Learning situation large model construction method based on double-engine collaborative framework and academic early warning system

Abstract

The invention relates to a learning situation large model construction method based on a double-engine collaborative framework and an academic early warning system, and the method comprises the steps: S1, carrying out the preprocessing and dynamic triple filling of multi-source heterogeneous data, and constructing a high-fidelity student learning behavior holographic data cube; s2, constructing a scoring engine based on causal enhancement; s3, constructing a report generation engine of knowledge distillation and DPO, and generating a structured diagnosis report with temperature; and S4, closed-loop feedback and adaptive evolution are carried out. Through a full-chain technical framework of multi-source heterogeneous data construction, a causal scoring engine, a semantic report engine and closed-loop feedback, the method has the following advantages of high-fidelity data reduction capability, accurate early risk capture capability, semantic alignment and interpretability conforming to educational ethics, and a vitality self-evolution mechanism.

Description

Technical Field

[0001] This invention belongs to the field of educational informatization and artificial intelligence technology, specifically involving a method for constructing a large learning status model based on a dual-engine collaborative framework and an academic early warning system. Background Technology

[0002] With the popularization of smart campus construction, massive amounts of data have been generated throughout the education process. However, existing academic early warning and evaluation systems face the following significant technical bottlenecks in practical applications: 1) The data is diverse, heterogeneous, and severely incomplete, lacking a mechanism to repair spatiotemporal relationships. Learning data is scattered across isolated systems such as the academic affairs system, learning management system, library, and campus card system. Due to interference from numerous factors (such as missed attendance or equipment malfunction), it exhibits significant randomness and distinct periodic characteristics (such as weekly timetable patterns). Existing technologies often employ simple mean-filling or zero-value-filling, which disrupts the temporal and periodic characteristics of the data, leading to distorted data profiles.

[0003] 2) Lack of perception of "intermediate states" and "hidden risks". Traditional early warning models are mostly based on rules or simple supervised learning, and can usually only identify explicit problems that have already occurred. However, students often go through a complex intermediate process from a normal state to an academic crisis. For example, if a student studies hard but uses the wrong methods, it may lead to low efficiency, which in turn leads to student anxiety and ultimately causes a decline in student grades. Existing models lack the ability to model such non-linear and dynamically changing learning situations, and are extremely lacking in such negative samples (repeating a grade / dropping out of school), resulting in a serious class imbalance problem during model training.

[0004] 3) The scoring and explanation are disconnected, lacking pedagogical interpretability. Existing AI models are often a "black box," outputting a risk probability value without explaining the specific judgment criteria. Furthermore, suggestions generated based on general models are often too generalized (e.g., read more books, study diligently), lacking attribution analysis for specific knowledge points and behavioral patterns, making it difficult to guarantee the educational ethical safety of the generated content (e.g., an overly harsh tone may damage students' self-esteem).

[0005] 4) Lack of a closed-loop evolutionary mechanism of "perception-intervention-feedback". After the academic warning is issued, the teacher's intervention behavior and the student's subsequent state changes are not fed back to the model in a structured way. Moreover, the model parameters are fixed once deployed and cannot adapt to the dynamic migration of external environments such as the adjustment of teaching syllabus and the change of exam difficulty in the educational context. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings and deficiencies of existing technologies by providing a method for constructing a large learning situation model based on a dual-engine collaborative framework and an academic early warning system. It proposes a full-chain technical framework from "multi-source heterogeneous data construction → causal scoring engine → semantic reporting engine → closed-loop feedback".

[0007] The technical problem solved by this invention is achieved through the following technical solution: A method for constructing a large learning situation model based on a dual-engine collaborative framework, the steps of which are as follows: S1. Multi-source heterogeneous data preprocessing and dynamic triple filling to construct a high-fidelity holographic data cube of student learning behavior; S2. Construct a scoring engine based on causal reinforcement to address the problems of "lack of negative samples" and "difficulty in defining intermediate states," and achieve accurate mapping from behavioral features to learning status scores (0-100 points). S3. Build a knowledge distillation and DPO report generation engine to generate structured and humane diagnostic reports; S4, Closed-loop feedback and adaptive evolution.

