Group teaching strategy generation method and device based on intelligent learning cloud platform

Abstract

The application is suitable for the technical field of group teaching, and particularly relates to a group teaching strategy generation method and device based on a smart learning cloud platform, which comprises the following steps: obtaining teaching behavior data, biological data and teaching progress data of multiple classes through the smart learning cloud platform; performing space-time alignment processing on the teaching stages of the multiple classes according to the teaching progress data of the multiple classes to obtain a teaching stage sequence; performing group anomaly detection on the multiple classes based on each teaching stage of the teaching stage sequence to obtain an anomaly detection result; and generating a group teaching strategy based on the teaching behavior data and the biological data of the teaching stage in the anomaly detection result. The method can discover group problems in time and correct them, help teachers make more accurate teaching decisions, and thus improve the overall teaching quality and learning effect.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of group teaching, and particularly relates to a group teaching strategy generation method and device based on a smart learning cloud platform. BACKGROUND

[0002] With the development of science and technology, online education platforms and network education resources are increasing, teachers can understand the learning situation of students through the education platform, and can realize a group teaching mode through the education platform. In the mode, teachers and students are connected through the platform, teachers can face multiple students for teaching at the same time, and students can participate in interactive learning through the platform. The platform usually provides functions such as video live broadcast, courseware sharing, and discussion area, to promote communication and cooperation between teachers and students and between students.

[0003] In the prior art, in the teaching environment of multiple classes or across school districts, student backgrounds and learning progress are different. Although the group teaching based on the education platform can enable teachers to access more data, it is still a challenge to systematically identify common problems among different classes and different school districts. Although the education platform realizes online interaction and courseware sharing, teachers still need to spend a lot of time analyzing the learning situation of students to design a teaching plan. Even if sampling checking student homework or test scores can save time, it is easy to cause teachers to overlook some implicit misunderstandings and confusion of knowledge points.

[0004] In summary, when using the education platform to teach students in groups, there is a problem of lagging discovery of group learning blind spots, which leads to a decrease in teaching quality. SUMMARY

[0005] The embodiments of the application provide a group teaching strategy generation method and device based on a smart learning cloud platform, which can solve the problem of lagging discovery of group learning blind spots when using the education platform to teach students in groups in the related art, which leads to a decrease in teaching quality.

[0006] In a first aspect, the embodiments of the application provide a group teaching strategy generation method based on a smart learning cloud platform, including:

[0007] obtaining teaching behavior data, biological data, and teaching progress data of multiple classes through the smart learning cloud platform; wherein the teaching progress data includes taught knowledge points and corresponding teaching dates;

[0008] performing space-time alignment processing on teaching stages of the multiple classes according to the teaching progress data of the multiple classes to obtain a teaching stage sequence; wherein the teaching stage sequence includes multiple teaching stages, and each teaching stage includes teaching behavior data and biological data of the multiple classes in the same teaching stage;

[0009] perform group anomaly detection on the multiple classes based on each teaching stage of the teaching stage sequence to obtain an anomaly detection result; wherein the anomaly detection result comprises a teaching stage in which a group anomaly exists;

[0010] generate a group teaching strategy based on the teaching behavior data and the biological data of the teaching stage in the anomaly detection result.

[0011] The technical solutions described above in the embodiments of the present application have at least the following technical effects:

[0012] The group teaching strategy generation method based on the smart learning cloud platform provided in the present application first acquires teaching behavior data, biological data, and teaching progress data (taught knowledge points and corresponding teaching dates) of multiple classes through the smart learning cloud platform, then performs spatiotemporal alignment processing on the teaching stages of the multiple classes according to the teaching progress data of the multiple classes to obtain a teaching stage sequence (the teaching stage sequence comprises multiple teaching stages, and each teaching stage comprises teaching behavior data and biological data of the multiple classes in the same teaching stage), then performs group anomaly detection on the multiple classes based on each teaching stage of the teaching stage sequence to obtain an anomaly detection result (a teaching stage in which a group anomaly exists), and finally generates a group teaching strategy based on the teaching behavior data and the biological data of the teaching stage in the anomaly detection result. This method can timely identify a teaching stage in which a group learning blind spot or learning difficulty exists, and then adjust the teaching strategy, which is conducive to timely finding problems and correcting them, and avoiding a decline in teaching quality. This method not only can improve the learning progress and understanding of individual students, but also can optimize the learning atmosphere and interaction of the entire class. This optimization is targeted at collective learning blind spots and difficulties, avoids some students being ignored in group learning, and promotes the learning progress of all students. The generated group teaching strategy can better cope with different learning needs of different classes and students in the same teaching stage, help teachers make more accurate teaching decisions, and thus improve the overall teaching quality and learning effect.

[0013] In a second aspect, the embodiments of the present application provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the method in any of the embodiments of the first aspect when executing the computer program.

[0014] It can be understood that the beneficial effects of the above-mentioned second aspect can be referred to the related description in the above-mentioned first aspect, which will not be repeated here. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0016] Figure 1 is a flowchart of a group teaching strategy generation method based on a smart learning cloud platform provided by an embodiment of the present application;

[0017] Figure 2 is an implementation flowchart of group anomaly detection in the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiment of the present application;

[0018] Figure 3 is a schematic diagram of the overall architecture of the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiment of the present application;

[0019] Figure 4 is an example diagram of an equivalent knowledge point table in the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiment of the present application. DETAILED DESCRIPTION

[0020] In the following description, specific details such as specific system structures, techniques, etc. are presented in order to thoroughly understand the embodiments of the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits and methods are omitted so as not to obscure the description of the present application with unnecessary details.

[0021] It should be understood that when used in the present application specification and the appended claims, the term "comprising" indicates the presence of the described features, whole, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, whole, steps, operations, elements, components and / or sets thereof.

[0022] In the related art, in a multi-class or cross-campus teaching environment, students have different backgrounds and learning progress. Although group teaching based on an education platform can enable teachers to access more data, it is still a challenge to systematically identify common problems among different classes and different campuses. Especially when there is a lack of unified data analysis and management tools, teachers have difficulty in conducting comprehensive learning analysis through a unified platform, making it difficult to find common problems among students in different areas. Although the education platform realizes online interaction and courseware sharing, teachers still need to spend a lot of time analyzing students' learning to design teaching plans. Even if sampling inspection of student homework or test scores can save time, it is easy to cause teachers to overlook some implicit misunderstandings and confusion of knowledge points. For example, although some students answer correctly, they may have conceptual confusion or not fully understand some details, which are often difficult to be discovered and corrected in time when only sampling inspection is used.

