Search

CN-122023082-A - Learning early warning and personalized intervention method based on multi-mode data processing

CN122023082ACN 122023082 ACN122023082 ACN 122023082ACN-122023082-A

Abstract

The invention provides a learning early warning and personalized intervention method based on multi-mode data processing, which is characterized in that the comprehensive real-time acquisition of teaching process data is firstly carried out, and after a learning state prediction model is formed by constructing multi-dimensional data, accurate dynamic prediction early warning is realized, so that the early warning accuracy is high, the false alarm rate is low, meanwhile, a corresponding intervention scheme is generated based on the root of a problem and learning characteristics of different students, dynamic adjustment optimization is realized, the intervention effect is obvious, the intervention homogeneity is effectively reduced, and the accurate support is effectively provided for teaching management decision by constructing a macroscopic-mesoscopic-microscopic three-level data analysis system, so that the problems of incomplete data acquisition, inaccurate early warning prediction, homogeneous intervention measures and insufficient management decision support can be solved.

Inventors

  • LI MIANSHENG
  • YI XIAOXIN
  • LI YOUHONG
  • CHEN ZHUAN
  • GAO LE
  • CHEN LEI

Assignees

  • 广州华立学院
  • 中国电信股份有限公司广州分公司

Dates

Publication Date
20260512
Application Date
20260209

Claims (10)

