AI-assisted teaching method and platform based on multi-agent collaboration

Abstract

This invention discloses an AI-assisted teaching method and platform based on multi-agent collaboration, belonging to the field of AI-assisted teaching technology. The method includes: collecting multi-dimensional static features of learners through a multi-agent teaching cognitive adaptation model; capturing dynamic learning status data through an agent-based learning assessment and analysis system; generating a cognitive adaptation matrix through cross-dimensional feature fusion; dynamically generating a learning content sequence by inputting it into a personalized learning path generation model combined with a subject knowledge graph; outputting a knowledge mastery trend vector through a learning time-series prediction and collaborative optimization algorithm; and finally dynamically adjusting teaching parameters and strategies based on real-time assessment results to form a closed-loop control. This invention achieves organic integration of data integration, feature analysis, path adaptation, and dynamic control through multi-agent collaboration, comprehensively considering individual learner differences and dynamic learning states, improving teaching adaptability and accuracy, and providing an efficient solution for personalized AI-assisted teaching.

Description

Technical Field

[0001] This invention relates to the field of AI-assisted teaching technology, and in particular to an AI-assisted teaching method and platform based on multi-agent collaboration. Background Technology

[0002] Against the backdrop of the deepening digital transformation of education, traditional uniform teaching models are no longer suitable for the individual differences in learners' cognitive characteristics, learning pace, and behavioral habits. Accurately capturing dynamic changes in the learning process and achieving efficient matching of teaching resources with learners' needs have become core demands in the current education field. With the widespread application of intelligent technologies in educational scenarios, single-function teaching aids can no longer meet the needs of multi-dimensional data processing and dynamic control in complex teaching processes. There is an urgent need to build an intelligent teaching system that can integrate multi-dimensional learning data and achieve cross-stage collaborative operation to solve problems such as fragmented teaching data and static teaching strategies. This will drive the transformation of teaching models from mass supply to precise adaptation, providing technical support for personalized teaching and efficient learning management.

[0003] Existing technologies in the field of intelligent assisted teaching have two significant shortcomings: First, they lack a systematic collaborative operation mechanism, and the data flow and functional connection between various teaching assistance links are relatively loose, failing to achieve comprehensive integration and in-depth utilization of learning data. This results in insufficient perception of learners' status during the teaching process, making it difficult to form a coherent and precise teaching control logic. Second, they lack the ability to adapt to learners' personalized needs, failing to fully consider the dynamic changes in the learning process and failing to achieve dynamic matching of teaching resources, teaching strategies, and learners' status. This limits the effectiveness of teaching assistance, making it difficult to effectively meet the differentiated development needs of different learners and achieve precise and personalized teaching goals. Summary of the Invention

[0004] In order to overcome the shortcomings and deficiencies of existing technologies, this invention provides an AI-assisted teaching method and platform based on multi-agent collaboration.

[0005] The technical solution adopted in this invention is an AI-assisted teaching method based on multi-agent collaboration, comprising the following steps: S1, collecting learners' cognitive styles, knowledge mastery, learning behavior sequences, and interactive response features through the distributed nodes of the multi-agent teaching cognitive adaptation model to construct a multi-dimensional static feature vector set of learners; S2, dynamically capturing resource access trajectories, task completion behaviors, question-and-answer feedback content, and physiological signal fluctuation parameters during the learning process through the real-time data interaction interface of the intelligent agent learning assessment and analysis system to form a time-series learning state dataset; S3, calling the multi-agent teaching cognitive adaptation model to process the static features... S4. The vector set and the temporal learning state dataset are fused across dimensions to generate a learner cognitive fit matrix. S5. The cognitive fit matrix is ​​input into the personalized learning path generation model, and combined with the node association weights of the subject knowledge graph, a multi-branch learning content sequence is dynamically generated. S6. The learning progress temporal prediction collaborative optimization algorithm is used to perform rolling prediction on the state data during the execution of the learning path, and outputs a knowledge mastery trend vector. S7. Based on the trend vector and the real-time evaluation results of the agent learning progress evaluation and analysis system, the multi-agent collaborative unit dynamically adjusts the teaching cognitive fit parameters, learning path node configuration and interactive feedback strategy.

[0006] Furthermore, the expression for the multi-agent teaching cognitive adaptation model is:

[0007] in, For the cognitive fit between the i-th learner and the j-th teaching content, For multi-agent weight coefficients, For learner cognitive feature vectors, For teaching content attribute vectors, For the cognition-content matching matrix, For learners' physiological signal feature vectors, Adapting vectors for teaching scenarios The physiological-scene correlation matrix For learner interaction behavior vectors, For teaching feedback feature vectors, For behavior-feedback mapping matrix, This is the Sigmoid activation function.

[0008] Furthermore, the expression for the personalized learning path generation model is:

[0009] in, The optimal learning path for the k-th learner. Generate weight coefficients for the path. Let be the mastery level of the k-th learner at the t-th knowledge node. Let be the pre-association weight between the t-th node and the n-th node. Let be the learning efficiency of the k-th learner at the t-th node. Let be the temporal correlation weight between the t-th node and the n-th node. Let be the interaction depth of the k-th learner at the t-th node. The weight is the association weight between the content of the t-th node and the n-th node.

[0010] Furthermore, the expression for the collaborative optimization algorithm for learning time series prediction is:

[0011] in, For the first The predicted value of learning progress at any given time. Let m be the measured value of the student's learning progress at time m. These are the fusion coefficients of the time series model. To optimize the adjustment coefficient, Let be the cognitive state feature vector at time m. Let m be the behavioral interaction feature vector at time m. Let be the physiological signal feature vector at time m. For prediction functions of Long Short-Term Memory (LSTM) networks, For the prediction function of the gated recurrent unit, To collaboratively optimize the objective function.

[0012] Furthermore, the evaluation model expression of the intelligent agent learning assessment and analysis system is as follows:

[0013] in, Let p be the overall learning assessment value of the learner. To evaluate the dimensional weight coefficients, For knowledge testing feature vectors, For knowledge standard feature vectors, For knowledge assessment matrix, This is a feature vector representing behavioral characteristics. For behavioral standard feature vectors, For the behavior assessment matrix, This is a feature vector representing the physiological state. For physiological standard feature vectors, This is a physiological assessment matrix. It is the activation function for the exponential linear unit.

[0014] Furthermore, the fusion model expression for wearable physiological signal acquisition and AI intelligent health risk assessment is as follows:

[0015] in, Let q be the health risk assessment value of the learner. To integrate the weighting coefficients, The time decay factor, For the qth learner Multimodal physiological signal sequences at different times. For signal convolution kernel, For pooling window parameters, For normalization parameters, For convolution operations, For pooling operations, For standardized calculations.

[0016] Further, S3 includes the following sub-steps: S31, the feature extraction unit of the multi-agent teaching cognitive adaptation model performs dimensional mapping on the cognitive style and knowledge mastery parameters in the static feature vector set to form a standardized cognitive feature matrix; S32, the model's temporal processing unit performs sliding window sampling on the behavioral sequences and interaction response parameters in the temporal learning state dataset to generate a dynamic behavioral feature tensor; S33, the model's cross-modal fusion unit uses an attention mechanism to perform weight allocation and element-wise multiplication on the standardized cognitive feature matrix and the dynamic behavioral feature tensor to obtain a fused feature matrix; S34, the model's adaptation calculation unit performs nonlinear transformation and normalization on the fused feature matrix to output a dimension-matched learner cognitive adaptation matrix.