[0008] Moreover, S1 specifically refers to: S101. Data collection is achieved by accessing the campus database through ETL tools, collecting more than 30 core indicators in real time. The indicators are divided into the following three categories: (a) High-frequency time series data: including daily login time of the learning system, timestamps of library access gates, canteen consumption patterns (used to analyze the regularity of life), and dormitory return time; (b) Low-frequency discrete data: the advance of homework submission time relative to the deadline, test scores, class ranking percentage, and absence records; (c) Unstructured text: course forum posts and counselor conversation records. S102. By designing differentiated imputation strategies based on the missing characteristics of different types of data, a dynamic triple imputation algorithm for educational scenarios is implemented. The specific strategies are as follows: Strategy 1: Imputation based on stratified statistics is suitable for numerical indicators with a missing rate of less than or equal to 15%. Instead, it divides students into three levels: L1 (excellent academic performance), L2 (average academic performance), and L3 (poor academic performance) based on historical GPA information, and uses the median of all students in the student's level on the indicator for imputation. Strategy 2: Weighted KNN interpolation based on academic distance, suitable for numerical indicators with missing values ​​greater than 15%, constructing a KNN (K-Nearest Neighbors) model, defining the academic distance D between students. ij K=7; ; Where: w kThe information gain weight of indicator k is calculated from the feature importance of random forest and interpolated using the weighted average of the seven most similar students for this indicator. Strategy 3: Time series imputation based on STL decomposition, suitable for time series indicators with continuous missing values. Targeting the significant weekly seasonality of student behavior, the STL (Seasonal-Trend decomposition using LOESS) algorithm is used to imput the time series data Y... t Decomposed into: Y t = T t (Semester Trends) +S t (periodic periodic term) + R t (Residual term); For the missing time period, retain S. t (Weekly cycle term) means retaining the daily routine from Monday to Friday, and only performing linear interpolation on Tt and Rt to ensure that the filled data matches the students' life rhythm, with the error rate controlled within 5%; S103. Use the BERT-edu model, fine-tuned for the education field, to extract text features, and combine TF-IDF to extract words such as "anxiety," "give up," and "can't understand," concatenating them into a 768+100 dimensional hybrid feature vector; construct a sliding window for time-series features, setting the window to 14 days and the length to 1 day, and calculate behavioral stability features (such as the coefficient of variation (CV) of learning duration), trend features, and lag features (Lag-1, Lag-7); it is recommended to use a four-dimensional knowledge base in JSON format, including: indicator standard name, dynamic calculation formula, reasonable threshold range (such as setting the upper limit of "daily reading time" to 12 hours to prevent interference from idle data), and grade level weight configuration (first-year students focus on attendance rate, and senior students focus on thesis progress).

[0009] Moreover, S2 specifically refers to: S201. Fine-tuning of the basic model: Qwen3-32B is used as the base model. LoRA (Low-Rank Adaptation) technology is used to fine-tune the model. The rank r of LoRA is set to 16 and α is set to 32. The target module covers all linear layers, so that the model can accurately classify the current status of students (excellent / normal / warning / high risk) based on the input feature vector. S202. Sample expansion based on causal reasoning: Granger causality test is introduced to construct a DAG (Directed Acyclic Graph) between indicators: a causal chain is constructed by identifying critical paths (e.g., late return to dormitory (cause) → absence from early class (intermediary) → failure to submit homework (intermediary) → risk of failing the course (effect)); then Gaussian perturbation (mean taken as the sample median, standard deviation dynamically adjusted) is applied to the upstream nodes (e.g., frequency of late return to dormitory) on the causal chain, and the changes in the downstream nodes are observed, thereby generating a large number of synthetic samples of critical states between "normal" and "failing the course"; finally, the synthetic samples are mapped to a two-dimensional space through the T-SNE test to ensure that their distribution fills the blank area between positive and negative samples, rather than overlapping. S203, rating mapping and OOD (out-of-distribution) detection: Extract the hidden states from the last layer of the large model, normalize them using a LayerNorm layer, and then input them into the MLP network. Set the Dropout value in the network to 0.1, and map them to a continuous scalar value from 0 to 100. During the inference phase, calculate the Mahalanobis distance D between the input sample and x and the training set distribution in real time. M (x): ; Where μ is the mean of the training set, and Σ is the covariance matrix; When D M When (x)>3σ (e.g., encountering a "behavioral pattern caused by a sudden family change" that has never been seen before), the system will not output a score, but will trigger a manual review signal to prevent the model from outputting content with low credibility.

