CN-121998519-A - Double-engine collaborative framework-based learning situation large model construction method and academic early warning system

CN121998519ACN 121998519 ACN121998519 ACN 121998519ACN-121998519-A

Abstract

The invention relates to a learning situation large model construction method and a learning industry early warning system based on a double-engine collaborative framework, comprising the steps of S1, preprocessing multi-source heterogeneous data and dynamically triple filling, 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, generating a structured and temperature diagnosis report, and S4, performing closed-loop feedback and self-adaptive evolution. The invention has the advantages of high-fidelity data reduction capability, accurate early risk capturing capability, semantic alignment and interpretability according with education ethics and self-evolution mechanism with vitality through a full-chain technical framework of multi-source heterogeneous data construction, causality scoring engine, semantic reporting engine and closed loop feedback.

Inventors

WANG JIANRONG
WANG ZIMING
LI QI
ZHOU CHENGJIE

Assignees

天津大学

Dates

Publication Date: 20260508
Application Date: 20251225

Claims (7)

1. A learning situation big model construction method based on a double-engine collaborative framework is characterized by comprising the following steps: S1, preprocessing multi-source heterogeneous data and dynamically filling three times to construct a high-fidelity holographic data cube for learning behaviors of students; S2, constructing a scoring engine based on causal enhancement, so as to realize the problems of 'lack of negative samples' and 'difficulty in defining intermediate states', and realize the accurate mapping from behavior characteristics to learning situation scores (0-100 minutes); s3, constructing a report generation engine of knowledge distillation and DPO, and generating a diagnosis report which is structured and has temperature; S4, closed loop feedback and self-adaptive evolution.
2. The learning situation big model construction method based on the double-engine collaborative framework of claim 1 is characterized in that the S1 specifically comprises: S101, accessing a campus database through an ETL tool to realize data acquisition, and acquiring thirty items of core indexes in real time, wherein the indexes are divided into three types, namely (a) high-frequency time sequence data comprising daily learning system login time, library gate entry and exit time stamps, canteen consumption rules (used for analyzing living regularity) and dormitory time, (b) low-frequency discrete data comprising the advance of operation submission time relative to deadline, test results, class ranking percentage and absences records, (c) unstructured text comprising course forum speaking and counselor talking records; S102, designing differentiated filling strategies aiming at missing features of different types of data to realize a dynamic triple filling algorithm aiming at educational scenes, wherein the specific strategies are as follows: Strategy 1, filling based on hierarchical statistics, wherein the filling is applicable to numerical indexes with the loss rate less than or equal to 15%, students are divided into three layers of L 1 (excellent learning performance), L 2 (medium learning performance) and L 3 (poor learning performance) according to historical performance point information, and the filling is carried out by using the median of all students in the layers to which the students belong on the indexes; Strategy 2, based on weighted KNN interpolation of the academic distance, the method is suitable for numerical indexes with the deficiency value larger than 15%, a KNN (K-Nearest Neighbors) model is built, and the academic distance D ij between students is defined, wherein K=7; ; W k is the information gain weight of index k, calculated by the random forest feature importance, and interpolation is carried out by using the weighted average value of the index of 7 most similar students; Strategy 3, based on STL decomposition time sequence filling, is suitable for continuous missing time sequence indexes, and adopts STL (seal-Trend decomposition using LOESS) algorithm to decompose time sequence data Y t into: Y t = T t (learning trend) +s t (weekly term) +r t (residual term); In the missing time period, reserving S t (cycle term), namely reserving the work and rest law of monday to friday, and carrying out linear interpolation on Tt and Rt only, so as to ensure that the filling data accords with the life rhythm of students, and controlling the error rate within 5%; S103, extracting text features by using a BERT-edu model finely tuned for the education field, extracting words of anxiety, giving up and not understanding by integrating TF-IDF, splicing into mixed feature vectors of 768+100 dimensions, constructing a sliding window of time sequence features, setting the window as 14 days, setting the cloth length as 1 day, calculating behavior stability features (such as variation coefficient CV of learning duration), trend features and hysteresis features (Lag-1 and Lag-7), and suggesting a four-dimensional knowledge base of json format, wherein the four-dimensional knowledge base comprises index standard names, dynamic calculation formulas, reasonable threshold intervals (such as upper limit of single-day reading duration is set as 12 hours, and preventing on-hook data interference) and school weight configuration (large side heavy duty rate and large four-side heavy progress).