[0023] To solve the above problems, the embodiment of the present application provides a group teaching strategy generation method and device based on a smart learning cloud platform. In the method, first, the teaching behavior data, biological data, and teaching progress data (taught knowledge points and corresponding teaching dates) of multiple classes are obtained through the smart learning cloud platform. Then, according to the teaching progress data of the multiple classes, the teaching stages of the multiple classes are processed for space-time alignment to obtain a teaching stage sequence (the teaching stage sequence includes multiple teaching stages, and each teaching stage includes the teaching behavior data and biological data of the multiple classes in the same teaching stage). Then, based on each teaching stage of the teaching stage sequence, group anomaly detection is performed on the multiple classes to obtain an anomaly detection result (a teaching stage with group anomaly). Finally, based on the teaching behavior data and biological data of the teaching stage in the anomaly detection result, a group teaching strategy is generated. The method can timely identify a teaching stage with group learning blind spots or learning difficulties, and then adjust the teaching strategy, which is beneficial to timely find and correct problems and avoid a decline in teaching quality. The method not only improves the learning progress and understanding of individual students, but also optimizes the learning atmosphere and interaction of the entire class. This optimization targets collective learning blind spots and difficulties, avoids neglecting some students in group learning, and promotes the learning progress of all students. The generated group teaching strategy can better cope with different learning needs of different classes and students in the same teaching stage, help teachers make more accurate teaching decisions, and thus improve the overall teaching quality and learning effect.

[0024] The group teaching strategy generation method based on the smart learning cloud platform provided by the embodiment of the present application can be applied to an electronic device. At this time, the electronic device is the execution subject of the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiment of the present application, and the embodiment of the present application does not make any limitation on the specific type of the electronic device.

[0025] For example, the electronic device can be a mobile phone, a tablet computer, a notebook computer, a netbook, a desktop computer, a smart large screen, a smart television, a computer, a laptop computer, etc.

[0026] In order to better understand the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiments of the present application, the specific implementation process of the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiments of the present application is exemplarily introduced below.

[0027] Figure 1 The schematic flowchart of the group teaching strategy generation method based on the smart learning cloud platform provided by the embodiments of the present application is shown, and the group teaching strategy generation method based on the smart learning cloud platform comprises:

[0028] S100, obtaining teaching behavior data, biological data and teaching progress data of multiple classes through a smart learning cloud platform. The teaching progress data comprises taught knowledge points and corresponding teaching dates.

[0029] It can be understood that the teaching behavior data can comprise test records of students, interaction trajectories with the learning platform, social interaction data. The test records can comprise the scores and answering conditions of students in various online tests, and can reflect the mastery of students on knowledge points; the interaction trajectories with the learning platform can comprise that students click course materials, browse learning resources, participate in online discussions, etc.; the social interaction data can comprise speeches and comments of students in the discussion area, or interaction conditions in collaborative tasks, and the social interaction data can reflect the communication conditions of students with classmates or teachers, and is helpful to understand the cooperation spirit, expression ability and participation degree of group cooperation of students.

[0030] Exemplarily, the teaching behavior data can be automatically collected through the smart learning cloud platform, and the smart learning cloud platform can record each learning activity of students, and record the data of each student interacting with the platform through a database.

[0031] It can be understood that biological data can be collected by biosensor devices or other sensing tools, which can provide information about students' physiological responses, emotions and psychological states, helping to understand the stress, attention, emotional fluctuations and other situations that students may encounter during the learning process. Biological data can include heart rate variability (HRV), galvanic skin response (GSR), facial expression recognition, and microphone speech prosody. Heart rate variability (HRV) can reflect students' emotional fluctuations, stress levels and physical health conditions by measuring changes in students' heart rate; galvanic skin response (GSR) can assess students' emotional responses through changes in skin conductance, and GSR can monitor emotional changes such as anxiety, excitement or tension; facial expression recognition can assess students' emotional responses during the learning process through facial expression analysis technology; microphone speech prosody can assess students' emotional state and psychological response by analyzing the speech characteristics (such as pitch, speed, tone change, etc.) of students when speaking.

[0032] Exemplarily, biological data can be collected in real time by specific hardware devices (such as smart watches, health monitoring devices, facial recognition cameras, etc.) and transmitted to the smart learning cloud platform for storage and analysis through Bluetooth, Zig-Bee, WiFi and other wireless transmission methods. For example, heart rate monitoring can be done by wearing a smart bracelet or watch, skin galvanic response can be recorded by wearing a sensor, facial expression recognition can be analyzed in real time by camera and computer vision technology, and speech prosody data can be captured by microphone and analyzed by speech signal analysis.

[0033] It can be understood that teaching progress data is a record of the teaching process of teachers, including the knowledge points taught, teaching content and the corresponding teaching date of each knowledge point, which can help teaching managers and teachers track course progress and arrangement, and ensure that the teaching content of all classes can be covered on time.

[0034] Exemplarily, teaching progress data can be manually input by teachers or automatically recorded by the smart cloud learning platform, and the teaching date of each knowledge point can be marked and the progress data can be updated. It can be displayed in the form of course calendar or course progress table, and can also be stored and accessed through system-generated reports.

[0035] The teaching behavior data, biological data and teaching progress data stored through the smart learning cloud platform not only enable a comprehensive understanding of students' learning status and emotional changes, but also help teachers to conduct personalized teaching, optimize group teaching strategies, and provide accurate data support for teaching decisions.

[0036] S200, according to the teaching progress data of multiple classes, the teaching stages of multiple classes are spatio-temporally aligned to obtain a teaching stage sequence. The teaching stage sequence includes multiple teaching stages, and each teaching stage includes teaching behavior data and biological data of multiple classes in the same teaching stage.

[0037] It can be understood that the teaching stage can be divided according to the knowledge point teaching date in the teaching progress data, and the start and end of each stage can be defined by a series of teaching content teaching dates. For example, a teaching stage can include teaching content from knowledge point A to knowledge point B.

[0038] It can be understood that the teaching stage can be divided according to the knowledge point teaching date in the teaching progress data, and the start and end of each stage can be defined by a series of teaching content teaching dates. For example, a teaching stage can include teaching content from knowledge point A to knowledge point B.

[0039] By way of example, for the teaching progress of different classes, the teaching start date of different classes can be adjusted due to the progress difference between classes. Assuming that the starting point of the first stage of all classes is aligned based on a certain standard date (for example, the first stage of all classes starts from the first week), and then each class synchronizes the time of each teaching stage according to its actual progress arrangement.

[0040] Each class can have different teaching content and order, and spatial alignment means aligning the teaching content of each class according to its teaching progress arrangement into a unified standard sequence. For example, a certain class may teach knowledge point A in the second stage, but another class may teach knowledge point B in the second stage. Therefore, the teaching content of each class can be uniformly divided into stages according to the teaching progress data, so that each class teaches similar content in the same teaching stage.

[0041] When the spatio-temporal alignment process is completed, the teaching behavior data and biological data of each class are grouped and classified according to each teaching stage to form an overall teaching stage sequence, and each teaching stage contains the teaching behavior data and biological data of all classes in the teaching stage. Each teaching stage can form a data matrix or sequence, where each element represents the data of a class in that stage, which will be used for subsequent analysis, learning progress evaluation and personalized learning intervention.

[0042] This step can help teachers and managers understand the learning dynamics of each class in different teaching stages, and also provides strong data support for personalized learning, teaching strategy optimization and learning situation analysis.

