A consistency detection and abnormal intervention method and device for mixed teaching

Abstract

This invention belongs to the field of intelligent education, and specifically relates to a method and device for consistency detection and anomaly intervention in blended learning. The method includes: S1: Constructing a preset teaching path model for each course based on the curriculum syllabus and storing it in a directed graph structure; S2: Collecting learning process data corresponding to the target course and converting it into a learning behavior sequence sorted by timestamp; S3: Comparing the learning behavior sequence with the preset teaching path model and calculating a learning consistency index, which includes sequence edit distance, common subsequence length ratio, and probability difference; S4: Determining students' abnormal behaviors based on the learning behavior sequence and the learning consistency index and generating abnormal behavior information for students; S5: Executing corresponding intervention actions according to predefined intervention rules. This addresses the problems of lagging monitoring of the learning process, superficial analysis, and intervention relying on manual intervention and lacking specificity.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education, and in particular relates to a method and device for consistency detection and anomaly intervention in blended learning. Background Technology

[0002] Blended learning is widely used in higher education and vocational training, organically integrating online self-study with face-to-face teaching. However, the separation of time and space in the teaching process makes it difficult for teachers to grasp the actual process and status of students during online self-study in a real-time and comprehensive manner. Offline teaching and online learning are prone to disconnect, resulting in teaching effectiveness falling short of expectations. Therefore, it is necessary to manage and monitor blended learning.

[0003] Currently, the management and monitoring solutions for blended learning mainly rely on learning management systems to provide superficial statistical reports such as access frequency, online time, and assignment scores, or to score and categorize students' learning effectiveness or engagement. This management method depends on lagging data such as tests, which cannot detect early deviations in the learning process or reveal in depth the learning path and behavioral patterns. Moreover, this management method based on superficial statistical reports still requires teachers to manually analyze the causes and decide on intervention measures. It cannot automatically trigger targeted remedial actions based on specific types of process anomalies, resulting in an insufficiently timely and accurate response. Summary of the Invention

[0004] This invention provides a method and apparatus for consistency detection and anomaly intervention in blended learning, which solves the problems of lagging monitoring of the learning process, superficial analysis, and intervention relying on manual intervention and lacking specificity.

[0005] This invention provides a basic solution: a consistency detection and anomaly intervention method for blended learning, specifically including the following steps: S1: Construct a pre-set teaching path model for each course based on the course syllabus and store it in a directed graph structure; S2: Collect learning process data corresponding to the target course and convert the learning process data into a learning behavior sequence sorted by timestamp; S3: Compare the learning behavior sequence with the preset teaching path model and calculate the learning consistency index, which includes sequence edit distance, common subsequence length ratio, and probability difference based on state transition. S4: Determine students' abnormal behavior based on learning behavior sequences and learning consistency indicators, and generate students' abnormal behavior information based on the determination results; S5: In response to the detection of abnormal behavior information, trigger and execute the corresponding intervention action according to the predefined intervention rules.

[0006] Preferably, in the directed graph structure of the preset teaching path model, nodes represent an independent learning activity unit in the course, and edges represent logical constraint relationships between learning activity units, including prerequisite relationships and recommended order relationships.

[0007] Preferably, in step S2, the learning process data includes video viewing logs, quiz submission records, discussion forum posting records, and resource download records collected from the online learning management platform, as well as classroom attendance records, classroom interaction records, and physical experiment operation records collected from the offline teaching environment; The process of converting learning process data into a sequence of learning behaviors includes: extracting student identifiers, activity types, activity object identifiers, and timestamps from each record, arranging and merging them in ascending order of timestamps to form a sequence with learning activity units as the basic elements.

[0008] Preferably, the sequence edit distance, the proportion of common subsequence lengths, and the probability difference based on state transitions are calculated as follows: The sequence edit distance is calculated by taking the learning behavior sequence as the first sequence and at least one standard path sequence derived from the preset teaching path model as the second sequence. The minimum number of single-step operations required to convert the first sequence into the second sequence is calculated using a dynamic programming algorithm, and the minimum number of single-step operations is used as the sequence edit distance value. The single-step operation includes inserting, deleting or replacing an activity unit. The common subsequence length ratio is obtained by calculating the longest common subsequence length between the learning behavior sequence and the standard path sequence, and dividing this length by the length of the standard path sequence. The probability difference degree based on state transition is calculated by constructing an ideal state transition probability matrix based on the connection relationship between nodes in the preset teaching path model; statistically analyzing the actual state transition frequency matrix based on the sequential relationship of activity units in the learning behavior sequence; and calculating the probability distribution difference value between the ideal state transition probability matrix and the actual state transition frequency matrix as the probability difference degree.

