Multi-dimensional learning condition statistical analysis method based on deep learning

Abstract

The invention discloses a multi-dimensional learning condition statistical analysis method based on deep learning, and relates to the technical field of learning condition statistical analysis, and the method captures the learning performance of students in hearing, speaking, reading and writing dimensions, analyzes the learning behavior learning condition, measures the learning progress and quality of the students from multiple aspects, and improves the learning efficiency. The learning ability of students is graded by using Kmeans clustering and SVM, learning behavior changes of the students are tracked in real time in combination with an LSTM time sequence model, ability grading of the students is dynamically adjusted according to performance of the students in a specific learning stage, and a personalized learning report is generated for each student based on a multi-dimensional grading result of the students. According to the method and the system, the ability analysis and behavior data of the students are combined, a customized learning route map is conveniently provided for the students, meanwhile, the LSTM time sequence model is combined with the answer data and video learning behavior data of the students, subtle changes in the learning process are captured, the grading results of the students are updated in real time, and real-time and personalized learning support is provided for the students.

Description

Technical Field

[0001] The present invention relates to the technical field of student situation statistical analysis, and in particular to a multi-dimensional student situation statistical analysis method based on deep learning. Background Art

[0002] The current analysis of English learning status in junior and senior high schools mainly relies on data such as grades, homework completion and class participation to assess students' learning status. However, these data lack details and it is difficult to capture the specific changes in students' listening, speaking, reading and writing skills. Since these data are usually collected and analyzed regularly, there is a certain lag and it is difficult to reflect the fluctuations that may occur in students' learning process in real time.

[0003] To solve this problem, some learning analysis solutions introduce multi-dimensional tabular statistics or report data to provide learning feedback from different angles. For example, dimensions such as homework and after-class exercises are added, or semester grades are broken down into various test modules. Some solutions update learning data through periodic tests and conduct periodic comparative analysis.

[0004] However, these solutions still rely on teachers or systems to summarize and analyze large amounts of data, resulting in long analysis cycles and delayed feedback. In the study of oral and listening skills, personalized feedback and immediate correction are particularly important, but the delayed feedback of traditional solutions will affect the timeliness of teaching intervention. In addition, facing the diverse needs of different students, these static reporting solutions can only provide unified summary information and it is difficult to provide personalized suggestions for specific students. Therefore, a multi-dimensional learning situation statistical analysis method based on deep learning is urgently needed to solve such problems. Summary of the invention

[0005] In view of the above existing problems, the present invention is proposed.

[0006] The present invention provides a multi-dimensional learning situation statistical analysis method based on deep learning to solve the problem of data processing lag in traditional English learning situation statistical analysis methods. It is difficult for teachers to adjust teaching strategies according to real-time learning conditions and cannot meet the instant feedback needs of English subjects. At the same time, it lacks dynamic tracking capabilities and cannot effectively identify students' specific difficulties and progress at different learning stages.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0008] The present invention provides a multi-dimensional learning situation statistical analysis method based on deep learning, which comprises:

[0009] Step S1: Listening, speaking, reading and writing data collection.

[0010] Capture students’ learning process behavior data in real time;

[0011] Step S2, deep feature extraction and encoding,

[0012] Preprocess and clean the behavioral data to eliminate noise and outliers, and use a recurrent neural network to extract learning features; encode the learning features into a feature matrix, and represent each student's learning status and behavior pattern in the form of a feature matrix;

[0013] Learning characteristics include listening, speaking, reading and writing ability characteristics and learning behavior characteristics, specifically:

[0014] Based on natural language processing (NLP) and speech recognition technology, we extract students’ oral pronunciation, vocabulary usage, online answer error types and answer time characteristics to form listening, speaking, reading and writing ability characteristics.

[0015] Analyze students' answering time, error types, video viewing frequency, stay time, and video playback progress behavior characteristics to build a learning behavior model that reflects learning behavior characteristics;

[0016] Step S3, dynamic learning situation grading,

[0017] Based on the feature matrix of step S2, the long short-term memory network (LSTM) temporal model is used to grade the students’ learning process.

[0018] Step S4, personalized feedback,

[0019] According to the ability grading result in step S3, a personalized learning report is generated for each student.

[0020] Furthermore, behavioral data includes voice data, answer data, and video learning data, specifically:

[0021] Voice data is the student's spoken pronunciation and vocabulary usage data collected using voice recognition technology.

[0022] Answering data refers to students’ answering habits on the online platform, including the answering time and error types for each question.

[0023] Video learning data includes students’ video viewing frequency, duration of stay, and video playback progress data on the learning platform.