[0017] Further, S4 includes the following sub-steps: S41, the knowledge graph parsing unit of the personalized learning path generation model traverses the node attributes and relationships in the subject knowledge graph, extracting the knowledge node weights and prerequisite dependencies; S42, the model's fit mapping unit matches the elements in the cognitive fit matrix with the corresponding knowledge node weights to generate a node fit priority sequence; S43, the model's path construction unit generates an initial multi-branch learning path set based on the node fit priority sequence and prerequisite dependencies using a greedy algorithm; S44, the model's path optimization unit combines learner historical learning behavior data to prune redundant nodes and strengthen labeled nodes in the initial path set, outputting an optimized personalized learning content sequence.

[0018] Further, S5 includes the following sub-steps: S51, the dataset partitioning unit of the learning path execution process divides the state data into training set and validation set according to timestamps, and determines the rolling prediction window length; S52, the feature engineering unit of the algorithm performs feature screening and dimensionality reduction on the state data in the training set, and extracts a subset of features strongly correlated with the knowledge mastery trend; S53, the prediction model training unit of the algorithm uses the feature subset to iteratively train the time-series prediction model, adjusts the model hyperparameters through the validation set, and determines the optimal prediction parameter configuration; S54, the collaborative optimization unit of the algorithm compares the output result of the prediction model with the real-time data of the intelligent agent learning assessment and analysis system, corrects the prediction deviation through gradient descent, and outputs an accurate knowledge mastery trend vector.

[0019] This AI-assisted teaching platform, based on multi-agent collaboration, is applied to AI-assisted teaching methods based on multi-agent collaboration. It includes: a multimodal learner data acquisition and preprocessing module, a distributed multi-agent teaching cognitive adaptation computation module, a knowledge graph-driven personalized learning path generation module, a time-series learning prediction and collaborative optimization module, an agent dynamic learning assessment and feedback module, and a multi-agent collaborative teaching regulation and resource scheduling module. The multimodal learner data acquisition and preprocessing module collects physiological signals, behavioral data, and cognitive feature parameters through a wearable device interface and an interactive terminal, outputting a standardized dataset. The distributed multi-agent teaching cognitive adaptation computation module receives the standardized dataset and processes it through multi-agent parallel processing. A cognitive fit matrix is ​​calculated and generated, and a data transmission link is established with the personalized learning path generation module. The personalized learning path generation module generates a learning path sequence based on the cognitive fit matrix and knowledge graph data, and sends it to the time-series learning prediction and collaborative optimization module. The time-series learning prediction and collaborative optimization module predicts the path execution status and outputs a trend vector to the agent dynamic learning assessment and feedback module. The agent dynamic learning assessment and feedback module generates control instructions based on real-time assessment data and transmits them to the multi-agent collaborative teaching control and resource scheduling module. The multi-agent collaborative teaching control and resource scheduling module adjusts the parameter configuration and resource allocation scheme of each module according to the control instructions, forming a closed-loop collaborative operation architecture.

[0020] Beneficial Effects: This invention proposes an AI-assisted teaching method and platform based on multi-agent collaboration. It integrates multi-dimensional learner data collection and dynamic state capture mechanisms, breaking down data fragmentation barriers. Through cross-dimensional feature fusion and comprehensive evaluation analysis, it achieves a comprehensive perception of learners' cognitive characteristics, behavioral habits, and physiological states, solving the problem of incomplete learner state perception in traditional technologies. Relying on core functional modules that operate collaboratively, it constructs a closed-loop link from data collection, cognitive adaptation, path generation to learning prediction and dynamic control, enabling real-time data interaction and parameter linkage at each stage, replacing the original loosely structured model where each function operated independently. This approach significantly enhances the continuity and precision of teaching control; it dynamically generates suitable learning paths based on individual learner characteristics and real-time status, and continuously optimizes teaching strategies through time-series prediction and dynamic adjustment mechanisms. It fully considers the differentiated needs of different learners and the dynamic changes in the learning process, making up for the shortcomings of traditional platforms in terms of insufficient personalized adaptation capabilities and lack of targeted path adjustments; through multi-module collaborative scheduling and resource optimization, it promotes the transformation of teaching models from static and uniform to dynamic and precise, significantly improving the efficiency of teaching resource utilization and learning assistance effects, and providing reliable technical support for personalized teaching and efficient learning management. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S3 of the present invention; Figure 3 This is a flowchart of method step S4 of the present invention; Figure 4 This is a flowchart of step S5 of the method of the present invention; Figure 5 This is a diagram showing the platform module composition of the present invention. Detailed Implementation

[0022] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] like Figure 1As shown, the AI-assisted teaching method based on multi-agent collaboration includes the following steps: S1, collecting learners' cognitive styles, knowledge mastery, learning behavior sequences, and interactive response characteristics through the distributed nodes of the multi-agent teaching cognitive adaptation model to construct a multi-dimensional static feature vector set of learners; S2, dynamically capturing resource access trajectories, task completion behaviors, question-and-answer feedback content, and physiological signal fluctuation parameters during the learning process through the real-time data interaction interface of the intelligent agent learning assessment and analysis system to form a time-series learning state dataset; S3, calling the multi-agent teaching cognitive adaptation model to apply the static feature vector set to the time-series learning state dataset. S4. The sequential learning state dataset is fused across dimensions to generate a learner cognitive fit matrix; S5. The cognitive fit matrix is ​​input into the personalized learning path generation model, and combined with the node association weights of the subject knowledge graph, a multi-branch learning content sequence is dynamically generated; S6. The state data during the execution of the learning path is predicted in a rolling manner through a learning progress time-series prediction collaborative optimization algorithm, and a knowledge mastery trend vector is output; S7. Based on the trend vector and the real-time evaluation results of the agent learning progress evaluation and analysis system, the multi-agent collaborative unit dynamically adjusts the teaching cognitive fit parameters, learning path node configuration, and interactive feedback strategies.

[0024] Step S1 utilizes the 32 distributed data collection nodes of the multi-agent teaching cognitive adaptation model to simultaneously collect multi-dimensional static characteristic data of learners. The collection frequency is set to once every 24 hours, covering 14 calendar days from the initial enrollment stage to the stable learning stage, ensuring the comprehensiveness and representativeness of the data. The core parameters collected include eight dimensions related to cognitive style, specifically classification parameters such as field dependence-field independence and impulsive-reflective types, each divided into 10 quantitative levels; knowledge mastery focuses on core subject knowledge points, divided into 20 knowledge units according to course modules, with each unit quantitatively assessed through 15 test questions, recording the number of correct answers and the time spent answering; learning behavior sequences include 12 behavioral parameters such as learning duration, learning time distribution, and resource type preference over the past 30 days, with learning duration accurate to the minute and time distribution divided into 6 time periods; interaction response characteristics include nine indicators such as question-and-answer response duration, interaction frequency, and feedback content relevance, with response duration accurate to the second, interaction frequency counted daily, and feedback content relevance quantified through semantic matching. By collecting, organizing, and encoding the above 44 specific parameters across 4 categories, a multi-dimensional static feature vector set of learners with a dimension of 44 times the number of learners is constructed, providing basic data support for subsequent cognitive adaptation analysis.

