Fusion plan rationality adaptive learning path planning method and ai teaching terminal

Abstract

The application discloses a fusion plan rationality adaptive learning path planning method and an AI teaching terminal, belongs to the intelligent teaching field, S1, three types of basic constraint data of standardized learning teaching resource library, directed acyclic course knowledge graph and quantitative teaching constraint set are constructed; S2, a learner comprehensive portrait data set is generated; S3, a double-layer heterogeneous multi-agent collaborative interaction architecture is built, and an unchecked initial path set is output; S4, the unchecked initial path set is evaluated quantitatively, and path item scores and comprehensive rationality evaluation results are output; S5, a multi-level candidate adaptive path set is obtained through iteration; S6, a space cognition-knowledge progression-regulation compliance multi-objective optimization mathematical model is constructed, and a unique global optimal learning path and a complete planning explanation document are obtained; and S7, a closed loop is updated. The fusion plan rationality adaptive learning path planning method and the AI teaching terminal can generate an adaptive learning path with global optimality, teaching scientificity and explainability, and can be widely applied to personalized learning recommendation and intelligent learning guidance scenes of a hybrid classroom AI teaching terminal.

Description

Technical Field

[0001] This invention relates to the field of intelligent teaching technology, and in particular to a method for adaptive learning path planning based on integrated planning rationality and an AI teaching terminal. Background Technology

[0002] With the popularization of blended learning models and the widespread application of AI teaching terminals, adaptive learning path planning has become a core technology of intelligent learning guidance systems. Learning path planning aims to dynamically sort learning and teaching resources according to learners' knowledge base, learning ability, and personalized needs, and construct the optimal learning trajectory, thereby improving learning efficiency and teaching quality.

[0003] Existing learning path planning techniques mainly suffer from the following four types of shortcomings: First, the planning model is too simplistic, making it difficult to balance optimality and interpretability. One type of solution uses reinforcement learning, genetic algorithms, etc., for global optimization, but the model is a black box structure and cannot provide the teaching logic for path generation, making it difficult for learners and teachers to trust the recommendation results. Another type of solution uses a predefined engine or large language model to generate interpretable paths, but it lacks global optimization capabilities and is prone to problems such as redundant teaching resources and constraint conflicts, making it difficult to guarantee the overall optimality of the path.

[0004] Second, the evaluation of the rationality of learning plans is based on a single dimension and lacks a systematic quantitative standard. Existing solutions mostly only verify the correctness of the dependencies between knowledge points, without constructing a multi-dimensional evaluation system based on educational theory. They are unable to quantitatively assess core teaching indicators such as cognitive load suitability and the rationality of difficulty progression, resulting in learning plans that often have problems such as cognitive overload and leaps in difficulty that violate the laws of learning.

[0005] Third, the dynamic adaptive capability is insufficient. Most solutions adopt a periodic batch update path, which cannot trigger replanning based on learners' real-time learning status, changes in teaching resources, and adjustments to regulations. This makes it difficult to adapt to the real-time interaction and dynamic adjustment needs of AI teaching terminals in blended classrooms. Summary of the Invention

[0006] The purpose of this invention is to provide an adaptive learning path planning method and an AI teaching terminal for integrating planning rationality, thereby solving the aforementioned technical problems.

[0007] To achieve the above objectives, this invention provides an adaptive learning path planning method for integrating plan rationality, comprising the following steps: S1. Collect course knowledge graph, teaching resource metadata, and teaching management regulations, and sequentially complete multi-dimensional feature annotation of teaching resources, construction of knowledge point topology relationships, and structured coding of constraint regulations. Output three types of basic constraint data: standardized learning and teaching resource library, directed acyclic course knowledge graph, and quantitative teaching constraint set. S2. Collect raw data of multimodal interaction from AI teaching terminals, complete data cleaning and standardization, and calculate academic risk markers, full-dimensional knowledge point mastery vectors, and individual cognitive capacity indicators through the MFFEP multi-feature fusion academic risk prediction model, SIKT sequence knowledge tracking model, and cognitive load three-component model, respectively. Integrate learners' basic attributes to generate standardized learner feature vectors and output a comprehensive learner profile dataset. S3. Construct a two-layer heterogeneous multi-agent collaborative interaction architecture. The lower-layer learner agent, teaching resource agent, and teaching regulation agent respectively load the learner comprehensive profile output by S2, the standardized learning and teaching resource library output by S1, and the quantitative teaching constraint set and synchronize their status in real time. The upper-layer learning analysis agent receives the lower-layer aggregated global status and generates a standardized learning diagnosis report. The path planning agent uses a multi-agent joint deep Q network as the bottom-layer global optimization module and an educational fine-tuning large language model as the upper-layer fine-tuning module. Based on the S1 course knowledge graph, it generates an initial learning path and supporting interpretable recommendation text, and outputs an unverified initial path set. S4. Based on cognitive load theory, zone of proximal development theory, and teaching compliance criteria, a three-dimensional fusion plan rationality quantification system is constructed. A unified standardized calculation formula is used to calculate the three sub-indicators of cognitive load matching degree, knowledge sequence progression, and teaching resource compliance. The weighted fusion is used to obtain the comprehensive rationality score. The unverified initial path set output from S3 is quantitatively evaluated one by one, and the path sub-scores and comprehensive rationality evaluation results are output. S5. Deploy the reflection and verification agent to read all path sub-scores and comprehensive rationality evaluation results output by S4, compare them with the preset comprehensive rationality threshold to complete the standard determination. If the path does not meet the standard, locate the weak dimension with low score and generate a targeted iterative modification instruction to be sent back to the path planning agent of S3 for cyclic regeneration, and output a multi-level candidate adaptation path set after the iteration ends. S6. Construct a multi-objective optimization mathematical model of spatial cognition-knowledge progression-regulation compliance, introduce the NSGA-II multi-objective optimization algorithm to perform global Pareto optimization on the multi-level candidate adaptation path set output by S5, and simultaneously use AD axiom design to complete the verification of teaching logic compatibility, and select the unique global optimal learning path and complete planning explanation document. S7. Compile the optimal learning path output by S6 into a standardized guided learning parameter file for AI terminal visualization and deploy it. Collect new interactive behavior data of terminal learners in real time and feed it back to S2 to update the comprehensive profile of learners. Iterate the training parameters to dynamically adjust the model and periodically update the modular learning teaching resource parameter library of S1 to achieve full-cycle adaptive closed-loop update.

[0008] Therefore, the beneficial effects of the above-mentioned adaptive learning path planning method and AI teaching terminal based on the rationality of the fusion plan are as follows: 1. By utilizing the triple embedding encoding of answer sequence and the dual-branch attention network of sequence-knowledge point interaction, we can perform deep feature mining on learners' scattered and sparse answer interaction behaviors. Combined with the knowledge graph topological prior constraints, we can accurately model the dynamic knowledge mastery status. By using the exponential moving average smoothing algorithm to filter out accidental answer noise such as careless wrong answers and random guessing, we can output a stable, continuous, and quantifiable full-dimensional knowledge point mastery vector. This enables objective, quantifiable, and highly robust intelligent diagnosis of individual learning, solving the technical defects of traditional learning assessment, which is highly subjective and interferes with the judgment of single abnormal answer results.

[0009] 2. Based on learners' individual cognitive capacity and knowledge mastery profiles, a hierarchical teaching resource screening mechanism is constructed. Candidate learning and teaching resources are filtered step by step from multiple dimensions such as teaching resource availability, knowledge point topological relevance, cognitive load threshold, and matching degree of learning difficulty. It also supports dynamic refresh triggered by learning behavior and updates to the teaching resource library, so as to realize the precise teaching resource push of thousands of questions for learners with different weak knowledge points and different cognitive capacity limits, avoiding the recommendation of learning content that is beyond the scope of the curriculum, overloaded, or mismatched in difficulty, and conforming to the cognitive load carrying capacity of the human brain.

[0010] 3. By using a differentiated weight bias strategy in the reinforcement learning reward function, multiple candidate learning paths with different orientations such as cognitive load priority, knowledge progression priority, and policy compliance priority are generated in batches. Targeted path optimization is achieved for high-risk students with weak cognition, students with weak knowledge gaps, and standardized teaching management scenarios. This overcomes the limitation that a single optimal path cannot adapt to different learning groups and teaching objectives, and improves the scenario adaptability of the path scheme.

[0011] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0012] Figure 1 This is a flowchart of the adaptive learning path planning method for integrating planning rationality according to the present invention. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0014] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0015] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0016] like Figure 1 As shown, the adaptive learning path planning method for integrating plan rationality includes the following steps: S1. Collect course knowledge graph, teaching resource metadata, and teaching management regulations, and sequentially complete multi-dimensional feature annotation of teaching resources, construction of knowledge point topology relationships, and structured coding of constraint regulations. Output three types of basic constraint data: standardized learning and teaching resource library, directed acyclic course knowledge graph, and quantitative teaching constraint set. S2. Collect raw data of multimodal interaction from AI teaching terminals, complete data cleaning and standardization, and calculate academic risk markers, full-dimensional knowledge point mastery vectors, and individual cognitive capacity indicators through the MFFEP multi-feature fusion academic risk prediction model, SIKT sequence knowledge tracking model, and cognitive load three-component model, respectively. Integrate learners' basic attributes to generate standardized learner feature vectors and output a comprehensive learner profile dataset. S3. Construct a two-layer heterogeneous multi-agent collaborative interaction architecture. The lower-layer learner agent, teaching resource agent, and teaching regulation agent respectively load the learner comprehensive profile output by S2, the standardized learning and teaching resource library output by S1, and the quantitative teaching constraint set and synchronize their status in real time. The upper-layer learning analysis agent receives the lower-layer aggregated global status and generates a standardized learning diagnosis report. The path planning agent uses a multi-agent joint deep Q network as the bottom-layer global optimization module and an educational fine-tuning large language model as the upper-layer fine-tuning module. Based on the S1 course knowledge graph, it generates an initial learning path and supporting interpretable recommendation text, and outputs an unverified initial path set. S4. Based on cognitive load theory, zone of proximal development theory, and teaching compliance criteria, a three-dimensional fusion plan rationality quantification system is constructed. A unified standardized calculation formula is used to calculate the three sub-indicators of cognitive load matching degree, knowledge sequence progression, and teaching resource compliance. The weighted fusion is used to obtain the comprehensive rationality score. The unverified initial path set output from S3 is quantitatively evaluated one by one, and the path sub-scores and comprehensive rationality evaluation results are output. S5. Deploy the reflection and verification agent to read all path sub-scores and comprehensive rationality evaluation results output by S4, compare them with the preset comprehensive rationality threshold to complete the standard determination. If the path does not meet the standard, locate the weak dimension with low score and generate a targeted iterative modification instruction to be sent back to the path planning agent of S3 for cyclic regeneration, and output a multi-level candidate adaptation path set after the iteration ends. S6. Construct a multi-objective optimization mathematical model of spatial cognition-knowledge progression-regulation compliance, introduce the NSGA-II multi-objective optimization algorithm to perform global Pareto optimization on the multi-level candidate adaptation path set output by S5, and simultaneously use AD axiom design to complete the verification of teaching logic compatibility, and select the unique global optimal learning path and complete planning explanation document. S7. Compile the optimal learning path output by S6 into a standardized guided learning parameter file for AI terminal visualization and deploy it. Collect new interactive behavior data of terminal learners in real time and feed it back to S2 to update the comprehensive profile of learners. Iterate the training parameters to dynamically adjust the model and periodically update the modular learning teaching resource parameter library of S1 to achieve full-cycle adaptive closed-loop update.