[0009] More preferably, in step S4, the abnormal behavior determination rules include: The direct judgment rule is that if there is an activity unit sequence in the learning behavior sequence that violates the prerequisite relationship in the preset teaching path model, then it is directly judged that there is an abnormal behavior of prerequisite violation. The threshold determination rule, based on preset abnormal thresholds for sequence edit distance, common subsequence length ratio, and probability difference, identifies abnormal behavior and further determines the abnormality type; the abnormality type determination strategy is as follows: When the sequence edit distance is greater than the corresponding abnormal threshold and the proportion of common subsequences is lower than the corresponding abnormal threshold, it is judged as an abnormal behavior of severe deviation of path order. When the proportion of common subsequences is lower than the corresponding abnormal threshold, it is judged as an abnormal behavior of missing core content; When the probability difference is greater than the corresponding abnormal threshold, it is judged as an abnormal behavior of abnormal learning rhythm. Sequence analysis judgment rules analyze the sequence characteristics of learning behavior. When the effective learning time ratio or the first attempt accuracy is lower than its corresponding threshold, it is judged as shallow participation type abnormal behavior.

[0010] Preferably, in step S4, the abnormal behavior is determined using a model, specifically including: The sequence edit distance, the proportion of common subsequence lengths, and the probability dissimilarity are combined to form a feature vector; The feature vector is input into a pre-trained abnormal behavior classification model, which outputs a determination result representing whether abnormal behavior exists and the corresponding type of abnormal behavior. The abnormal behavior types include those that violate prerequisites, those that deviate significantly from the path order, those that lack core content, those that exhibit abnormal learning pace, and those that involve superficial participation.

[0011] More preferably, the abnormal behavior classification model is a cascaded hybrid structure model, which includes: The fast filtering layer employs an unsupervised anomaly detection model built on the isolated forest algorithm or a single-class support vector machine to quickly screen out a high-risk subset of suspected abnormal behavior from all students. The fine-grained discrimination layer employs a supervised ensemble learning model, taking the high-risk subset output from the previous stage and its corresponding multi-dimensional feature vectors as input to perform the final abnormal behavior classification; the supervised ensemble learning model is either a gradient boosting decision tree model or a random forest model.

[0012] More preferably, the model training employs a phased, differentiated training strategy: For the fast filtering layer, unsupervised training is performed using unlabeled learning consistency index data of all historical students. For the fine-grained discrimination layer, supervised training is performed using training data labeled by experts in the field according to the type of abnormal behavior, and a cost-sensitive learning strategy is adopted during the training process.

[0013] More preferably, in step S5, the predefined intervention rules are stored as an abnormal behavior type-intervention action mapping table. The intervention actions include sending targeted learning tips and remedial resource links to the student's client, temporarily locking one or more subsequent learning activity nodes in the preset teaching path model, generating an early warning report containing abnormal details and suggestions, and pushing it to the teacher's management terminal.

[0014] Another basic solution provided by this invention: a device for consistency detection and anomaly intervention in blended learning, used to run the above method, comprising: The path model construction and storage module is used to construct and store preset teaching path models in a directed graph structure. The behavior sequence generation module is used to collect learning process data and generate learning behavior sequences sorted by timestamps. The consistency index calculation module is used to compare the learning behavior sequence with the preset teaching path model and calculate the sequence edit distance, the proportion of common subsequence length, and the probability difference based on state transition. The abnormal behavior determination module determines students’ abnormal behavior based on the learning behavior sequence and the calculated learning consistency index. The intervention execution module is used to trigger and execute corresponding intervention actions for students who are determined to have abnormal behavior, based on predefined intervention rules.