[0010] Moreover, S3 specifically refers to: S301, Teacher-Student Distillation Framework: The teacher model uses Deepseek-R1, which has strong reasoning capabilities, to construct 5000 sets of triplets consisting of "data indicators + reasoning paths + diagnostic conclusions". For example, "Data shows that the average time for homework submission has been delayed by 4 hours in the past two weeks (indicator) → Combined with the student's reduced library attendance records, it is inferred that there may be time management problems or knowledge gaps (reasoning) → It is recommended to prioritize reviewing the content of Chapter 3 and adjust the schedule (conclusion)". The student model uses a model with a smaller amount of material and a faster response, and distillation is performed by minimizing the KL divergence with the output of the teacher model. S302, Reinforcement Learning Direct Preference Optimization (DPO): To ensure the final generated report conforms to educational ethics (primarily encouraging, avoiding discriminatory language), the DPO algorithm is used for alignment training; a preference dataset is constructed, with positive samples y. w The report, polished by a senior counselor, features clear logic, accurate causal attribution, tactful tone, and specific recommendations. (Negative sample y) lReports containing logical breaks, accusatory language, and vague suggestions (such as "You must work harder") are addressed using the following DPO loss function: ; Wherein: the β temperature coefficient is set to 0.1, and this process forces the model to learn the preference distribution of human education experts; S303, Triple Verification and Output: The generated report needs to undergo triple verification by the rule engine: First, the logical consistency of the report is verified, and whether the report score and the final conclusion conflict; then, the standardization of the terminology in the report is verified, and whether the report contains professional terms such as "cognitive load" and "formative assessment" and is not colloquial; finally, the resource accessibility of the report is verified, and whether the courses and activities mentioned in the report actually exist in the school's current resource library.

[0011] Moreover, S4 specifically refers to: S401. Semantic Density Index (SDI) monitoring: The system calculates the semantic density index of the generated report in real time. If the SDI is lower than 0.7 (meaning that the model starts to have a lot of repetitive clichés), the regeneration mechanism is automatically triggered and the sample is marked for subsequent optimization. S402, Incremental learning mechanism, establishes a data closed loop of "early warning-intervention-effect", records the intervention measures taken by teachers (such as conversations) after the early warning is issued and the changes in students' behavior in the following week. Cases with effective intervention are used as positive samples and cases with ineffective intervention are used as negative samples. Incremental learning is performed once a month so that the model can adapt to the characteristics of different semesters and different student groups.

[0012] An academic early warning system, characterized in that it includes: Data processing module: Used to perform dynamic triple padding and feature extraction; Scoring engine module: Deploys a trained scoring model to output a learning status score and risk level; Report generation module: Deployed with an optimized generation model for outputting structured diagnostic reports; Closed-loop feedback module: Used to record intervention measures and effects, and to perform incremental learning.

[0013] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method.

[0014] The advantages and beneficial effects of this invention are as follows: 1. High-fidelity data restoration capability: An innovative dynamic triple imputation algorithm for educational scenarios is proposed. By introducing STL decomposition to extract periodic terms and adaptively switching between hierarchical statistical imputation and KNN weighted interpolation based on academic distance under different missing rates, the problem of spatiotemporal fragmentation of educational data is effectively solved, enabling accurate restoration of students' real life and learning rhythms even when the data missing rate is high.

[0015] 2. Precise early risk capture capability: A scoring engine based on causal reinforcement was constructed. A causal graph of indicators was built using Granger causality test, and a large number of synthetic samples covering the "critical state" were generated through Gaussian perturbation. Without relying on a large number of real negative samples, the intermediate process of academic deterioration was successfully simulated, which significantly improved the model's ability to identify early hidden risks and its robustness.

[0016] 3. Ethical Semantic Alignment and Interpretability: The scoring engine and reporting engine are decoupled and coordinated. Knowledge distillation and DPO (Direct Preference Optimization) techniques are employed to ensure that the generated diagnostic reports are not only logically rigorous and accurately attributive, but also conform to the educational psychology norm of "encouragement as the primary focus." Combined with a triple verification mechanism (logic, terminology, and resources), AI illusions are effectively avoided, providing practical and personalized intervention suggestions.