[0017] Step S1 specifically includes the following steps: S11. Collect all original metadata of teaching resources, perform multi-dimensional feature annotation and group cognitive load calibration, and build a standardized learning and teaching resource library; S111. Collect the original metadata of all learning and teaching resources within the curriculum system, covering four types of teaching resources: videos, exercises, documents, and practical training projects. Extract six basic attributes: unique identifier of teaching resources, type of teaching resources, content description, chapter to which they belong, and original duration, to form a set of original teaching resource metadata. S112. Perform multi-dimensional feature annotation on each type of teaching resource. The annotation dimensions include four teaching attributes: corresponding knowledge point number, list of prerequisite knowledge points, difficulty level label, and teaching objective level. In this embodiment, the difficulty level is divided into 1 to 5 levels according to the cognitive objective. Level 1 corresponds to basic content of memorization and understanding, and level 5 corresponds to advanced content of comprehensive application. S113. Based on the historical learning data of all learners, calculate the average learning time for each teaching resource as the standard cognitive load for that teaching resource: ; In the formula, For the first Standard cognitive load of each teaching resource; For the first Learners learn the first Effective learning time for each teaching resource; The total number of learners who have studied this teaching resource for history; S114. Structure-encode the basic attributes, teaching characteristics, and standard cognitive load of all teaching resources, and establish an index with the unique identifier of the teaching resource as the primary key to generate a standardized learning and teaching resource database. S12. Extract the knowledge point entities and prerequisite dependencies of the course to construct a hierarchical directed acyclic course knowledge graph; S121. Extract all core knowledge point entities from the course syllabus, chapter lesson plans, and assessment instructions, assign a unique number to each knowledge point, and record the three basic attributes of knowledge point name, module to which it belongs, and assessment weight to form a set of knowledge point nodes. S122. Based on the teaching logic, define the prerequisite dependencies between knowledge points and generate a set of directed edges; if learning knowledge points... It is essential to master the knowledge points Given this premise, then establish from point to A directed edge represents a successor-precedence relationship; S123. Based on the topological sorting algorithm, knowledge points are hierarchically divided, and the quantitative difficulty level of each knowledge point is calculated in combination with the assessment weight: ; In the formula, For the first The quantitative difficulty level of each knowledge point, with a value range of [value range missing]. ; For the first The hierarchical number of each knowledge point in the topological sorting, with the bottom-level basic knowledge point being level 1; For the first The assessment weight for each knowledge point, with a value range of [value range missing]. ; This is the weighting adjustment coefficient, with a default value of 0.8; S124. Verify the acyclicity of the knowledge point topology to ensure there are no directed cycles with circular dependencies. Finally, generate a directed acyclic course knowledge graph with knowledge points as nodes, prerequisite relationships as directed edges, and difficulty level as node attributes. ; S13. Analyze the provisions of teaching management regulations, implement the structured coding and quantitative conversion of the constraint provisions, and construct a set of quantitative teaching constraints; S131. Extract three core constraints related to the learning path from the three types of regulations: talent training program, course syllabus, and assessment management method: prerequisite compliance constraint, credit achievement constraint, and total time constraint. S132. Transform prior compliance constraints into calculable, quantifiable provisions: ; In the formula, First, adjust the constraint satisfaction rate; The number of teaching resources in the path that meet the prerequisite requirements; This represents the total number of teaching resources included in the path. S133. Transform credit achievement constraints into calculable quantitative regulations: The sum of the weights of the assessment knowledge points covered by the learning path must reach the minimum percentage of the total course credit requirement; the credit achievement rate is defined as the ratio of the sum of the assessment weights of the knowledge points covered by the path to the total assessment weight, calculated using the following formula: ; In the formula, For the credit achievement rate; The set of knowledge points included in the learning path; For the first The assessment weight of each knowledge point; This represents the total number of all knowledge points in the course. S134. Transform the total duration constraint into a computable quantification: ; In the formula, For duration compliance rate; This is the maximum study time limit; Total study time; S135. Assign preset weights to the three types of constraints to form a structured set of quantitative teaching constraints, and store the constraint thresholds and weight parameters simultaneously. This set of constraints provides the basis for the lower-level teaching regulations in step S3, and provides an evaluation standard for the compliance calculation in step S4. S14. Perform consistency linkage verification of teaching resources-map-constraints and output three types of standardized basic constraint data; S141. Verify the consistency between the teaching resource base and the knowledge graph: Verify that all knowledge points labeled in the teaching resources exist in the knowledge graph node set, and there are no knowledge point labels outside the graph; S142. Verify the consistency between the constraint set and the knowledge graph: Verify that the knowledge points referenced by the prerequisite constraints and credit constraints correspond one-to-one with the nodes in the knowledge graph, and that the constraint thresholds are consistent with the requirements of the teaching syllabus. S143. After completing all verifications, output three types of data simultaneously: a standardized learning and teaching resource library, a directed acyclic course knowledge graph, and a set of quantitative teaching constraints.

[0018] Step S2 specifically includes the following steps: S21. Collect raw multimodal interaction data from AI teaching terminals using a unified timestamp, and integrate them using the learner's unique identifier and timestamp as the primary key to form a raw behavior dataset. Then, sequentially perform missing value imputation, 3σ outlier removal, and max-min normalization on the raw behavior data to obtain normalized feature values. The system outputs a standardized behavior dataset. The original multimodal interaction data of the AI ​​teaching terminal includes video viewing data, answer data, forum interaction data, and terminal operation data. For missing data in the learning duration category, the mean of the same teaching resource group is used for imputation. For missing data in the interaction frequency category, zero-value imputation is used. S22. Based on the standardized behavioral dataset obtained in step S21, calculate the weekly academic risk probability of learners using the MFFEP multi-feature fusion academic risk prediction model, and generate high-risk learner labels. The MFFEP multi-feature fusion academic risk prediction model includes, from top to bottom, a weekly window multi-source heterogeneous data access and sample slicing sub-module, a unified feature preprocessing and numerical coding sub-module, a four-way parallel deep extraction of academic features sub-module, an attention mechanism adaptive feature weighted fusion sub-module, a Stacking two-layer integrated learning risk probability prediction sub-module, and a multi-level risk threshold determination and labeling output and closed-loop write-back sub-module. The weekly window multi-source heterogeneous data access and sample slicing submodule segments the full learner interaction data using a fixed sliding window of the natural week, establishing a one-to-one mapping between learner unique IDs and weekly behavior subsets. It also connects in parallel with offline static ledger data, including three types of structured data: knowledge point mastery vectors output by the SIKT model, calculated individual cognitive load three-component parameters, and historical weekly academic risk labels, completing the data alignment and aggregation of streaming dynamic behavioral data and static learning indicators. A sample validity judgment rule is set, directly removing learner samples with fewer than three valid learning records in a single week to avoid prediction bias caused by sparse samples. A unified feature preprocessing and numerical encoding submodule is used to normalize and map continuous features such as learning duration, answer accuracy, cognitive load deviation rate, and task completion rate to... The interval eliminates the differences in the dimensions of different indicators, and converts the qualitative features such as learning and teaching resource preferences, self-jumping behavior types, and error types into model-computable numerical vectors through one-hot encoding, and finally outputs a standardized feature matrix with regular dimensions and consistent dimensions. The four-path parallel learning feature deep extraction submodule is configured with a static feature extraction branch for knowledge point mastery, a dynamic feature extraction branch for time-series learning behavior, a feature extraction branch for cognitive load suitability, and an additional feature extraction branch for learning execution compliance, which respectively output... , , , The static feature extraction branch for knowledge point mastery is used to extract the mastery value of each knowledge point within the week, the number of prerequisite knowledge points that were not passed, the number of consecutive weak knowledge points, and the credit completion progress achievement rate. The dynamic feature extraction branch for time-series learning behavior uses a single-layer LSTM to encode the daily effective learning time, the frequency of task exits, and the fluctuation sequence of answering time in the 7-day sliding window within the week. It captures the dynamic evolution of learning perseverance decay, procrastination behavior, and continuous decline in attention, and outputs time-series latent vectors to represent the risk of procrastination in process behavior. The cognitive load fit feature extraction branch is used to extract the number of times the cognitive load exceeds the limit on a single day within the week, the weekly average value of path load matching degree, and the number of times high-load teaching resources are repeatedly reviewed. The additional feature extraction branch for learning execution compliance is used to extract the number of times the learner deviates from the planned path, the weekly learning task compliance completion rate, and the proportion of subjective negative reviews of learning and teaching resources. The attention mechanism adaptive feature weighting fusion submodule assigns feature importance weights based on the risk root causes of each learner, and obtains the final global fused feature vector based on these weights. : ; ; In the formula, For the first The branch features are dynamically weighted to satisfy the following conditions. ; The model can be trained using an attention weight matrix; For the first The original feature vectors of each branch; For the first The original feature vectors of the branches, and ; The Stacking two-layer ensemble learning risk probability prediction submodule consists of a bottom-level parallel weak learner layer and an upper-level meta-learner fusion layer. The bottom-level parallel weak learner layer deploys three types of heterogeneous base learners—LightGBM, Random Forest, and Logistic Regression—in parallel, respectively processing the fused feature vectors. The system performs an initial independent risk probability prediction, and uses the outputs of the three models as secondary input features. These features are then fed into the upper-level meta-learner fusion layer to complete the result stacking and fusion, ultimately outputting a weekly academic risk probability value that takes a continuous value in the 0-1 range. ; The multi-level risk threshold determination and labeling output and closed-loop write-back submodule is used to determine the weekly academic risk probability value. Risk level classification: ; Learners with high-risk levels are tagged with a weekly high-risk learner label, and risk root cause labels are bound to them based on the weight of feature importance (i.e., the branch corresponding to the feature with the highest feature importance weight is used as the root cause label). The root cause labels are divided into four categories: knowledge gap risk, cognitive overload risk, learning behavior decline risk, and insufficient compliance risk. The weekly risk probability value, risk level, and risk root cause label are simultaneously written into the learner's comprehensive profile. S23. Based on the standardized answer data obtained in step S21, the knowledge status is iteratively updated through the SIKT sequence interaction knowledge tracing model, and the full-dimensional knowledge point mastery vector is output. S231. Sort the standardized answer data by learner ID and time sequence to construct an ordered answer sequence; each answer record includes the corresponding knowledge point number. Learner ID and the answer results ,in This indicates that the answer is correct. This indicates an incorrect answer; S232. Input the ordered answer sequence into the SIKT sequence-interactive knowledge tracing model, and iteratively update the learner's knowledge status according to time steps; Let... For learners exist Always keep the knowledge points The degree of mastery, with values ​​ranging from [0,1], where higher values ​​indicate stronger mastery; the formula for iteratively updating the knowledge state is: ; In the formula, For learners Knowledge points at the current moment The degree of mastery; For learners The knowledge points at the previous moment The degree of mastery; The learning gain coefficient represents the increase in knowledge mastery when a correct answer is given, and its value ranges from [value missing]. ; The forgetting decay coefficient represents the degree of knowledge loss when an answer is incorrect, and its value ranges from [value missing]. ; S233, Traverse all the knowledge points included in the course to generate learners. exist Moment Mastering the knowledge points of dimension vector ,in This represents the total number of knowledge points included in the course. S24. Based on the standardized behavioral data obtained in step S21, the cognitive load three-component model is used to calculate the individual cognitive capacity of learners and output the upper limit index of cognitive load of a single teaching resource. S241. Calculate intrinsic cognitive load: Intrinsic cognitive load is characterized by learners' historical average learning time per instructional resource. (This is based on statistical analysis of learners'...) Effective learning time for all completed history learning resources: ; In the formula, For learners Internal cognitive load; For learners Learning the first The effective duration of each teaching resource; For learners The total number of completed history teaching resources; S242. Calculate the associated cognitive load: The associated cognitive load is represented by the learner's average daily forum interaction time. The daily forum participation time of learners within the statistical period is as follows: ; In the formula, For learners Related cognitive load; For learners In the Forum interaction duration per day; The total number of days in the statistical period; S243. Define the external cognitive load; the external cognitive load is determined by the operational complexity and interface interaction cost of the AI ​​teaching terminal, and is set as a constant for a fixed AI teaching terminal type. For tablet-based AI teaching terminals, The value is set at 0.8 minutes per class hour; S244, Total cognitive capacity of synthetic learners: ; In the formula, For learners Total cognitive capacity; S25. Integrate learner basic attributes, high-risk markers from step S22, knowledge point mastery vectors from step S23, and total cognitive capacity indicators from step S24 to generate standardized learner feature vectors and output a comprehensive learner profile dataset. S251. Extract learner basic attribute features, including grade, major, and learning preference type, and convert them into numerical features using one-hot encoding; S252. Align the basic attribute features, knowledge point mastery vector, total cognitive capacity index, and high-risk markers dimensionally and concatenate them into a unified structure. Original feature vector; S253. Perform Z-score standardization on the original feature vector to eliminate the dimensional differences between different dimensions; S254. The standardized feature vectors of all learners constitute the learner comprehensive profile dataset.