[0025] Step S2 uses the real-time data interaction interface built into the intelligent agent learning assessment and analysis system. This interface uses the WebSocket protocol to achieve real-time data transmission, with a transmission rate of 100Mbps and data latency controlled within 50 milliseconds to ensure the timeliness and accuracy of dynamic learning status data capture. The core content of the data capture includes resource access trajectory, specifically recording eight parameters such as the type of teaching resources accessed by learners, access order, dwell time, and number of repeated visits. Resource types are categorized into six types, including videos, documents, exercises, and case studies, with dwell time accurate to the second. Task completion behavior includes seven parameters such as completion progress, submission time, and error type distribution for five types of tasks, including assignments, quizzes, and practical tasks. Completion progress is quantified as a percentage, and error types are categorized into nine types, including conceptual errors, calculation errors, and logical errors. Question and answer feedback includes six parameters such as the number of questions initiated by learners, question type, completeness of question expression, and attitude towards teacher feedback. Question types are categorized into four types, including knowledge questions, methodological consultations, and equipment malfunctions. Physiological signal fluctuation parameters are collected through wearable devices, including five physiological indicators such as heart rate, respiratory rate, and skin conductivity, collected 20 times per minute, recording the real-time values ​​and fluctuation amplitudes of these indicators. Through continuous capture and real-time storage of these four categories of a total of 26 dynamic parameters, a time-series learning status dataset is formed, providing real-time data support for dynamic learning analysis.

[0026] Step S3 invokes the multi-agent teaching cognitive adaptation model to perform cross-dimensional feature fusion operations on the constructed multi-dimensional learner static feature vector set and the temporal learning state dataset. The model runs using a distributed parallel computing architecture, with 16 computing nodes processing data simultaneously, increasing processing efficiency to 12 times that of single-node computation. First, the static feature vector set is standardized by normalizing the parameters of each dimension to the [0,1] interval to eliminate the influence of differences in the magnitude of different parameters. The temporal learning state dataset is processed using a sliding window method with a window length of 30 minutes. Six statistical features, including mean, variance, and extreme values, are extracted from the dynamic parameters within the window to generate dynamic feature sub-vectors. Subsequently, an attention mechanism is used to assign weights to the standardized static feature vector and dynamic feature sub-vectors. Based on the importance of features to cognitive fit, basic weights ranging from 0.01 to 0.05 are assigned to the 44 parameters in the static feature vector, and dynamic weights ranging from 0.02 to 0.06 are assigned to the 26 parameters in the dynamic feature sub-vector. The weight values ​​are continuously optimized using gradient descent during model training. Next, element-wise multiplication of cross-dimensional features is performed, multiplying the static and dynamic features according to their corresponding dimensions to obtain the initial values ​​of the fused feature matrix. Finally, the initial fused feature matrix is ​​mapped using the model's built-in nonlinear transformation function, and then normalized to ensure that the matrix elements are within the range of [0,1]. The final result is a learner cognitive fit matrix with dimensions of number of learners × number of teaching content items. Each element in the matrix directly reflects the degree of fit between the corresponding learner and the teaching content.

[0027] Step S4 inputs the generated learner cognitive fit matrix into the personalized learning path generation model. The model has a built-in subject knowledge graph database, encompassing the knowledge systems of 12 grades and 9 subjects from elementary to high school. The knowledge graph includes 1200 core knowledge nodes, each labeled with 6 attributes such as knowledge point difficulty coefficient, suggested learning time, and prerequisite knowledge dependencies. First, the association weights of each node in the knowledge graph are extracted, including direct and indirect association weights. Direct association weights are divided into 9 levels (0.1-0.9) based on the logical dependence between knowledge points. Indirect association weights are calculated using the shortest path algorithm, with values ​​ranging from 0.01 to 0.5. Then, each element in the cognitive fit matrix is ​​multiplied by the association weight of its corresponding knowledge node to obtain a comprehensive node fit score. The knowledge nodes are then sorted from highest to lowest score to generate a node fit priority sequence. Based on the pre-dependencies between the sequence and knowledge nodes, a greedy algorithm is used to generate an initial set of multi-branch learning paths. Each learner generates 3-5 candidate paths, each containing 8-12 knowledge nodes. The path length is dynamically adjusted according to the subject module and the learner's learning progress. Finally, the initial path set is optimized by analyzing 20 key indicators from the learner's historical learning behavior data to identify redundant and key nodes in the paths. Redundant nodes are pruned, and key nodes have their learning time and practice intensity configured. The final output is an optimized personalized learning content sequence, which clearly indicates the learning order, learning time, supporting resource type, and assessment method for each knowledge node.

[0028] Step S5 utilizes a collaborative optimization algorithm for learning progress time-series prediction to perform rolling predictions on the state data during the learning path execution process. The algorithm operates in an online learning mode, updating the prediction model parameters every 30 minutes to ensure the timeliness and accuracy of the prediction results. First, the state data during the learning path execution process is filtered, selecting 32 core state parameters across three categories: cognitive state, behavioral interaction, and physiological signals, as the prediction input data. These are divided into training and validation sets according to timestamp order, with the training set accounting for 70% and the validation set for 30%. The rolling prediction window length is set to 2 hours, meaning that the learning progress changes for the next hour are predicted based on the state data from the previous 2 hours. Subsequently, feature filtering and dimensionality reduction are performed on the input data in the training set. Correlation analysis is used to remove redundant features with correlation coefficients below 0.3, reducing the input feature dimension from 32 to 18, thus improving the computational efficiency of the prediction model. Next, a temporal prediction model was constructed, combining the advantages of Long Short-Term Memory (LSTM) networks and gated recurrent units (GRUs). The model had four hidden layers with 64 neurons per layer. An adaptive learning rate optimization algorithm was used for model training, with an initial learning rate of 0.001, dynamically adjusted based on the prediction error on the validation set. Then, a collaborative optimization unit compared the prediction model's output with real-time evaluation data from the intelligent agent learning assessment and analysis system, calculating the absolute and relative errors between the predicted and measured values. When the relative error exceeded 10%, gradient descent was used to correct the prediction model's weight parameters, continuously optimizing prediction accuracy. The final output is a knowledge mastery trend vector with dimensions of number of learners × number of knowledge nodes. Each element in the vector represents the predicted future mastery level of the corresponding learner at a specific knowledge node.