[0024] Furthermore, in step S1, the data collection step includes:

[0025] Construct a voice data model to capture voice data, collect students' pronunciation information, and analyze pronunciation fluency, pronunciation frequency, and grammar usage. The voice data model:

[0026] Among them, V srepresents the student's speech data score, which represents the student's performance in the speech task. i represents the weight associated with the ith pronunciation node, indicating pronunciation clarity, and f p (x i ) is the pronunciation clarity function, which determines the pronunciation accuracy based on the frequency domain features extracted from the audio signal, and b i represents the weight associated with the ith grammar node, indicating the correctness of the grammar usage, g s (y i ) is a grammatical error detection function, which determines whether the grammar is correct based on the grammatical structure analysis after speech recognition. n represents the number of pronunciations and grammatical nodes, representing the number of dimensions of analysis;

[0027] Capture answer data, including answer time, correctness, and error type of each question. Use the answer data model to capture answer data and establish a time series to dynamically track students' answer performance. Answer data model:

[0028] Among them, T q It represents the comprehensive score of the students in the task of answering questions, that is, the performance of the students on each question. It represents the time taken to answer the jth question, reflecting the student's response speed to the question. represents the error type coefficient of the jth question, representing the difficulty of the question type or the student's error pattern. represents the accuracy of the jth answer, i.e., the correct score, and m represents the total number of questions;

[0029] The video learning data model is used to capture video learning data. The video learning data model:

[0030] Among them, P v represents the comprehensive learning progress of students in the video learning task, that is, the learning efficiency of students, represents the weight associated with the kth video, represents the viewing time of the kth video, represents the duration of the kth video. represents the playback progress of the kth video, and p represents the total number of videos.

[0031] Furthermore, in step S2, the behavior data captured in step S1 is preprocessed and cleaned, and then the recurrent neural network RNN ​​is used to extract learning features and encode them into a feature matrix:

[0032] For each data item, noise removal and standardization are performed. Noise removal:

[0033] in, represents the data after noise removal, d i is the original data, Noise(d i ) represents the noise recognized by the filter;

[0034] Outlier processing: If but

[0035] like but in, For the final processed data, is the data after noise removal, μ d and σ d are the mean and standard deviation of the data set d, respectively, and k is a constant that controls the tolerance of outliers.

[0036] Furthermore, in step S2, the feature extraction step includes:

[0037] Natural language processing (NLP) and speech recognition technology are used to extract students' listening, speaking, reading and writing characteristics, and four ability characteristics are defined, including listening ability, speaking ability, reading ability and writing ability;

[0038] Extract listening ability features: Among them, h i is the listening ability characteristic of the i-th student, T is the total time period of the listening test, correct(t) is the number of correct answers to the listening test in time period t, and total(t) is the total number of questions in time period t;

[0039] Extracting oral ability features: Among them, s i is the oral ability characteristic of the i-th student, T is the total duration of the oral assessment, accuracy(t) is the pronunciation accuracy in time period t, and fluency(t) is the pronunciation fluency in time period t;

[0040] Extract reading ability features: Among them, r i is the reading ability characteristic of the i-th student, T is the total period of reading assessment, words_read(t) is the number of words read in time period t, and reading_time(t) is the reading time in time period t;

[0041] Extract writing ability features: Among them, w i is the writing ability characteristic of the i-th student, T is the total duration of the writing task, correct_sentences(t) is the number of correct sentences in time period t, and total_sentences(t) is the total number of sentences in time period t.

[0042] Furthermore, in step S2, the learning behavior pattern of each student is analyzed to extract characteristics of answering habits, learning time, and learning attitude. Specifically:

[0043] Extract answering behavior features: Among them, a i is the answering behavior characteristic of the i-th student, T is the total time period for answering questions, time_spent(t) is the time spent answering questions in time period t, and error_type(t) is the error type in time period t;

[0044] Extract video learning behavior features: Among them, v i is the video learning behavior feature of the i-th student, T is the total time period of video viewing, view_time(t) is the viewing time in time period t, and total_video_time(t) is the total duration of all videos in time period t;

[0045] The LSTM long short-term memory network is used to encode the extracted features, and each student's listening, speaking, reading and writing ability characteristics and learning behavior characteristics are combined to form a high-dimensional feature matrix: Set the feature matrix F i For the learning state and behavior pattern of student i, the LSTM model is used to model the sequence data, and the final output state vector is used as the feature vector of the student;

[0046] Feature matrix encoding: F i =[h i ,s i ,r i ,w i ,a i ,v i ], where F i is the characteristic matrix of student i, h i is the listening ability characteristic, s i is the oral ability feature, r i is the reading ability characteristic, w i For writing ability characteristics, a i is the answering behavior characteristic, v i Learning behavioral features for videos.

[0047] Furthermore, the ability grading methods include:

[0048] Based on K-means clustering and support vector machine (SVM), the listening, speaking, reading and writing ability characteristics are graded from four perspectives: listening, speaking, reading and writing.

[0049] And based on the LSTM time series model, the learning behavior characteristic data of students is tracked, and the learning behavior time series is automatically generated according to the students' historical learning behavior characteristic data. Based on this sequence, dynamic grading labels are assigned to students at a specific learning stage to achieve automatic grading adjustment.