[0019] The SIKT sequence interaction knowledge tracking model described in S23 includes a sequentially arranged question-answering timing embedding layer, a course knowledge graph topology constraint injection layer, a timing-knowledge point interaction dual-branch attention encoding layer, a mastery probability smoothing decoding layer, and a multi-module interface adaptation layer. The answer sequence embedding layer is used to uniformly map the three types of non-numerical symbolic information in the standardized answer sequence data—discrete knowledge point numbers, correct / incorrect answers, and the order of answers—into a low-dimensional dense computable vector, which is then concatenated to obtain the original input vector of a single interactive record. , , , They are respectively The system binds the knowledge point number embedding vector, the correct / incorrect result embedding vector, and the position encoding vector to the questions answered at any time. This is a vector concatenation operation; the positional encoding preserves the sequential relationship of the answer sequence, represents the temporal order of knowledge learning, completes the standardized transformation of the original business data into the model feature space, and ensures that this model is completely consistent with the upstream learner data cleaning specifications. The topological constraint injection layer reads the adjacency matrix of the directed acyclic course knowledge graph constructed in step S1, generates a binary mask matrix, and uses the generated binary mask matrix as a hard-constraint embedding model. During computation, it forces constraints to determine whether the mastery status of prior knowledge points has a positive or negative impact on the prediction of the mastery probability of subsequent related knowledge points. This follows the objective, step-by-step teaching and cognitive laws, avoiding prediction biases caused by purely data-driven models that ignore the logical sequence of knowledge points. ; ; In the formula, This is a binary engineering mask matrix; The basic topological critical matrix; The temporal-knowledge point interaction dual-branch attention encoding layer includes a temporal self-attention encoding branch, a knowledge point interaction attention encoding branch, and a dual-branch feature weighting fusion unit. The temporal self-attention encoding branch uses a multi-head Transformer structure to globally model long-cycle answer sequences, capturing the temporal evolution characteristics of learners over time, such as knowledge forgetting, periodic practice for consolidation, and fluctuations in learning state. This reconstructs the dynamic changes in ability during the real learning process and outputs a temporal hidden state vector. The knowledge point interaction attention branch is based on a binary engineering mask matrix to calculate the mutual transfer influence weights between different knowledge points. It quantifies the positive ability transfer of mastered knowledge points to subsequent content and the dragging effect of weak knowledge points on subsequent learning, reflecting the intrinsic connection between knowledge points. ; ; ; In the formula, For knowledge points Knowledge points The mutual migration influence weights; After adding knowledge graph dependency constraints, knowledge points Knowledge points Corrected score; After adding knowledge graph dependency constraints, knowledge points Knowledge points Corrected score; For knowledge points without the addition of course topological constraints Knowledge points The original association score; For the first A single-line query vector corresponding to each knowledge point; For the first Transpose of the Key vector of each knowledge point; For vector dimensions; The dual-branch feature weighted fusion unit uses trainable weights to weight and fuse the two latent vectors, resulting in a deep fusion representation vector that takes into account both dynamic changes over time and logical coupling of knowledge points. ; In the formula, This is the fused deep feature vector; and All are trainable balanced weights for the model; For the reason The knowledge point topological coupling implicit vector obtained by weighted aggregation; Mastering the probability smoothing decoding layer and fusing the deep feature vectors Nonlinear decoding is performed, and the original mastery probability value of a single knowledge point in the 0-1 interval is output through the Sigmoid activation function. A first-order exponential moving average algorithm is introduced to smooth the prediction results for continuous periods, filtering out drastic jumps in mastery values ​​caused by accidental noise such as careless mistakes or random guesses, ensuring the stability and continuity of the output results. Finally, a one-dimensional structured, full-dimensional knowledge point mastery vector is output. , For the first The final probability value of mastering each knowledge point after applying an exponential moving average smoothing filter. This represents the total number of all knowledge points in the course knowledge graph of S2. The multi-module interface adaptation layer is used to standardize the format of the knowledge point mastery vector across all dimensions.

[0020] Step S3 specifically includes the following steps: S31. Initialize the two-layer heterogeneous multi-agent collaborative interaction architecture and establish a decoupled interaction mechanism between the upper and lower layers; S311. Construct a two-layer heterogeneous multi-agent collaborative interaction architecture: a subject constraint layer and a function execution layer. The lower layer is the subject constraint layer, which deploys three types of entity agents: learner agent, teaching resource agent, and teaching regulation agent, responsible for domain state maintenance, constraint verification, and data uploading. The upper layer is the function execution layer, which deploys three types of functional agents: learning analysis agent, path planning agent, and reflection and verification agent, responsible for the division of labor and execution of the entire path planning task. The two layers interact with each other through a standardized JSON message interface to achieve decoupling and collaboration between state maintenance and task execution, thus solving the defects of single-layer multi-agent architectures such as incomplete constraint coverage and insufficient iteration capabilities. S312. Define global state data specifications and interaction protocols: specify the state data format transmitted from lower-level agents to upper-level agents, including three types of fields: entity type, state dimension, and timestamp; specify the control command format transmitted from upper-level agents to lower-level agents, including three types of fields: command type, parameter set, and execution cycle; unify the minimum time step for data interaction as a single teaching day to ensure the consistency of state synchronization across the entire system. S313. Initialize the operating parameters of each agent and intelligent agent: The lower-level agent sets the state update trigger conditions, including three types of events: changes in learner behavior data, updates to teaching resource library content, and adjustments to teaching regulations; The upper-level intelligent agent loads the basic weights and structured prompt word templates of the educational fine-tuning large language model and completes the initialization of the operating environment and parameters; Among them, the educational fine-tuning large language model uses curriculum topology regulations, cognitive load parameters, regulations, and high-risk learning cases to construct instruction-answer pair fine-tuning samples to complete the training. The training input is structured instructions with constraints, and the supervision label is the standard path scheme and correction logic that conforms to educational regulations and system quantitative indicators; S32, the lower-level main constraint layer has multi-agent state perception and global state aggregation, providing full-dimensional constraint input for the upper functional layer; S321. The learner agent loads the learner comprehensive profile dataset output in step S2 and maintains the learner's real-time status, including four core status dimensions: knowledge point mastery vector, total cognitive capacity, current learning progress, and learning style preference. When new learning behavior data is received from the terminal, a status update is triggered to ensure that the status data is synchronized with the learner's actual situation in real time. S322. The standardized learning and teaching resource library output by the teaching resource agent loading step S1 maintains the availability status of teaching resources, including four status dimensions: teaching resource metadata, standard cognitive load, knowledge point difficulty level, and online availability markers. When the teaching resource library experiences content additions, removals, or capacity adjustments, the teaching resource status is updated synchronously, and a set of currently available candidate teaching resources is selected. The selection criteria are as follows: Hard pre-filtering of availability: online teaching resources marked as unavailable, invalid, outdated, or with abnormal links are removed, and only valid teaching resources marked as normally accessible are retained, forming the first set of available teaching resources. Knowledge graph topological association constraint filtering: Based on the first round of available teaching resource subsets, read the set of knowledge points of the learning target to be planned this time, and based on the pre-dependency relationship of the directed acyclic course knowledge graph constructed in step 1, retain only the associated teaching resources that are directly bound to the target knowledge points or the pre-required knowledge points bound to the target knowledge points, and remove irrelevant teaching resources that have no topological logical relationship with the current learning task to obtain the second round of associated teaching resource subsets; Cognitive load coarse screening and pre-classification: learner individual cognitive capacity based on S2 output Standard cognitive load of teaching resources within the second-round associated teaching resource subset Make a judgment: For Teaching resources are directly included in the selection pool; for However, teaching resources that do not exceed 1.8 times the individual cognitive capacity limit are marked and temporarily stored, to be used in the S351 stage after being broken down into 2-3 low-load continuous learning segments based on the internal logic of the knowledge points; Overloaded teaching resources that exceed the threshold and cannot be split to ensure the integrity of teaching are directly eliminated, forming a third set of compliant teaching resources. Refined selection based on the difficulty of knowledge points and their matching with students' learning progress: Retrieving learners' information from the S233 output. Mastering the knowledge points of time and degree vector Weakness assessment was conducted on the knowledge points bound to the subset of teaching resources compliant with the third round of workload assessment: knowledge point mastery. If identified as a weak point, prioritize retaining the corresponding low or medium difficulty level. Consolidation-type teaching resources; if the mastery of knowledge points meets the standard, medium and high-level improvement-type teaching resources are retained, and finally the final set of candidate teaching resources that can be called upon in this round of path planning is output; S323. The quantitative set of teaching constraints output by the teaching regulations agent loading step S1 maintains the effective status of the regulations constraints, including three types of constraint parameters: prerequisite constraints, credit achievement thresholds, and single-cycle duration limits. When the teaching management regulations are adjusted, the constraint thresholds and weight parameters are updated synchronously to ensure that the regulations are consistent with the current regulations. S324 and the three types of agents respectively upload their current state data to the state aggregation module, which then concatenates the data according to the dimensional order to form... global state vector at time step The formula for concatenating the global state vector is: ; In the formula, for The learner state vector output by the learner agent at each moment has dimensions consistent with the learner's comprehensive profile; for The teaching resource state vector output by the teaching resource agent at any time contains the feature matrix of the candidate teaching resources; for The time-based teaching specifies the constraint vector output by the agent, which includes the thresholds and weights of three types of constraints; The aggregated global state vector serves as the unified input for all functional agents in the upper layer. S33. The upper-level learning situation analysis agent receives the global status and generates a standard learning situation diagnosis report; S331. Learning Situation Analysis Agent Extracts Global State Vector Learner state components Combining the directed acyclic course knowledge graph output in step S1, the educational fine-tuning large language model is invoked, and the structured prompt word template for learning diagnosis is loaded. The prompt word template contains four fixed modules: role definition, task objectives, dynamic input information, and standardized output format, and dynamically embeds real-time parameters such as learner knowledge mastery, cognitive capacity, and weak knowledge points. S332, The Educational Fine-tuning Large Language Model performs learning situation inference based on cue word constraints: First, it identifies the learner's set of weak knowledge points and set of knowledge points that have been mastered; second, it assesses the learner's cognitive load capacity limit and suitable learning pace; and finally, it provides targeted learning direction and content focus suggestions. S333, Output a standardized JSON format learning progress diagnostic report. It includes four fields: overall evaluation, detailed list of weaknesses, cognitive adaptation suggestions, and learning direction guidance; S34. The path planning agent uses a multi-agent joint deep Q network to perform global optimization and output a coarse-grained optimal teaching resource sequence. S341. Define the state space, action space, and reward function for joint reinforcement learning, and construct a unified optimization objective with multiple constraints: State space: the global state vector obtained in step S324 As input states, they cover three constraint dimensions: learner ability, supply of teaching resources, and regulations. Action space: Define joint actions The selection and sequence arrangement of teaching resources for the learning path involves selecting several teaching resources from the candidate teaching resource set and arranging them according to the learning order; the feasible domain of the action is limited by the prerequisite constraints of the course knowledge graph output in step S1. Reward function: The comprehensive rationality score serves as the immediate reward, directly transforming the optimization goal of the learning plan's rationality into the maximization goal of reinforcement learning. A higher reward value indicates stronger overall path rationality; and the reward function expression is: ; In the formula, for The instant reward obtained after performing a combined action at any time, with a value range of [value missing]. ; The cognitive load mismatch rate for the current path. For the sake of knowledge sequence consistency, Regulations on compliance for teaching resources; These are the weights for cognitive load matching, knowledge sequence progression, and compliance with teaching resource regulations, respectively, to meet the following requirements. ; S342. Construct a multi-agent joint deep Q-network, which includes two fully connected neural networks with the same structure: a main Q-network and a target Q-network. The main Q-network is used to evaluate the current state action value in real time, and the target Q-network is used to calculate the target Q-value. The parameters of the target Q-network are updated synchronously from the main Q-network at fixed training steps to suppress training fluctuations and improve convergence stability. S343. Update the parameters of the main Q-network using the Bellman iterative equation: ; In the formula, Main Q network output The moment-state action value is the expected cumulative reward for performing the corresponding joint action in the current state. The learning rate has a range of values. The step size for controlling parameter updates; The immediate reward obtained at the current moment; This is the discount factor, and its value range is... This is used to measure the importance of future rewards; The next state action value output by the target Q-network; This is a candidate joint action for the next moment; The maximum Q value corresponding to all candidate actions in the next state; S344. After the model converges through pre-training, input the current global state. By employing a greedy strategy to select the joint action that maximizes the Q-value, a coarse-grained optimal teaching resource sequence is output. The sequence satisfies the knowledge graph prerequisite constraints and the basic constraints, and has global approximate optimality, but lacks refined cognitive load adaptation and interpretability description. S35. The upper layer of the path planning agent uses an educational fine-tuning large language model to perform fine-tuning and interpretable generation. S351. Cognitive Load Adaptation and Adjustment: Read the standard cognitive load of each instructional resource in the coarse-grained instructional resource sequence. With the learner's total cognitive capacity Compare them one by one; if there are teaching resources that meet the requirements... If necessary, the teaching resource should be divided into 2 to 3 consecutive learning segments according to the teaching constraints, ensuring that the cognitive load of each segment is lower than the learner's capacity limit; after the division, the teaching integrity of the original knowledge points should be maintained to avoid content fragmentation; S352. Difficulty Smooth Transition Adjustment: Based on the course knowledge graph output in step S1, calculate the difficulty difference between corresponding knowledge points of adjacent teaching resources in the sequence; if the difficulty level difference between corresponding knowledge points of adjacent teaching resources is... Then, transitional teaching resources of corresponding difficulty level are inserted between the two to fill the gap in difficulty and ensure that the learning progression conforms to the zone of proximal development theory; the formula for calculating the difficulty difference is: ; In the formula, For the first in the path The and the first The difference in difficulty between individual knowledge points; For the first The quantitative difficulty level of each knowledge point; S353, Explainable Text Generation: Call the educational fine-tuning large language model to generate localized recommendation reasons for each teaching resource in the path, explaining the correspondence between the teaching resource and the learner's weak knowledge points, learning objectives and expected benefits; S354, Output the refined initial learning path It includes four types of information: teaching resource number, learning segment division, knowledge point correspondence, and recommendation reason text; S36. Construct a multi-directional initial path candidate set and output the set of unvalidated initial paths; S361. For a single learner, by adjusting the weight coefficients of the reward function, 3 to 5 initial paths with different focuses are generated, corresponding to three optimization directions: cognitive load priority, knowledge progression priority, and regulatory compliance priority, forming a diverse set of initial path candidates; among them, the cognitive load priority path (mainly recommended for learners with high risk and low cognitive capacity) is preferred. The values ​​are 0.6, 0.2, and 0.2 respectively; the knowledge progression priority path is recommended for learners with weak foundations or gaps in knowledge. The values ​​are 0.2, 0.6, and 0.2 respectively; the compliance priority path is specified (for scenarios with strict academic management and rigid teaching performance indicators). The values ​​are 0.2, 0.2, and 0.6 respectively. S362. Perform structured encoding on all paths within the set, uniformly store four types of data for each path: teaching resource sequence, total duration statistics, difficulty sequence, and explanatory text, and output the unvalidated initial path set.