[0017] 4. A robust self-evolutionary mechanism: A complete closed loop has been established, from "perception" to "intervention" and then to "feedback." Abnormal patterns are identified through Out-of-Distribution (OOD) detection using Mahalanobis distance, and effective intervention cases are positively fed back based on an incremental learning mechanism. This allows the system to continuously adapt to new teaching models (such as blended online and offline teaching) and changes in student demographics, thus extending the technology's lifespan. Attached Figure Description

[0018] Figure 1 This is a diagram illustrating the architecture and data flow of the present invention. Detailed Implementation

[0019] The present invention will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present invention.

[0020] A method for constructing a large-scale learning situation model based on a dual-engine collaborative framework is innovative in that the method comprises the following steps: S1. Multi-source heterogeneous data preprocessing and dynamic triple imputation to construct a high-fidelity holographic data cube of student learning behavior; specifically: S101. Data collection is achieved by accessing the campus database through ETL tools, collecting more than 30 core indicators in real time. The indicators are divided into the following three categories: (a) High-frequency time series data: including daily login time of the learning system, timestamps of library access gates, canteen consumption patterns (used to analyze the regularity of life), and dormitory return time; (b) Low-frequency discrete data: the advance of homework submission time relative to the deadline, test scores, class ranking percentage, and absence records; (c) Unstructured text: course forum posts and counselor conversation records. S102. By designing differentiated imputation strategies based on the missing characteristics of different types of data, a dynamic triple imputation algorithm for educational scenarios is implemented. The specific strategies are as follows: Strategy 1: Imputation based on stratified statistics is suitable for numerical indicators with a missing rate of less than or equal to 15%. Instead, it divides students into three levels: L1 (excellent academic performance), L2 (average academic performance), and L3 (poor academic performance) based on historical GPA information, and uses the median of all students in the student's level on the indicator for imputation. Strategy 2: Weighted KNN interpolation based on academic distance, suitable for numerical indicators with missing values ​​greater than 15%, constructing a KNN (K-Nearest Neighbors) model, defining the academic distance D between students. ij K=7; ; Where: w k The information gain weight of indicator k is calculated from the feature importance of random forest and interpolated using the weighted average of the seven most similar students for this indicator. Strategy 3: Time series imputation based on STL decomposition, suitable for time series indicators with continuous missing values. Targeting the significant weekly seasonality of student behavior, the STL (Seasonal-Trend decomposition using LOESS) algorithm is used to imput the time series data Y... t Decomposed into: Y t = T t (Semester Trends) +S t (periodic periodic term) + R t (Residual term); For the missing time period, retain S. t (Weekly cycle term) means retaining the daily routine from Monday to Friday, and only performing linear interpolation on Tt and Rt to ensure that the filled data matches the students' life rhythm, with the error rate controlled within 5%; S103. Use the BERT-edu model, fine-tuned for the education field, to extract text features, and combine TF-IDF to extract words such as "anxiety," "give up," and "can't understand," concatenating them into a 768+100 dimensional hybrid feature vector; construct a sliding window for time-series features, setting the window to 14 days and the length to 1 day, and calculate behavioral stability features (such as the coefficient of variation (CV) of learning duration), trend features, and lag features (Lag-1, Lag-7); it is recommended to use a four-dimensional knowledge base in JSON format, including: indicator standard name, dynamic calculation formula, reasonable threshold range (such as setting the upper limit of "daily reading time" to 12 hours to prevent interference from idle data), and grade level weight configuration (first-year students focus on attendance rate, and senior students focus on thesis progress).