[0043] In a possible implementation, S200, according to the teaching progress data of the multiple classes, the teaching stages of the multiple classes are spatiotemporally aligned to obtain a teaching stage sequence, including:

[0044] S210, comparing the knowledge points in the teaching progress data of the multiple classes with the knowledge points in the standard teaching framework, if the knowledge points are inconsistent, replacing the inconsistent knowledge points in the teaching progress data with the corresponding knowledge points in the standard teaching framework according to the equivalent knowledge point table. The equivalent knowledge point table includes the equivalent relationship between the same or similar knowledge points in different teaching materials and the knowledge points in the standard teaching framework, and the equivalent knowledge point table is obtained through semantic similarity calculation and manual rule library supplement. The standard teaching framework includes each teaching stage ID, the knowledge points contained in each teaching stage, and the expected teaching duration of each teaching stage.

[0045] It can be understood that each class may use different teaching materials (which can be different schools using the smart learning cloud platform), and the same knowledge point may be expressed differently in different teaching materials (such as straight line equation in teaching material A, and once function in teaching material B). The corresponding knowledge points in the standard teaching framework of different expressions can be the same or equivalent. The equivalent knowledge point table is to solve the difference between the knowledge points in different teaching materials and the corresponding knowledge points in the standard teaching framework.

[0046] Exemplarily, the teaching material knowledge points can be extracted from various teaching materials (different publishers and different versions). The extraction method can include: using an education field keyword library (such as theorems, formulas, functions, laws, etc.) to identify possible knowledge point phrases; using named entity recognition (NER) technology to identify knowledge point entities in a specific field; using TF-IDF to extract high-frequency appearing, educationally characteristic word groups (such as Ohm's law, quadratic function vertex formula, etc.) in the teaching materials. The standard teaching framework can be formulated by an education management department or a platform party.

[0047] For each type of teaching material, a pre-trained model (such as BERT, RoBERTa, etc.) can be used to convert the teaching material knowledge points and the knowledge points in the standard teaching framework into teaching material vectors and standard vectors, respectively. The semantic similarity between the teaching material vectors and the standard vectors is calculated using cosine similarity or Euclidean distance, and the teaching material knowledge points with a similarity higher than a similarity threshold (which can be adjusted according to actual conditions, such as 0.8) are determined as the knowledge points equivalent to the knowledge points in the standard teaching framework.

[0048] The semantic similarity calculation is effective, but there may be different teaching material idioms in the education scene (such as straight line equation and linear function image), and implicit knowledge points (the textbook does not explicitly write, but there is a separate item in the standard framework). Therefore, an artificial rule base can be constructed for supplementation, which can be constructed by education experts, textbook writers or front-line teachers. For example, artificial rules can include synonym rules, artificially defined synonym pairs; superordinate and subordinate rules, such as parallelogram area formula belonging to polygon area formula. When the semantic similarity is low but the rule base exists, the rule base mapping is given priority; when the semantic similarity is high and the rule base has no conflict, the similarity calculation result is given priority.

[0049] Through the above semantic similarity calculation and artificial rule base mapping, the equivalent textbook knowledge points in the standard teaching framework can be retrieved in various textbooks, and an equivalent knowledge point table is constructed based on this. For equivalent knowledge point representation, please refer to Figure 4 .

[0050] When aligning the class progress, the textbook knowledge points used in the actual teaching of the class (each knowledge point in the teaching progress data of each class) can be detected to see if they are completely consistent with the knowledge points in the standard teaching framework. If it is completely consistent with the knowledge points in the standard teaching framework, there is no need to replace; if the knowledge points in the textbook used by the class are not completely consistent with the knowledge points in the standard teaching framework, they can be replaced with the corresponding knowledge points in the standard teaching framework according to the equivalent knowledge point table. For example, a certain knowledge point in the textbook used by class A is a straight line equation, and the equivalent knowledge point table is found to correspond to a linear function in the standard teaching framework, so when aligning, the straight line equation learned by class A is replaced by the linear function in the standard teaching framework. Similarly, if the Ohm's law calculation in the textbook used by class B is not completely consistent with the circuit law in the standard teaching framework, then according to the equivalent knowledge point table, the Ohm's law calculation of class B will also be replaced by the knowledge point in the standard teaching framework, which is conducive to the consistency of alignment.

[0051] By using the equivalent knowledge point table, it can be ensured that no matter what textbook is used by the class, the teaching progress can be accurately mapped to the corresponding knowledge point in the standard teaching framework, which is conducive to the alignment of the progress of different classes, and can eliminate the differences between textbooks and unify the representation of knowledge points.

[0052] S220, calculate the distance matrix between the knowledge points in the teaching progress data of each class and the knowledge points in the standard teaching framework; based on each distance matrix, calculate the optimal alignment path between the teaching progress of each class and the standard teaching framework using the dynamic time warping algorithm; according to the optimal alignment path between the teaching progress of each class and the standard teaching framework, map the teaching progress of each class to each teaching stage in the standard teaching framework to obtain a teaching stage sequence.

[0053] It can be understood that the dynamic time warping algorithm (DTW) is an algorithm for comparing the similarity of two sequences, which can process sequences of different lengths and speeds, and find the best alignment path. In the alignment of teaching progress, DTW can bend the teaching progress sequences of different classes to make up for the difference in time, even if the progress of the classes is different, DTW can find the alignment path that minimizes the cost, and match the knowledge points of each class to the corresponding knowledge points in the standard teaching framework. DTW allows the stretching of the sequence, even if a class may learn some knowledge points in advance or delay, DTW can align these differences by inserting or deleting steps, so that the progress sequences of different classes coincide at some knowledge points.

[0054] Illustratively, the distance matrix between the knowledge points in the teaching progress of each class and the knowledge points in the standard teaching framework can be calculated, which can be based on timestamp differences, similarity of knowledge points (such as semantic distance), and other factors. The distance matrix reflects the matching degree between the teaching progress of the class and the standard teaching framework.

[0055] The DTW algorithm can be used to find the optimal alignment path of the teaching progress of each class and the standard teaching framework, which is obtained by calculating the minimum cost path of each point in the distance matrix. DTW allows time stretching, that is, the progress of the class can be advanced or delayed to achieve the minimum cost, that is, the most reasonable alignment. For example, class A may have explained the definition of function at a certain time, while class B has not reached this part, and DTW can flexibly adjust the progress of class B to lengthen or delay the teaching time of function definition.

[0056] After finding the optimal alignment path, the teaching progress of each class will be mapped to each teaching stage of the standard teaching framework, and the DTW algorithm can generate a teaching stage sequence to show which stage of the standard teaching framework corresponds to the actual teaching time of each class. According to the timestamps in the teaching behavior data and biological data of each class, the teaching behavior data and biological data of each class can be grouped and classified according to each teaching stage.

[0057] Through the DTW algorithm, flexible alignment between multiple classes can be achieved, progress differences can be handled, and the teaching progress of multiple classes can be effectively aligned to the standard teaching framework, which is beneficial to the reasonable handling of progress differences between different classes, and enables the data between classes to be comparable.

[0058] In one possible implementation, the group teaching strategy generation method based on the smart learning cloud platform further includes:

[0059] S201, extracting knowledge points from the teaching content based on text analysis techniques. The teaching content is pre-stored in the intelligent learning cloud platform.