[0021] Step S4 specifically includes the following steps: S41. Initialize the rationality assessment system for the three-dimensional fusion plan, and load the path to be evaluated and basic constraint data; S411. Load the unverified initial path set output by step S3, and extract the three types of core assessment basic data for each path: teaching resource sequence, corresponding knowledge point number of teaching resource, and standard cognitive load of teaching resource. S412. Synchronously load related basic data: Load the total cognitive capacity index in the learner comprehensive profile output in step S1 and the three types of constraint thresholds and sub-item weight parameters in the quantitative teaching constraint set, and synchronously load the knowledge point difficulty level data in the directed acyclic course knowledge graph output in step S2. S413. Initialize the global parameters and calculation rules of the evaluation system: Set the weight coefficients, score ranges, and output data formats for the three evaluation dimensions to ensure that all paths are calculated using a unified standard; This three-dimensional system comprehensively quantifies the rationality of the plan from three dimensions: learner cognition, teaching laws, and management regulations, breaking through the limitations of traditional single-dimensional evaluation and providing a unified quantitative target for path optimization. S42. Calculate the cognitive load matching index based on cognitive load theory to quantify the degree of fit between path load and learner cognitive capacity. S421. Extract the path to be evaluated. Based on the standard cognitive load of each teaching resource specified in step S213, the average cognitive load per teaching resource for this path is calculated; the calculation formula is as follows: ; In the formula, For the first The average cognitive load of a single teaching resource in each path to be evaluated; For the first The total number of learning and teaching resources included in each path to be evaluated; S422. Read the learner's total cognitive capacity calculated in step S24. Calculate the cognitive load mismatch rate for learners along a single path: ; In the formula, The cognitive load mismatch rate for a single learner and a single path; when When the path load is lower than the learner's capacity, the plan is too simple; when This indicates that the path load exceeds the learner's capacity, posing a risk of cognitive overload. For learners The total cognitive load accumulated across the entire generated learning path; S423. Convert the mismatch rate into a cognitive load matching score. The higher the score, the better the fit between the path load and the learner's cognitive ability. ; In the formula, The cognitive load matching score has a range of values. ; It represents the absolute value of the cognitive load mismatch rate; this indicator transforms the cognitive load theory into a calculable quantitative value, solving the problem that traditional path planning cannot quantitatively assess cognitive adaptability, and can effectively avoid two unreasonable situations: cognitive overload or underload. S43. Based on the theory of the zone of proximal development, calculate the progressive indicators of the knowledge sequence to quantify the rationality of teaching with progressively increasing difficulty of the path; S431. Based on the directed acyclic course knowledge graph output in step S1, extract the quantitative difficulty level of each knowledge point corresponding to each teaching resource in the path to be evaluated, and form a difficulty sequence according to the learning order of the path. ; S432. Using a strict incremental judgment rule, calculate the proportion of adjacent knowledge points in a path whose difficulty strictly increases, and then calculate the consistency of the knowledge sequence for a single path: ; In the formula, For the first The knowledge sequence consistency score of the path to be evaluated ranges from [value range missing]. A higher score indicates a smoother progression in the difficulty of the knowledge points along the path, which aligns with the teaching principles of the zone of proximal development. This is an indicator function; the function takes the value 1 when the condition inside the parentheses is true, and takes the value 0 when the condition is false. S44. Calculate the compliance indicators of teaching resource regulations based on teaching management guidelines, and quantify the degree to which the path meets the teaching regulations. S441. Calculate the prerequisite constraint satisfaction rate: Traverse each teaching resource in the path and verify whether all prerequisite knowledge points for its corresponding knowledge point have been mastered by the learner or appear in the preceding learning sequence of that teaching resource; count the proportion of teaching resources that satisfy the prerequisite requirements out of the total number of teaching resources to obtain the prerequisite constraint satisfaction rate. ; S442. Calculate the credit pass rate: Sum the assessment weights of all knowledge points covered by the statistical path, and calculate the ratio of the assessment weights to the total assessment weights of the course to obtain the credit pass rate. ; S443. Calculate the time compliance rate: Compare the total learning time of the statistical path with the maximum learning time stipulated in the single-cycle teaching plan, and calculate the time compliance rate using a piecewise function. ; S444. The three sub-indicators are weighted and summed according to the preset sub-weights to obtain the comprehensive score for compliance with teaching resource regulations: ; In the formula, For the first The compliance score for the teaching resources regulations for each path is specified, and the range of values ​​is as follows: ; All are weighting coefficients, satisfying The default values ​​are 0.5, 0.3, and 0.2. S45. Weighted fusion of three-dimensional indicators to calculate the overall rationality score of the learning plan; S451. Construct a dynamic weight adjustment model adapted to different teaching scenarios, and dynamically adjust the weight ratio according to the emphasis of the teaching scenario: ; In the formula, The cognitive load matching weight is dynamically adjusted. The progressive weights of the dynamically adjusted knowledge sequence; A compliance weight is assigned to the dynamically adjusted teaching resources; These are the maximum adjustment values ​​for the cognitive load matching weight, the knowledge sequence progression weight, and the teaching resource compliance weight, respectively, to satisfy... ; The cognitive adaptation scenario preference coefficient has a set value. Positive values ​​are taken for enrichment scenarios, focusing on adapting to learners' cognitive patterns; The knowledge progression scenario preference coefficient has a value of [value missing]. For exam-oriented reinforcement scenarios, positive values ​​are taken, emphasizing the effect of knowledge gradient progression; To define the preference coefficient for compliant scenarios, the value is set as follows: The system standardizes teaching management scenarios by taking positive values, emphasizing compliance with teaching regulations; through this set of constraint equations, adaptive dynamic allocation of weights can be achieved in multiple scenarios, while ensuring that the weight normalization constraint remains constant. S452. The weighted sum of the scores for each of the three dimensions is used to obtain the overall reasonableness score of the path to be evaluated: ; In the formula, For the first The overall reasonableness score of the path to be evaluated, with a value range of [value range missing]. The higher the score, the stronger the overall rationality of the learning plan in the three dimensions of cognitive adaptation, teaching logic, and management compliance; this score also serves as the immediate reward function for the reinforcement learning module in step S34, achieving a unity between the evaluation criteria and the optimization goals. This three-dimensional fusion quantitative system constructs a rationality evaluation framework that integrates learner cognition, teaching laws, and management regulations. It transforms the abstract rationality of a plan into calculable, comparable, and optimizable quantitative indicators, providing a scientific quality measurement standard for adaptive path planning. S46. Batch path evaluation and structured result output; S461. Traverse all paths to be evaluated in the unverified initial path set, calculate the three sub-scores and the comprehensive rationality score for each path, and form a mapping data table that corresponds one-to-one between path identifiers and scores. S462. Sort the path set in descending order according to the comprehensive rationality score, and mark the strengths and weaknesses of each path. S463. Output a structured dataset containing path item scores and comprehensive rationality evaluation results.