[0021] S2. Construct a scoring engine based on causal reinforcement to address the problems of "lack of negative samples" and "difficulty in defining intermediate states," achieving a precise mapping from behavioral features to learning state scores (0-100 points); specifically: S201. Fine-tuning of the basic model: Qwen3-32B is used as the base model. LoRA (Low-Rank Adaptation) technology is used to fine-tune the model. The rank r of LoRA is set to 16 and α is set to 32. The target module covers all linear layers, so that the model can accurately classify the current status of students (excellent / normal / warning / high risk) based on the input feature vector. S202. Sample expansion based on causal reasoning: Granger causality test is introduced to construct a DAG (Directed Acyclic Graph) between indicators: a causal chain is constructed by identifying critical paths (e.g., late return to dormitory (cause) → absence from early class (intermediary) → failure to submit homework (intermediary) → risk of failing the course (effect)); then Gaussian perturbation (mean taken as the sample median, standard deviation dynamically adjusted) is applied to the upstream nodes (e.g., frequency of late return to dormitory) on the causal chain, and the changes in the downstream nodes are observed, thereby generating a large number of synthetic samples of critical states between "normal" and "failing the course"; finally, the synthetic samples are mapped to a two-dimensional space through the T-SNE test to ensure that their distribution fills the blank area between positive and negative samples, rather than overlapping. S203, rating mapping and OOD (out-of-distribution) detection: Extract the hidden states from the last layer of the large model, normalize them using a LayerNorm layer, and then input them into the MLP network. Set the Dropout value in the network to 0.1, and map them to a continuous scalar value from 0 to 100. During the inference phase, calculate the Mahalanobis distance D between the input sample and x and the training set distribution in real time. M (x): ; Where μ is the mean of the training set, and Σ is the covariance matrix; When DM When (x)>3σ (e.g., encountering a "behavioral pattern caused by a sudden family change" that has never been seen before), the system will not output a score, but will trigger a manual review signal to prevent the model from outputting content with low credibility.

[0022] S3. Build a knowledge distillation and DPO report generation engine to generate structured and empathetic diagnostic reports; specifically: S301, Teacher-Student Distillation Framework: The teacher model uses Deepseek-R1, which has strong reasoning capabilities, to construct 5000 sets of triplets consisting of "data indicators + reasoning paths + diagnostic conclusions". For example, "Data shows that the average time for homework submission has been delayed by 4 hours in the past two weeks (indicator) → Combined with the student's reduced library attendance records, it is inferred that there may be time management problems or knowledge gaps (reasoning) → It is recommended to prioritize reviewing the content of Chapter 3 and adjust the schedule (conclusion)". The student model uses a model with a smaller amount of material and a faster response, and distillation is performed by minimizing the KL divergence with the output of the teacher model. S302, Reinforcement Learning Direct Preference Optimization (DPO): To ensure the final generated report conforms to educational ethics (primarily encouraging, avoiding discriminatory language), the DPO algorithm is used for alignment training; a preference dataset is constructed, with positive samples y. w The report, polished by a senior counselor, features clear logic, accurate causal attribution, tactful tone, and specific recommendations. (Negative sample y) l Reports containing logical breaks, accusatory language, and vague suggestions (such as "You must work harder") are addressed using the following DPO loss function: ; Wherein: the β temperature coefficient is set to 0.1, and this process forces the model to learn the preference distribution of human education experts; S303, Triple Verification and Output: The generated report needs to undergo triple verification by the rule engine: First, the logical consistency of the report is verified, and whether the report score and the final conclusion conflict; then, the standardization of the terminology in the report is verified, and whether the report contains professional terms such as "cognitive load" and "formative assessment" and is not colloquial; finally, the resource accessibility of the report is verified, and whether the courses and activities mentioned in the report actually exist in the school's current resource library.

[0023] S4, closed-loop feedback and adaptive evolution, specifically: S401. Semantic Density Index (SDI) monitoring: The system calculates the semantic density index of the generated report in real time. If the SDI is lower than 0.7 (meaning that the model starts to have a lot of repetitive clichés), the regeneration mechanism is automatically triggered and the sample is marked for subsequent optimization. S402, Incremental learning mechanism, establishes a data closed loop of "early warning-intervention-effect", records the intervention measures taken by teachers (such as conversations) after the early warning is issued and the changes in students' behavior in the following week. Cases with effective intervention are used as positive samples and cases with ineffective intervention are used as negative samples. Incremental learning is performed once a month so that the model can adapt to the characteristics of different semesters and different student groups.

[0024] An academic early warning system, its innovation lies in: including: Data processing module: Used to perform dynamic triple padding and feature extraction; Scoring engine module: Deploys a trained scoring model to output a learning status score and risk level; Report generation module: Deployed with an optimized generation model for outputting structured diagnostic reports; Closed-loop feedback module: Used to record intervention measures and effects, and to perform incremental learning.