[0060] For example, the teaching content exists in the form of structured or semi-structured text. The text data can be pre-processed for subsequent analysis. The long teaching content can be split into paragraphs and sentences, so that the subsequent analysis can be more clearly targeted at specific content; Chinese word segmentation tools such as HanLP, LTP, Jieba, etc. can be used to divide the text into smaller units such as words or phrases; stop words in the text can be removed to reduce noise and improve analysis efficiency, stop words are words that frequently appear in text analysis but have no actual meaning (such as "of", "is", etc.); the part of speech of each word in the text can be classified (such as noun, verb, adjective, etc.), and when identifying knowledge points, noun phrases (such as "physical law", "triangle") are preferred as potential knowledge points.

[0061] Key entities in the text can be identified through named entity recognition (NER) technology. For example, core concepts (such as Newton's law, particle motion) and methods (such as derivative rule) in physics, mathematics, chemistry, etc. Topic modeling techniques such as Latent Dirichlet Allocation (LDA) can be used to automatically discover potential topics or core concepts from teaching content, thereby helping to identify knowledge points. Important keywords in the teaching content can be extracted through TF-IDF, TextRank, etc. These keywords may be core knowledge points.

[0062] S202, using a relationship extraction algorithm to analyze the context in the teaching content and identify the dependency relationship between knowledge points to construct a knowledge point dependency graph.

[0063] For example, dependency syntax analysis is an important tool for relationship extraction, which can identify the dependency relationship between words. Each sentence forms a syntax tree, where each word is a node and the edge represents the dependency relationship between words. For example, in the relationship between acceleration and force, there is a dependency relationship between acceleration and force, indicating that acceleration depends on force. Dependency analysis can identify the grammatical association between knowledge points and build a higher level of understanding.

[0064] In addition to syntactic relationships, semantic analysis can be used to identify more complex relationships between knowledge points. Semantic relationship extraction involves understanding the deep meaning of the text, not just the grammatical relationship. For example, causal relationships (such as heat increases causing the temperature of an object to rise), premise relationships (such as mastering basic mathematics can understand advanced mathematics), and progressive relationships (such as deriving Newton's second law from Newton's first law).

[0065] Once the dependency relationships between knowledge points are identified, a knowledge point dependency graph can be constructed using the dependency relationships. The knowledge point dependency graph is a directed graph, where the nodes represent knowledge points and the edges represent the dependency relationships between knowledge points, i.e., prerequisite knowledge point -> current knowledge point.

[0066] By analyzing the context in the teaching content using the relation extraction algorithm, the dependency relationships between the knowledge points in the teaching content can be identified. These dependency relationships include not only grammatical connections but also semantic relationships such as causality, premise, progression, and the like. The knowledge point dependency graph can organize the knowledge points into an ordered structure, helping to implement tasks such as teaching design, learning recommendation, and course optimization.

[0067] S203, according to the knowledge point dependency graph, divide the knowledge points into various teaching stages, and label an ID for each teaching stage.

[0068] Illustratively, the knowledge point dependency graph can be topologically sorted. The role of topological sorting is to determine the teaching order of the knowledge points, which facilitates the teaching of all knowledge points that depend on a certain knowledge point before the certain knowledge point is taught. The in-degree of each node (i.e., the number of edges pointing to the node) is calculated, and all nodes with an in-degree of 0 are added to a queue (these nodes have no prerequisite dependencies and can be taught first). Nodes are taken out of the queue one by one, and the in-degree of the next node pointed to by the taken-out node (i.e., the taken-out node is a prerequisite dependency of the next node) is reduced by 1. If the in-degree of the next node becomes 0, the node is added to the queue. The above steps are continued until all nodes are processed. Through topological sorting, an ordered list that satisfies the knowledge point dependency relationships can be obtained.

[0069] When the topological sorting is completed, the simpler knowledge points and the more complex knowledge points can be allocated to different stages according to the difficulty of the knowledge points or the teaching objectives. For example, basic concepts, laws, formulas, and the like can be allocated to early stages, while complex derivation processes, application methods, and the like can be allocated to later stages. According to the order of topological sorting, the knowledge points can be divided into multiple teaching stages according to the dependency relationships. Knowledge points with more dependencies can be arranged in later stages, while basic knowledge points can be arranged in early stages.

[0070] When the teaching stage division is completed, a unique ID can be assigned to each teaching stage. The ID assignment method can use simple numerical numbering. For example, teaching stage S1 contains knowledge points K1, K2, teaching stage S2 contains knowledge points K3, K4, K5, and teaching stage S3 contains knowledge points K6, K7.

[0071] The division of teaching stages facilitates students to gradually master knowledge, and the dependency relationships between knowledge points ensure that the teaching content of each stage is built on the basis of the previous stage.

[0072] S204, based on the number of knowledge points in each teaching stage, allocating the expected teaching duration of each teaching stage.

[0073] Illustratively, the average teaching time required for each knowledge point can be calculated according to the historical teaching data, the teaching time required for all knowledge points in each teaching stage is calculated according to the average teaching time required for each knowledge point, and the teaching time required for all knowledge points in each teaching stage is determined as the corresponding expected teaching duration.

[0074] S205, arranging each teaching stage ID, the knowledge points contained in each teaching stage and the expected teaching duration of each teaching stage according to the dependency relationship between the knowledge points to obtain a standard teaching framework.

[0075] Illustratively, according to the topological sorting result of the knowledge point dependency graph, an ordered teaching stage list (standard teaching framework) is output, including each teaching stage ID, the knowledge points contained in each teaching stage and the expected teaching duration of each teaching stage.

[0076] Through the above steps, support can be provided for subsequent class teaching progress alignment.

[0077] In one possible implementation, the group teaching strategy generation method based on the wisdom learning cloud platform further comprises:

[0078] S2001, in the space-time alignment process, using dynamic time warping algorithm to detect whether there is a progress jump feature. Wherein, the progress jump feature is that the knowledge point corresponding to the teaching progress of a class in the standard teaching framework is not covered by the teaching progress data of the class.

[0079] Illustratively, when using the DTW algorithm to align the teaching progress of the class, if a class skips a knowledge point (such as directly entering the equation solution without learning the property of equation), the DTW algorithm can find that the knowledge point contained in the standard teaching framework is not covered by the actual progress of the class.

[0080] For example, in the alignment path, the knowledge point (such as the property of equation) corresponding to the class progress in the standard teaching framework does not match any teaching time point of the class. The actual progress of the class directly skips the prerequisite knowledge point (such as the property of equation) and starts to learn the content of the subsequent stage (such as the equation solution).

[0081] In order to handle the progress jump, a virtual teaching stage or a missing knowledge point can be inserted.

[0082] S2002, if the progress jump feature exists, a virtual knowledge point is inserted at the position of the knowledge point not covered in the teaching stage sequence to fill in the not covered knowledge point. Or, if the progress jump feature exists, the not covered knowledge point is marked in the teaching stage sequence, and the missing not covered knowledge point is noted.

