustering algorithm; The cluster analysis submodule is used to perform cluster analysis on the candidate features of students using the K-means clustering algorithm, assign data points to the group with the closest cluster center, then recalculate the center of each group, and iterate continuously until the cluster center no longer changes or reaches the preset number of iterations, and output the results of the K-means clustering algorithm, where the preset number of iterations is 100 times; The classification prediction submodule is used to classify and predict students' learning needs based on the clustering results of the K-means clustering algorithm using a decision tree algorithm to obtain the results of the students' learning needs analysis.

[0045] Specifically, the classification prediction submodule in this embodiment includes a construction unit, a calculation unit and an input unit, wherein: Building blocks for building decision trees: ; In the formula, represents a decision tree, Represents input data, Indicates that the decision tree parameters are determined by the leaf node area With number Calculated, Indicated by and The collection composed of Indicates the division area. Represents the decision value of the partition area, and the final decision tree is obtained after several iterations; Computational unit for splitting by calculating information gain: ; In the formula, represents the improved information gain, Representation dataset The original entropy of Representation characteristics The set of all possible values ​​of Indicates when the feature The value is The data subset at time Representation Subset The number of samples, Representation dataset The total number of samples, Representation Subset The entropy of represents the number of features, Indicates The important weight coefficient of the feature, is an indicator function, if the feature With features In the value The following is relevant, =1, otherwise 0; The input unit is used to input the clustering results obtained by the K-means clustering algorithm into the constructed decision tree, starting from the root node, and traversing downward step by step according to the splitting conditions of each node until reaching the leaf node to obtain the classification results of students' learning needs.

[0046] The course recommendation module 402 is used to build a course recommendation model based on the fusion of LSTM and attention mechanism, and recommend personalized course content to students through the course recommendation model according to the learning needs analysis results; Specifically, the course recommendation module in this embodiment includes an information input submodule, a calculation submodule, a normalization processing submodule, a weighting submodule and a fusion submodule, wherein: The information input submodule is used to obtain student demand data based on the results of learning demand analysis, input the sequence information of student demand data into the course recommendation model, and capture the time series characteristics of student demand changes at different learning stages through the forget gate, input gate, and output gate of the LSTM layer; The calculation submodule is used to calculate the hidden state containing the student demand sequence characteristics output by the LSTM layer , the score of each time step is calculated through the attention layer : ; In the formula, represents the weight matrix of the attention layer, represents the hyperbolic tangent function; Normalization submodule, used to use the softmax function to score all time steps Normalize and get the attention weight: ; In the formula, represents the total number of time steps, Indicates The score of each task; The weighted submodule is used to transform the hidden state of each time step According to the attention weight Weighted to get the final context vector : ; The fusion submodule is used to fuse the obtained context vector with the course content feature vector through the fully connected layer, input the fused vector into the linear regression layer, output a recommendation score, and rank and recommend courses according to the scores to obtain personalized course content recommended to students.

[0047] A real-time question-answering module 403 is used to answer questions for students in real time through natural language processing technology, knowledge graph and BERT model when students are learning personalized course content; Specifically, the real-time question answering module in this embodiment includes a question analysis submodule, an extraction submodule, an association submodule, and an encoding and decoding submodule, wherein: The question analysis submodule is used to use natural language processing technology to perform lexical analysis, syntactic analysis and semantic analysis on questions raised by students in the process of learning personalized course content, and obtain the question analysis results; The extraction submodule is used to recognize named entities using the CRF conditional random field model to extract the key concepts in the question. Let the text sequence of the question be X and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ; In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ; ; In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; The association submodule is used to associate the extracted key concepts with the pre-built knowledge graph, and extract information related to the question based on the associated knowledge graph nodes and edges, where the knowledge graph includes various knowledge points in the subject area and the relationships between them, and the key concepts include at least the subject name and knowledge points; The encoding and decoding submodule is used to fuse the question analysis results obtained by natural language processing with the information in the knowledge graph. The fused text is input into the BERT model, and the BERT model is used to encode and decode the input text to generate answer text for questions raised by students, so as to provide real-time answers to students' questions.