[0015] The principles and advantages of this invention are as follows: 1. By comparing the learning behavior sequence with the preset teaching path in real time, the learning trajectory of students can be presented in a timely manner, enabling teachers to clearly understand whether students are progressing in accordance with the teaching objectives. It can immediately identify situations such as students deviating from the path, learning stagnation, and abnormal pace, providing teachers with early warnings, avoiding the accumulation of problems, and overcoming the lag of traditional results-based data.

[0016] 2. By introducing multiple consistency indicators such as sequence edit distance, proportion of common subsequences, and state transition probability difference, the consistency of learning paths can be comprehensively evaluated from the perspectives of structure, order, and probability distribution. This can deeply quantify the degree of deviation between the learning process and the instructional design, and identify specific abnormal behavior patterns, rather than just performing superficial statistics.

[0017] 3. A rule mapping between abnormality types and intervention measures has been established, which can automatically trigger graded and classified intervention actions based on diagnostic results, improving the timeliness and accuracy of interventions, reducing the manual burden on teachers, freeing their energy from massive monitoring, and allowing them to focus on high-value human guidance, which greatly improves the efficiency and personalization of teaching support.

[0018] Definitions: Blended learning: A teaching model that organically combines online digital learning with traditional offline classroom teaching, aiming to leverage the advantages of both forms to improve learning effectiveness and flexibility.

[0019] Consistency in blended learning refers to the consistency between the pre-designed teaching path and the actual learning path, and also includes the consistency between learning behaviors, process data, and the achievement of the final learning objectives.

[0020] Pre-designed learning path: The ideal learning process designed by the teacher based on the teaching objectives. Anomalies in blended learning: behavioral patterns that deviate significantly from effective learning models and may lead to poor learning outcomes. Attached Figure Description

[0021] Figure 1 This is a flowchart of the present invention; Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0022] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: Example 1 See Figure 1 A consistency detection and anomaly intervention method for blended learning includes the following steps: S1: Construct a pre-set teaching path model for each course based on the course syllabus and store it in a directed graph structure; In the directed graph structure of the pre-defined teaching path model, nodes represent an independent learning activity unit in the course, and edges represent logical constraints between learning activity units. The logical constraints include prerequisite relationships and recommended order relationships. Specifically, the prerequisite relationship means that the learning activity corresponding to the source node must be completed before the learning activity corresponding to the target node is allowed to proceed. The recommended order relationship indicates that, in instructional design, it is suggested to perform the learning activities corresponding to the source node first, and then perform the learning activities corresponding to the target node. The independent learning activity unit refers to a basic teaching event that can be independently identified, recorded, and evaluated during the teaching process. The learning activity unit can be an online multimedia content learning event, an online assessment and practice event, an online or offline interactive discussion and practical operation event, as well as a learning summary and reflection event. Each learning activity unit has a unique identifier, a clear type attribute, and the specific knowledge content associated with it.

[0023] S2: Collect learning process data corresponding to the target course and convert the learning process data into a learning behavior sequence sorted by timestamp; In step S2, the learning process data includes video viewing logs, quiz submission records, discussion forum posting records, and resource download records collected from the online learning management platform, as well as classroom attendance records, classroom interaction records, and physical experiment operation records collected from the offline teaching environment. The process of converting learning process data into a sequence of learning behaviors includes: extracting student identifiers, activity types, activity object identifiers, and timestamps from each record, arranging and merging them in ascending order of timestamps to form a sequence with learning activity units as the basic elements.