[0083] Exemplarily, when the DTW algorithm detects that some classes have progress jump situations, the skipped knowledge points can be marked, a virtual knowledge point is inserted at the position of the skipped knowledge points in the teaching stage sequence, and the knowledge point is marked as to be supplemented. The virtual knowledge point is used to fill in the missing content and does not represent the actual teaching content, which can remind the subsequent teaching arrangement to supplement this part of knowledge.

[0084] For example, class A directly enters the equation solution and skips the equality property, so class A in the teaching stage sequence can be as shown in Table 1:

[0085]

[0086] Table 1

[0087] Another method of processing progress jump is to mark the missing knowledge points, mark the skipped knowledge points in the teaching stage sequence, and note the missing. Teachers can supplement the missing knowledge points in the subsequent teaching plan according to the mark, which is beneficial to the students to master the complete course content.

[0088] For example, class A skips the equality property and directly starts the equation solution, so class A in the teaching stage sequence can be as shown in Table 2:

[0089]

[0090] Table 2

[0091] Through the missing knowledge points in the space-time alignment result, a reminder information can be generated to prompt the teacher to supplement the missing knowledge points in the subsequent teaching process, and the teacher can check the reminder information in the intelligent learning cloud platform. Not only helps the teacher to find the missing, but also is beneficial to the integrity of the teaching stage sequence.

[0092] The teaching stage alignment converts the scattered and heterogeneous teaching data into space-time comparable standardized input through three core technologies of standard teaching framework construction, flexible progress mapping and knowledge point semantic matching, which provides a reliable basis for subsequent group anomaly detection and strategy generation. This process not only retains the original teaching characteristics, but also eliminates the analysis deviation caused by progress noise.

[0093] S300, performing group anomaly detection on the plurality of classes based on each teaching stage of the teaching stage sequence to obtain an anomaly detection result. The anomaly detection result includes a teaching stage in which a group anomaly exists.

[0094] It can be understood that the group anomaly detection refers to identifying an abnormal stage based on the teaching behavior data and the biological data of all classes in each teaching stage. By comparing the teaching behavior data and the biological data of all classes, it can be found that which teaching stage has an anomaly at the group level, thereby providing a basis for adjusting the teaching strategy.

[0095] Exemplarily, there can be missing data in the teaching stage sequence. For example, the biological data or the teaching behavior data of some classes can not be recorded in some stages. The missing values can be processed by interpolation, filling or deletion; if some data is obviously abnormal (such as extreme values of biological data, or high-frequency clicks in teaching behavior data that do not conform to logic), it can be removed or corrected.

[0096] The teaching behavior data and the biological data of different classes can be standardized or normalized by using Z-score standardization or Min-Max normalization method, so that data of different dimensions can be compared on the same scale.

[0097] The Z-score of the teaching behavior data and the biological data of each class in each teaching stage can be calculated, and whether it is abnormal can be judged based on the Z-score value. If the z-score of a class is greater than a certain threshold (which can be ±2 or ±3), the class can have an anomaly in that stage.

[0098] The classes can be divided into different groups by using a clustering algorithm (such as K-means, DBSCAN, etc.), and a group that is significantly different from other groups is identified. Through clustering analysis, abnormal behavior or abnormal patterns within the group can be detected.

[0099] The teaching behavior data and the biological data of all classes can be clustered by using a clustering algorithm (such as K-means, DBSCAN, etc.), which can capture common patterns within the group and identify outliers that are significantly different from most classes. For example, if the behavior patterns of some classes are significantly different from other classes in a certain teaching stage, these classes can belong to an abnormal group.

[0100] The teaching stage that deviates from the normal trend in the group can be identified by the Z-Score method or the clustering algorithm, the abnormal teaching stage can be marked, and whether there are some common problems can be analyzed through the biological data and teaching behavior data of the abnormal teaching stage. For example, it may be that a certain teaching method is not effective, a certain knowledge point is too complex to cause students to be anxious, or the teaching rhythm is not suitable to cause students to lose attention.

[0101] The purpose of group anomaly detection is to find abnormal performance in the teaching stage by analyzing the teaching behavior data and biological data of multiple classes, and to provide basis for adjusting the teaching strategy. By reasonably selecting the detection method, combining data preprocessing and anomaly evaluation, potential problems in the teaching process can be effectively identified.

[0102] In one possible implementation, please refer to FIG. 2, S300, based on each teaching stage of the teaching stage sequence, group anomaly detection is performed on multiple classes to obtain an anomaly detection result, including:

[0103] S310, the teaching behavior data and biological data of multiple classes in each teaching stage are extracted from the teaching stage sequence.

[0104] Exemplarily, the group anomaly detection is for each teaching stage, which is used to detect whether there is a group anomaly situation (such as a high error rate of multiple classes on the same problem) in each teaching stage. Each teaching stage in the teaching stage sequence already contains the teaching behavior data and biological data of all classes, the teaching behavior data and biological data in each teaching stage can be analyzed to detect whether there is a group anomaly in the teaching stage.

[0105] The following processing is performed for each teaching stage:

[0106] S320, common error patterns are extracted from the teaching behavior data of multiple classes; a cognitive network graph is constructed based on the common error patterns, and a group knowledge fracture point is identified based on the cognitive network graph; the common error patterns and the group knowledge fracture point are integrated to obtain group cognitive characteristics. The cognitive network graph is used to represent the mastery of students in the learning process and the correlation between different knowledge points.

[0107] Exemplarily, the common error mode refers to the error type that repeatedly occurs in most students in multiple classes and the knowledge point associated with the error type. The answer records of the students can be extracted from the test records of the teaching behavior data, the error questions can be screened from the answer records, and structured features can be constructed for each error question, which can include the knowledge point associated with the error question, the error type, and the error option. The structured features can provide a basis for subsequent common error mode mining. The association rule mining algorithm (such as Apriori, FP-Growth) can be used to mine high-frequency combinations of error types and knowledge points from the structured features of all error questions. If a certain error type and a certain knowledge point appear simultaneously in most students, it is marked as a common error mode. For example, the unit conversion error and the circuit law frequently appear, forming a common error mode.

[0108] The structured features of all error questions can be clustered by clustering analysis (such as K-Means, hierarchical clustering), and the error modes (error types) that students often encounter in the learning process can be identified. Each cluster is a common error mode (error type and associated knowledge point). By clustering the error data in the teaching behavior data by knowledge point, it can be identified which knowledge points have universal errors in multiple classes. For example, the concept confusion between mass and weight can be identified by analyzing the homework, tests, and question-and-answer records of the students, and further deduced which knowledge points may be the root cause of these errors.