[0048] A learning evaluation module 404 is used to obtain students' real-time learning behavior data and evaluate the real-time learning behavior data using a multimodal model; Specifically, the learning evaluation module in this embodiment includes a data acquisition submodule, a data processing submodule, an image feature extraction submodule, a text feature extraction submodule, a feature fusion submodule and an analysis and evaluation submodule, wherein: The data acquisition submodule is used to acquire text data of students during the learning process, and collect image data of students during the learning process through the camera system to obtain students' real-time learning behavior data, where the text data at least includes the content of questions, discussion speeches, and homework answers, and the image data at least includes students' facial expressions, postures, learning environment, and pictures of textbooks and notes that students are viewing; A data processing submodule, used for preprocessing both text data and image data in the real-time learning behavior data, and using the preprocessed real-time learning behavior data as input data of the multimodal model; The image feature extraction submodule is used to detect image features through YOLOv5 in the multimodal model, introduce the attention mechanism into the feature extraction network of YOLOv5, and adopt the GIoU loss function, where the attention mechanism includes channel attention and spatial attention; The text feature extraction submodule is used to extract text features through the Transformer in the multimodal model. Sine position encoding is used in the Transformer to add position information to the text features. The number of heads in the multi-head attention mechanism is increased to capture more different text feature subspaces. The attention weights of different heads are dynamically adjusted according to the learning behavior data. The number of heads in the multi-head attention mechanism is 8. The feature fusion submodule is used to fuse the image features extracted by YOLOv5 and the text features extracted by Transformer, and dynamically adjust the fusion weights using the attention mechanism; The analysis and evaluation submodule is used to classify and evaluate the fused features using the fully connected layer and softmax function of the multimodal model to determine the students' learning behavior status and obtain the evaluation results.

[0049] 405. A course adjustment module is used to adjust personalized course content using the Kuhn-Munkres algorithm based on the evaluation results output by the multimodal model, so as to complete the online teaching task for students through the teaching robot.

[0050] Specifically, the course adjustment module in this embodiment includes a state learning submodule, an association matching submodule and a state evaluation submodule, wherein: The state learning submodule is used to obtain the student's learning state set and the course content feature set of personalized course content based on the evaluation results output by the multimodal model, and construct the student's learning state Course content characteristics The correlation matrix ; The association matching submodule is used to match using the Kuhn-Munkres algorithm to maximize the sum of the matched elements in the association matrix and obtain the optimal matching solution. The objective function is: ; In the formula, Indicates learning status Course content characteristics The degree of correlation, is a decision variable, representing the learning state Is it related to the course content characteristics? Match, if match then If it does not match, ; The constraints of the Kuhn-Munkres algorithm are: For each learning state , can only match one course content feature at most, i.e. ; For each course content feature , can only match one learning state at most, i.e. ; The status evaluation submodule is used to determine the course content characteristics corresponding to each student's learning status according to the optimal matching scheme, so as to adjust the personalized course content, continuously collect students' learning behavior data, re-evaluate students' learning status, and use the Kuhn-Munkres algorithm to adjust the course content.

[0051] Through the implementation of the above scheme, the AI ​​computing chip in the robot uses the K-means clustering algorithm and decision tree algorithm to process and analyze the student learning history data, which can deeply understand the students' learning behavior patterns and knowledge mastery, obtain accurate learning needs analysis results, provide data support for subsequent personalized education services, and help educators better understand the characteristics and individual differences of the student group, so as to formulate more targeted teaching strategies. The course recommendation model based on the fusion of LSTM and attention mechanism can accurately recommend personalized course content that meets the learning progress, interests and abilities of each student according to the results of the student's learning needs analysis, improve students' acceptance of courses and learning enthusiasm, avoid students getting lost in a large number of irrelevant courses, improve learning efficiency and learning effects, and promote students' personalized development. In the process of students' learning, using natural language processing technology, knowledge graphs and BERT models to answer students' questions in real time can solve students' problems in learning in a timely manner, reduce learning obstacles, maintain the coherence and fluency of learning, enhance students' learning confidence and autonomous learning ability, create a good learning atmosphere, and improve learning satisfaction. Using a multimodal model to evaluate students' real-time learning behavior data can comprehensively and objectively reflect students' learning status, including information such as concentration and learning emotions, and provide educators with richer and more detailed feedback so that they can promptly identify problems that students may have in the learning process and take corresponding intervention measures. Finally, based on the evaluation results output by the multimodal model, the Kuhn-Munkres algorithm is used to adjust the personalized course content, which can dynamically optimize the course content according to the students' real-time learning situation, ensure that the course is always closely matched with the students' learning needs, further improve the quality and effect of education and teaching, realize the efficient use of educational resources, and promote the intelligent and personalized development of education.