[0050] Furthermore, in step S3, K-means clustering and support vector machine (SVM) are used to classify the four ability characteristics of listening, speaking, reading and writing. At the same time, the LSTM time series model is used to track the learning behavior characteristics of students and generate a learning behavior time series. Specifically:

[0051] Based on the characteristics of students' listening, speaking, reading and writing abilities, the K-means algorithm is used to perform preliminary grading of students. Assuming that in the feature matrix, the eigenvalues ​​of each student's listening, speaking, reading and writing are recorded as h, s, r, and w respectively, the K-means grading process formula is:

[0052] Among them, C j represents the cluster center of the jth group, N j is the number of students in the jth category, h i ,s i ,r i ,w i are the listening, speaking, reading and writing ability characteristics of student i, represents the mean of the listening, speaking, reading, and writing features of the j-th category center, and iteratively updates the distance between each student and the group center. K-means divides students into multiple categories with similar abilities to achieve preliminary grading of the four dimensions of listening, speaking, reading, and writing.

[0053] Based on the K-means grouping, SVM is used to classify the characteristics of students in each category, and the students in each category are divided into two groups according to their abilities:

[0054] Among them, α i is the Lagrange multiplier, N is the number of students in the category, y i is the classification label of student i, K(x i ,x j ) is the kernel function, defined as K(x i ,x j )=(h i ·h j +s i ·s j +r i ·r j +w i ·w j ), used to measure the similarity between student features, x i =[hi ,s i ,r i ,w i ] is the feature vector of student i.

[0055] Furthermore, in step S3, the LSTM time series model is used to capture the behavior changes of each student's learning behavior characteristics and generate a learning behavior time series:

[0056] Assume that the learning behavior features include the answering time, error type and video learning progress, recorded as t, e, v respectively, and the unit state update and output formula of the LSTM network is:

[0057] f t =σ(W f ·[t t-1 ,x t ]+b f ), where f t is the forget gate, which controls the retention ratio of the previous moment information at the current moment, W f is the weight matrix of the forget gate, t t-1 is the time taken to answer the question at the previous moment, x t =[t,e,v] is the behavior feature vector at the current moment, b f is the bias of the forget gate;

[0058] i t =σ(W i ·[t t-1 ,x t ]+b i ), where i t is the input gate, which controls the introduction ratio of the current behavior feature, W i is the weight matrix of the input gate, b i is the bias of the input gate;

[0059] C t =f t ·C t-1 +i t tanh(W c ·[t t-1 ,x t ]+b c ), where C t is the memory cell state at the current moment, C t-1 is the memory state of the previous moment, W c is the weight matrix of the memory cell state, b c is the bias of the memory unit;

[0060] o t =σ(W o ·[t t-1,x t ]+b o ), where o t is the output gate, which is used to control the output ratio of the current state, W o is the weight matrix of the output gate, b o is the bias of the output gate;

[0061] h t =o t tanh(C t ), where h t is the output of LSTM, that is, the encoding result of the state of learning behavior at the current moment. The student's learning behavior time series is encoded by LSTM, and the generated behavior feature sequence {h t}, which is used to dynamically assign grading labels. The model adjusts the students’ grading labels at each time point t according to the changes in behavioral characteristics, thereby achieving real-time and dynamic grading.

[0062] Furthermore, in step S4, the method of generating a personalized learning report is:

[0063] Based on the grading results of each student, ability evaluation scores are generated for the four dimensions of listening, speaking, reading and writing. A weighted scoring formula is used to integrate the characteristics of each ability to generate ability evaluation scores:

[0064] Among them, A i represents the overall ability evaluation score of student i, w k represents the weight of dimension k, f ik represents the ability score of student i on dimension k;

[0065] Based on the dynamic behavior characteristics of students, learning behavior analysis is performed and behavior scores are estimated using a time series model: Among them, B i represents the behavior score of student i, represents the overall performance of learning behavior, T represents the total number of time periods, and v t represents the video learning duration of the tth time period, t t represents the time taken to answer the questions in the tth time period, e t represents the error rate in the tth time period, α, β, γ represent the weight coefficients used to balance different behavior characteristics;

[0066] For the time series of grading results, the autoregressive integrated moving average ARIMA model is applied to predict the trend of each capability dimension. The trend prediction model is:

[0067] P i,k =φ1f i,k (t-1)+φ2f i,k (t-2)+…+φ pf i,k (tp)+θ0, where P i,k represents the predicted score of student i in the next stage on dimension k, f i,k (tj) represents the historical score of student i at the tjth moment on dimension k, φ j Represents the autoregressive coefficient of the ARIMA model, which is fitted based on historical data, and θ0 represents the constant term of the model, i.e., the basic score level;

[0068] Combine the ability scores, behavior scores and prediction results to generate personalized learning suggestions and learning path optimization formula: Among them, L i represents the comprehensive learning path score of student i, δ represents the ability evaluation weight, λ represents the behavior performance weight, μ represents the trend prediction weight, and A i ,B i ,P i,k represents the ability evaluation score, behavior score and prediction score of student i respectively.

[0069] The beneficial effects of the present invention are:

[0070] The present invention captures students' learning performance in listening, speaking, reading and writing, analyzes learning behavior and learning situation, measures students' learning progress and quality from multiple aspects, provides more comprehensive and accurate student learning situation, uses Kmeans clustering and SVM to grade students' learning ability, and combines LSTM time series model to track students' learning behavior changes in real time, and dynamically adjusts their ability classification according to their performance at a specific learning stage.