[0109] All knowledge points can be extracted from the common error patterns and taken as nodes in the cognitive network graph, with one node representing one knowledge point. The relationship between knowledge points can be identified by analyzing the common error patterns. It can be determined whether two or more knowledge points are included in the common error patterns. If there is only one knowledge point, no edge is generated. If two or more knowledge points are included, an edge can be established between the knowledge points (nodes), i.e., two knowledge points (one knowledge point pair) exist in the common error patterns, and an edge is established between the two knowledge points (nodes). If there are more than two knowledge points (e.g., three knowledge points, three knowledge point pairs) in the common error patterns, an edge is established between every two knowledge points (nodes). The weight of the edge can be calculated according to the relationship between the knowledge points and the common error patterns. The greater the weight, the stronger the connection between the knowledge points, and the student's mastery of one knowledge point may affect the learning of other knowledge points. The number of occurrences of each knowledge point pair in all common error patterns can be counted, and the number of occurrences is taken as the weight of the edge between the two knowledge points (knowledge point pairs). According to the above method, an undirected cognitive network graph can be constructed, in which each node represents a knowledge point, and the weight of each edge represents the student's mastery of the relevant knowledge points (i.e., the more the number of occurrences of the knowledge point pair, the lower the student's mastery of the two knowledge points). The cognitive network graph can be automatically updated each time a new round of test or learning data is fed back, and the student's mastery of knowledge points at different stages can affect the weight and connection relationship of the knowledge points in the cognitive network graph.

[0110] The number of occurrences of each knowledge point in all common error patterns can be calculated to obtain the error concentration of each knowledge point. Based on the cognitive network graph, the degree centrality of each node (knowledge point) can be calculated, i.e., the number of edges contained by each node; the weighted degree of each node can be calculated, i.e., the total weight of all connected edges of each node; and the average degree centrality and the average weighted degree can be calculated based on the degree centrality and the weighted degree of all nodes. A concentration threshold can be preset, and the knowledge points with an error concentration higher than the concentration threshold, and a degree centrality and a weighted degree higher than the average degree centrality and the average weighted degree, respectively, are determined as group knowledge breakpoints.

[0111] In addition, in the cognitive network graph, knowledge points with low mastery degree can be marked as potential breaking points, which are knowledge points that students generally do not master well in the learning process and may be the core area of cognitive breaking. The influence of low mastery degree knowledge points on other related knowledge points can be calculated through a propagation algorithm. For example, if students generally make mistakes on a certain knowledge point (such as basic formulas of mechanics), and this mistake affects the understanding of subsequent knowledge points (such as the law of conservation of energy), then the knowledge point and the subsequent knowledge point can be marked as a group knowledge breaking point. The most influential breaking points in the group, i.e. those that have a wide impact on the learning path, can be calculated. By analyzing the connectivity of the cognitive network graph and the correlation between knowledge points, it can be determined which knowledge points are not mastered and lead to subsequent cognitive fault lines.

[0112] The identified common error patterns can be combined with group knowledge breaking points to form group cognitive characteristics. Through group cognitive characteristics, it can be known that students have common errors on certain knowledge points, common understanding obstacles, and which knowledge points are learning bottlenecks for the entire group, providing a basis for optimizing group teaching strategies.

[0113] S330, based on the teaching behavior data of multiple classes, performing frequency analysis on the learning behavior of the students using Fourier transform to obtain a behavior frequency spectrum; calculating a resonance matching degree between the behavior frequency spectrum and an ideal learning path frequency spectrum; identifying a non-explicit behavior pattern based on the resonance matching degree to obtain a group behavior characteristic.

[0114] It can be understood that step S330 is the implementation process of the cross-modal behavior resonance analysis method. By combining physical field theory, the teaching behavior data of the students is regarded as a vibration signal in an energy field, and then the matching degree with the ideal learning path is analyzed. In traditional behavior analysis, the operation of the students is regarded as a simple activity or task, while the cross-modal behavior resonance analysis regards the behavior of the students as a dynamic system by borrowing the concepts of energy field and vibration frequency in physical field theory, and uses Fourier transform and other tools to convert the behavior pattern into a frequency spectrum, so that the behavior resonance phenomenon of the students in the learning process can be identified, and on this basis, the group teaching strategy can be optimized.

[0115] The behavior frequency spectrum is similar to the frequency spectrum of a physical vibration system, and the operation rhythm of the students can be analyzed through Fourier transform to extract the main frequency component in the operation.

[0116] The resonance matching degree is a comparison between the frequency spectrum of the current behavior pattern of the students and the ideal learning path, and a calculation of the harmonic coincidence degree, i.e. whether the behavior pattern is resonantly matched with the ideal path. The resonance matching degree reflects whether the learning state of the students is consistent with the ideal learning state.

[0117] Exemplarily, various behaviors of students on the learning platform (such as clicking, browsing, speaking, etc.) can be regarded as events occurring at moments, and by time slicing, intensity quantification and sequence construction of these events, a continuous and unified structure of behavior time sequence on the time axis is formed. A fixed granularity time unit can be set, such as every 1 minute, every 5 minutes, every 10 minutes as a time slice (depending on the platform granularity). For example, the learning time is 2 hours a day, and every 5 minutes is selected as the granularity, a total of 24 time slices. In each time slice, based on the interaction track and social interaction data of each student's teaching behavior data with the learning platform, the occurrence times of various behaviors are counted and behavior quantification processing is performed to obtain the behavior intensity of each time slice. For example, in a certain time slice, a student clicks the course material 3 times and participates in the discussion 1 time, and the behavior intensity is 4. Each time slice can be arranged in time sequence to form a behavior time sequence . Through the above method, a behavior time sequence for each student can be constructed.

[0118] The time domain data (behavior time sequence) can be converted into frequency domain data by Fourier transform, that is, by analyzing the frequency components of the signal to capture periodic behavior patterns. The one-dimensional discrete Fourier transform (DFT) can be performed on the behavior time sequence of each student to obtain its frequency domain representation , that is , where represents the total length of the behavior time sequence, represents the frequency, represents the amplitude complex number at the frequency , representing the intensity and phase information of the frequency component. By calculating the modulus of the frequency domain , the amplitude distribution of the frequency spectrum is obtained, that is . Through the above Fourier transform calculation, the amplitude distribution of the frequency spectrum of each student can be obtained, that is, the behavior frequency spectrum of each student. The behavior frequency spectrum obtained after Fourier transform represents different frequency components of the student's learning behavior, and the behavior frequency spectrum reveals the periodicity of the student's learning behavior. High-frequency components represent short-term rapid change behaviors in the learning process (such as temporary anxiety, sudden attention fluctuations, etc.), and high frequency may be related to short-term behavior response, task completion, etc.; low-frequency components represent long-term stable behaviors in the learning process (such as regular learning, long-term attention persistence, etc.), and low-frequency components reflect the long-term trend or stable behavior pattern of the student's learning.

[0119] An ideal learning path refers to an expected pattern of learning behaviors defined based on instructional design, course objectives, or excellent learning performance, reflecting a reasonable distribution of instructional content, pacing, and cognitive load over time. The ideal learning path can be a teacher-predefined learning pace, such as what should be done at which stage, how many times each behavior should occur, how active the student should be each day / hour, etc. The ideal learning path can be converted into an ideal time series using the above-mentioned method of constructing a behavior time series. The ideal time series can be Fourier transformed to obtain the ideal learning path frequency spectrum using the above-mentioned method of calculating a behavior frequency spectrum.