[0022] Step S5 specifically includes the following steps: S51. Reflect on and verify the initialization of the intelligent agent, and load the evaluation results and verification parameters; S511. Load all path sub-scores and comprehensive rationality evaluation results output from step S4, and synchronously associate them with the teaching resource sequence, difficulty distribution, and duration statistics corresponding to each path to establish a one-to-one mapping relationship between path identifiers and evaluation data. S512. Set a comprehensive rationality compliance threshold. and the qualified threshold of each item dimension ; S513, Initialize the iteration counter Set the maximum number of iterations. Establish an iterative path archive to record the parameter configuration, sub-item scores, and overall score changes for each round of the path; S52. Comprehensive rationality assessment, distinguishing between qualified paths and paths that need optimization; S521. Extract the comprehensive rationality score of the current path to be verified. , and the preset threshold Perform numerical comparison; If satisfied If the path is deemed to meet the criteria, it is directly marked as a qualified path and stored in the candidate path set, without needing to enter the iterative optimization process; If satisfied If the path is deemed not to meet the standard, the process of identifying weaknesses and performing targeted iterative optimization will begin. S53, Low-scoring weak dimension location and targeted iterative modification instruction generation; S531. Extract the three sub-scores of the non-compliant paths. , , , respectively with the sub-item qualification threshold Compare and count the number of dimensions below the threshold and the degree of bias; S532. Calculate the relative deviation rate of each item's score: ; In the formula, For the first The relative deviation rate of each dimension; a larger value indicates a more severe defect in that dimension; subscript These correspond to three dimensions: cognitive load matching degree, knowledge sequence progression, and compliance of teaching resource regulations; The scores are for each item in the corresponding dimension; S533. Select the dimension with the largest relative deviation rate as the core weak dimension, and generate targeted iterative modification instructions in the corresponding direction: If the core weak dimension is cognitive load matching, the instructions are: adjust the distribution of teaching resource duration, split teaching resources whose standard cognitive load exceeds the learner's capacity, optimize the balance between the load of a single teaching resource and the overall load, and control the average load of the path to be close to the learner's cognitive capacity limit. If the core weak dimension is the progression of the knowledge sequence, the instructions are: supplement transitional teaching resources at the difficulty jump nodes, smooth the difficulty difference between adjacent knowledge points, and ensure that the difficulty progression gradient meets the requirements of the zone of proximal development theory. If the core weak dimension is the compliance of teaching resources, the instructions are as follows: Adjust the range of teaching resources to prioritize the coverage of knowledge points with high assessment weight, verify prerequisite dependencies and the total time constraints of a single cycle, and comprehensively improve the degree of compliance with the regulations. S534. Synchronously generate the reward weight adjustment coefficients for the corresponding dimensions: ; In the formula, For dimension The original weighting coefficients, The adjusted weighting coefficients; by amplifying the reward weights of weak dimensions, the next round of path planning is guided to specifically address the defects. Compared with random regeneration of fixed weights, this can improve the iteration convergence speed by more than 40%. S54, Iterative modification command feedback and path regeneration; S541. The targeted modification instructions and the adjusted reward weight parameters are sent back to the path planning agent in step S3 through a standardized JSON message interface. S542. After receiving the instruction, the path planning agent keeps the global state and the range of candidate teaching resources unchanged. Based on the adjusted reward weight and modified constraints, it re-executes the underlying joint DQN optimization and upper-level large model fine-tuning process to output a new version of the optimized path. S543, Iteration Counter Increment by 1, send the new version of the path to step S4 to re-execute the three-dimensional rationality assessment, and update the sub-item scores and comprehensive score data; After each iteration, two termination conditions are checked simultaneously: Condition 1: Overall rationality score of the current path To meet the standards; Condition 2: Iteration Counter The preset maximum number of iterations has been reached. If condition one is met, the iteration is determined to be converged, the current path is marked as a qualified path, and stored in the multi-level candidate adaptation path set; If the target is still not met even if condition 2 is met, the version with the highest comprehensive score in the three rounds of iteration is taken as the final candidate path, marked as the path to be optimized, and stored in the secondary adaptation level of the candidate set. S55, Construction and structured output of multi-level candidate adaptation path sets; S551. All paths that have completed the iterative process are categorized into three levels based on their comprehensive score and achievement status: First-level adaptation path: The path that meets the criteria during the iteration process and has a comprehensive score of ≥0.8 is the most reasonable and optimal path, and is suitable for the formal learning plan push scenario; Secondary adaptation path: The path that meets the criteria during the iteration process and has a comprehensive score between 0.75 and 0.8 is a good and reasonable adaptation path, suitable for the rapid learning plan generation scenario; Level 3 adaptation path: This is the optimal version for iteration but does not meet the standards. It is a basic adaptation alternative path and an adaptation reference scenario for emergency learning. S552. Attach an iterative optimization file to each path, recording the iteration rounds, score change trajectory, and core optimization adjustment points; S553, outputs a structured multi-level candidate adaptation path set.

[0023] In step S6, the multi-objective optimization mathematical model includes the following three objective functions: First objective function: ; In the formula, Decision variables The corresponding cognitive load matching degree, and its value range. The larger the value, the better the adaptability; Total cognitive capacity of learners; For path number Selected teaching resources Standard cognitive load; This represents the total number of teaching resources included in the path. ; Second objective function: ; In the formula, Decision variables The corresponding knowledge sequence consistency has a value range of [0,1], and the larger the value, the smoother the progression. For the path sequence, the first The difficulty level of the knowledge points attached to each teaching resource; This is an indicator function that takes the value 1 if the condition is true, and 0 otherwise. For the path sequence, the first The difficulty level of the knowledge points attached to each teaching resource; The third objective function: ; In the formula, Decision variables The corresponding teaching resources are subject to compliance regulations and their value range. The larger the value, the better the compliance. , , These are the prerequisite constraint satisfaction rate, credit achievement rate, and duration compliance rate, respectively. , , These are the weighting coefficients corresponding to the prerequisite constraint satisfaction rate, credit achievement rate, and duration compliance rate, respectively.

[0024] An AI teaching terminal is configured to execute an adaptive learning path planning method for fusion plan rationality.

[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. An adaptive learning path planning method that integrates planning rationality, characterized by: Includes the following steps: S1. Collect course knowledge graph, teaching resource metadata, and teaching management regulations, and sequentially complete multi-dimensional feature annotation of teaching resources, construction of knowledge point topology relationships, and structured coding of constraint regulations. Output three types of basic constraint data: standardized learning and teaching resource library, directed acyclic course knowledge graph, and quantitative teaching constraint set. S2. Collect raw data of multimodal interaction from AI teaching terminals, complete data cleaning and standardization, and calculate academic risk markers, full-dimensional knowledge point mastery vectors, and individual cognitive capacity indicators through the MFFEP multi-feature fusion academic risk prediction model, SIKT sequence knowledge tracking model, and cognitive load three-component model, respectively. Integrate learners' basic attributes to generate standardized learner feature vectors and output a comprehensive learner profile dataset. S3. Construct a two-layer heterogeneous multi-agent collaborative interaction architecture. The lower-layer learner agent, teaching resource agent, and teaching regulation agent respectively load the learner comprehensive profile output by S2, the standardized learning and teaching resource library output by S1, and the quantitative teaching constraint set and synchronize their status in real time. The upper-layer learning analysis agent receives the lower-layer aggregated global status and generates a standardized learning diagnosis report. The path planning agent uses a multi-agent joint deep Q network as the bottom-layer global optimization module and an educational fine-tuning large language model as the upper-layer fine-tuning module. Based on the S1 course knowledge graph, it generates an initial learning path and supporting interpretable recommendation text, and outputs an unverified initial path set. S4. Based on cognitive load theory, zone of proximal development theory, and teaching compliance criteria, a three-dimensional fusion plan rationality quantification system is constructed. A unified standardized calculation formula is used to calculate the three sub-indicators of cognitive load matching degree, knowledge sequence progression, and teaching resource compliance. The weighted fusion is used to obtain the comprehensive rationality score. The unverified initial path set output from S3 is quantitatively evaluated one by one, and the path sub-scores and comprehensive rationality evaluation results are output. S5. Deploy the reflection and verification agent to read all path sub-scores and comprehensive rationality evaluation results output by S4, compare them with the preset comprehensive rationality threshold to complete the standard determination. If the path does not meet the standard, locate the weak dimension with low score and generate a targeted iterative modification instruction to be sent back to the path planning agent of S3 for cyclic regeneration, and output a multi-level candidate adaptation path set after the iteration ends. S6. Construct a multi-objective optimization mathematical model of spatial cognition-knowledge progression-regulation compliance, introduce the NSGA-II multi-objective optimization algorithm to perform global Pareto optimization on the multi-level candidate adaptation path set output by S5, and simultaneously use AD axiom design to complete the verification of teaching logic compatibility, and select the unique global optimal learning path and complete planning explanation document. S7. Compile the optimal learning path output by S6 into a standardized guided learning parameter file for AI terminal visualization and deploy it. Collect new interactive behavior data of terminal learners in real time and feed it back to S2 to update the comprehensive profile of learners. Iterate the training parameters to dynamically adjust the model and periodically update the modular learning teaching resource parameter library of S1 to achieve full-cycle adaptive closed-loop update.