[0052] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited by the above embodiments. The above embodiments and descriptions are only preferred examples of the present invention and are not intended to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A robot online teaching method based on artificial intelligence, characterized in that: The robot online teaching method based on artificial intelligence comprises the following steps: Sending a data collection instruction to the teaching robot, the teaching robot collects the student's learning history data after receiving the data collection instruction, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; A course recommendation model is constructed based on the fusion of LSTM and attention mechanism, and personalized course content is recommended to students through the course recommendation model according to the learning needs analysis results; While students are learning personalized course content, they can answer questions in real time through natural language processing technology, knowledge graphs, and BERT models; Acquire students' real-time learning behavior data, and evaluate the real-time learning behavior data using a multimodal model; Based on the evaluation results output by the multimodal model, the Kuhn-Munkres algorithm is used to adjust the personalized course content so that the teaching robot can complete the online teaching task for students.

2. The method for online teaching of robots based on artificial intelligence as claimed in claim 1, characterized in that: The data collection instruction is sent to the teaching robot, and the teaching robot collects the student's learning history data after receiving the data collection instruction, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results, including: Collect various learning history data of students on the online learning platform, wherein the learning history data at least includes course selection records, learning progress data, homework and test scores, and interactive behavior data; Performing data cleaning and data normalization processing on the learning history data to obtain preprocessed learning history data; Extracting candidate features from the preprocessed learning history data, and using the candidate features as data points input into the K-means clustering algorithm; Use K-means clustering algorithm to cluster the candidate features of students, assign data points to the group with the closest cluster center, then recalculate the center of each group, and iterate until the cluster center no longer changes or reaches the preset number of iterations, and output the results of K-means clustering algorithm, where the preset number of iterations is 100; The clustering results based on the K-means clustering algorithm use the decision tree algorithm to classify and predict the students' learning needs and obtain the results of the students' learning needs analysis.

3. The method for online teaching of robots based on artificial intelligence as claimed in claim 2, characterized in that: The clustering results based on the K-means clustering algorithm use a decision tree algorithm to classify and predict the students' learning needs, and obtain the students' learning needs analysis results, including: Build a decision tree: ； In the formula, represents a decision tree, Represents input data, Indicates that the decision tree parameters are determined by the leaf node area With number Calculated, Indicated by and The collection composed of Indicates the division area. Represents the decision value of the partition area, and the final decision tree is obtained after several iterations; Split by calculating information gain: ； In the formula, represents the improved information gain, Representation dataset The original entropy of Representation characteristics The set of all possible values ​​of Indicates when the feature The value is The data subset at time Representation Subset The number of samples, Representation dataset The total number of samples, Representation Subset The entropy of represents the number of features, Indicates The important weight coefficient of the feature, is an indicator function, if the feature With features In the value The following is relevant, =1, otherwise 0; The clustering results obtained by the K-means clustering algorithm are input into the constructed decision tree. Starting from the root node, it is traversed downward step by step according to the splitting conditions of each node until it reaches the leaf node to obtain the classification results of students' learning needs.