[0025] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method.

[0026] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. A method for constructing a large-scale learning situation model based on a dual-engine collaborative framework, characterized in that: The steps of the method are as follows: S1. Multi-source heterogeneous data preprocessing and dynamic triple filling to construct a high-fidelity holographic data cube of student learning behavior; S2. Construct a scoring engine based on causal reinforcement to address the problems of "lack of negative samples" and "difficulty in defining intermediate states," and achieve accurate mapping from behavioral features to learning status scores (0-100 points). S3. Build a knowledge distillation and DPO report generation engine to generate structured and humane diagnostic reports; S4, Closed-loop feedback and adaptive evolution.

2. The method for constructing a large learning situation model based on a dual-engine collaborative framework according to claim 1, characterized in that: Specifically, S1 is: S101. Data collection is achieved by accessing the campus database through ETL tools, collecting more than 30 core indicators in real time. The indicators are divided into the following three categories: (a) High-frequency time series data: including daily login time of the learning system, timestamps of library access gates, canteen consumption patterns (used to analyze the regularity of life), and dormitory return time; (b) Low-frequency discrete data: the advance of homework submission time relative to the deadline, test scores, class ranking percentage, and absence records; (c) Unstructured text: course forum posts and counselor conversation records. S102. By designing differentiated imputation strategies based on the missing characteristics of different types of data, a dynamic triple imputation algorithm for educational scenarios is implemented. The specific strategies are as follows: Strategy 1: Imputation based on stratified statistics is suitable for numerical indicators with a missing rate of less than or equal to 15%. Instead, it divides students into three levels: L1 (excellent academic performance), L2 (average academic performance), and L3 (poor academic performance) based on historical GPA information, and uses the median of all students in the student's level on the indicator for imputation. Strategy 2: Weighted KNN interpolation based on academic distance, suitable for numerical indicators with missing values ​​greater than 15%, constructing a KNN (K-Nearest Neighbors) model, defining the academic distance D between students. ij K=7; ； Where: w k The information gain weight of indicator k is calculated from the feature importance of random forest and interpolated using the weighted average of the seven most similar students for this indicator. Strategy 3: Time series imputation based on STL decomposition, suitable for time series indicators with continuous missing values. Targeting the significant weekly seasonality of student behavior, the STL (Seasonal-Trend decomposition using LOESS) algorithm is used to imput the time series data Y... t Decomposed into: Y t = T t (Semester Trends) +S t (periodic periodic term) + R t (Residual term); For the missing time period, retain S. t (Weekly cycle term) means retaining the daily routine from Monday to Friday, and only performing linear interpolation on Tt and Rt to ensure that the filled data matches the students' life rhythm, with the error rate controlled within 5%; S103. Use the BERT-edu model, fine-tuned for the education field, to extract text features, and combine TF-IDF to extract words such as "anxiety," "give up," and "can't understand," concatenating them into a 768+100 dimensional hybrid feature vector; construct a sliding window for time-series features, setting the window to 14 days and the length to 1 day, and calculate behavioral stability features (such as the coefficient of variation (CV) of learning duration), trend features, and lag features (Lag-1, Lag-7); it is recommended to use a four-dimensional knowledge base in JSON format, including: indicator standard name, dynamic calculation formula, reasonable threshold range (such as setting the upper limit of "daily reading time" to 12 hours to prevent interference from idle data), and grade level weight configuration (first-year students focus on attendance rate, and senior students focus on thesis progress).