[0029] Step S6, based on the knowledge mastery trend vector output by the learning progress time-series prediction collaborative optimization algorithm, and combined with the real-time evaluation results of the intelligent agent learning progress assessment and analysis system, involves dynamic control operations performed by a multi-agent collaborative unit. This collaborative unit comprises three sub-units: teaching strategy control, resource allocation control, and feedback interaction control. Each sub-unit communicates in real-time via a data bus, with the control command response time controlled within 100 milliseconds. First, the knowledge mastery trend vector and the real-time evaluation results are fused and analyzed, calculating their similarity coefficient. When the similarity coefficient is higher than 0.8, the trend vector is used as the primary control basis; when the similarity coefficient is lower than 0.6, the real-time evaluation results are used; and when the similarity coefficient is between 0.6 and 0.8, a weighted average method is used to fuse the two as the control basis, with weights set to 0.4 and 0.6 respectively. Subsequently, the teaching cognition adaptation parameters are dynamically adjusted, including adjusting the weight coefficients of each feature dimension in the multi-agent teaching cognition adaptation model. The feature weights corresponding to knowledge modules showing weak mastery in the trend vector are increased by 20%-30%, while the feature weights corresponding to modules showing good mastery are decreased by 10%-20%. Simultaneously, the configuration of learning path nodes was adjusted. For knowledge nodes predicted to be difficult to grasp, preparatory nodes were added or node content was split, extending the learning time of the original nodes by 30%-50%. For nodes predicted to be easy to grasp, the learning time was appropriately shortened or adjacent nodes were merged. Furthermore, the interactive feedback strategy was adjusted. For learners lagging behind, the frequency of real-time Q&A was increased, with the interval shortened to once every 15 minutes, and feedback primarily consisting of guiding prompts. For learners ahead of schedule, extended resources and in-depth exploration tasks were provided, with feedback primarily consisting of encouraging evaluations. Through multi-dimensional dynamic adjustment, a closed-loop teaching control chain was formed to ensure that the teaching process continuously adapts to the learners' real-time status and developmental needs.

[0030] Preferably, the expression for the multi-agent teaching cognitive adaptation model is:

[0031] in, For the cognitive fit between the i-th learner and the j-th teaching content, For multi-agent weight coefficients, For learner cognitive feature vectors, For teaching content attribute vectors, For the cognition-content matching matrix, For learners' physiological signal feature vectors, Adapting vectors for teaching scenarios The physiological-scene correlation matrix For learner interaction behavior vectors, For teaching feedback feature vectors, For behavior-feedback mapping matrix, This is the Sigmoid activation function.

[0032] Specifically, the multi-agent teaching cognitive adaptation model is based on the distributed feature fusion concept of multi-agent collaboration. It integrates three core data dimensions—cognitive features, physiological signals, and interactive behaviors—and is constructed using a combination of linear weighting and nonlinear activation. The core influencing factors of cognitive adaptation are clearly defined, with the matching degree between learners' cognitive features and teaching content attributes, the adaptation degree between physiological signals and teaching scenarios, and the fit between interactive behaviors and teaching feedback as the three core components. To reflect the differences in importance of each component, three weight coefficients are introduced, and their value ranges are determined through iterative training with 1000 sets of sample data. The cognitive-content matching weight ranges from 0.3 to 0.5, the physiological-scenario adaptation weight ranges from 0.2 to 0.4, and the behavioral-feedback fit weight ranges from 0.1 to 0.3, with a sum of 1. To quantify the matching relationship of each dimension, cosine similarity is used to calculate the linear correlation between cognitive features and content attributes. The nonlinear fluctuation characteristics of physiological signals are processed using the hyperbolic tangent function, and the correlation of interactive behavior feedback is mapped to the 0-1 interval using the sigmoid function. Finally, matrix operations are used to achieve the alignment and fusion of feature dimensions. During implementation, the model relies on 16 distributed computing nodes to process data in parallel. Each node is responsible for calculating one type of feature. The calculation results are aggregated through a data bus and then weighted and fused. This formula combines the independence of multi-dimensional data with the deep integration of cross-modal features. The values ​​have been verified by a large number of samples to ensure stability. It can accurately quantify the degree of fit between learners and teaching content, providing a core basis for the subsequent formulation of personalized teaching strategies.

[0033] Preferably, the expression for the personalized learning path generation model is:

[0034] in, The optimal learning path for the k-th learner. Generate weight coefficients for the path. Let be the mastery level of the k-th learner at the t-th knowledge node. Let be the pre-association weight between the t-th node and the n-th node. Let be the learning efficiency of the k-th learner at the t-th node. Let be the temporal correlation weight between the t-th node and the n-th node. Let be the interaction depth of the k-th learner at the t-th node. The weight is the association weight between the content of the t-th node and the n-th node.

[0035] Specifically, the personalized learning path generation model focuses on the node association characteristics of the knowledge graph and learner suitability. Based on the optimization idea of ​​a greedy algorithm, it constructs an objective function through weighted summation of multi-dimensional indicators. Three core evaluation indicators for learning path optimization are clearly defined: knowledge node mastery, learning efficiency, and interaction depth. These three are coupled with the weights of prior associations, temporal associations, and content associations between knowledge nodes, respectively, forming the core components of path evaluation. To balance the influence of each indicator, weight coefficients are determined through simulation experiments of 500 different learning scenarios. The weights for knowledge mastery are set at 0.4-0.6, learning efficiency at 0.2-0.3, and interaction depth at 0.1-0.2, ensuring that path generation prioritizes adaptation to the learner's knowledge base. During the formula establishment process, the argmax function is used to select the optimal path, and the association evaluation values ​​of all nodes in the path are integrated through summation operations to achieve quantitative comparison of multi-branch paths. In implementation, the model first analyzes 1200 core nodes and their relationships in the subject knowledge graph, extracts various relationship weights and quantifies them into a numerical range of 0.1-0.9. Then, it combines learner real-time adaptability data to calculate the comprehensive score of each path, and finally selects the path with the highest score as the optimal solution. This formula fully combines the structure of the knowledge system with the individual differences of learners. The parameter values ​​have been validated in various scenarios, and it can dynamically generate personalized learning sequences adapted to different learners, improving the targeting and efficiency of the learning process.

[0036] Preferably, the expression for the collaborative optimization algorithm for learning time series prediction is:

[0037] in, For the first The predicted value of learning progress at any given time. Let m be the measured value of the student's learning progress at time m. These are the fusion coefficients of the time series model. To optimize the adjustment coefficient, Let be the cognitive state feature vector at time m. Let m be the behavioral interaction feature vector at time m. Let be the physiological signal feature vector at time m. For prediction functions of Long Short-Term Memory (LSTM) networks, For the prediction function of the gated recurrent unit, To collaboratively optimize the objective function.

[0038] Specifically, the collaborative optimization algorithm for learning data time-series prediction integrates the temporal prediction advantages of Long Short-Term Memory (LSTM) networks and gated recurrent units (GRUs), introducing a collaborative optimization term to construct a hybrid prediction model. First, the temporal characteristics of learning data are analyzed, identifying historical learning data, cognitive state characteristics, behavioral interaction characteristics, and physiological signal characteristics as core input variables. LTM networks are used to capture long-term dependencies in the learning data, while GRUs are used to improve the response speed of short-term data. A weighted coefficient balances the prediction results of both, with the coefficient's value range of 0.3-0.7 determined through training with 200 sets of time-series data. To reduce prediction bias, a collaborative optimization term is introduced. An optimization objective function is constructed based on gradient descent. The prediction results are corrected by calculating the gradient of the deviation between predicted and measured values, with the optimization adjustment coefficient ranging from 0.05 to 0.15 to ensure the stability of the optimization process. When establishing the formula, a linear weighting method is used to integrate the three types of prediction outputs, preserving the predictive ability of the time-series model while improving prediction accuracy through the collaborative optimization term. In implementation, the algorithm updates model parameters every 30 minutes. Input data is divided into training and validation sets in a 7:3 ratio. Weighting and optimization coefficients are dynamically adjusted using the validation set. The prediction window length is set to 2 hours, enabling rolling predictions of learning progress changes over the next hour. This formula fully considers the complexity and dynamism of learning data, and the parameter values ​​have been validated with time-series data. It accurately outputs knowledge mastery trends, providing a forward-looking basis for teaching adjustments.