4. The method for online teaching of robots based on artificial intelligence as claimed in claim 1, characterized in that: The course recommendation model is constructed based on the fusion of LSTM and attention mechanism, and personalized course content is recommended to students through the course recommendation model according to the learning needs analysis result, including: Obtain student demand data according to the learning demand analysis results, input the sequence information of the student demand data into the course recommendation model, and capture the time series characteristics of the changes in student demand at different learning stages through the forget gate, input gate and output gate of the LSTM layer; According to the hidden state of the LSTM layer output containing the characteristics of the student demand sequence , the score of each time step is calculated through the attention layer : ； In the formula, represents the weight matrix of the attention layer, represents the hyperbolic tangent function; Use the softmax function to score all time steps Normalize and get the attention weight: ； In the formula, represents the total number of time steps, Indicates The score of each task; The hidden state at each time step According to the attention weight Weighted to get the final context vector : ； The context vector obtained is fused with the course content feature vector through the fully connected layer, and the fused vector is input into the linear regression layer to output a recommendation score. The courses are ranked and recommended according to the scores to obtain personalized course content recommended to students.

5. The method for online teaching of robots based on artificial intelligence as claimed in claim 1, characterized in that: In the process of students learning personalized course content, natural language processing technology, knowledge graph and BERT model are used to provide students with real-time questions and answers, including: When students ask questions during their personalized course learning, natural language processing technology is used to perform lexical analysis, syntactic analysis, and semantic analysis of the questions to obtain the analysis results. The CRF conditional random field model is used for named entity recognition to extract the key concepts in the question. Let the text sequence of the question be X and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ； In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ； ； In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; Associating the extracted key concepts with a pre-built knowledge graph, and extracting information related to the question based on the associated knowledge graph nodes and edges, wherein the knowledge graph includes various knowledge points in the subject area and the relationships between them, and the key concepts include at least the subject name and knowledge points; The question analysis results obtained by natural language processing are integrated with the information in the knowledge graph, and the integrated text is input into the BERT model. The BERT model is used to encode and decode the input text to generate answer text for students' questions, so as to provide real-time answers to students' questions.

6. The method for online teaching of robots based on artificial intelligence as claimed in claim 1, characterized in that: The step of obtaining the real-time learning behavior data of the students and evaluating the real-time learning behavior data using a multimodal model includes: Acquire text data of students during the learning process, and collect image data of students during the learning process through a camera system to obtain real-time learning behavior data of students, wherein the text data at least includes question content, discussion speeches, and homework answers, and the image data at least includes students' facial expressions, postures, learning environment, and pictures of textbooks and notes that students are viewing; Preprocessing the text data and image data in the real-time learning behavior data, and using the preprocessed real-time learning behavior data as input data of the multimodal model; The image features are detected by YOLOv5 in the multimodal model, and the attention mechanism is introduced into the feature extraction network of YOLOv5. The GIoU loss function is adopted, where the attention mechanism includes channel attention and spatial attention. The text features are extracted through the Transformer in the multimodal model. Sine position encoding is used in the Transformer to add position information to the text features. The number of heads in the multi-head attention mechanism is increased to capture more different text feature subspaces. The attention weights of different heads are dynamically adjusted according to the learning behavior data. The number of heads in the multi-head attention mechanism is 8. The image features extracted by YOLOv5 and the text features extracted by Transformer are fused, and the fusion weights are dynamically adjusted using the attention mechanism; The fully connected layer and softmax function of the multimodal model are used to classify and evaluate the fused features to determine the students’ learning behavior status and obtain the evaluation results.

7. The method for online teaching of robots based on artificial intelligence as claimed in claim 1, characterized in that: The evaluation results based on the multimodal model output are used to adjust the personalized course content using the Kuhn-Munkres algorithm, including: Based on the evaluation results output by the multimodal model, the student's learning status set and the course content feature set of personalized course content are obtained to construct the student's learning status Course content characteristics The correlation matrix ; Use the Kuhn-Munkres algorithm to match so that the sum of the matched elements in the incidence matrix is ​​maximized and the optimal matching solution is obtained. The objective function is: ； In the formula, Indicates learning status Course content characteristics The degree of correlation, is a decision variable, representing the learning state Is it related to the course content characteristics? Match, if match then If it does not match, ; The constraints of the Kuhn-Munkres algorithm are: For each learning state , can only match one course content feature at most, i.e. ; For each course content feature , can only match one learning state at most, i.e. ; According to the optimal matching scheme, the course content characteristics corresponding to each student's learning status are determined to adjust the personalized course content, continuously collect students' learning behavior data, re-evaluate students' learning status, and use the Kuhn-Munkres algorithm to adjust the course content.