3. The method for constructing a large learning situation model based on a dual-engine collaborative framework according to claim 1, characterized in that: Specifically, S2 is: S201. Fine-tuning of the basic model: Qwen3-32B is used as the base model. LoRA (Low-Rank Adaptation) technology is used to fine-tune the model. The rank r of LoRA is set to 16 and α is set to 32. The target module covers all linear layers, so that the model can accurately classify the current status of students (excellent / normal / warning / high risk) based on the input feature vector. S202. Sample expansion based on causal reasoning: Granger causality test is introduced to construct a DAG (Directed Acyclic Graph) between indicators: a causal chain is constructed by identifying critical paths (e.g., late return to dormitory (cause) → absence from early class (intermediary) → failure to submit homework (intermediary) → risk of failing the course (effect)); then Gaussian perturbation (mean taken as the sample median, standard deviation dynamically adjusted) is applied to the upstream nodes (e.g., frequency of late return to dormitory) on the causal chain, and the changes in the downstream nodes are observed, thereby generating a large number of synthetic samples of critical states between "normal" and "failing the course"; finally, the synthetic samples are mapped to a two-dimensional space through the T-SNE test to ensure that their distribution fills the blank area between positive and negative samples, rather than overlapping; S203, rating mapping and OOD (out-of-distribution) detection: Extract the hidden states from the last layer of the large model, normalize them using a LayerNorm layer, and then input them into the MLP network. Set the Dropout value in the network to 0.1, and map them to a continuous scalar value from 0 to 100. During the inference phase, calculate the Mahalanobis distance D between the input sample and x and the training set distribution in real time. M (x): ； Where μ is the mean of the training set, and Σ is the covariance matrix; When D M When (x) > 3σ (e.g., encountering a "behavioral pattern caused by a sudden family change" that has never been seen before), the system will not output a score, but will trigger a manual review signal to prevent the model from outputting content with low credibility.

4. The method for constructing a large learning situation model based on a dual-engine collaborative framework according to claim 1, characterized in that: Specifically, S3 is: S301, Teacher-Student Distillation Framework: The teacher model uses Deepseek-R1, which has strong reasoning capabilities, to construct 5000 sets of triplets consisting of "data indicators + reasoning paths + diagnostic conclusions". For example, "Data shows that the average time for homework submission has been delayed by 4 hours in the past two weeks (indicator) → Combined with the student's reduced library attendance records, it is inferred that there may be time management problems or knowledge gaps (reasoning) → It is recommended to prioritize reviewing the content of Chapter 3 and adjust the schedule (conclusion)". The student model uses a model with a smaller amount of material and a faster response, and distillation is performed by minimizing the KL divergence with the output of the teacher model. S302, Reinforcement Learning Direct Preference Optimization (DPO): To ensure the final generated report conforms to educational ethics (primarily encouraging, avoiding discriminatory language), the DPO algorithm is used for alignment training; a preference dataset is constructed, with positive samples y. w The report, polished by a senior counselor, features clear logic, accurate causal attribution, tactful tone, and specific recommendations. (Negative sample y) l Reports containing logical breaks, accusatory language, and vague suggestions (such as "You must work harder") are addressed using the following DPO loss function: ； Wherein: the β temperature coefficient is set to 0.1, and this process forces the model to learn the preference distribution of human education experts; S303, Triple Verification and Output: The generated report needs to undergo triple verification by the rule engine: First, verify the logical consistency of the report and whether the report score and the final conclusion conflict; then verify the standardization of the report's terminology and whether the report contains professional terms such as "cognitive load" and "formative assessment" and is not just colloquial language; finally, verify the resource accessibility of the report and whether the courses and activities mentioned in the report actually exist in the school's current resource library.

5. The method for constructing a large learning situation model based on a dual-engine collaborative framework according to claim 1, characterized in that: Specifically, S4 is: S401. Semantic Density Index (SDI) monitoring: The system calculates the semantic density index of the generated report in real time. If the SDI is lower than 0.7 (meaning that the model starts to have a lot of repetitive clichés), the regeneration mechanism is automatically triggered and the sample is marked for subsequent optimization. S402, Incremental learning mechanism, establishes a data closed loop of "early warning-intervention-effect", records the intervention measures taken by teachers (such as conversations) after the early warning is issued and the changes in students' behavior in the following week. Cases with effective intervention are used as positive samples and cases with ineffective intervention are used as negative samples. Incremental learning is performed once a month so that the model can adapt to the characteristics of different semesters and different student groups.

6. An academic early warning system, characterized in that: include: Data processing module: Used to perform dynamic triple padding and feature extraction; Scoring engine module: Deploys a trained scoring model to output a learning status score and risk level; Report generation module: Deployed with an optimized generation model for outputting structured diagnostic reports; Closed-loop feedback module: Used to record intervention measures and effects, and to perform incremental learning.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 5.

Images