3. The learning situation big model construction method based on the double-engine collaborative framework of claim 1 is characterized in that the S2 specifically comprises: S201, fine tuning a basic model, namely fine tuning the model by adopting Qwen-32B as a base model and using LoRA (Low-Rank Adaption) technology, wherein the rank r of LoRA is set to 16, the alpha is set to 32, and a target module covers all Linear layers, so that the model can accurately classify the current state (excellent/normal/early warning/high risk) of students according to the input feature vectors; S202, sample expansion based on causal reasoning, a Granger causal test is introduced to construct DAG (directed acyclic graph) among indexes, wherein the construction of a causal chain is realized by identifying a critical path (such as dormitory late return (cause), early class absences (intermediaries), job non-intersection (intermediaries), hanging department risks (results)), gaussian disturbance (mean value sample median, standard differential state adjustment) is then applied to an upstream node (such as dormitory late return frequency) on the causal chain, and the change of a downstream node is observed, so that a large number of critical state synthetic samples between normal and hanging department are generated; S203, scoring mapping and OOD (out-of-distribution) detection, extracting a last layer HIDDEN STATES (hidden layer state) of the large model, inputting the large model into an MLP network after LayerNorm layers of normalization, setting Dropout in the network to be 0.1, mapping the Dropout into a continuous scalar of 0-100, and calculating the Marsh distance D M (x) between an input sample and x and the distribution of a training set in real time in an reasoning stage: ; Mu is the training set mean value, and sigma is the covariance matrix; When D M (x) > 3 sigma (if a behavior pattern caused by sudden household accident which is never seen is encountered), the system does not output a score, but triggers a manual review signal, so that the model is prevented from outputting contents with lower credibility.
4. The learning situation big model construction method based on the double-engine collaborative framework of claim 1 is characterized in that the S3 specifically comprises: S301, a teacher-student distillation frame, wherein a teacher model adopts Deepseek-R1 with extremely strong reasoning capability to construct 5000 groups of triplets of data index, reasoning path and diagnosis conclusion, for example, data display is delayed by 4 hours (index) when the average time of two weeks of job submission is delayed, time management problems possibly exist or knowledge point is inferred, priority review of the content of a third chapter is suggested, and work and rest (conclusion) are supposed, a student model adopts a model with smaller material quantity and faster response, and distillation is performed by minimizing KL divergence output by the teacher model; S302, reinforcement learning Direct Preference Optimization (DPO) is to perform alignment training by adopting a DPO algorithm in order to ensure that the finally generated report accords with education ethics (encouragement is mainly and discrimination is avoided), and to construct a preference data set, wherein a positive sample y w is a report which is wetted by a senior citizen, the content logic is clear, attribution is accurate, the mind is graceful, a specific negative sample y l is a logic fracture, a report which comprises responsibility language and advice hollows (such as 'you must struggle'), and a specific DPO loss function is as follows: ; wherein the beta temperature coefficient is set to 0.1, which forces the model to learn the preference distribution of human educational specialists; s303, triple verification and output are carried out on the generated report, wherein the generated report needs to be subjected to triple verification of a rule engine, namely, firstly, the logical consistency of the report is verified, whether the report score conflicts with a final conclusion, then, the term normalization of the report is verified, whether the report contains professional terms of 'cognitive load', 'formative evaluation', which are not spoken, and finally, the resource accessibility of the report is verified, and whether courses appear in the advice of the report and activities really exist in the current resource library of the school.
5. The method for building the large learning situation model based on the double-engine collaborative framework according to claim 1, wherein the step S4 is specifically as follows: S401, monitoring a semantic Density Index (SEMANTIC DENSITY Index, SDI), calculating the semantic Density Index of a report generated by the system in real time, automatically triggering a regeneration mechanism if the SDI is lower than 0.7 (meaning that a model starts to appear a large number of repetitions polite), and marking the sample for subsequent optimization; S402, an incremental learning mechanism is used for establishing a data closed loop of early warning-intervention-effect, recording intervention measures (such as talking) taken by teachers and behavior changes of students in the following week after early warning is sent, taking cases with effective intervention as positive samples and cases with ineffective intervention as negative samples, and executing incremental learning once a month, so that the model can adapt to the characteristics of different learning periods and different sources.
6. The academic early warning system is characterized by comprising: The data processing module is used for executing dynamic triple filling and feature extraction; The scoring engine module is deployed with a trained scoring model and is used for outputting learning situation scores and risk grades; the report generation module is deployed with an optimized generation model and is used for outputting a structured diagnosis report; and the closed loop feedback module is used for recording the intervention measures and effects and executing increment learning.
7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of claims 1-5 when executing the program.