2. The adaptive learning path planning method for fusion plan rationality according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Collect all original metadata of teaching resources, perform multi-dimensional feature annotation and group cognitive load calibration, and build a standardized learning and teaching resource library; S111. Collect the original metadata of all learning and teaching resources within the curriculum system, covering four types of teaching resources: videos, exercises, documents, and practical training projects. Extract six basic attributes: unique identifier of teaching resources, type of teaching resources, content description, chapter to which they belong, and original duration, to form a set of original teaching resource metadata. S112. Perform multi-dimensional feature annotation for each type of teaching resource. The annotation dimensions include four teaching attributes: corresponding knowledge point number, list of prerequisite knowledge points, difficulty level label, and teaching objective level. S113. Based on the historical learning data of all learners, calculate the average learning time for each teaching resource as the standard cognitive load for that teaching resource: ； In the formula, For the first Standard cognitive load of each teaching resource; For the first Learners learn the first Effective learning time for each teaching resource; The total number of learners who have studied this teaching resource for history; S114. Structure-encode the basic attributes, teaching characteristics, and standard cognitive load of all teaching resources, and establish an index with the unique identifier of the teaching resource as the primary key to generate a standardized learning and teaching resource database. S12. Extract the knowledge point entities and prerequisite dependencies of the course to construct a hierarchical directed acyclic course knowledge graph; S121. Extract all core knowledge point entities from the course syllabus, chapter lesson plans, and assessment instructions, assign a unique number to each knowledge point, and record the three basic attributes of knowledge point name, module to which it belongs, and assessment weight to form a set of knowledge point nodes. S122. Based on the teaching logic, define the prerequisite dependencies between knowledge points and generate a set of directed edges; if learning knowledge points... It is essential to master the knowledge points Given this premise, then establish from point to A directed edge represents a successor-precedence relationship; S123. Based on the topological sorting algorithm, knowledge points are hierarchically divided, and the quantitative difficulty level of each knowledge point is calculated in combination with the assessment weight: ； In the formula, For the first The quantitative difficulty level of each knowledge point; For the first The hierarchical numbering of each knowledge point in topological sorting; For the first The assessment weight of each knowledge point; This is the weighting adjustment coefficient; S124. Verify the acyclicity of the knowledge point topology to ensure there are no directed cycles with circular dependencies. Finally, generate a directed acyclic course knowledge graph with knowledge points as nodes, prerequisite relationships as directed edges, and difficulty level as node attributes. ; S13. Analyze the provisions of teaching management regulations, implement the structured coding and quantitative conversion of the constraint provisions, and construct a set of quantitative teaching constraints; S131. Extract three core constraints related to the learning path from the three types of regulations: talent training program, course syllabus, and assessment management method: prerequisite compliance constraint, credit achievement constraint, and total time constraint. S132. Transform prior compliance constraints into calculable, quantifiable provisions: ； In the formula, First, adjust the constraint satisfaction rate; The number of teaching resources in the path that meet the prerequisite requirements; This represents the total number of teaching resources included in the path. S133. Transform credit achievement constraints into calculable quantitative regulations: The sum of the weights of the assessment knowledge points covered by the learning path must reach the minimum percentage of the total course credit requirement; the credit achievement rate is defined as the ratio of the sum of the assessment weights of the knowledge points covered by the path to the total assessment weight, calculated using the following formula: ； In the formula, For the credit achievement rate; The set of knowledge points included in the learning path; For the first The assessment weight of each knowledge point; This represents the total number of all knowledge points in the course. S134. Transform the total duration constraint into a computable quantification: ； In the formula, For duration compliance rate; This is the maximum study time limit; Total study time; S135. Assign preset weights to the three types of constraints to form a structured set of quantitative teaching constraints, and store the constraint thresholds and weight parameters simultaneously. This set of constraints provides the basis for the lower-level teaching regulations in step S3, and provides an evaluation standard for the compliance calculation in step S4. S14. Perform consistency linkage verification of teaching resources-map-constraints and output three types of standardized basic constraint data; S141. Verify the consistency between the teaching resource base and the knowledge graph: Verify that all knowledge points labeled in the teaching resources exist in the knowledge graph node set, and there are no knowledge point labels outside the graph; S142. Verify the consistency between the constraint set and the knowledge graph: Verify that the knowledge points referenced by the prerequisite constraints and credit constraints correspond one-to-one with the nodes in the knowledge graph, and that the constraint thresholds are consistent with the requirements of the teaching syllabus. S143. After completing all verifications, output three types of data simultaneously: a standardized learning and teaching resource library, a directed acyclic course knowledge graph, and a set of quantitative teaching constraints.

3. The adaptive learning path planning method for fusion plan rationality according to claim 2, characterized in that: Step S2 specifically includes the following steps: S21. Collect raw multimodal interaction data from AI teaching terminals using a unified timestamp, and integrate them using the learner's unique identifier and timestamp as the primary key to form a raw behavior dataset. Then, sequentially perform missing value imputation, 3σ outlier removal, and max-min normalization on the raw behavior data to obtain normalized feature values. The system outputs a standardized behavior dataset. The original multimodal interaction data of the AI ​​teaching terminal includes video viewing data, answer data, forum interaction data, and terminal operation data. For missing data in the learning duration category, the mean of the same teaching resource group is used for imputation. For missing data in the interaction frequency category, zero-value imputation is used. S22. Based on the standardized behavioral dataset obtained in step S21, calculate the weekly academic risk probability of learners using the MFFEP multi-feature fusion academic risk prediction model, and generate high-risk learner labels. The MFFEP multi-feature fusion academic risk prediction model includes, from top to bottom, a weekly window multi-source heterogeneous data access and sample slicing sub-module, a unified feature preprocessing and numerical coding sub-module, a four-way parallel deep extraction of academic features sub-module, an attention mechanism adaptive feature weighted fusion sub-module, a Stacking two-layer integrated learning risk probability prediction sub-module, and a multi-level risk threshold determination and labeling output and closed-loop write-back sub-module. The weekly window multi-source heterogeneous data access and sample slicing submodule segments the full learner interaction data using a fixed sliding window based on the natural week, establishing a one-to-one mapping between learner unique IDs and weekly behavior subsets; the unified feature preprocessing and numerical encoding submodule is used to normalize and map continuous features to... The interval is defined, and the qualitative features of the classification are one-hot encoded into numerical vectors that can be computed by the model. Finally, a standardized feature matrix with regular dimensions and consistent units is output. The four-path parallel learning feature deep extraction submodule is configured with a static feature extraction branch for knowledge point mastery, a dynamic feature extraction branch for time-series learning behavior, a feature extraction branch for cognitive load suitability, and an additional feature extraction branch for learning execution compliance, which respectively output... , , , The static feature extraction branch for knowledge point mastery is used to extract the mastery value of each knowledge point within the week, the number of prerequisite knowledge points that were not passed, the number of consecutive weak knowledge points, and the credit completion progress achievement rate. The dynamic feature extraction branch for time-series learning behavior uses a single-layer LSTM to encode the daily effective learning time, the frequency of task exits, and the fluctuation sequence of answering time in the 7-day sliding window within the week. It captures the dynamic evolution of learning perseverance decay, procrastination behavior, and continuous decline in attention, and outputs time-series latent vectors to represent the risk of procrastination in process behavior. The cognitive load fit feature extraction branch is used to extract the number of times the cognitive load exceeds the limit on a single day within the week, the weekly average value of path load matching degree, and the number of times high-load teaching resources are repeatedly reviewed. The additional feature extraction branch for learning execution compliance is used to extract the number of times the learner deviates from the planned path, the weekly learning task compliance completion rate, and the proportion of subjective negative reviews of learning and teaching resources. The attention mechanism adaptive feature weighting fusion submodule assigns feature importance weights based on the risk root causes of each learner, and obtains the final global fused feature vector based on these weights. : ； ； In the formula, For the first The branch features are dynamically weighted to satisfy the following conditions. ; The model can be trained using an attention weight matrix; For the first The original feature vectors of each branch; For the first The original feature vectors of the branches, and ; The Stacking two-layer ensemble learning risk probability prediction submodule consists of a bottom-level parallel weak learner layer and an upper-level meta-learner fusion layer. The bottom-level parallel weak learner layer deploys three types of heterogeneous base learners—LightGBM, Random Forest, and Logistic Regression—in parallel, respectively processing the fused feature vectors. The system performs an initial independent risk probability prediction, and uses the outputs of the three models as secondary input features. These features are then fed into the upper-level meta-learner fusion layer to complete the result stacking and fusion, ultimately outputting a weekly academic risk probability value that takes a continuous value in the 0-1 range. ; The multi-level risk threshold determination and labeling output and closed-loop write-back submodule is used to determine the weekly academic risk probability value. Risk level classification: ； Learners with high risk levels are tagged with a weekly high-risk learner exclusive label, and risk root cause labels are bound to them according to the weight of feature importance. The root cause labels are divided into four categories: knowledge gap risk, cognitive overload risk, learning behavior decline risk, and insufficient compliance risk. The weekly risk probability value, risk level, and risk root cause label are simultaneously written into the learner's comprehensive profile. S23. Based on the standardized answer data obtained in step S21, the knowledge status is iteratively updated through the SIKT sequence interaction knowledge tracing model, and the full-dimensional knowledge point mastery vector is output. S231. Sort the standardized answer data by learner ID and time sequence to construct an ordered answer sequence; each answer record includes the corresponding knowledge point number. Learner ID and the answer results ,in This indicates that the answer is correct. This indicates an incorrect answer; S232. Input the ordered answer sequence into the SIKT sequence-interactive knowledge tracing model, and iteratively update the learner's knowledge status according to time steps; Let... For learners exist Always keep the knowledge points The degree of mastery, with a value range of [0,1]; the knowledge state iterative update formula is: ； In the formula, For learners Knowledge points at the current moment The degree of mastery; For learners The knowledge points at the previous moment The degree of mastery; The learning gain coefficient; This is the forgetting decay coefficient; S233, Traverse all the knowledge points included in the course to generate learners. exist Moment Mastering the knowledge points of dimension vector ,in This represents the total number of knowledge points included in the course. S24. Based on the standardized behavioral data obtained in step S21, the cognitive load three-component model is used to calculate the individual cognitive capacity of learners and output the upper limit index of cognitive load of a single teaching resource. S241. Calculate intrinsic cognitive load: Intrinsic cognitive load is characterized by learners' historical average learning time per instructional resource. (This is based on statistical analysis of learners'...) Effective learning time for all completed history learning resources: ； In the formula, For learners Internal cognitive load; For learners Learning the first The effective duration of each teaching resource; For learners The total number of completed history teaching resources; S242. Calculate the associated cognitive load: The associated cognitive load is represented by the learner's average daily forum interaction time. The daily forum participation time of learners within the statistical period is as follows: ； In the formula, For learners Related cognitive load; For learners In the Forum interaction duration per day; The total number of days in the statistical period; S243. Define the external cognitive load; the external cognitive load is determined by the operational complexity and interface interaction cost of the AI ​​teaching terminal, and is set as a constant for a fixed AI teaching terminal type. For tablet-based AI teaching terminals, The value is set at 0.8 minutes per class hour; S244, Total cognitive capacity of synthetic learners: ； In the formula, For learners Total cognitive capacity; S25. Integrate learner basic attributes, high-risk markers from step S22, knowledge point mastery vectors from step S23, and total cognitive capacity indicators from step S24 to generate standardized learner feature vectors and output a comprehensive learner profile dataset. S251. Extract learner basic attribute features, including grade, major, and learning preference type, and convert them into numerical features using one-hot encoding; S252. Align the basic attribute features, knowledge point mastery vector, total cognitive capacity index, and high-risk markers dimensionally and concatenate them into a unified structure. Original feature vector; S253. Perform Z-score standardization on the original feature vector to eliminate the dimensional differences between different dimensions; S254. The standardized feature vectors of all learners constitute the learner comprehensive profile dataset.