[0039] Preferably, the evaluation model expression of the intelligent agent learning assessment and analysis system is as follows:

[0040] in, Let p be the overall learning assessment value of the learner. To evaluate the dimensional weight coefficients, For knowledge testing feature vectors, For knowledge standard feature vectors, For knowledge assessment matrix, This is a feature vector representing behavioral characteristics. For behavioral standard feature vectors, For the behavior assessment matrix, This is a feature vector representing the physiological state. For physiological standard feature vectors, This is a physiological assessment matrix. It is the activation function for the exponential linear unit.

[0041] Specifically, the intelligent agent learning assessment and analysis system's evaluation model is based on a comprehensive evaluation concept of multi-dimensional learning data, integrating three evaluation dimensions—knowledge testing, behavioral performance, and physiological state—to construct a fusion model. The core characteristic indicators for each evaluation dimension are clearly defined: the knowledge testing dimension includes 8 core indicators, the behavioral performance dimension includes 7 core indicators, and the physiological state dimension includes 5 core indicators. Standardization processes map these indicators to a unified quantification range. In formula development, the softmax function is used to probabilistically process the knowledge testing results, the ReLU function enhances the non-linear expression of behavioral performance features, and the ELU function optimizes the adaptability to fluctuations in physiological state data. The three processing results are coupled with their corresponding evaluation matrices. To reflect the importance of each dimension, a comparative experiment with 300 sets of evaluation data was conducted to determine the weighting coefficients: the knowledge testing weight is 0.5-0.7, the behavioral performance weight is 0.2-0.3, and the physiological state weight is 0.1-0.2. During implementation, the system simultaneously collects real-time data across three dimensions, calculates the evaluation values ​​for each dimension through 16 parallel processing nodes, and then weights and merges them according to weight coefficients to generate a comprehensive learning assessment value in the range of 0-1. This formula comprehensively covers the core dimensions of learning assessment, and the parameter values ​​have been verified in multiple scenarios. It can objectively and comprehensively quantify the learner's overall learning status, providing precise data support for adjusting teaching strategies.

[0042] Preferably, the fusion model expression for the wearable physiological signal acquisition and AI intelligent health risk assessment is as follows:

[0043] in, The health risk assessment value for the q-th learner To integrate the weighting coefficients, The time decay factor, For the qth learner Multimodal physiological signal sequences at different times. For signal convolution kernel, For pooling window parameters, For normalization parameters, For convolution operations, For pooling operations, For standardized calculations.

[0044] Specifically, the wearable physiological signal acquisition and AI-powered intelligent health risk assessment fusion model is based on the temporal characteristics of physiological signals and the concept of multimodal fusion. It integrates convolution, pooling, and standardization operations to construct the risk assessment model. First, the dynamic change patterns of physiological signals are analyzed. Integral operations are used to process the continuous time-dimension physiological signal sequence, and an exponential decay factor is introduced to characterize the temporal decay characteristics of physiological signals. The decay factor ranges from 0.01 to 0.05 to ensure that recent signals dominate the assessment results. When establishing the formula, local features of physiological signals are extracted through convolution operations, with the convolution kernel size set to 3×3. Max pooling is used to compress the feature dimension, with the pooling window parameter set to 2×2. Standardization maps the feature data to the [-1, 1] interval, improving the model's generalization ability. To balance the influence of each feature extraction stage, the fusion weight coefficients are determined through training with 150 sets of physiological signal data. The integral operation weight ranges from 0.5 to 0.6, the pooling operation weight ranges from 0.2 to 0.3, and the standardization operation weight ranges from 0.1 to 0.2. In practice, wearable devices collect physiological signals 20 times per minute, which are then transmitted to the assessment model and sequentially processed through convolution, pooling, and standardization. Finally, a weighted fusion is used to generate a health risk assessment value. This formula fully adapts to the dynamic characteristics and multimodal attributes of physiological signals, and the parameter values ​​have been validated with clinical data. It can accurately assess learners' health risk status, providing technical support for health assurance during the teaching process.

[0045] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, the feature extraction unit of the multi-agent teaching cognitive adaptation model performs dimensional mapping on the cognitive style and knowledge mastery parameters in the static feature vector set to form a standardized cognitive feature matrix; S32, the model's temporal processing unit performs sliding window sampling on the behavioral sequences and interaction response parameters in the temporal learning state dataset to generate a dynamic behavioral feature tensor; S33, the model's cross-modal fusion unit uses an attention mechanism to perform weight allocation and element-wise multiplication on the standardized cognitive feature matrix and the dynamic behavioral feature tensor to obtain a fused feature matrix; S34, the model's adaptation calculation unit performs nonlinear transformation and normalization on the fused feature matrix to output a dimension-matched learner cognitive adaptation matrix.

[0046] Specifically, step S3 consists of four sub-steps, which constitute the feature fusion process of the multi-agent teaching cognitive adaptation model. S31 is executed by the model's feature extraction unit. For the two core parameters—cognitive style and knowledge mastery—in the static feature vector set, a dimension mapping algorithm is used to convert the raw data into a standardized matrix. During the mapping process, the feature dimension is uniformly set to 64 dimensions, and the quantization range of each dimension is controlled between 0 and 1 to ensure the comparability of different types of features. S32 uses the model's temporal processing unit to perform sliding window sampling on the dynamic learning state data. The window length is set to 15 minutes, and the step size is 5 minutes. Statistical calculations are performed on the behavioral sequences and interaction response parameters within each window to generate a 32×64 dynamic behavioral feature tensor, effectively preserving the local correlations of the temporal data. The process involves: S33, which invokes the cross-modal fusion unit and employs an attention mechanism to assign weights to the standardized cognitive feature matrix and the dynamic behavioral feature tensor. The basic weights for cognitive features account for 40%, and those for dynamic behavioral features account for 60%. Deep fusion of the two types of features is then achieved through element-wise multiplication. During the fusion process, the data processing rate is maintained at 1024 data entries per second to ensure real-time performance. S34 utilizes the fitness calculation unit to perform a nonlinear transformation on the fused feature matrix. The transformation uses a piecewise function and is subsequently normalized to concentrate the matrix element values ​​between 0.1 and 0.9. The final output is a cognitive fitness matrix whose dimensions match the number of learners and the amount of teaching content. This process, through four refinement steps, achieves comprehensive fusion of static and dynamic features. The parameter settings have been validated through 100 sets of experiments to ensure the accuracy of the fitness calculation, providing reliable data support for subsequent learning path generation.

[0047] Preferred, such as Figure 3 As shown, S4 includes the following sub-steps: S41, the knowledge graph parsing unit of the personalized learning path generation model traverses the node attributes and relationships in the subject knowledge graph, extracting the knowledge node weights and pre-dependencies; S42, the model's fit mapping unit matches the elements in the cognitive fit matrix with the corresponding knowledge node weights to generate a node fit priority sequence; S43, the model's path construction unit generates an initial multi-branch learning path set based on the node fit priority sequence and pre-dependencies using a greedy algorithm; S44, the model's path optimization unit combines learner historical learning behavior data to prune redundant nodes and strengthen labeled nodes in the initial path set, outputting an optimized personalized learning content sequence.