[0024] S3: Compare the learning behavior sequence with the preset teaching path model and calculate the learning consistency index, which includes sequence edit distance, common subsequence length ratio, and probability difference based on state transition. The sequence edit distance, the proportion of common subsequence lengths, and the probability difference based on state transitions are calculated as follows: The sequence edit distance is calculated by taking the learning behavior sequence as the first sequence and at least one standard path sequence derived from the preset teaching path model as the second sequence. The minimum number of single-step operations required to convert the first sequence into the second sequence is calculated using a dynamic programming algorithm, and the minimum number of single-step operations is used as the sequence edit distance value. The single-step operation includes inserting, deleting or replacing an activity unit. Specifically, taking the learning behavior sequence as the first sequence and at least one standard path sequence derived from the preset teaching path model as the second sequence, a two-dimensional matrix is ​​constructed using dynamic programming. The number of rows in this matrix corresponds to the length of the first sequence plus one, and the number of columns corresponds to the length of the second sequence plus one. The value of each position in the matrix represents the minimum number of editing operations required to convert the first few elements of the first sequence into the first few elements of the second sequence. The editing operations are limited to inserting an active unit, deleting an active unit, or replacing an active unit. Starting from the top left corner of the matrix, the matrix is ​​filled recursively in row-major order. If the two sequence elements corresponding to the current position are the same, the current value is equal to the value at the top left diagonal position; if they are different, the current value is equal to the minimum of the values ​​at the position directly above (representing the insertion operation), the position directly to the left (representing the deletion operation), and the value at the top left diagonal position (representing the replacement operation) plus one. The value at the bottom right corner of the matrix represents the minimum number of single-step operations required, which serves as the sequence edit distance value.

[0025] For example, in this embodiment, a student's actual learning behavior sequence (sequence A) is: [watch video V1, take quiz Q1, watch video V2], and the corresponding compliant path sequence (sequence B) derived from the preset path model is: [watch video V1, watch video V2, take quiz Q1], which stipulates that V1 and V2 should be watched first, and then Q1 should be taken; The calculation process for sequence edit distance is as follows: Initialize a matrix with row and column lengths equal to the lengths of sequences A and B plus 1; Fill the matrix from top left to bottom right, where the value of each cell represents the minimum edit cost required to transform the first i elements of A into the first j elements of B; The filling rule is as follows: if A[i] == B[j], the cost is 0 (no operation required); otherwise, the cost is 1 + min(top value (insertion), left value (deletion), top-left value (replacement)). The value in the bottom right corner of the matrix is ​​the sequence edit distance. A distance of 0 indicates complete consistency; the larger the distance, the greater the difference between the actual behavior and the compliant path, and the more steps are needed to "modify".

[0026] The common subsequence length ratio is obtained by calculating the longest common subsequence length between the learning behavior sequence and the standard path sequence, and dividing this length by the length of the standard path sequence. Specifically, taking the learning behavior sequence as the first sequence and the standard path sequence as the second sequence, a two-dimensional matrix is ​​constructed using dynamic programming. The number of rows and columns of the matrix is ​​one more than the length of the two sequences. The value of each position in the matrix represents the length of the longest common subsequence between the corresponding prefix subsequences of the two sequences. If the current elements of the two sequences are the same, the value of the current position is equal to the value of the upper left corner plus one; if they are different, the value of the current position is the larger of the value of the position directly above it and the value of the position directly to its left. After filling, the value of the lower right corner of the matrix is ​​the length of the longest common subsequence. Finally, divide this length value by the length of the standard path sequence to obtain the proportion of the common subsequence length.

[0027] In this embodiment, the specific calculation process for the proportion of the length of the common subsequence between sequence A and sequence B is as follows: Calculate the length of the longest common subsequence LCS[i][j] between sequence A and sequence B. Specifically, if A[i] == B[j], then LCS[i][j] = LCS[i-1][j-1] + 1; otherwise, LCS[i][j] = max(LCS[i-1][j], LCS[i][j-1]). The proportion of common subsequence length is calculated based on the length of the longest common subsequence, specifically as LCS length / length of compliant sequence B or LCS length / max(length of sequence A, length of sequence B); A percentage of 1 indicates that the actual sequence fully contains the core order of the compliant path; the smaller the percentage, the further it deviates from the core path.