4. The adaptive learning path planning method for fusion plan rationality according to claim 3, characterized in that: The SIKT sequence interaction knowledge tracking model described in S23 includes a sequentially arranged question-answering timing embedding layer, a course knowledge graph topology constraint injection layer, a timing-knowledge point interaction dual-branch attention encoding layer, a mastery probability smoothing decoding layer, and a multi-module interface adaptation layer. The answer sequence embedding layer is used to uniformly map the three types of non-numerical symbolic information in the standardized answer sequence data—discrete knowledge point numbers, correct / incorrect answers, and the order of answers—into a low-dimensional dense computable vector, which is then concatenated to obtain the original input vector of a single interactive record. , , , They are respectively The system binds the knowledge point number embedding vector, the correct / incorrect result embedding vector, and the position encoding vector to the questions answered at any time. This is a vector concatenation operation; The topological constraint injection layer reads the adjacency matrix of the directed acyclic course knowledge graph constructed in step S1, generates a binary mask matrix, and uses the generated binary mask matrix as a hard-constraint embedding model. During computation, it forces constraints to influence the prediction of the mastery probability of subsequent related knowledge points, thus positively or negatively affecting the mastery status of these prior knowledge points. ； ； In the formula, This is a binary engineering mask matrix; The basic topological critical matrix; The temporal-knowledge point interaction dual-branch attention encoding layer includes a temporal self-attention encoding branch, a knowledge point interaction attention encoding branch, and a dual-branch feature weighting fusion unit. The temporal self-attention encoding branch uses a multi-head Transformer structure to globally model long-cycle answer sequences, capturing the temporal evolution features of learners over time and outputting a temporal hidden state vector. The knowledge point interaction attention branch is based on a binary engineering mask matrix to calculate the mutual transfer influence weights between different knowledge points: ； ； ； In the formula, For knowledge points Knowledge points The mutual migration influence weights; After adding knowledge graph dependency constraints, knowledge points Knowledge points Corrected score; After adding knowledge graph dependency constraints, knowledge points Knowledge points Corrected score; For knowledge points without the addition of course topological constraints Knowledge points The original association score; For the first A single-line query vector corresponding to each knowledge point; For the first Transpose of the Key vector of each knowledge point; For vector dimensions; The dual-branch feature weighted fusion unit uses trainable weights to weight and fuse the two latent vectors, resulting in a deep fusion representation vector that takes into account both dynamic changes over time and logical coupling of knowledge points. ； In the formula, This is the fused deep feature vector; and All are trainable balanced weights for the model; For the reason The knowledge point topological coupling implicit vector obtained by weighted aggregation; Mastering the probability smoothing decoding layer and fusing the deep feature vectors Nonlinear decoding is performed, and the original mastery probability value of a single knowledge point in the 0-1 interval is output through the Sigmoid activation function. A first-order exponential moving average algorithm is then introduced to smooth the prediction results for consecutive periods, filtering out drastic jumps in mastery values ​​caused by random answer noise. Finally, a one-dimensional structured, full-dimensional knowledge point mastery vector is output. , For the first The final probability value of mastering each knowledge point after applying an exponential moving average smoothing filter. This represents the total number of all knowledge points in the course knowledge graph of S2. The multi-module interface adaptation layer is used to standardize the format of the knowledge point mastery vector across all dimensions.

5. The adaptive learning path planning method for fusion plan rationality according to claim 3 or 4, characterized in that: Step S3 specifically includes the following steps: S31. Initialize the two-layer heterogeneous multi-agent collaborative interaction architecture and establish a decoupled interaction mechanism between the upper and lower layers; S311. Construct a two-layer heterogeneous multi-agent collaborative interaction architecture: a subject constraint layer and a function execution layer. The lower layer is the subject constraint layer, which deploys three types of entity agents: learner agent, teaching resource agent, and teaching regulation agent, responsible for domain state maintenance, constraint verification, and data uploading. The upper layer is the function execution layer, which deploys three types of functional agents: learning analysis agent, path planning agent, and reflection and verification agent, responsible for the division of labor and execution of the entire path planning task. The two layers interact with each other through a standardized JSON message interface to achieve decoupling and collaboration between state maintenance and task execution. S312. Define global state data specifications and interaction protocols: specify the state data format transmitted from lower-level agents to upper-level agents, including three types of fields: entity type, state dimension, and timestamp; specify the control command format transmitted from upper-level agents to lower-level agents, including three types of fields: command type, parameter set, and execution cycle; unify the minimum time step for data interaction as a single teaching day. S313. Initialize the operating parameters of each agent and intelligent agent: The lower-level agent sets the state update trigger conditions, including three types of events: changes in learner behavior data, updates to teaching resource library content, and adjustments to teaching regulations; The upper-level intelligent agent loads the basic weights and structured prompt word templates of the educational fine-tuning large language model and completes the initialization of the operating environment and parameters; Among them, the educational fine-tuning large language model uses curriculum topology regulations, cognitive load parameters, regulations, and high-risk learning cases to construct instruction-answer pair fine-tuning samples to complete the training. The training input is structured instructions with constraints, and the supervision label is the standard path scheme and correction logic that conforms to educational regulations and system quantitative indicators; S32, the lower-level main constraint layer has multi-agent state perception and global state aggregation, providing full-dimensional constraint input for the upper functional layer; S321. The learner agent loads the learner comprehensive profile dataset output in step S2 and maintains the learner's real-time status, including four core status dimensions: knowledge point mastery vector, total cognitive capacity, current learning progress, and learning style preference. When new learning behavior data is received from the terminal, a status update is triggered. S322. The standardized learning and teaching resource library output by the teaching resource agent loading step S1 maintains the availability status of teaching resources, including four status dimensions: teaching resource metadata, standard cognitive load, knowledge point difficulty level, and online availability markers. When the teaching resource library experiences content additions, removals, or capacity adjustments, the teaching resource status is updated synchronously, and a set of currently available candidate teaching resources is selected. The selection criteria are as follows: Hard pre-filtering of availability: online teaching resources marked as unavailable, invalid, outdated, or with abnormal links are removed, and only valid teaching resources marked as normally accessible are retained, forming the first set of available teaching resources. Knowledge graph topological association constraint filtering: Based on the first round of available teaching resource subsets, read the set of knowledge points of the learning target to be planned this time, and based on the pre-dependency relationship of the directed acyclic course knowledge graph constructed in step 1, retain only the associated teaching resources that are directly bound to the target knowledge points or the pre-required knowledge points bound to the target knowledge points, and remove irrelevant teaching resources that have no topological logical relationship with the current learning task to obtain the second round of associated teaching resource subsets; Cognitive load coarse screening and pre-classification: learner individual cognitive capacity based on S2 output Standard cognitive load of teaching resources within the second-round associated teaching resource subset Make a judgment: For Teaching resources are directly included in the selection pool; for However, teaching resources that do not exceed 1.8 times the individual cognitive capacity limit are marked and temporarily stored, to be used in the S351 stage after being broken down into 2-3 low-load continuous learning segments based on the internal logic of the knowledge points; Overloaded teaching resources that exceed the threshold and cannot be split to ensure the integrity of teaching are directly eliminated, forming a third set of compliant teaching resources. Refined selection based on the difficulty of knowledge points and their matching with students' learning progress: Retrieving learners' information from the S233 output. Mastering the knowledge points of time and degree vector Weakness assessment was conducted on the knowledge points bound to the subset of teaching resources compliant with the third round of workload assessment: knowledge point mastery. If identified as a weak point, prioritize retaining the corresponding low or medium difficulty level. Consolidation-type teaching resources; if the mastery of knowledge points meets the standard, medium and high-level improvement-type teaching resources are retained, and finally the final set of candidate teaching resources that can be called upon in this round of path planning is output; S323. The quantitative set of teaching constraints output by the teaching regulations agent loading step S1 maintains the effective status of the regulations constraints, including three types of constraint parameters: prerequisite constraints, credit achievement thresholds, and single-cycle duration limits; when the teaching management regulations are adjusted, the constraint thresholds and weight parameters are updated synchronously. S324 and the three types of agents respectively upload their current state data to the state aggregation module, which then concatenates the data according to the dimensional order to form... global state vector at time step The formula for concatenating the global state vector is: ； In the formula, for The learner state vector output by the learner agent at each moment has dimensions consistent with the learner's comprehensive profile; for The teaching resource state vector output by the teaching resource agent at any given time; for The time-bound teaching specification defines the constraint vector for the agent's output. This is the aggregated global state vector; S33. The upper-level learning situation analysis agent receives the global status and generates a standard learning situation diagnosis report; S331. Learning Situation Analysis Agent Extracts Global State Vector Learner state components Combining the directed acyclic course knowledge graph output in step S1, the educational fine-tuning large language model is invoked, and the structured prompt word template for learning diagnosis is loaded. The prompt word template contains four fixed modules: role definition, task objectives, dynamic input information, and standardized output format, with real-time parameters dynamically embedded. S332, The Educational Fine-tuning Large Language Model performs learning situation inference based on cue word constraints: First, it identifies the learner's set of weak knowledge points and set of knowledge points that have been mastered; second, it assesses the learner's cognitive load capacity limit and suitable learning pace; and finally, it provides targeted learning direction and content focus suggestions. S333, Output a standardized JSON format learning progress diagnostic report. It includes four fields: overall evaluation, detailed list of weaknesses, cognitive adaptation suggestions, and learning direction guidance; S34. The path planning agent uses a multi-agent joint deep Q network to perform global optimization and output a coarse-grained optimal teaching resource sequence. S341. Define the state space, action space, and reward function for joint reinforcement learning, and construct a unified optimization objective with multiple constraints: State space: the global state vector obtained in step S324 As input states, they cover three constraint dimensions: learner ability, supply of teaching resources, and regulations. Action space: Define joint actions The selection and sequence arrangement of teaching resources for the learning path; the feasible domain of actions is limited by the prerequisite constraints of the course knowledge graph output in step S1. Reward function: The comprehensive rationality score serves as the immediate reward, directly transforming the optimization goal of the learning plan's rationality into the maximization goal of reinforcement learning. A higher reward value indicates stronger overall path rationality; and the reward function expression is: ； In the formula, for The instant reward obtained after performing a combined action at any time, with a value range of [value missing]. ; The cognitive load mismatch rate for the current path. For the sake of knowledge sequence consistency, Regulations on compliance for teaching resources; These are the weights for cognitive load matching, knowledge sequence progression, and compliance with teaching resource regulations, respectively, to meet the following requirements. ; S342. Construct a multi-agent joint deep Q-network, which includes two fully connected neural networks with the same structure: a master Q-network and a target Q-network. The master Q-network is used to evaluate the current state action value in real time, and the target Q-network is used to calculate the target Q-value. S343. Update the parameters of the main Q-network using the Bellman iterative equation: ； In the formula, Main Q network output Time-state action value; The learning rate; The immediate reward obtained at the current moment; Discount factor; The next state action value output by the target Q-network; This is a candidate joint action for the next moment; The maximum Q value corresponding to all candidate actions in the next state; S344. After the model converges through pre-training, input the current global state. By employing a greedy strategy to select the joint action that maximizes the Q-value, a coarse-grained optimal teaching resource sequence is output. ; S35. The upper layer of the path planning agent uses an educational fine-tuning large language model to perform fine-tuning and interpretable generation. S351. Cognitive Load Adaptation and Adjustment: Read the standard cognitive load of each instructional resource in the coarse-grained instructional resource sequence. With the learner's total cognitive capacity Compare them one by one; if there are teaching resources that meet the requirements... If so, the teaching resource is divided into 2 to 3 consecutive learning segments according to the teaching constraints, ensuring that the cognitive load of each segment is lower than the learner's capacity limit; S352. Difficulty Smooth Transition Adjustment: Based on the course knowledge graph output in step S1, calculate the difficulty difference between corresponding knowledge points of adjacent teaching resources in the sequence; if the difficulty level difference between corresponding knowledge points of adjacent teaching resources is... If a transitional teaching resource of the corresponding difficulty level is inserted between the two, it will fill the gap in difficulty; the formula for calculating the difficulty difference is: ； In the formula, For the first in the path The and the first The difference in difficulty between individual knowledge points; For the first The quantitative difficulty level of each knowledge point; S353, Explainable Text Generation: Call the educational fine-tuning large language model to generate localized recommendation reasons for each teaching resource in the path, explaining the correspondence between the teaching resource and the learner's weak knowledge points, learning objectives and expected benefits; S354, Output the refined initial learning path It includes four types of information: teaching resource number, learning segment division, knowledge point correspondence, and recommendation reason text; S36. Construct a multi-directional initial path candidate set and output the set of unvalidated initial paths; S361. For a single learner, by adjusting the weight coefficients of the reward function, 3 to 5 initial paths with different focuses are generated, corresponding to three optimization directions: cognitive load priority, knowledge progression priority, and regulatory compliance priority, forming a diverse set of initial path candidates; among them, the cognitive load priority path... The values ​​are 0.6, 0.2, and 0.2 respectively; Knowledge progression priority path. The values ​​are set to 0.2, 0.6, and 0.2 respectively; a compliance priority path is specified. The values ​​are 0.2, 0.2, and 0.6 respectively. S362. Perform structured encoding on all paths within the set, uniformly store four types of data for each path: teaching resource sequence, total duration statistics, difficulty sequence, and explanatory text, and output the unvalidated initial path set.