[0048] Specifically, step S4 consists of four sub-steps for constructing and optimizing the personalized learning path generation model. S41 involves the model's knowledge graph parsing unit traversing the subject knowledge graph using a depth-first search algorithm to extract attribute information and relationships from 1200 core knowledge nodes. The pre-dependency annotation accuracy for each node reaches 100%, and the association weights between nodes are calculated, with weight values ​​ranging from 0.1 to 0.9, directly reflecting the logical tightness of the knowledge nodes. S42 uses the model's fitness mapping unit to match the cognitive fitness matrix with the knowledge node weights. The matching uses a cosine similarity calculation method, with a similarity threshold set to 0.6. Nodes with similarity values ​​higher than the threshold are prioritized. The algorithm generates a node adaptation priority list of length 20-30. Step S43 uses a path construction unit to generate an initial set of multi-branch learning paths based on the priority sequence and prerequisite dependencies, employing a greedy algorithm. Each path includes 8-12 knowledge nodes, and the number of paths is set to 5 to ensure diversity in path selection. Step S44 uses a path optimization unit to prune redundant nodes and strengthen key nodes in the initial paths. The pruning criteria are nodes with an adaptation degree below 0.3 and no prerequisite dependencies. The strengthening method is to increase the number of supporting resources for key nodes to 3-5 types and extend the learning time to 1.5 times the basic time. This algorithm, through four refined steps, achieves a deep integration of knowledge graph and learner adaptation. The parameter settings conform to the laws of subject teaching, and the generated learning paths satisfy both the logic of the knowledge system and adapt to individual learner differences.

[0049] Preferred, such as Figure 4 As shown, S5 includes the following sub-steps: S51, the dataset partitioning unit of the learning path execution process divides the state data into training set and validation set according to timestamps, and determines the rolling prediction window length; S52, the feature engineering unit of the algorithm performs feature screening and dimensionality reduction on the state data in the training set, and extracts a subset of features strongly correlated with the knowledge mastery trend; S53, the prediction model training unit of the algorithm uses the feature subset to iteratively train the time series prediction model, adjusts the model hyperparameters through the validation set, and determines the optimal prediction parameter configuration; S54, the collaborative optimization unit of the algorithm compares the output result of the prediction model with the real-time data of the intelligent agent learning assessment and analysis system, corrects the prediction deviation through gradient descent, and outputs an accurate knowledge mastery trend vector.

[0050] Specifically, step S5 consists of four sub-steps for predicting and optimizing the learning progress time-series prediction collaborative optimization algorithm. S51 involves splitting the state data by timestamp using the algorithm's dataset partitioning unit, with a partition ratio of 70% for the training set and 30% for the validation set. The rolling prediction window length is set to 2 hours, and the number of state data samples included within the window is no less than 1000 to ensure sufficient training data. S52 uses the feature engineering unit to filter features in the training set data, employing correlation analysis to calculate the correlation coefficient between each feature and the knowledge mastery trend. A screening threshold of 0.3 is set, retaining 18 core features with correlation coefficients higher than the threshold, reducing the input dimension from 32 to 18, and improving model computational efficiency. S53 calls the prediction model for training. The training unit iteratively trains the time-series prediction model using a subset of core features, with 500 iterations and an initial learning rate of 0.001, decreasing by 50% every 100 iterations. Prediction errors are monitored in real-time using a validation set; training stops when the error falls below 5%, determining the optimal hyperparameter configuration. The S54 co-optimization unit compares the prediction results with real-time data from the intelligent agent's learning assessment and analysis system, calculating absolute and relative errors. When the relative error exceeds 10%, gradient descent is used to correct the prediction model's weight parameters with a step size of 0.0001. After 3-5 rounds of correction, a knowledge mastery trend vector is output. This four-step implementation achieves both accuracy and dynamic optimization in learning prediction. The parameter settings balance computational efficiency and prediction accuracy, enabling advance prediction of learners' knowledge mastery and providing a forward-looking basis for instructional adjustments.

[0051] This invention addresses the pain points of traditional teaching methods—namely, the use of diverse indicator parameters and complex multi-step calculations—to adapt to the highly differentiated characteristics of individual learners. It overcomes the limitations of traditional teaching methods, such as relying on single data sources, providing only partial understanding, and rigid adaptation. This lays a solid data and algorithmic foundation for precise and dynamic AI-assisted teaching. A learner's learning state is the result of the combined effects of multiple dimensions, including cognitive ability, knowledge reserves, behavioral habits, emotional fluctuations, and physiological responses. A single or limited set of indicators cannot fully depict an individual's overall picture. Basic data such as answer accuracy and learning time only provide a superficial reflection of learning outcomes, completely ignoring crucial, implicit factors that influence learning effectiveness, such as focus, fatigue, cognitive preferences, and emotional state. Therefore, this invention breaks down learning from three core dimensions: static inherent traits, dynamic real-time behavior, and internal physiological state. It further refines these into dozens of sub-indicators, including cognitive style, knowledge mastery, learning behavior sequence, interactive response characteristics, resource access trajectory, task completion behavior, question-and-answer feedback content, and indicators such as heart rate, respiratory rate, and skin conductivity, covering the entire learning process and all its characteristics. These indicators are by no means redundant designs, but rather the core units for building accurate learner profiles. Static indicators anchor an individual's long-term stable learning characteristics, dynamic indicators capture real-time changes during the learning process, and physiological indicators map physical and mental states that are difficult to perceive through external behavior. Multiple indicators complement and cross-verify each other, fundamentally avoiding misjudgments in teaching caused by single data biases, and providing comprehensive, solid, and reliable data support for subsequent personalized adaptation.

[0052] This invention employs multi-step complex computations to process data and parameters because the raw data collected for various indicators suffers from inherent defects such as disorganized formats, inconsistent dimensions, discrete temporal sequences, noise interference, and weak correlations. Simple addition, subtraction, multiplication, and division operations cannot effectively transform the data, let alone achieve cross-dimensional fusion, temporal pattern mining, and accurate trend prediction. Therefore, a series of progressively complex computations are necessary to clean, fuse, transform, and optimize the data. In the raw data, static data dimensions such as cognitive style and knowledge mastery are scattered, while dynamic data such as heart rate and learning behavior are disordered and fluctuate frequently. Physiological signals also contain a large amount of environmental interference noise; direct use of such data would lead to distorted results. Therefore, this invention first performs dimensional mapping and standardization on static data to eliminate differences in the magnitude of different indicators; then, it performs sliding window sampling on dynamic time-series data to extract time-series features and preserve data correlations; next, it achieves cross-modal data fusion through an attention mechanism, rationally allocating weights for different indicators and highlighting the influence of key features; subsequently, it undergoes multiple operations such as nonlinear transformation, normalization, greedy algorithms, time-series model training, and gradient descent to correct biases, gradually transforming the scattered raw data into core results that can directly guide teaching, such as cognitive fit matrices, optimal learning paths, knowledge mastery trend vectors, and comprehensive learning assessment values. These operations are interconnected and indispensable, with each step precisely designed to address data deficiencies and teaching needs. This ensures both data dimensional matching and effective feature extraction, while reducing errors through multiple rounds of optimization, achieving a closed-loop transformation from raw data to precise teaching strategies. This is something that simple calculations cannot achieve and is the core technical support for ensuring the accuracy and dynamism of teaching.