[0028] The probability difference degree based on state transition is calculated by constructing an ideal state transition probability matrix based on the connection relationship between nodes in the preset teaching path model; statistically analyzing the actual state transition frequency matrix based on the sequential relationship of activity units in the learning behavior sequence; and calculating the probability distribution difference value between the ideal state transition probability matrix and the actual state transition frequency matrix as the probability difference degree. Specifically, the steps for calculating the probability dissimilarity are as follows: 1) Derive the ideal transition probability matrix P from the preset path model. Treat each learning activity node in the model as a state. Based on the connection relationship and attributes of the directed edges between nodes in the model, define the ideal transition probability of each state (node) to all possible successor states. For example, if a node has only one recommended outgoing edge, the transition probability can be set to 1.0. If there are multiple parallel or optional outgoing edges, they can be assigned equal or weighted probability values ​​according to the instructional design. The sum of the probabilities of all outgoing edges is 1, thus obtaining a matrix, denoted as matrix P, where the element P(i, j) represents the ideal probability of transitioning from state i to state j. The matrix constructed in this way is denoted as the ideal state transition probability matrix P, where the element P(i, j) represents the ideal probability of transitioning from state i to state j. In this embodiment, according to the preset path graph, assign probabilities to the outgoing edges of each node. From the "Video V1" node, there is a 100% probability of pointing to the "Test Q1" node, so P(V1->Q1) = 1.0.

[0029] 2) Construct an actual transition frequency matrix Q based on the student behavior sequence. Iterate through the student's actual learning behavior sequence, counting the frequency of each pair of adjacent learning activity units (i.e., the previous state and the next state). Normalize all the counted transition frequencies. For each state i, divide the transition frequency from state i to each state j by the total transition frequency from state i to obtain the actual transition frequency. This yields another matrix, denoted as matrix Q, where the element Q(i, j) represents the actual observed frequency of transitioning from state i to state j. The constructed matrix is ​​denoted as the actual state transition frequency matrix Q, where the element Q(i, j) represents the actual observed frequency of transitioning from state i to state j. For example, traversing the student behavior sequence and counting the frequency of transitions from each state to the next, after "Video V1", there are 2 transitions to "Test Q1" and 1 transition to "Video V2". Therefore, Q(V1->Q1) = 0.67, Q(V1->V2) = 0.33.

[0030] 3) Calculate the KL divergence or JS divergence between the two matrices as the probability distribution dissimilarity. In this embodiment, KL divergence is used. The KL divergence formula is: D_KL(Q || P) = ΣQ(i) log(Q(i) / P(i)), where Q(i) is the actual transition frequency, P(i) is the ideal transition probability, i represents a possible state transition pair, and Σ represents the summation of all possible state transitions.

[0031] When the actual transition frequency distribution is exactly the same as the ideal transition probability distribution, the difference is 0. The larger the difference value, the further the student's actual learning path deviates from the preset ideal model in terms of state transition mode, and the more random, erratic or disordered the learning process may be.

[0032] S4: Determine students' abnormal behavior based on learning behavior sequences and learning consistency indicators, and generate students' abnormal behavior information based on the determination results; In step S4, the abnormal behavior is determined using a model, specifically including: The sequence edit distance, the proportion of common subsequence lengths, and the probability dissimilarity are combined to form a feature vector; The feature vector is input into a pre-trained abnormal behavior classification model, which outputs a determination result representing whether abnormal behavior exists and the corresponding type of abnormal behavior. The abnormal behavior types include those that violate prerequisites, those that deviate significantly from the path order, those that lack core content, those that exhibit abnormal learning pace, and those that involve superficial participation.

[0033] The abnormal behavior classification model is a cascaded hybrid structure model, which includes: The fast filtering layer employs an unsupervised anomaly detection model based on the Isolation Forest algorithm or a single-class support vector machine to quickly screen out a high-risk subset of students suspected of abnormal behavior. In this embodiment, the fast filtering layer uses an unsupervised anomaly detection model based on the Isolation Forest algorithm (Isolation Forest model). It constructs multiple isolation trees using the feature vectors (i.e., consistency index) of all historical students. For a new student feature vector, its anomaly score is evaluated by calculating the path length from the root node to the isolated leaf node in each tree. The shorter the path, the easier it is for the feature vector to be isolated, and the higher its probability of being abnormal. By setting a score threshold, students with high anomaly scores can be quickly screened out to form a high-risk subset. The fine-grained discriminant layer employs a supervised ensemble learning model, taking the high-risk subset output from the previous stage and its corresponding multi-dimensional feature vectors as input for final abnormal behavior classification. The supervised ensemble learning model is either a gradient boosting decision tree model or a random forest model. In this embodiment, the fine-grained discriminant layer uses a gradient boosting decision tree model (GBDT). This model takes the high-risk subset of students and their feature vectors output from the fast filtering layer as input for training and inference. It iteratively constructs a series of decision trees to optimize the prediction results. Each tree learns the residuals of the prediction results from all previous trees. This model is trained as a multi-classifier, and its output is the probability of belonging to each predefined abnormality type. The training process employs a cost-sensitive learning strategy, that is, assigning higher misclassification weights to minority class samples in the loss function to alleviate data imbalance and improve the ability to identify various abnormalities. The optimization algorithm uses gradient descent to minimize the weighted multi-class cross-entropy loss function.