[0071] The present invention generates a personalized learning report for each student based on the students' multi-dimensional grading results, and proposes specific learning suggestions in combination with the students' ability analysis and behavior data, so as to provide students with a tailored learning roadmap. At the same time, the LSTM time series model is used in combination with the students' answer data and video learning behavior data to capture subtle changes in the learning process, update the students' grading results in real time, and dynamically generate a learning behavior time series.

[0072] The present invention combines the LSTM time series model with a dynamic grading mechanism to adjust students' grading labels in real time during the learning process, flexibly respond to the learning rhythm and changes of different students, has a high degree of adaptability, and provides students with real-time and personalized learning support. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0074] Figure 1 This is a flow chart of the multi-dimensional learning situation statistical analysis method based on deep learning of the present invention. DETAILED DESCRIPTION

[0075] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0076] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0077] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0078] Example 1, reference Figure 1 This embodiment provides a multi-dimensional learning situation statistical analysis method based on deep learning, including the following steps:

[0079] Step S1: Listening, speaking, reading and writing data collection.

[0080] Capture students’ learning process behavior data in real time;

[0081] Behavioral data includes voice data, answer data, and video learning data. Specifically:

[0082] Voice data is the student's spoken pronunciation and vocabulary usage data collected using voice recognition technology.

[0083] Answering data refers to students’ answering habits on the online platform, including the answering time and error types for each question.

[0084] Video learning data includes students’ video viewing frequency, duration of stay, and video playback progress data on the learning platform.

[0085] In step S1, the data collection step includes:

[0086] Construct a voice data model to capture voice data, collect students' pronunciation information, and analyze pronunciation fluency, pronunciation frequency, and grammar usage. The voice data model:

[0087] Among them, V s represents the student's speech data score, which represents the student's performance in the speech task. i represents the weight associated with the ith pronunciation node, indicating pronunciation clarity, and f p (x i ) is the pronunciation clarity function, which determines the pronunciation accuracy based on the frequency domain features extracted from the audio signal, and b i represents the weight associated with the ith grammar node, indicating the correctness of the grammar usage, g s (y i ) is a grammatical error detection function, which determines whether the grammar is correct based on the grammatical structure analysis after speech recognition. n represents the number of pronunciations and grammatical nodes, representing the number of dimensions of analysis;

[0088] Capture answer data, including answer time, correctness, and error type of each question. Use the answer data model to capture answer data and establish a time series to dynamically track students' answer performance. Answer data model:

[0089] Among them, T q It represents the comprehensive score of the students in the task of answering questions, that is, the performance of the students on each question. It represents the time taken to answer the jth question, reflecting the student's response speed to the question. represents the error type coefficient of the jth question, representing the difficulty of the question type or the student's error pattern. represents the accuracy of the jth answer, i.e., the correct score, and m represents the total number of questions;

[0090] The video learning data model is used to capture video learning data. The video learning data model:

[0091] Among them, P v represents the comprehensive learning progress of students in the video learning task, that is, the learning efficiency of students, represents the weight associated with the kth video, represents the viewing time of the kth video, represents the duration of the kth video. represents the playback progress of the kth video, and p represents the total number of videos;

[0092] Specifically, this method comprehensively captures the students' learning process, including multiple behavioral data such as voice data, answering data and video learning data, covering students' language ability, listening, speaking, reading, writing and learning behavior answering habits, video learning, learning time and error types in two important dimensions, to more accurately evaluate students' actual learning status and development potential.

[0093] Step S2, deep feature extraction and encoding,

[0094] Preprocess and clean the behavioral data to eliminate noise and outliers, and use a recurrent neural network to extract learning features; encode the learning features into a feature matrix, and represent each student's learning status and behavior pattern in the form of a feature matrix;

[0095] Learning characteristics include listening, speaking, reading and writing ability characteristics and learning behavior characteristics, specifically:

[0096] Based on natural language processing (NLP) and speech recognition technology, we extract students’ oral pronunciation, vocabulary usage, online answer error types and answer time characteristics to form listening, speaking, reading and writing ability characteristics.

[0097] Analyze students' answering time, error types, video viewing frequency, stay time, and video playback progress behavior characteristics to build a learning behavior model that reflects learning behavior characteristics;

[0098] In step S2, the behavior data captured in step S1 is preprocessed and cleaned, and then the recurrent neural network RNN ​​is used to extract learning features and encode them into a feature matrix:

[0099] For each data item, noise removal and standardization are performed. Noise removal:

[0100] in, represents the data after noise removal, d i is the original data, Noise(d i ) represents the noise recognized by the filter;

[0101] Outlier processing: If but

[0102] like but in, For the final processed data, is the data after noise removal, μ d and σ d are the mean and standard deviation of the data set d, respectively, and k is a constant that controls the tolerance of outliers;

[0103] In step S2, the feature extraction step includes:

[0104] Natural language processing (NLP) and speech recognition technology are used to extract students' listening, speaking, reading and writing characteristics, and four ability characteristics are defined, including listening ability, speaking ability, reading ability and writing ability;

[0105] Extract listening ability features: Among them, h i is the listening ability characteristic of the i-th student, T is the total time period of the listening test, correct(t) is the number of correct answers to the listening test in time period t, and total(t) is the total number of questions in time period t;