[0053] The five physiological indicators (heart rate, respiratory rate, skin conductivity, etc.) collected in this invention are direct quantitative representations of learners' internal physical and mental state, effectively filling the perceptual blind spots of cognitive and behavioral data. The transformation process spans the entire process of data collection, feature extraction, cross-dimensional fusion, model calculation, and teaching regulation, covering multiple core aspects such as learning assessment, health warning, and strategy optimization. At the data transformation level, physiological data is first captured in real-time at high frequency using wearable devices, forming a continuous time-series physiological dataset. Then, the original physiological signals are processed through convolution, pooling, and standardization to filter environmental noise and extract signal fluctuation features, transforming the disordered time-series data into standardized physiological feature vectors. Subsequently, these physiological feature vectors are deeply integrated with cognitive and behavioral features, participating in core model calculations such as cognitive fit calculation, comprehensive learning assessment, and knowledge mastery trend prediction. These become key parameters for teaching adaptation, learning assessment, and trend prediction, and are linked with data from other dimensions throughout the process, jointly influencing teaching decision-making outcomes. In practical terms, physiological indicators can accurately reflect states that are difficult to detect through external behavior: abnormal heart rate and respiratory rate can directly reflect learning fatigue, anxiety, or excessive tension, while fluctuations in skin conductivity can accurately capture changes in concentration. These data can be used to conduct real-time health risk assessments and provide early warnings of potential physical and mental health risks caused by over-learning. At the same time, they provide core basis for dynamic teaching regulation, such as automatically reducing learning difficulty and extending rest time when fatigue is detected, and adjusting the feedback rhythm and optimizing the content presentation when insufficient concentration is detected. Ultimately, this balances learning efficiency and physical and mental health, greatly improving the comprehensiveness and humanization of teaching adaptation, so that personalized teaching not only matches the learner's knowledge level but also deeply adapts to their real-time physical and mental state.

[0054] like Figure 5As shown, an AI-assisted teaching platform based on multi-agent collaboration is applied to AI-assisted teaching methods based on multi-agent collaboration. The platform includes: a multimodal learner data acquisition and preprocessing module, a distributed multi-agent teaching cognitive adaptation computation module, a knowledge graph-driven personalized learning path generation module, a time-series learning prediction and collaborative optimization module, an agent dynamic learning assessment and feedback module, and a multi-agent collaborative teaching control and resource scheduling module. The multimodal learner data acquisition and preprocessing module collects physiological signals, behavioral data, and cognitive feature parameters through a wearable device interface and an interactive terminal, outputting a standardized dataset. The distributed multi-agent teaching cognitive adaptation computation module receives the standardized dataset and, through multi-agent collaboration... The system calculates and generates a cognitive fit matrix, establishing a data transmission link with the personalized learning path generation module. The personalized learning path generation module generates a learning path sequence based on the cognitive fit matrix and knowledge graph data, and sends it to the time-series learning prediction and collaborative optimization module. The time-series learning prediction and collaborative optimization module predicts the path execution status and outputs a trend vector to the agent dynamic learning assessment and feedback module. The agent dynamic learning assessment and feedback module combines real-time assessment data to generate control commands, which are transmitted to the multi-agent collaborative teaching control and resource scheduling module. The multi-agent collaborative teaching control and resource scheduling module adjusts the parameter configurations and resource allocation schemes of each module according to the control commands, forming a closed-loop collaborative operation architecture.

[0055] This invention integrates different scalar and vector parameters for unified calculation. Through standardization, dimension alignment, and weight adaptation mechanisms, it eliminates parameter type differences and constructs a logically consistent quantitative calculation system. First, for scalar parameters such as learning time and answer accuracy (single-value quantitative indicators), linear transformations are used to map them to a unified range of 0-1, ensuring consistency with the numerical range of vector parameters. For vector parameters such as cognitive feature vectors and behavioral sequence vectors (multi-dimensional data), matrix transposition and dimension expansion are used to achieve dimensional matching with scalar parameters. For example, the 8-dimensional vector parameter of learner cognitive style is expanded into a dimension that matches the scalar parameter of knowledge mastery through matrix operations. Second, weight coefficients are introduced to quantify the importance of different types of parameters. For example, in the multi-agent teaching cognitive adaptation model, the weight coefficient of cognitive feature vectors is set to 0.3-0.5, and the weight coefficient of physiological signal scalars is set to 0.2-0.4. Weighted fusion achieves the organic combination of different types of parameters. Meanwhile, by leveraging activation functions and matrix operations to handle the nonlinear relationships between parameters—for example, using the sigmoid function to map the coupling results of vectors and scalars to a reasonable range—the validity of the calculation results is ensured. Furthermore, all parameters undergo consistency verification during preprocessing. Scalar parameters are expanded according to vector dimensions, and vector parameters are adjusted to the target dimension through dimensionality reduction or expansion. For instance, the 12-dimensional vector of the learning behavior sequence and the scalar of the answer accuracy rate are adapted together to a 64-dimensional calculation dimension. Ultimately, this achieves collaborative calculation of different types of parameters in the same formula, accurately quantifying the comprehensive impact of each parameter on core objectives such as teaching adaptation, path generation, and learning prediction.

[0056] This AI-assisted teaching method and platform, based on multi-agent collaboration, comprehensively captures learners' static characteristics and dynamic learning states through a multi-dimensional data acquisition mechanism. It breaks through the limitations of fragmented data in traditional technologies, achieving systematic integration and cross-dimensional analysis of cognitive, behavioral, and physiological data. Relying on a collaborative operational functional system, it constructs a closed-loop chain from data acquisition, feature fusion, path generation to learning prediction and dynamic adjustment, replacing the previous loosely structured model where each function operated independently. This allows for close linkage among all aspects of teaching, significantly improving the coherence and accuracy of teaching decisions. This design effectively overcomes the shortcomings of existing technologies in comprehensively perceiving learners' states, enabling teaching strategies to be more tailored to learners' actual situations through multi-dimensional data integration.

[0057] This invention's platform dynamically generates highly adaptable learning content sequences based on learner individual characteristics and real-time learning status. Simultaneously, through temporal prediction and dynamic adjustment mechanisms, it continuously optimizes teaching resource allocation and interactive feedback strategies, fully considering the cognitive differences, learning pace, and dynamically changing needs of different learners. A multi-module collaborative scheduling mechanism enables efficient allocation of teaching resources and precise control of teaching strategies, driving the transformation of teaching models from uniform and static to personalized and dynamic. This design successfully overcomes the shortcomings of traditional platforms, such as insufficient personalization and lack of targeted path adjustments, making teaching assistance more flexible and adaptable, providing customized teaching support for learners with different needs, and significantly improving the overall teaching assistance effect.