[0034] The cascaded hybrid structure model is trained using a phased differentiation strategy: For the fast filtering layer, unsupervised training is performed using unlabeled learning consistency index data of all historical students. For the fine discriminative layer, supervised training is performed using training data labeled by experts in the field according to the type of abnormal behavior, and a cost-sensitive learning strategy is adopted during the training process to optimize the classification performance of minority class abnormal samples.

[0035] S5: In response to the detection of abnormal behavior information, trigger and execute the corresponding intervention action according to the predefined intervention rules.

[0036] In step S5, the predefined intervention rules are stored as an abnormal behavior type-intervention action mapping table, which defines the abnormal behavior types and their corresponding intervention actions: the intervention actions include sending targeted learning tips and remedial resource links to the student's client, temporarily locking one or more subsequent learning activity nodes in the preset teaching path model, generating an early warning report containing abnormal details and suggestions, and pushing it to the teacher management terminal.

[0037] In this embodiment, the intervention actions corresponding to the prerequisite violation type are immediate blocking and guidance (automatically locking the subsequent learning activity nodes that the student is currently trying to access but do not meet the prerequisites) and (precise resource push: automatically sending a prompt message to the student's client and attaching the learning resource link corresponding to the missing prerequisite knowledge); the intervention actions corresponding to the serious deviation of the path order type are learning path reconstruction suggestions (pushing personalized learning path adjustment suggestions to the student's client, highlighting the missing or out-of-order core node sequence) and teacher key warning (generating a warning report containing detailed deviation analysis and pushing it to the teacher's management terminal, prompting the teacher to pay special attention to the student or provide one-on-one tutoring), etc.; the intervention actions corresponding to the core content missing type are core content list reminders (sending reminders to the student's client, The system provides a list of potentially missing core learning activity units and a remedial learning package (automatically combining and pushing a learning resource package containing all missing core units). For students with abnormal learning pace, the corresponding intervention actions are progress monitoring and planning (sending learning progress reminders to the student's client and providing an adjusted, feasible learning time plan suggestion) and progress synchronization with the teacher (briefly informing the teacher's management console of the student's abnormal pace). For students with shallow participation, the corresponding intervention actions are mild prompts and guidance (sending encouraging prompts to the student's client, emphasizing the importance of deep learning, and possibly recommending more engaging learning resources) and mild marking on the teacher's end (marking the student as "to be observed" in the class view on the teacher's management console, without issuing a strong warning). After determining the type of abnormality and the specific context, by querying this mapping table, one or more corresponding intervention action combinations can be matched and executed to achieve automated and differentiated teaching intervention.

[0038] Example 2 The difference from Example 1 is that in step S4, the rules for determining abnormal behavior include: The direct judgment rule is that if there is an activity unit sequence in the learning behavior sequence that violates the prerequisite relationship in the preset teaching path model, then it is directly judged that there is an abnormal behavior of prerequisite violation. The threshold determination rule, based on preset abnormal thresholds for sequence edit distance, common subsequence length ratio, and probability difference, identifies abnormal behavior and further determines the abnormality type; the abnormality type determination strategy is as follows: When the sequence edit distance is greater than the corresponding abnormal threshold and the proportion of common subsequences is lower than the corresponding abnormal threshold, it is judged as an abnormal behavior of severe deviation of path order. When the proportion of common subsequences is lower than the corresponding abnormal threshold, it is judged as an abnormal behavior of missing core content; When the probability difference is greater than the corresponding abnormal threshold, it is judged as an abnormal behavior of abnormal learning rhythm. Sequence analysis judgment rules analyze the sequence characteristics of learning behavior. When the effective learning time ratio or the first attempt accuracy is lower than its corresponding threshold, it is judged as shallow participation type abnormal behavior.