  1. 1. A learning early warning and personalized intervention method based on multi-mode data processing is characterized by comprising the following steps: s1, acquiring multidimensional data through a multidimensional data acquisition module, wherein the multidimensional data comprises student classroom interaction data, resource use data and experimental operation data; S2, preprocessing multidimensional data to determine learning characteristics, learning problems and student groups of students, dividing the students into different groups, predicting learning trend based on a learning state prediction model, judging corresponding early warning categories and early warning grades according to the learning trend prediction, the different student groups and the learning early warning model, and dynamically updating the weight of the learning state prediction model and the threshold value of the learning early warning model through preset time intervals; S3, constructing a mapping relation between learning problems and root tags in a learning state prediction model, classifying the root tags of the learning problems, determining an intervention mode and a resource type according to the root tags and learning characteristics, screening out final intervention resources through quality sequencing, collecting feedback data in real time after intervention, realizing dynamic adjustment of an intervention scheme according to the feedback data, and forming optimal intervention resources after comprehensive dynamic adjustment to form a personalized chemogenetic intervention scheme.
  2. 2. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 1, wherein step S1 comprises steps S1.1-S1.3, S1.1, collecting student classroom interaction data: In a classroom, face or limb data of a student are collected in real time through shooting equipment arranged in a teaching environment, the classroom speaking content of the student is recorded through sound collecting equipment, the data are converted into text data through voice recognition, and indexes of speaking frequency, speaking duration and language logic are analyzed, and then the data are transmitted to a data processing server; s1.2, collecting resource use data: The method comprises the steps of recording a resource access address, access time, stay time and operation behavior of students when the students access teaching resources through a data acquisition plug-in real time, recording the time, accuracy and wrong question type of the students finishing online homework, and recording resource use data; s1.3, collecting experimental operation data: The key parameters in the experimental process are acquired in real time through the sensor arranged on the experimental equipment, the experimental operation process of students is recorded through the video equipment, and the completion sequence of experimental steps, whether the operation is standard and whether the experimental data recording is timely and accurate are identified through the video analysis technology.
  3. 3. The learning early warning and personalized intervention method based on multi-mode data processing according to claim 1, wherein in step S2, the method further comprises steps S2.1-S2.3; S2.1, preprocessing the collected multidimensional data, S2.2, determining learning characteristics, learning problems and student groups for the preprocessed multidimensional data, constructing a learning state prediction model and a learning early warning model, and judging early warning grades according to the learning state prediction model and the learning early warning model; s2.3, dynamically updating the weight of the learning state prediction model and the threshold value of the learning early warning model through preset time intervals; step S2.1 includes S2.1.1-S2.1.3, S2.1.1, eliminating extreme abnormal data through a 3 sigma principle, and complementing the missing data through a multiple interpolation method to obtain data with different dimensions; s2.1.2, carrying out characteristic standardization on data of different dimensions through a formula (1), Z = (X - μ) / σ(1), Z is the standard deviation multiple of the deviation of the data to be evaluated from the mean value, X is the data of single different dimension to be standardized, mu is the characteristic mean value, sigma is the characteristic standard deviation, and then the data value to be evaluated is mapped to a section with the mean value of 0 and the standard deviation of 1; S2.1.3, determining the association tightness degree of the data to be evaluated and the learning problem label through a mutual information algorithm, measuring the collinearity between the data to be evaluated through a variance expansion factor, and eliminating multiple collinearity characteristics.
  4. 4. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 3, wherein step S2.2 comprises steps S2.2.1-S2.2.3, S2.2.1, partitioning student populations by K-means clustering: Three types of characteristic data, namely a learning style score, a past learning period average score and a resource use preference, are selected as clustering dimensions, then the clustering quantity k is determined through an elbow rule (Elbow Method), k clustering centers are initialized, and the minimum intra-cluster Sum of Squares (SSE) is determined through more than two iterations so as to divide students into corresponding groups and output the characteristic average value of each group; S2.2.2, predicting learning trend through LSTM unit: Carrying out format conversion, dimension regulation and standardization verification on characteristic data of students in different groups, processing the characteristic data into a standard format (batch_size, T, D), wherein the batch_size in the standard format is the number of samples, T is the time sequence length of each sample, D is the characteristic dimension on each time step, then inputting the standard format (batch_size, T, D) into a hidden layer consisting of two LSTM units and Dropout layers for characteristic data fitting, and then outputting a learning state predicted value for n days in the future through a full-connection layer, wherein the output dimension is (batch_size, n); s2.2.3, judging the early warning category through a random forest classifier: And determining the deviation rate of the characteristic value of the data to be evaluated and the group mean value and the deviation of the LSTM unit predicted value and the normal threshold value, and then inputting the deviation rate of the group mean value and the deviation of the normal threshold value into a random forest classifier and outputting the early warning category and the early warning grade.
  5. 5. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 3, wherein step S2.3 comprises steps S2.3.1-S2.3.2, S2.3.1, collecting new student learning data features at preset time intervals, updating data of the random forest classifier in an incremental learning mode, and updating decision tree weights in the random forest classifier by using newly added learning data samples through weighted voting; s2.3.2, redetermining the characteristic mean values of different student groups through newly added learning data, redetermining the normal threshold range in the groups based on the latest characteristic mean values of the different student groups, and simultaneously combining early warning correction results fed back by teachers to adjust the classification threshold value of the random forest classifier.
  6. 6. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 1, wherein step S3 comprises steps S3.1-S3.4, S3.1, constructing a mapping relation of a learning problem and a root tag, and dividing the early warning problem into different major categories and minor categories so that each category of early warning problem corresponds to the root tag; s3.2, constructing a multi-dimensional matching matrix based on the root tag by combining three features of learning style, knowledge mastering level and time availability of students, and determining an intervention mode and a resource type; s3.3, establishing a resource label based on resources in the intervention resource library, determining the matching degree of student features and the resource label through a cosine similarity algorithm, screening a matching candidate resource pool, carrying out quality sequencing on the candidate resources by combining with resource use data, and selecting a final intervention resource; S3.4, collecting feedback data of students in the process of executing intervention tasks in real time, performing staged evaluation after completing preset intervention tasks, and then combining manual adjustment of teachers to realize dynamic adjustment of intervention schemes.
  7. 7. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 6, wherein step S3.2 comprises steps S3.2.1-S3.2.3, S3.2.1, respectively matching different learning resources for students of visual type, auditory type and kinesthesia type, and then respectively executing an intervention mode mainly comprising 'autonomous watching+visual exercise', mainly comprising 'audio learning+voice feedback', and mainly comprising 'actual exercise+instant feedback'; S3.2.2, respectively matching intervention resources which are mainly based on basic concept types, mainly based on knowledge consolidation and expansion, and mainly based on high expansion types for students according to the knowledge basic weak type, the knowledge medium level type and the knowledge excellent level type of student groups; s3.2.3, respectively making systematic, simplified and fragmented intervention schemes for students according to time-abundant, time-medium and time-tense student groups.
  8. 8. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 6, wherein step S3.3 comprises steps S3.3.1-S3.3.3, S3.3.1, constructing a resource label for each resource in the intervention resource library, wherein the resource label comprises a knowledge field label, a resource type label, a difficulty level label and a learning style adaptation label; s3.3.2, calculating the matching degree of learning features and resource labels by a cosine similarity algorithm based on the problem root labels, learning style labels and knowledge mastering level labels of students in the root labels, wherein the calculation formula is shown as follows, (2), In order for the degree of matching to be achieved, For the root tag weight to be the same, Weighting the resource labels, and screening intervention resources with matching degree more than or equal to 0.8 to enter a candidate resource pool; S3.3.3, quality sorting is carried out on the intervention resources of the candidate resources by combining with the resource use data of the students, and the intervention resources with scores larger than or equal to a preset score value and test accuracy improvement amplitude larger than or equal to 15% are selected as final intervention resources.
  9. 9. The learning pre-warning and personalized intervention method based on multi-modal data processing as set forth in claim 6, wherein step S3.4 comprises steps S3.4.1-S3.4.3, S3.4.1, collecting feedback data in real time in the process of executing the intervention task by the student, and triggering an intervention scheme adjustment mechanism when the feedback data triggers a preset adjustment threshold value; S3.4.2, after the students finish the preset intervention tasks, carrying out staged evaluation, adjusting the subsequent intervention schemes according to the evaluation results, wherein the evaluation results are characterized as evaluation scores, if the evaluation scores are improved by more than or equal to 20%, the intervention schemes are effective, the intervention difficulty is improved, or the intervention range is widened; S3.4.3, checking the execution condition and feedback data of the student intervention scheme through a teacher end management platform, manually adjusting intervention resources, intervention duration or intervention modes if the student intervention scheme has deviation, and marking adjustment reasons to realize dynamic adjustment of the intervention scheme.
  10. 10. The learning early warning and personalized intervention method based on multi-mode data processing according to claim 1, which is characterized by further comprising the steps of S4, taking three-level data of macroscopic teaching quality analysis, mesoscopic teaching process analysis and microscopic student individual intervention analysis as a core, combining with a multi-dimensional data analysis and diversification visualization component, realizing presentation and management decision of teaching process data, wherein the step S4 comprises the steps of S4.1-S4.3, S4.1, carrying out multi-dimensional data analysis by taking a macroscopic-mesoscopic-microscopic three-level data visual angle as a core through macroscopic teaching quality analysis, mesoscopic teaching process analysis and microscopic student individual analysis on a dynamic display module of a display screen; S4.2, the presentation and management decision of the teaching process 'macro-mesoscopic-microcosmic' three-level data are realized on the dynamic display module through the macroscopic data visualization component, the mesoscopic data visualization component and the microcosmic data visualization component; s4.3, acquiring multi-dimensional data selection information through an interaction module of the display screen, and correspondingly displaying corresponding early warning data.