6. The adaptive learning path planning method for fusion plan rationality according to claim 5, characterized in that: Step S4 Specifically, the following steps are included: S41. Initialize the rationality assessment system for the three-dimensional fusion plan, and load the path to be evaluated and basic constraint data; S411. Load the unverified initial path set output by step S3, and extract the three types of core assessment basic data for each path: teaching resource sequence, corresponding knowledge point number of teaching resource, and standard cognitive load of teaching resource. S412. Synchronously load related basic data: Load the total cognitive capacity index in the learner comprehensive profile output in step S1 and the three types of constraint thresholds and sub-item weight parameters in the quantitative teaching constraint set, and synchronously load the knowledge point difficulty level data in the directed acyclic course knowledge graph output in step S2. S413. Initialize the global parameters and calculation rules of the evaluation system: Set the weight coefficients, score ranges, and output data formats for the three evaluation dimensions to ensure that all paths are calculated using a unified standard. S42. Calculate the cognitive load matching index based on cognitive load theory to quantify the degree of fit between path load and learner cognitive capacity. S421. Extract the path to be evaluated. Based on the standard cognitive load of each teaching resource specified in step S213, the average cognitive load per teaching resource for this path is calculated; the calculation formula is as follows: ； In the formula, For the first The average cognitive load of a single teaching resource in each path to be evaluated; For the first The total number of learning and teaching resources included in each path to be evaluated; S422. Read the learner's total cognitive capacity calculated in step S24. Calculate the cognitive load mismatch rate for learners along a single path: ； In the formula, The cognitive load mismatch rate for a single learner and a single path; when When the path load is lower than the learner's capacity, the plan is too simple; when This indicates that the path load exceeds the learner's capacity, posing a risk of cognitive overload. For learners The total cognitive load accumulated across the entire generated learning path; S423. Convert the mismatch rate into a cognitive load matching score: ； In the formula, The cognitive load matching score has a range of values. ; This represents the absolute value of the cognitive load mismatch rate. S43. Based on the theory of the zone of proximal development, calculate the progressive indicators of the knowledge sequence to quantify the rationality of teaching with progressively increasing difficulty of the path; S431. Based on the directed acyclic course knowledge graph output in step S1, extract the quantitative difficulty level of each knowledge point corresponding to each teaching resource in the path to be evaluated, and form a difficulty sequence according to the learning order of the path. ; S432. Using a strict incremental judgment rule, calculate the proportion of adjacent knowledge points in a path whose difficulty strictly increases, and then calculate the consistency of the knowledge sequence for a single path: ； In the formula, For the first Knowledge sequence consistency score of the path to be evaluated; This is an indicator function; the function takes the value 1 when the condition inside the parentheses is true, and takes the value 0 when the condition is false. S44. Calculate the compliance indicators of teaching resource regulations based on teaching management guidelines, and quantify the degree to which the path meets the teaching regulations. S441. Calculate the prerequisite constraint satisfaction rate: Traverse each teaching resource in the path and verify whether all prerequisite knowledge points for its corresponding knowledge point have been mastered by the learner or appear in the preceding learning sequence of that teaching resource; count the proportion of teaching resources that satisfy the prerequisite requirements out of the total number of teaching resources to obtain the prerequisite constraint satisfaction rate. ; S442. Calculate the credit pass rate: Sum the assessment weights of all knowledge points covered by the statistical path, and calculate the ratio of the assessment weights to the total assessment weights of the course to obtain the credit pass rate. ; S443. Calculate the time compliance rate: Compare the total learning time of the statistical path with the maximum learning time stipulated in the single-cycle teaching plan, and calculate the time compliance rate using a piecewise function. ; S444. The three sub-indicators are weighted and summed according to the preset sub-weights to obtain the comprehensive score for compliance with teaching resource regulations: ； In the formula, For the first The compliance score for the teaching resources regulations for each path is specified, and the range of values ​​is as follows: ; All are weighting coefficients, satisfying ; S45. Weighted fusion of three-dimensional indicators to calculate the overall rationality score of the learning plan; S451. Construct a dynamic weight adjustment model adapted to different teaching scenarios, and dynamically adjust the weight ratio according to the emphasis of the teaching scenario: ； In the formula, The cognitive load matching weight is dynamically adjusted. The progressive weights of the dynamically adjusted knowledge sequence; A compliance weight is assigned to the dynamically adjusted teaching resources; These are the maximum adjustment values ​​for the cognitive load matching weight, the knowledge sequence progression weight, and the teaching resource compliance weight, respectively, to satisfy... ; For cognitive adaptation scenario preference coefficient; For the knowledge progression scenario preference coefficient; To define the preference coefficient for compliant scenarios; S452. The weighted sum of the scores for each of the three dimensions is used to obtain the overall reasonableness score of the path to be evaluated: ； In the formula, For the first The overall reasonableness score of the path to be evaluated, with a value range of [value range missing]. ; S46. Batch path evaluation and structured result output; S461. Traverse all paths to be evaluated in the unverified initial path set, calculate the three sub-scores and the comprehensive rationality score for each path, and form a mapping data table that corresponds one-to-one between path identifiers and scores. S462. Sort the path set in descending order according to the comprehensive rationality score, and mark the strengths and weaknesses of each path. S463. Output a structured dataset containing path item scores and comprehensive rationality evaluation results.

7. The adaptive learning path planning method for fusion plan rationality according to claim 6, characterized in that: Step S5 specifically includes the following steps: S51. Reflect on and verify the initialization of the intelligent agent, and load the evaluation results and verification parameters; S511. Load all path sub-scores and comprehensive rationality evaluation results output from step S4, and synchronously associate them with the teaching resource sequence, difficulty distribution, and duration statistics corresponding to each path to establish a one-to-one mapping relationship between path identifiers and evaluation data. S512. Set a comprehensive rationality compliance threshold. and the qualified threshold of each item dimension ; S513, Initialize the iteration counter Set the maximum number of iterations. Establish an iterative path archive to record the parameter configuration, sub-item scores, and overall score changes for each round of the path; S52. Comprehensive rationality assessment, distinguishing between qualified paths and paths that need optimization; S521. Extract the comprehensive rationality score of the current path to be verified. , and the preset threshold Perform numerical comparison; If satisfied If the path is deemed to meet the criteria, it is directly marked as a qualified path and stored in the candidate path set, without needing to enter the iterative optimization process; If satisfied If the path is deemed not to meet the standard, the process of identifying weaknesses and performing targeted iterative optimization will begin. S53, Low-scoring weak dimension location and targeted iterative modification instruction generation; S531. Extract the three sub-scores of the non-compliant paths. , , , respectively with the sub-item qualification threshold Compare and count the number of dimensions below the threshold and the degree of bias; S532. Calculate the relative deviation rate of each item's score: ； In the formula, For the first The relative deviation rate of each dimension; a larger value indicates a more severe defect in that dimension; subscript These correspond to three dimensions: cognitive load matching degree, knowledge sequence progression, and compliance of teaching resource regulations; The scores are for each item in the corresponding dimension; S533. Select the dimension with the largest relative deviation rate as the core weak dimension, and generate targeted iterative modification instructions in the corresponding direction: If the core weak dimension is cognitive load matching, the instructions are: adjust the distribution of teaching resource duration, split teaching resources whose standard cognitive load exceeds the learner's capacity, optimize the balance between the load of a single teaching resource and the overall load, and control the average load of the path to be close to the learner's cognitive capacity limit. If the core weak dimension is the progression of the knowledge sequence, the instructions are: supplement transitional teaching resources at the difficulty jump nodes, smooth the difficulty difference between adjacent knowledge points, and ensure that the difficulty progression gradient meets the requirements of the zone of proximal development theory. If the core weak dimension is the compliance of teaching resources, the instructions are as follows: Adjust the range of teaching resources to prioritize the coverage of knowledge points with high assessment weight, verify prerequisite dependencies and the total time constraints of a single cycle, and comprehensively improve the degree of compliance with the regulations. S534. Synchronously generate the reward weight adjustment coefficients for the corresponding dimensions: ； In the formula, For dimension The original weighting coefficients, These are the adjusted weighting coefficients; S54, Iterative modification command feedback and path regeneration; S541. The targeted modification instructions and the adjusted reward weight parameters are sent back to the path planning agent in step S3 through a standardized JSON message interface. S542. After receiving the instruction, the path planning agent keeps the global state and the range of candidate teaching resources unchanged. Based on the adjusted reward weight and modified constraints, it re-executes the underlying joint DQN optimization and upper-level large model fine-tuning process to output a new version of the optimized path. S543, Iteration Counter Increment by 1, send the new version of the path to step S4 to re-execute the three-dimensional rationality assessment, and update the sub-item scores and comprehensive score data; After each iteration, two termination conditions are checked simultaneously: Condition 1: Overall rationality score of the current path To meet the standards; Condition 2: Iteration Counter The preset maximum number of iterations has been reached. If condition one is met, the iteration is determined to be converged, the current path is marked as a qualified path, and stored in the multi-level candidate adaptation path set; If the target is still not met even if condition 2 is met, the version with the highest comprehensive score in the three rounds of iteration is taken as the final candidate path, marked as the path to be optimized, and stored in the secondary adaptation level of the candidate set. S55, Construction and structured output of multi-level candidate adaptation path sets; S551. All paths that have completed the iterative process are categorized into three levels based on their comprehensive score and achievement status: First-level adaptation path: The path that meets the criteria during the iteration process and has a comprehensive score of ≥0.8 is the most reasonable and optimal path, and is suitable for the formal learning plan push scenario; Secondary adaptation path: The path that meets the criteria during the iteration process and has a comprehensive score between 0.75 and 0.8 is a good and reasonable adaptation path, suitable for the rapid learning plan generation scenario; Level 3 adaptation path: This is the optimal version for iteration but does not meet the standards. It is a basic adaptation alternative path and an adaptation reference scenario for emergency learning. S552. Attach an iterative optimization file to each path, recording the iteration rounds, score change trajectory, and core optimization adjustment points; S553, outputs a structured multi-level candidate adaptation path set.

8. The adaptive learning path planning method for fusion plan rationality according to claim 7, characterized in that: In step S6, the multi-objective optimization mathematical model includes the following three objective functions: First objective function: ; In the formula, Decision variables Corresponding cognitive load matching degree; Total cognitive capacity of learners; For path number Selected teaching resources Standard cognitive load; This represents the total number of teaching resources included in the path. ; Second objective function: ; In the formula, Decision variables The corresponding knowledge sequence consistency; For the path sequence, the first The difficulty level of the knowledge points attached to each teaching resource; This is an indicator function that takes the value 1 if the condition is true, and 0 otherwise. For the path sequence, the first The difficulty level of the knowledge points attached to each teaching resource; The third objective function: ; In the formula, Decision variables The corresponding teaching resources must comply with regulations. , , These are the prerequisite constraint satisfaction rate, credit achievement rate, and duration compliance rate, respectively. , , These are the weighting coefficients corresponding to the prerequisite constraint satisfaction rate, credit achievement rate, and duration compliance rate, respectively.

9. An AI teaching terminal is configured to execute the adaptive learning path planning method for fusion plan rationality as described in any one of claims 1-8.

Images