[0039] Example 3 See Figure 2 A device for consistency detection and anomaly intervention in blended learning, used to implement the above method, includes: The path model construction and storage module is used to construct and store preset teaching path models in a directed graph structure. The behavior sequence generation module is used to collect learning process data and generate learning behavior sequences sorted by timestamps. The consistency index calculation module is used to compare the learning behavior sequence with the preset teaching path model and calculate the sequence edit distance, the proportion of common subsequence length, and the probability difference based on state transition. The abnormal behavior determination module determines students’ abnormal behavior based on the learning behavior sequence and the calculated learning consistency index. The intervention execution module is used to trigger and execute corresponding intervention actions for students who are determined to have abnormal behavior, based on predefined intervention rules.

[0040] The device is deployed using a client-server architecture. The core logic of the path model construction and storage module, consistency index calculation module, abnormal behavior judgment module, and intervention execution module is deployed on the backend server. The behavior sequence generation module includes a data collection agent deployed on student learning terminals (such as personal computers and tablets) to collect local learning behavior logs and upload them to the server. This module also includes interfaces for connecting to the online learning management platform and offline attendance and interaction systems to automatically obtain relevant logs. The teacher management terminal and student client serve as front-end interactive interfaces, provided as web applications or mobile applications respectively. The teacher management terminal is used to visually monitor class consistency status, receive early warning reports, and view intervention records. The student client is used to receive personalized learning prompts and resources automatically pushed by the system. All modules exchange data and transmit instructions through a secure network communication protocol.

[0041] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for consistency detection and anomaly intervention in blended learning, characterized in that, Specifically, the following steps are included: S1: Construct a pre-set teaching path model for each course based on the course syllabus and store it in a directed graph structure; S2: Collect learning process data corresponding to the target course and convert the learning process data into a learning behavior sequence sorted by timestamp; S3: Compare the learning behavior sequence with the preset teaching path model and calculate the learning consistency index, which includes sequence edit distance, common subsequence length ratio, and probability difference based on state transition. S4: Determine students' abnormal behavior based on learning behavior sequences and learning consistency indicators, and generate students' abnormal behavior information based on the determination results; S5: In response to the detection of abnormal behavior information, trigger and execute the corresponding intervention action according to the predefined intervention rules.

2. The consistency detection and anomaly intervention method for blended learning according to claim 1, characterized in that: In the directed graph structure of the pre-defined teaching path model, nodes represent an independent learning activity unit in the course, and edges represent logical constraints between learning activity units. These logical constraints include prerequisite relationships and recommended order relationships.

3. The consistency detection and anomaly intervention method for blended learning according to claim 1, characterized in that: In step S2, the learning process data includes video viewing logs, quiz submission records, discussion forum posting records, and resource download records collected from the online learning management platform, as well as classroom attendance records, classroom interaction records, and physical experiment operation records collected from the offline teaching environment. The process of converting learning process data into a sequence of learning behaviors includes: extracting student identifiers, activity types, activity object identifiers, and timestamps from each record, arranging and merging them in ascending order of timestamps to form a sequence with learning activity units as the basic elements.

4. The consistency detection and anomaly intervention method for blended learning according to claim 1, characterized in that: The sequence edit distance, the proportion of common subsequence lengths, and the probability difference based on state transitions are calculated as follows: The sequence edit distance is calculated by taking the learning behavior sequence as the first sequence and at least one standard path sequence derived from the preset teaching path model as the second sequence. The minimum number of single-step operations required to convert the first sequence into the second sequence is calculated using a dynamic programming algorithm, and the minimum number of single-step operations is used as the sequence edit distance value. The single-step operation includes inserting, deleting or replacing an activity unit. The common subsequence length ratio is obtained by calculating the longest common subsequence length between the learning behavior sequence and the standard path sequence, and dividing this length by the length of the standard path sequence. The probability difference based on state transition is used to construct an ideal state transition probability matrix based on the connection relationship between nodes in the preset teaching path model. Based on the sequential relationship of activity units in the learning behavior sequence, the actual state transition frequency matrix is ​​statistically analyzed. Calculate the difference in probability distribution between the ideal state transition probability matrix and the actual state transition frequency matrix, and use this difference as the probability difference degree.