Description

Double-engine collaborative framework-based learning situation large model construction method and academic early warning system Technical Field The invention belongs to the technical field of education informatization and artificial intelligence, and particularly relates to a learning situation large model construction method and a learning early warning system based on a double-engine collaborative framework. Background Along with popularization of intelligent campus construction, massive data is generated in the whole education process, however, the existing academic early warning and evaluation system has the following remarkable technical bottlenecks in practical application: 1) The data are heterogeneous and seriously missing, and a space-time relationship repair mechanism is lacked. The learned learning data are scattered in the isolated systems such as educational administration system, learning management system, library, all-purpose card and the like, and have great artificial randomness due to the interference of a plurality of factors (such as missing card and equipment failure), and meanwhile, obvious periodic characteristics (such as Zhou Kebiao rule) exist. In the prior art, simple mean filling or zero filling is adopted, so that the time sequence periodic characteristics of data are destroyed, and the portrait is distorted. 2) Lacks the ability to perceive "intermediate states" and "invisible risks". Traditional early warning models are mostly based on rules or simple supervised learning, and usually only can identify dominant problems which have occurred, however students often go through a complex intermediate process from normal state to academic crisis state, such as student learning effort but the method is not pair and is likely to cause inefficiency, and then student anxiety, finally the student performance is caused to slide down, the existing models lack modeling capability for the nonlinear and dynamic learning situation, and the existing models lack such negative samples (class/leave school), so that serious class imbalance problems occur to the models during training. 3) Score and explain cleavage, lack educational interpretability. The existing AI model is usually a "black box" and outputs a risk probability value, which cannot explain the specific decision basis. While suggestions generated based on the general large model are often too generalized (e.g., multi-reading, earnest learning), lack of attribution analysis for specific knowledge points and behavior patterns, and difficulty in guaranteeing educational ethical safety of the generated content (e.g., too hard mood may frustrate students' self-esteem). 4) A closed-loop evolutionary mechanism that lacks "sense-intervention-feedback". After the early warning of the education industry is sent, the intervention behaviors of teachers and the subsequent state changes of students are not structured and flow back to the model, and once the model parameters are deployed, the model parameters are fixed, so that the model parameters cannot adapt to the dynamic migration of external environments such as adjustment of teaching outline, variation of examination difficulty and the like in an education scene. Disclosure of Invention The invention aims to overcome the defects and shortcomings of the prior art, and provides a learning situation large model construction method and a learning industry early warning system based on a double-engine collaborative framework, which provide a full-chain technical framework from multi-source heterogeneous data construction, causal scoring engine, semantic reporting engine and closed-loop feedback. The invention solves the technical problems by the following technical proposal: a learning situation big model construction method based on a double-engine collaborative framework comprises the following steps: S1, preprocessing multi-source heterogeneous data and dynamically filling three times to construct a high-fidelity holographic data cube for learning behaviors of students; S2, constructing a scoring engine based on causal enhancement, so as to realize the problems of 'lack of negative samples' and 'difficulty in defining intermediate states', and realize the accurate mapping from behavior characteristics to learning situation scores (0-100 minutes); s3, constructing a report generation engine of knowledge distillation and DPO, and generating a diagnosis report which is structured and has temperature; S4, closed loop feedback and self-adaptive evolution. Moreover, the S1 specifically includes: S101, accessing a campus database through an ETL tool to realize data acquisition, and acquiring thirty items of core indexes in real time, wherein the indexes are divided into three types, namely (a) high-frequency time sequence data comprising daily learning system login time, library gate entry and exit time stamps, canteen consumption rules (used for analyzing living regularity) and dormitory time, (b) l