[0058] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0059] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An AI-assisted teaching method based on multi-agent collaboration, characterized in that, Includes the following steps: S1, through the distributed nodes of the multi-agent teaching cognitive adaptation model, collect learners’ cognitive style, knowledge mastery, learning behavior sequence and interaction response characteristics, and construct a multi-dimensional learner static feature vector set; S2, through the real-time data interaction interface of the intelligent agent learning assessment and analysis system, dynamically captures the resource access trajectory, task completion behavior, question and answer feedback content and physiological signal fluctuation parameters during the learning process, forming a time-series learning state dataset; S3 calls the multi-agent teaching cognitive adaptation model to perform cross-dimensional feature fusion of the static feature vector set and the temporal learning state dataset to generate a learner cognitive adaptation matrix. S4. Input the cognitive fit matrix into the personalized learning path generation model, and combine it with the node association weights of the subject knowledge graph to dynamically generate a multi-branch learning content sequence. S5 uses a learning time-series prediction collaborative optimization algorithm to perform rolling prediction of the state data during the execution of the learning path and outputs a knowledge mastery trend vector. S6, based on the real-time evaluation results of the trend vector and the agent learning assessment and analysis system, dynamically adjusts the teaching cognitive adaptation parameters, learning path node configuration and interactive feedback strategies by the multi-agent collaborative unit.

2. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, The expression for the multi-agent teaching cognitive adaptation model is: ,in, The cognitive fit between the i-th learner and the j-th teaching content. For multi-agent weight coefficients, For learner cognitive feature vectors, For teaching content attribute vectors, For cognition-content matching matrix, For learners' physiological signal feature vectors, Adapting vectors for teaching scenarios The physiological-scene correlation matrix For learner interaction behavior vectors, For teaching feedback feature vectors, For behavior-feedback mapping matrix, This is the Sigmoid activation function.

3. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, The expression for the personalized learning path generation model is: ,in, The optimal learning path for the k-th learner. Generate weight coefficients for the path. Let be the mastery level of the k-th learner at the t-th knowledge node. Let be the pre-association weight between the t-th node and the n-th node. Let be the learning efficiency of the k-th learner at the t-th node. Let be the temporal correlation weight between the t-th node and the n-th node. Let be the interaction depth of the k-th learner at the t-th node. The weight is the association weight between the content of the t-th node and the n-th node.

4. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, The expression for the collaborative optimization algorithm for predicting learning progress over time is: ,in, For the first The predicted value of learning progress at any given time. Let m be the measured value of the student's learning progress at time m. These are the fusion coefficients of the time series model. To optimize the adjustment coefficient, Let be the cognitive state feature vector at time m. Let m be the behavioral interaction feature vector at time m. Let be the physiological signal feature vector at time m. For prediction functions of Long Short-Term Memory (LSTM) networks, For the prediction function of the gated recurrent unit, To collaboratively optimize the objective function.

5. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, The evaluation model expression of the intelligent agent learning assessment and analysis system is as follows: ,in, Let p be the overall learning assessment value of the learner. To evaluate the dimensional weight coefficients, For knowledge testing feature vectors, For knowledge standard feature vectors, For knowledge assessment matrix, This is a feature vector representing behavioral characteristics. For behavioral standard feature vectors, For the behavior assessment matrix, This is a feature vector representing the physiological state. For physiological standard feature vectors, This is a physiological assessment matrix. It is the activation function for the exponential linear unit.

6. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, The fusion model expression for wearable physiological signal acquisition and AI intelligent health risk assessment is as follows: ,in, Let q be the health risk assessment value of the learner. To integrate the weighting coefficients, The time decay factor, For the qth learner Multimodal physiological signal sequences at different times. For signal convolution kernel, For pooling window parameters, For normalization parameters, For convolution operations, For pooling operations, For standardized calculations.

7. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, S3 includes the following steps: S31, the feature extraction unit of the multi-agent teaching cognitive adaptation model performs dimensional mapping on the cognitive style and knowledge mastery parameters in the static feature vector set to form a standardized cognitive feature matrix; S32, through the model's temporal processing unit, performs sliding window sampling on the behavioral sequences and interaction response parameters in the temporal learning state dataset to generate a dynamic behavioral feature tensor; S33, the cross-modal fusion unit of the model is called to use the attention mechanism to perform weight allocation and element-wise multiplication on the standardized cognitive feature matrix and the dynamic behavior feature tensor to obtain the fusion feature matrix; S34 utilizes the model's fitness calculation unit to perform nonlinear transformation and normalization on the fused feature matrix, outputting a dimension-matched learner cognitive fitness matrix.

8. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, S4 includes the following sub-steps: S41, the knowledge graph parsing unit of the personalized learning path generation model traverses the node attributes and relationships in the subject knowledge graph and extracts the knowledge node weights and prerequisite dependencies. S42, the model's fitness mapping unit associates and matches the elements in the cognitive fitness matrix with the corresponding knowledge node weights to generate a node fitness priority sequence. S43, the initial multi-branch learning path set is generated by using a greedy algorithm based on the node adaptation priority sequence and the pre-dependency relationship through the path building unit of the model; S44 utilizes the model's path optimization unit, combined with learners' historical learning behavior data, to prune redundant nodes and reinforce labeled nodes in the initial path set, outputting an optimized personalized learning content sequence.

9. The AI-assisted teaching method based on multi-agent collaboration according to claim 1, characterized in that, S5 includes the following steps: S51, the dataset partitioning unit of the collaborative optimization algorithm for learning time series prediction divides the state data during the execution of the learning path into training set and validation set according to timestamp, and determines the length of the rolling prediction window; S52, the feature engineering unit of the algorithm performs feature filtering and dimensionality reduction on the state data in the training set, and extracts a subset of features that are strongly correlated with the trend of knowledge mastery. S53, the prediction model training unit of the algorithm uses a feature subset to iteratively train the time series prediction model, and adjusts the model hyperparameters through the validation set to determine the optimal prediction parameter configuration; S54 uses the algorithm's collaborative optimization unit to compare the output of the prediction model with the real-time data of the intelligent agent's learning assessment and analysis system, corrects the prediction deviation through gradient descent, and outputs an accurate knowledge mastery trend vector.

10. An AI-assisted teaching platform based on multi-agent collaboration, characterized in that: The platform is applied to the AI-assisted teaching method based on multi-agent collaboration as described in claim 1, including: a multimodal learner data acquisition and preprocessing module, a distributed multi-agent teaching cognitive adaptation calculation module, a knowledge graph-driven personalized learning path generation module, a time-series learning prediction and collaborative optimization module, an agent dynamic learning assessment and feedback module, and a multi-agent collaborative teaching regulation and resource scheduling module. The multimodal learner data acquisition and preprocessing module collects physiological signals, behavioral data, and cognitive feature parameters through a wearable device interface and interactive terminal, and outputs a standardized dataset. The distributed multi-agent teaching cognitive adaptation computing module receives the standardized dataset, generates a cognitive adaptation matrix through multi-agent parallel computing, and establishes a data transmission link with the personalized learning path generation module. The personalized learning path generation module generates a learning path sequence based on the cognitive adaptation matrix and knowledge graph data, and sends it to the time-series learning prediction and collaborative optimization module. The time-series learning prediction and collaborative optimization module predicts the path execution status and outputs a trend vector to the agent dynamic learning assessment and feedback module. The agent dynamic learning assessment and feedback module generates control instructions based on real-time assessment data and transmits them to the multi-agent collaborative teaching control and resource scheduling module. The multi-agent collaborative teaching control and resource scheduling module adjusts the parameter configuration and resource allocation scheme of each module according to the control instructions, forming a closed-loop collaborative operation architecture.

Images