5. The consistency detection and anomaly intervention method for blended learning according to claim 2, characterized in that: In step S4, the rules for determining abnormal behavior include: The direct judgment rule is that if there is an activity unit sequence in the learning behavior sequence that violates the prerequisite relationship in the preset teaching path model, then it is directly judged that there is an abnormal behavior of prerequisite violation. The threshold determination rule, based on preset abnormal thresholds for sequence edit distance, common subsequence length ratio, and probability difference, identifies abnormal behavior and further determines the abnormality type; the abnormality type determination strategy is as follows: When the sequence edit distance is greater than the corresponding abnormal threshold and the proportion of common subsequences is lower than the corresponding abnormal threshold, it is judged as an abnormal behavior of severe deviation of path order. When the proportion of common subsequences is lower than the corresponding abnormal threshold, it is judged as an abnormal behavior of missing core content; When the probability difference is greater than the corresponding abnormal threshold, it is judged as an abnormal behavior of abnormal learning rhythm. Sequence analysis judgment rules analyze the sequence characteristics of learning behavior. When the effective learning time ratio or the first attempt accuracy is lower than its corresponding threshold, it is judged as shallow participation type abnormal behavior.

6. The consistency detection and anomaly intervention method for blended learning according to claim 1, characterized in that: In step S4, the abnormal behavior is determined using a model, specifically including: The sequence edit distance, the proportion of common subsequence lengths, and the probability dissimilarity are combined to form a feature vector; The feature vector is input into a pre-trained abnormal behavior classification model, which outputs a determination result representing whether abnormal behavior exists and the corresponding type of abnormal behavior. The abnormal behavior types include those that violate prerequisites, those that deviate significantly from the path order, those that lack core content, those that exhibit abnormal learning pace, and those that involve superficial participation.

7. The consistency detection and anomaly intervention method for blended learning according to claim 6, characterized in that: The abnormal behavior classification model is a cascaded hybrid structure model, which includes: The fast filtering layer employs an unsupervised anomaly detection model built on the isolated forest algorithm or a single-class support vector machine to quickly screen out a high-risk subset of suspected abnormal behavior from all students. The fine-grained discrimination layer employs a supervised ensemble learning model, taking the high-risk subset and corresponding multi-dimensional feature vectors output from the previous stage as input to perform the final abnormal behavior classification; the supervised ensemble learning model is either a gradient boosting decision tree model or a random forest model.

8. The consistency detection and anomaly intervention method for blended learning according to claim 7, characterized in that, The model training employs a phased, differentiated training strategy: For the fast filtering layer, unsupervised training is performed using unlabeled learning consistency index data of all historical students. For the fine-grained discrimination layer, supervised training is performed using training data labeled by experts in the field according to the type of abnormal behavior, and a cost-sensitive learning strategy is adopted during the training process.

9. The consistency detection and anomaly intervention method for blended learning according to claim 5 or 6, characterized in that: In step S5, the predefined intervention rules are stored as an abnormal behavior type-intervention action mapping table. The intervention actions include sending targeted learning tips and remedial resource links to the student's client, temporarily locking one or more subsequent learning activity nodes in the preset teaching path model, generating an early warning report containing abnormal details and suggestions, and pushing it to the teacher's management terminal.

10. A device for consistency detection and anomaly intervention in blended learning, characterized in that, include: The path model construction and storage module is used to construct and store preset teaching path models in a directed graph structure. The behavior sequence generation module is used to collect learning process data and generate learning behavior sequences sorted by timestamps. The consistency index calculation module is used to compare the learning behavior sequence with the preset teaching path model and calculate the sequence edit distance, the proportion of common subsequence length, and the probability difference based on state transition. The abnormal behavior determination module determines students’ abnormal behavior based on the learning behavior sequence and the calculated learning consistency index. The intervention execution module is used to trigger and execute corresponding intervention actions for students who are determined to have abnormal behavior, based on predefined intervention rules.

Images