[0106] Extracting oral ability features: Among them, s i is the oral ability characteristic of the i-th student, T is the total duration of the oral assessment, accuracy(t) is the pronunciation accuracy in time period t, and fluency(t) is the pronunciation fluency in time period t;

[0107] Extract reading ability features: Among them, r i is the reading ability characteristic of the i-th student, T is the total period of reading assessment, words_read(t) is the number of words read in time period t, and reading_time(t) is the reading time in time period t;

[0108] Extract writing ability features: Among them, w i is the writing ability characteristic of the i-th student, T is the total duration of the writing task, correct_sentences(t) is the number of correct sentences in time period t, and total_sentences(t) is the total number of sentences in time period t;

[0109] In step S2, the learning behavior pattern of each student is analyzed to extract characteristics of answering habits, learning time, and learning attitude. Specifically:

[0110] Extract answering behavior features: Among them, a i is the answering behavior characteristic of the i-th student, T is the total time period for answering questions, time_spent(t) is the time spent answering questions in time period t, and error_type(t) is the error type in time period t;

[0111] Extract video learning behavior features: Among them, v iis the video learning behavior feature of the i-th student, T is the total time period of video viewing, view_time(t) is the viewing time in time period t, and total_video_time(t) is the total duration of all videos in time period t;

[0112] The LSTM long short-term memory network is used to encode the extracted features, and each student's listening, speaking, reading and writing ability characteristics and learning behavior characteristics are combined to form a high-dimensional feature matrix: Set the feature matrix F i For the learning state and behavior pattern of student i, the LSTM model is used to model the sequence data, and the final output state vector is used as the feature vector of the student;

[0113] Feature matrix encoding: F i =[h i ,s i ,r i ,w i ,a i ,v i ], where F i is the characteristic matrix of student i, h i is the listening ability characteristic, s i is the oral ability feature, r i is the reading ability characteristic, w i For writing ability characteristics, a i is the answering behavior characteristic, v i Learn behavioral features for videos;

[0114] Specifically, in terms of data processing, the LSTM model is used to perform deep feature extraction on students' learning features, effectively reducing data noise and bias and ensuring the accuracy of feature extraction. In addition, the LSTM network can reflect the long-term changing trends of students' learning behaviors by modeling time series data, capturing dynamic changes in the learning process, making the grading results more in line with the actual learning progress.

[0115] In particular, when analyzing learning behavior data, LSTM can effectively learn behavior patterns from historical data and encode them into feature vectors, effectively mining the temporal correlation of student behavior sequences. Compared with traditional static evaluation methods, it can provide real-time feedback while tracking changes in students' learning behaviors, and has higher timeliness and adaptability.

[0116] Step S3, dynamic learning situation grading,

[0117] Based on the feature matrix of step S2, the long short-term memory network (LSTM) temporal model is used to grade the students’ learning process.

[0118] The ability grading methods include:

[0119] Based on K-means clustering and support vector machine (SVM), the listening, speaking, reading and writing ability characteristics are graded from four perspectives: listening, speaking, reading and writing.

[0120] And based on the LSTM time series model, the learning behavior feature data of students is tracked, and the learning behavior time series is automatically generated according to the historical learning behavior feature data of students. Based on this sequence, dynamic grading labels are assigned to students at specific learning stages to achieve automatic grading adjustment;

[0121] In step S3, K-means clustering and support vector machine (SVM) are used to classify the four ability characteristics of listening, speaking, reading and writing. At the same time, the LSTM time series model is used to track the learning behavior characteristics of students and generate the learning behavior time series. Specifically:

[0122] Based on the characteristics of students' listening, speaking, reading and writing abilities, the K-means algorithm is used to perform preliminary grading of students. Assuming that in the feature matrix, the eigenvalues ​​of each student's listening, speaking, reading and writing are recorded as h, s, r, and w respectively, the K-means grading process formula is:

[0123] Among them, C j represents the cluster center of the jth group, N j is the number of students in the jth category, h i ,s i ,r i ,w i are the listening, speaking, reading and writing ability characteristics of student i, represents the mean of the listening, speaking, reading, and writing features of the j-th category center, and iteratively updates the distance between each student and the group center. K-means divides students into multiple categories with similar abilities to achieve preliminary grading of the four dimensions of listening, speaking, reading, and writing.

[0124] Based on the K-means grouping, SVM is used to classify the characteristics of students in each category, and the students in each category are divided into two groups according to their abilities:

[0125] Among them, α i is the Lagrange multiplier, N is the number of students in the category, y i is the classification label of student i, K(x i ,x j ) is the kernel function, defined as K(x i ,x j )=(h i ·h j +s i ·s j +ri ·r j +w i ·w j ), used to measure the similarity between student features, x i =[h i ,s i ,r i ,w i ] is the feature vector of student i;

[0126] In step S3, the LSTM time series model is used to capture the behavior changes of each student’s learning behavior characteristics and generate a learning behavior time series:

[0127] Assume that the learning behavior features include the answering time, error type and video learning progress, recorded as t, e, v respectively, and the unit state update and output formula of the LSTM network is:

[0128] f t =σ(W f ·[t t-1 ,x t ]+b f ), where f t is the forget gate, which controls the retention ratio of the previous moment information at the current moment, W f is the weight matrix of the forget gate, t t-1 is the time taken to answer the question at the previous moment, x t =[t,e,v] is the behavior feature vector at the current moment, b f is the bias of the forget gate;

[0129] i t =σ(W i ·[t t-1 ,x t ]+b i ), where i t is the input gate, which controls the introduction ratio of the current behavior feature, W i is the weight matrix of the input gate, b i is the bias of the input gate;

[0130] C t =f t ·C t-1 +i t tanh(W c ·[t t-1 ,x t ]+b c ), where C t is the memory cell state at the current moment, C t-1 is the memory state of the previous moment, W c is the weight matrix of the memory cell state, bc is the bias of the memory unit;

[0131] o t =σ(W o ·[t t-1 ,x t ]+b o ), where o t is the output gate, which is used to control the output ratio of the current state, W o is the weight matrix of the output gate, b o is the bias of the output gate;

[0132] h t =o t tanh(C t ), where h t is the output of LSTM, that is, the encoding result of the state of learning behavior at the current moment. The student's learning behavior time series is encoded by LSTM, and the generated behavior feature sequence {h t}, used to dynamically assign grading labels. The model adjusts the student's grading label at each time point t according to the changes in behavioral characteristics, thereby achieving real-time and dynamic grading;

[0133] Specifically, not only is the initial grading of students' listening, speaking, reading and writing abilities performed through Kmeans clustering and SVM, but dynamic ability adjustments are also performed through the LSTM time series model. The ability grading method based on behavioral data updates students' learning status in real time and adapts to students' different performances in the learning process. Through automated grading adjustments, students' ability grading is not only closely related to their current learning performance, but can also be dynamically adjusted according to the trend changes of their historical behavior patterns.

[0134] In addition, based on the ability grading results and behavior analysis, generating personalized learning reports not only provides students with comprehensive learning feedback, but also provides students with specific learning suggestions.

[0135] Step S4, personalized feedback,

[0136] Generate a personalized learning report for each student based on the ability grading result of step S3;

[0137] In step S4, the method of generating a personalized learning report is as follows:

[0138] Based on the grading results of each student, ability evaluation scores are generated for the four dimensions of listening, speaking, reading and writing. A weighted scoring formula is used to integrate the characteristics of each ability to generate ability evaluation scores:

[0139] Among them, A i represents the overall ability evaluation score of student i, w krepresents the weight of dimension k, f ik represents the ability score of student i on dimension k;

[0140] Based on the dynamic behavior characteristics of students, learning behavior analysis is performed and behavior scores are estimated using a time series model: Among them, B i represents the behavior score of student i, represents the overall performance of learning behavior, T represents the total number of time periods, and v t represents the video learning duration of the tth time period, t t represents the time taken to answer the questions in the tth time period, e t represents the error rate in the tth time period, α, β, γ represent the weight coefficients used to balance different behavior characteristics;

[0141] For the time series of grading results, the autoregressive integrated moving average ARIMA model is applied to predict the trend of each capability dimension. The trend prediction model is:

[0142] P i,k =φ1f i,k (t-1)+φ2f i,k (t-2)+…+φ p f i,k (tp)+θ0, where P i,k represents the predicted score of student i in the next stage on dimension k, f i,k (tj) represents the historical score of student i at the tjth moment on dimension k, φ j Represents the autoregressive coefficient of the ARIMA model, which is fitted based on historical data, and θ0 represents the constant term of the model, i.e., the basic score level;

[0143] Combine the ability scores, behavior scores and prediction results to generate personalized learning suggestions and learning path optimization formula: Among them, L i represents the comprehensive learning path score of student i, δ represents the ability evaluation weight, λ represents the behavior performance weight, μ represents the trend prediction weight, and A i ,B i ,P i,k represents the ability evaluation score, behavior score and prediction score of student i respectively.

[0144] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-dimensional learning situation statistical analysis method based on deep learning, characterized by: include, Step S1, listening, speaking, reading and writing data collection, capturing students' learning process behavior data in real time; Step S2, deep feature extraction and encoding, preprocessing and cleaning the behavior data, eliminating noise and outliers, and using a recurrent neural network to extract learning features; Encode learning characteristics into feature matrices, and represent each student's learning status and behavior pattern in the form of feature matrices; Learning characteristics include listening, speaking, reading and writing ability characteristics and learning behavior characteristics, specifically: Based on natural language processing (NLP) and speech recognition technology, we extract students’ oral pronunciation, vocabulary usage, online answer error types and answer time characteristics to form listening, speaking, reading and writing ability characteristics. Analyze students' answering time, error types, video viewing frequency, stay time, and video playback progress behavior characteristics to build a learning behavior model that reflects learning behavior characteristics; Step S3, dynamic learning situation grading, based on the feature matrix of step S2, using the long short-term memory network LSTM time series model to grade the students' learning process; Step S4, personalized feedback, generates a personalized learning report for each student based on the ability grading results of step S3.