[0120] The ideal learning path can be defined by the student's optimal learning pace, knowledge point mastery sequence, and depth and coherence of learning. For example, the ideal learning path is for the student to regularly switch tasks and gradually advance, maintain focus and deep thinking. The teaching behavior data of a portion of students with excellent performance and stable behavior can be selected, and the behavior time series of these students can be constructed based on the teaching behavior data of these students using the above-mentioned method of constructing a behavior time series. The ideal time series is obtained by averaging the behavior time series of these students. The frequency spectrum of the ideal learning path is relatively smooth and regular based on the above-mentioned method of calculating a behavior frequency spectrum.

[0121] The resonance matching degree between the student behavior frequency spectrum and the ideal learning path frequency spectrum can be calculated by cross-correlation analysis or similarity measurement, and the similarity between the behavior frequency spectrum and the ideal learning path frequency spectrum can be measured. Cross-correlation analysis can quantify the correlation between the student behavior frequency spectrum and the ideal learning path frequency spectrum by calculating the cross-correlation function of the two; similarity measurement can use methods such as cosine similarity, Euclidean distance, etc. to evaluate the similarity between the two. If the student's learning behavior frequency matches the ideal learning path frequency, it indicates that the student is in an ideal learning state; if the matching degree is low, it indicates that the student may be in an inefficient learning state and needs intervention. For example, the cosine similarity method can be used to calculate the resonance matching degree between the behavior frequency spectrum of each student and the ideal learning path frequency spectrum, and the resonance matching degree of each student is obtained, i.e. wherein, represents the resonance matching degree, represents the amplitude distribution of the ideal learning path frequency spectrum, represents the sum of the amplitude products of the behavior frequency spectrum and the ideal learning path frequency spectrum at each frequency, represents the norm (length) of the behavior frequency spectrum, represents the norm (length) of the ideal learning path frequency spectrum.

[0122] In addition to the cross-correlation analysis or similarity measure method, the resonance matching degree can also be calculated by harmonic coincidence, that is, the frequency spectrum of the ideal learning path is extracted The main frequency , a set of harmonic frequencies is constructed , and it is determined whether there is a frequency in the behavior frequency spectrum that is the same as the frequency in If there is, it is determined whether the amplitude of the frequency is greater than the amplitude threshold (which can be pre-set, such as 20% of the maximum amplitude in the behavior frequency spectrum), and if it is, it is determined that the frequency is in harmonic coincidence with the ideal learning path frequency spectrum. The frequencies that meet the above two conditions are counted, and the ratio between the number of frequencies that meet the above two conditions and the total number of frequencies in is calculated, which is the resonance matching degree

[0123] Non-explicit behavior patterns refer to behavior tendencies that are not easily discovered through traditional time domain observation, operation frequency statistics or performance analysis, but show specific structural characteristics, rhythm rules or potential cognitive biases in the frequency domain space. The behavior frequency of each student can be converted into an ordered amplitude vector The amplitude vector of each student is spliced with the resonance matching degree to form a comprehensive feature vector. Unsupervised clustering algorithms (such as K-means, hierarchical clustering, etc.) can be used, and the comprehensive feature vectors of all students are used as the input of the clustering algorithm. After clustering, multiple non-explicit behavior patterns are output, such as high resonance type, high resonance matching degree, and spectrum concentration; fragmented high frequency type, high frequency prominence, and low resonance matching degree; low frequency focus type, low frequency component, and medium-high resonance matching degree; and abnormal low activity type, low spectrum amplitude, and very low resonance matching degree. The proportion of each type of non-explicit behavior pattern in the entire student group can be calculated to obtain the group behavior pattern distribution. The average value of the behavior frequency spectrum amplitudes of all students can be calculated, and the group spectrum center (main frequency) can be calculated according to the average value, that is, , wherein represents the group spectrum center, represents the th frequency, represents the average amplitude of the frequency in the group. The average resonance matching degree can be calculated by calculating the average resonance matching degree of all students. The group behavior pattern distribution, group spectrum center, and average resonance matching degree can be used as the group behavior characteristics.

[0124] Based on the teaching behavior data of multiple classes, the behavior frequency spectrum of all students can be analyzed, and the resonance matching degree of each student can be calculated. Through clustering analysis or statistical methods, the behavior frequency spectrum of students is divided into several groups, and the overall learning mode of the group is identified. For students within the group, common behavior characteristics are identified, such as most students showing stable learning patterns (mainly low frequency spectrum), and some classes of students frequently appearing high-frequency behavior fluctuations, which may have cognitive or emotional problems.

[0125] Through this step, the student's attention dispersion, deep learning state, etc. can be identified, and the teaching strategy can be dynamically adjusted according to the student's behavior mode. Through the cross-modal behavior resonance analysis method, the learning state of the student can be more accurately understood, and targeted teaching support can be provided to improve the teaching effect.

[0126] S340, calculate the stress index according to the biological data of multiple classes to obtain the group emotional characteristics.

[0127] Exemplarily, the stress index of each student can be calculated by weighted summation of the four biological signals (heart rate variability (HRV), galvanic skin response (GSR), facial expression recognition, and microphone speech rhythm) of each student.

[0128] The mean or weighted average of the stress index of all students in each class can be calculated to obtain the overall stress index of each class, that is, wherein, represents the class stress index of the jth class, represents the stress index of the ith student in the class, and N represents the number of students in the class.

[0129] By calculating the stress index of all classes, the overall emotional fluctuations of the group (group emotional characteristics) can be obtained. The mean of the stress index of all classes can be calculated to reflect the group stress level, that is, wherein, P represents the group stress level (mean), and M represents the number of classes. The standard deviation of the stress index of all classes can be calculated to reflect the group emotional fluctuations, that is, wherein, E represents the group emotional fluctuations (standard deviation). The mean and standard deviation of the stress index of all classes are determined as the group emotional characteristics.

[0130] S350, feature fusion is performed on the group cognitive characteristics, group behavior characteristics and group emotional characteristics to obtain the fused features.

[0131] Exemplarily, the group cognitive characteristics, group behavior characteristics and group emotional characteristics can be assigned corresponding weights, and the group cognitive characteristics, group behavior characteristics and group emotional characteristics are weighted averaged according to the weights corresponding to the three characteristics to obtain the fused features.

[0132] If each feature dimension is high and redundant, principal component analysis (PCA) can be used for dimensionality reduction, combining multiple features into one or several principal components, which can reduce the dimensionality of the data while retaining most of the information.

[0133] A machine learning model can be trained to automatically learn the optimal feature fusion method. The model can automatically determine the fusion method and importance of different features based on the relationship in the data.

[0134] The fused features can be used as the overall learning state of the group, which can be used for subsequent anomaly detection and teaching strategy generation.

[0135] S360, compare the fused features with the dynamic anomaly threshold value, if the fused features are greater than the dynamic anomaly threshold value, determine that there is a group anomaly in this teaching stage, and obtain the anomaly detection result.

[0136] For example, compare the fused features with the dynamic anomaly threshold value, if the fused feature value is greater than the dynamic anomaly threshold value, determine that there is a group anomaly in this teaching stage. Or you can use statistical methods or machine learning models (such as anomaly detection algorithms) to identify group behavior patterns that exceed the normal range.