8. A robot online teaching system based on artificial intelligence, characterized in that: The robot online teaching system based on artificial intelligence includes a data analysis module, a course recommendation module, a real-time question-answering module, a learning evaluation module and a course adjustment module, wherein: The data analysis module is used to send data collection instructions to the teaching robot, and the teaching robot collects the student's learning history data after receiving the data collection instructions, and uses the K-means clustering algorithm and the decision tree algorithm to process and analyze the learning history data to obtain the student's learning needs analysis results; A course recommendation module is used to build a course recommendation model based on the fusion of LSTM and attention mechanism, and recommend personalized course content to students through the course recommendation model according to the learning needs analysis results; The real-time Q&A module is used to answer students' questions in real time through natural language processing technology, knowledge graphs and BERT models while they are learning personalized course content; A learning evaluation module is used to obtain students' real-time learning behavior data and evaluate the real-time learning behavior data using a multimodal model; The course adjustment module is used to adjust the personalized course content based on the evaluation results output by the multimodal model using the Kuhn-Munkres algorithm, so as to complete the online teaching tasks for students through the teaching robot.

9. The robot online teaching system based on artificial intelligence as claimed in claim 8, characterized in that: The course recommendation module includes an information input submodule, a calculation submodule, a normalization processing submodule, a weighting submodule and a fusion submodule, wherein: An information input submodule, used to obtain student demand data according to the learning demand analysis results, input the sequence information of the student demand data into the course recommendation model, and capture the time series characteristics of the changes in student demand at different learning stages through the forget gate, input gate and output gate of the LSTM layer; The calculation submodule is used to calculate the hidden state containing the student demand sequence characteristics output by the LSTM layer , the score of each time step is calculated through the attention layer : ； In the formula, represents the weight matrix of the attention layer, represents the hyperbolic tangent function; Normalization submodule, used to use the softmax function to score all time steps Normalize and get the attention weight: ； In the formula, represents the total number of time steps, Indicates The score of each task; The weighted submodule is used to transform the hidden state of each time step According to the attention weight Weighted to get the final context vector : ； The fusion submodule is used to fuse the obtained context vector with the course content feature vector through the fully connected layer, input the fused vector into the linear regression layer, output a recommendation score, and rank and recommend courses according to the scores to obtain personalized course content recommended to students.

10. The robot online teaching system based on artificial intelligence as claimed in claim 8, characterized in that: The real-time question answering module includes a question analysis submodule, an extraction submodule, an association submodule, and an encoding and decoding submodule, wherein: The question analysis submodule is used to use natural language processing technology to perform lexical analysis, syntactic analysis and semantic analysis on questions raised by students in the process of learning personalized course content, and obtain the question analysis results; The extraction submodule is used to recognize named entities using the CRF conditional random field model to extract the key concepts in the question. Let the text sequence of the question be X and the corresponding label sequence be Y. The score of the feature function is calculated as follows: ； In the formula, Indicates the length of the text sequence, represents the number of characteristic functions, Indicates The weight corresponding to the feature function is Represents the feature function, in the text sequence Middle, Location At, given the previous annotation and the current annotation The function value corresponding to the feature when Conditional probability of CRF conditional random field model for: ； ； In the formula, the conditional probability Indicates that in a given text sequence In the case of, the probability of label sequence Y, is the characteristic function score The exponential form is used to calculate the original probability of the label sequence Y. is the sum of all possible label sequences Y Sum so that the sum of the probabilities of all possible label sequences Y is 1; An association submodule is used to associate the extracted key concepts with a pre-built knowledge graph, and extract information related to the question based on the associated knowledge graph nodes and edges, wherein the knowledge graph includes various knowledge points in the subject area and the relationships between them, and the key concepts include at least the subject name and knowledge points; The encoding and decoding submodule is used to fuse the question analysis results obtained by natural language processing with the information in the knowledge graph. The fused text is input into the BERT model, and the BERT model is used to encode and decode the input text to generate answer text for questions raised by students, so as to provide real-time answers to students' questions.

Images