2. According to a multi-dimensional learning situation statistical analysis method based on deep learning according to claim 1, it is characterized in that: Behavioral data includes voice data, answer data, and video learning data. Specifically: Voice data is the student's spoken pronunciation and vocabulary usage data collected using voice recognition technology. Answering data refers to students’ answering habits on the online platform, including the answering time and error types for each question. Video learning data includes students’ video viewing frequency, duration of stay, and video playback progress data on the learning platform.

3. According to a multi-dimensional learning situation statistical analysis method based on deep learning according to claim 2, it is characterized in that: In step S1, the data collection step includes: Construct a voice data model to capture voice data, collect students' pronunciation information, and analyze pronunciation fluency, pronunciation frequency, and grammar usage. The voice data model: Among them, V s represents the student's speech data score, which represents the student's performance in the speech task. i represents the weight associated with the ith pronunciation node, indicating pronunciation clarity, and f p (x i ) is the pronunciation clarity function, which determines the pronunciation accuracy based on the frequency domain features extracted from the audio signal, and b i represents the weight associated with the ith grammar node, indicating the correctness of the grammar usage, g s (y i ) is a grammatical error detection function, which determines whether the grammar is correct based on the grammatical structure analysis after speech recognition. n represents the number of pronunciations and grammatical nodes, representing the number of dimensions of analysis; Capture answer data, including answer time, correctness, and error type for each question. Use the answer data model to capture answer data and establish a time series to dynamically track students' answer performance. Answer data model: Among them, T q It represents the comprehensive score of the students in the task of answering questions, that is, the performance of the students on each question. It represents the time taken to answer the jth question, reflecting the student's response speed to the question. represents the error type coefficient of the jth question, representing the difficulty of the question type or the student's error pattern. represents the accuracy of the jth answer, i.e., the correct score, and m represents the total number of questions; The video learning data model is used to capture video learning data. The video learning data model: Among them, P v represents the comprehensive learning progress of students in the video learning task, that is, the learning efficiency of students, represents the weight associated with the kth video, represents the viewing time of the kth video, represents the duration of the kth video. represents the playback progress of the kth video, and p represents the total number of videos.

4. According to a multi-dimensional learning situation statistical analysis method based on deep learning according to claim 3, it is characterized in that: In step S2, the behavior data captured in step S1 is preprocessed and cleaned, and then the recurrent neural network RNN ​​is used to extract learning features and encode them into a feature matrix: For each data item, noise removal and standardization are performed. Noise removal: in, represents the data after noise removal, d i is the original data, Noise(d i ) represents the noise identified by the filter; Outlier processing: If but like but in, For the final processed data, is the data after noise removal, μ d and σ d are the mean and standard deviation of the data set d, respectively, and k is a constant that controls the tolerance of outliers.

5. According to a multi-dimensional learning situation statistical analysis method based on deep learning according to claim 4, it is characterized in that: In step S2, the feature extraction step includes: Natural language processing (NLP) and speech recognition technology are used to extract students' listening, speaking, reading and writing characteristics, and four ability characteristics are defined, including listening ability, speaking ability, reading ability and writing ability; Extracting listening ability features: Among them, h i is the listening ability characteristic of the i-th student, T is the total time period of the listening test, correct(t) is the number of correct answers to the listening test in time period t, and total(t) is the total number of questions in time period t; Extracting oral ability features: Among them, s i is the oral ability characteristic of the i-th student, T is the total duration of the oral assessment, accuracy(t) is the pronunciation accuracy in time period t, and fluency(t) is the pronunciation fluency in time period t; Extract reading ability features: Among them, r i is the reading ability characteristic of the i-th student, T is the total period of reading assessment, words_read(t) is the number of words read in time period t, and reading_time(t) is the reading time in time period t; Extract writing ability features: Among them, w i is the writing ability characteristic of the i-th student, T is the total duration of the writing task, correct_sentences(t) is the number of correct sentences in time period t, and total_sentences(t) is the total number of sentences in time period t.

6. According to the multi-dimensional learning situation statistical analysis method based on deep learning in claim 5, it is characterized in that: In step S2, the learning behavior pattern of each student is analyzed to extract characteristics of answering habits, learning time, and learning attitude. Specifically: Extract answering behavior features: Among them, a i is the answering behavior characteristic of the i-th student, T is the total time period for answering questions, time_spent(t) is the answering time in time period t, and error_type(t) is the error type in time period t; Extract video learning behavior features: Among them, v i is the video learning behavior feature of the i-th student, T is the total time period of video viewing, view_time(t) is the viewing time in time period t, and total_video_time(t) is the total duration of all videos in time period t; The LSTM long short-term memory network is used to encode the extracted features, and each student's listening, speaking, reading and writing ability characteristics and learning behavior characteristics are combined to form a high-dimensional feature matrix: Set the feature matrix F i For the learning state and behavior pattern of student i, the LSTM model is used to model the sequence data, and the final output state vector is used as the feature vector of the student; Feature matrix encoding: F i =[h i ,s i ,r i ,w i ,a i ,v i ], where F i is the characteristic matrix of student i, h i is the listening ability characteristic, s i is the oral ability feature, r i is the reading ability characteristic, w i For writing ability characteristics, a i is the answering behavior characteristic, v i Learning behavioral features for videos.