Description

Learning early warning and personalized intervention method based on multi-mode data processing Technical Field The invention relates to the technical field of data processing methods, in particular to a learning early warning and personalized intervention method based on multi-mode data processing. Background Along with the rapid development of education informatization and big data service, the education informatization mainly optimizes the education process, improves the education quality and promotes education fairness through information technology means, and the deep fusion of the big data mining technology and the whole teaching process can realize accurate teaching and personalized learning, so that the comprehensive quality and innovation ability of students are improved, and the requirements of the personalized learning requirement of the students and the accurate improvement of the teaching quality are met. The method comprises the following steps of S1, synchronously collecting multi-mode data through a learning terminal built in an edge computing node, carrying out time-space alignment based on a time stamp, outputting integrated data, S2, dynamically distributing mode weights according to course types by the integrated data, carrying out multi-mode feature fusion according to a course type, generating fusion feature vectors, S3, constructing a cognitive state migration map based on the fusion feature vectors, dynamically updating knowledge node mastery degree by monitoring node state values in real time, S4, comparing the node state values with set dynamic threshold values, and triggering a hierarchical intervention strategy by combining multi-mode behavior combination according to comparison results. The above documents synchronously collect multi-mode data through edge computing nodes, monitor node state values in real time, dynamically update knowledge node mastery degree, and then determine hierarchical intervention strategies according to knowledge node mastery degree, namely, the knowledge node mastery degree is determined from multiple aspects, namely, analysis is performed from knowledge mastery conditions only, learning states cannot be determined from multiple data, such as real-time collection of procedural data of teaching resource use (e.g. video watching behavior, courseware labeling operation), experimental training operation (e.g. step normalization and data recording accuracy), learning dynamics of students cannot be reflected in time, the intervention mode effect of "one-cut" is poor according to the root cause of learning problems of students (e.g. concept understanding deviation and operation normalization deletion), learning styles (e.g. vision type and hearing type) and knowledge level formulation scheme, and therefore, individual problems of students are difficult to solve, the root cause of problems of different students can be determined according to different student groups, early warning and effective intervention can be truly realized, meanwhile, in the prior art, dynamic change of learning states cannot be adapted to by adopting unit early warning models for early warning, and timeliness is insufficient. Disclosure of Invention The invention aims to provide a learning early warning and personalized intervention method based on multi-mode data processing, which can realize the omnibearing real-time acquisition of teaching process data, the dynamic accurate early warning of student learning problems and the generation of a targeted personalized intervention scheme to provide systematic data support by processing multi-mode data through big data mining so as to solve the problems of incomplete data acquisition, inaccurate early warning prediction and homogeneous intervention measures. In order to achieve the above purpose, the invention provides a learning early warning and personalized intervention method based on multi-mode data processing, which comprises the following steps: s1, acquiring multidimensional data through a multidimensional data acquisition module, wherein the multidimensional data comprises student classroom interaction data, resource use data and experimental operation data; S2, preprocessing multidimensional data to determine learning characteristics, learning problems and student groups of students, dividing the students into different groups, predicting learning trend based on a learning state prediction model, judging corresponding early warning categories and early warning grades according to the learning trend prediction, the different student groups and the learning early warning model, and dynamically updating the weight of the learning state prediction model and the threshold value of the learning early warning model through preset time intervals; S3, constructing a mapping relation between the learning problem and the root tag, classifying the root tag of the learning problem, determining an intervention mode and a resource type according