7. The multi-dimensional student situation statistical analysis method based on deep learning according to claim 6 is characterized in that: The ability grading methods include: Based on K-means clustering and support vector machine (SVM), the listening, speaking, reading and writing ability characteristics are graded from four perspectives: listening, speaking, reading and writing. And based on the LSTM time series model, the students' learning behavior feature data is tracked, and the learning behavior time series is automatically generated according to the students' historical learning behavior feature data. Based on this sequence, dynamic grading labels are assigned to students at a specific learning stage.

8. The multi-dimensional student situation statistical analysis method based on deep learning according to claim 7 is characterized in that: In step S3, K-means clustering and support vector machine (SVM) are used to classify the four ability characteristics of listening, speaking, reading and writing. At the same time, the LSTM time series model is used to track the learning behavior characteristics of students and generate the learning behavior time series. Specifically: Based on the characteristics of students' listening, speaking, reading and writing abilities, the K-means algorithm is used to perform preliminary grading of students. Assuming that in the feature matrix, the eigenvalues ​​of each student's listening, speaking, reading and writing are recorded as h, s, r, and w respectively, the K-means grading process formula is: Among them, C j represents the cluster center of the jth group, N j is the number of students in the jth category, h i ,s i ,r i ,w i are the listening, speaking, reading and writing ability characteristics of student i, represents the mean of the listening, speaking, reading, and writing features of the j-th category center; Based on the K-means grouping, SVM is used to classify the characteristics of students in each category, and the students in each category are divided into two groups according to their abilities: Among them, α i is the Lagrange multiplier, N is the number of students in the category, y i is the classification label of student i, K(x i ,x j ) is the kernel function, defined as K(x i ,x j )=(h i ·h j +s i ·s j +r i ·r j +w i ·w j ) is used to measure the similarity between student features, x i =[h i ,s i , r i , w i ] is the feature vector of student i.

9. The multi-dimensional student situation statistical analysis method based on deep learning according to claim 8 is characterized in that: In step S3, the LSTM time series model is used to capture the behavior changes of each student’s learning behavior characteristics and generate a learning behavior time series: Assume that the learning behavior features include the answering time, error type and video learning progress, recorded as t, e, v respectively, and the unit state update and output formula of the LSTM network is: f t =σ(W f ·[t t-1 , x t ]+b f ), where f t is the forget gate, which controls the retention ratio of the previous moment information at the current moment, W f is the weight matrix of the forget gate, t t-1 is the time taken to answer the question at the previous moment, x t =[t, e, v] is the behavior feature vector at the current moment, b f is the bias of the forget gate; i t =σ(W i ·[t t-1 , x t ]+b i ), where i t is the input gate, which controls the introduction ratio of the current behavior feature, W i is the weight matrix of the input gate, b i is the bias of the input gate; C t =f t ·C t-1 +i t tanh(W c ·[t t-1 , x t ]+b c ), where C t is the memory cell state at the current moment, C t-1 is the memory state of the previous moment, W c is the weight matrix of the memory cell state, b c is the bias of the memory unit; o t =σ(W o ·[t t-1 , x t ]+b o ), where o t is the output gate, which is used to control the output ratio of the current state, W o is the weight matrix of the output gate, b o is the bias of the output gate; h t =o t tanh(C t ), where h t is the output of LSTM, that is, the encoding result of the state of the learning behavior at the current moment.

10. A multi-dimensional student situation statistical analysis method based on deep learning according to claim 9, characterized in that: In step S4, the method of generating a personalized learning report is as follows: Based on the grading results of each student, ability evaluation scores are generated for the four dimensions of listening, speaking, reading and writing. A weighted scoring formula is used to integrate the characteristics of each ability to generate ability evaluation scores: Among them, Ai represents the overall ability evaluation score of student i, w k represents the weight of dimension k, f ik represents the ability score of student i on dimension k; Based on the dynamic behavior characteristics of students, learning behavior analysis is performed and behavior scores are estimated using a time series model: Among them, B i represents the behavior score of student i, represents the overall performance of learning behavior, T represents the total number of time periods, and v t represents the video learning duration of the tth time period, t t represents the time taken to answer the questions in the tth time period, e t represents the error rate in the tth time period, α, β, γ represent the weight coefficients used to balance different behavior characteristics; For the time series of grading results, the autoregressive integrated moving average ARIMA model is applied to predict the trend of each capability dimension. The trend prediction model is: P i，k =φ1f i，k (t-1)+φ2f i，k (t-2)+...+φ p f i，k (tp)+θ0, where P i，k represents the predicted score of student i in the next stage on dimension k, f i，k (tj) represents the historical score of student i at the tjth moment on dimension k, φ j Represents the autoregressive coefficient of the ARIMA model, which is fitted based on historical data. θ0 represents the constant term of the model, i.e., the basic score level. Combine the ability scores, behavior scores and prediction results to generate personalized learning suggestions and learning path optimization formula: Among them, L i represents the comprehensive learning path score of student i, δ represents the ability evaluation weight, λ represents the behavior performance weight, μ represents the trend prediction weight, and A i , B i , P i，k represents the ability evaluation score, behavior score and prediction score of student i respectively.

Images