[0137] If the fused feature value exceeds the threshold value, the teaching stage can be marked as an abnormal stage, and a corresponding alarm or report can be generated. The abnormal teaching stage may include emotional fluctuations, cognitive error rate surges, or behavior deviations from the normal learning path of the student group.

[0138] The normal range of the group can be calculated by analyzing historical data, and the dynamic anomaly threshold value can be calculated by statistical methods (such as standard deviation, quantile, etc.). With the input of new teaching behavior data and biological data, the dynamic anomaly threshold value can be updated according to real-time data, so that the threshold value has self-adaptability.

[0139] Through this step, the abnormal situation of the group in the learning process can be identified and fed back in real time, thereby supporting personalized teaching and timely intervention.

[0140] In one possible implementation, the group teaching strategy generation method based on the intelligent learning cloud platform further includes:

[0141] MAML is a model-agnostic meta-learning algorithm that aims to train the learned model parameters through a small number of tasks so that they can quickly adapt to new tasks. MAML itself does not directly perform tasks. The key to MAML is to train the model on multiple tasks so that it can learn an initial parameter that can be quickly adapted to different tasks through a small number of gradient updates.

[0142] S301, obtain teaching behavior data of all classes in multiple historical teaching stages, and determine the teaching behavior data of all classes in each historical teaching stage as a meta task.

[0143] Exemplarily, the teaching behavior data of all classes of a certain grade in previous years can be obtained from the smart learning cloud platform. In the MAML framework, each task represents a learning task or learning situation, and the teaching behavior data of all classes in a teaching stage are regarded as the same meta task, and the meta task contains the teaching behavior data of all classes in the teaching stage. Each teaching stage corresponds to a meta task, but the data of different teaching stages will be trained and evaluated as different meta tasks. Training of all teaching stage data can jointly optimize the initial parameters of the model, so that the model can quickly adapt to all classes in each teaching stage and extract common behavior characteristics of all classes.

[0144] S302, for each meta task, divide the teaching behavior data of all classes into a training set and a test set; train the meta learning model based on the training set, and calculate the loss value of the meta learning model on the training set; update the model parameters of the meta learning model through the gradient descent algorithm according to the loss value; calculate the test loss based on the test set, and calculate the gradient of the test loss with respect to the model parameters based on the test loss.

[0145] Exemplarily, a neural network model can be defined as a meta learning model, which can learn the common behavior characteristics of all classes in a certain teaching stage. The initial parameters of the model will be optimized so that the model can quickly adapt to new class data through a small amount of updates. Each time the training is performed, MAML randomly selects a meta task (i.e., a teaching stage) from the data of all teaching stages, and uses all class data in the meta task for training and testing.

[0146] For each meta task (each teaching stage), part of the data in the teaching stage can be used as a training set, and the remaining data can be used as a test set. The meta learning model is trained using the training set in the meta task, forward propagation is performed on the training set, and the loss value of the model on the training set is calculated. According to the loss value, the model parameters are updated using the gradient descent algorithm, which is a local update, and the goal is to make the model adapt to the training data of the current meta task. After the update is completed, the updated model is evaluated using the test set in the task, the test loss of the meta task on the test set is calculated, and the gradient of the test loss with respect to the initial model parameters is calculated. The test loss reflects the adaptability of the meta learning model after training, and can be used to check whether the model can well adapt to new data.

[0147] S303, the average value of the gradient of all meta-tasks is calculated to obtain the meta-gradient. The model parameters of the meta-learning model are updated according to the meta-gradient to obtain the meta-learning model after training.

[0148] Exemplarily, according to the gradient of the test loss of all meta-tasks with respect to the initial model parameters , the average value of the gradients is calculated to obtain the meta-gradient, which reflects the generalization ability of the meta-learning model on multiple meta-tasks and helps the model learn an initial parameter that can adapt to all teaching stages. The initial parameters of the meta-learning model are updated by the meta-gradient , so that the initial parameters can quickly adapt to new tasks (such as new teaching stages) when facing new tasks.

[0149] S304, the teaching behavior data of multiple classes in each teaching stage is input into the meta-learning model after training, so that the meta-learning model after training extracts the behavior gradient and the balance coefficient of each teaching stage, and determines the behavior gradient and the balance coefficient of each teaching stage as the group behavior characteristics of each teaching stage.

[0150] Exemplarily, during the training process, the adjustment direction and amplitude of the initial model parameters for each meta-task can be obtained through the intra-meta-task gradient update. Through the meta-update, the model can capture the behavior gradient of each teaching stage, that is, the strategy adjustment direction of different classes in the learning process. By analyzing the task switching frequency, time allocation and the like in the teaching behavior data, the exploration-exploitation balance coefficient can be extracted, which represents the proportion of knowledge deepening and new field trying of students in the learning process. For example, if students repeatedly practice on some knowledge points, it indicates that the students tend to develop; if students frequently switch tasks, it indicates that the students tend to explore.

[0151] After the training of the meta-learning model is completed, the teaching behavior data of each teaching stage in the teaching stage sequence can be input into the meta-learning model, and the meta-learning model can extract the behavior gradient and the exploration-exploitation balance coefficient of each teaching stage, and determine the behavior gradient and the balance coefficient of each teaching stage as the group behavior characteristics of each teaching stage.

[0152] Using the MAML framework for training can extract shared group behavior characteristics from the teaching behavior data of multiple classes, and can identify the common laws of the behavior patterns of the group in the learning process. MAML achieves the goal of cross-task learning by optimizing shared parameters, so that the final group behavior characteristics can reflect the learning characteristics of all classes.

[0153] In one possible implementation, the group teaching strategy generation method based on the smart learning cloud platform further includes:

[0154] S3001, calculate the mean and standard deviation of the historical group data, and determine the sum of twice the mean and the standard deviation of the historical group data as the initial abnormal threshold. Wherein, the historical group data includes the fusion features of each teaching stage in the teaching stage sequence of the historical multiple periods.

[0155] Exemplarily, the mean and standard deviation of the fusion features of the teaching stage sequence in the historical multiple periods can be calculated, and the initial abnormal threshold can be determined according to the historical mean and the historical standard deviation, that is, , wherein, represents the initial abnormal threshold, represents the mean of the historical group data, represents the standard deviation of the historical group data.

[0156] S3002, based on the fusion features of all teaching stages of the teaching stage sequence, calculate the standard deviation of the teaching stage sequence.

[0157] Exemplarily, the standard deviation of the teaching stage sequence is calculated to understand the behavior and emotional fluctuations of the current group in the teaching stage. If the standard deviation is large, it means that the behavior fluctuation of the group is large, and there may be large abnormal fluctuations; if the standard deviation is small, it means that the behavior of the group is stable.

[0158] Similarly, according to the fusion features of each teaching stage in the teaching stage sequence, the mean of the teaching stage sequence is calculated, and the standard deviation of the teaching stage sequence is calculated based on the mean.

[0159] S3003, compare the standard deviation of the historical group data and the standard deviation of the teaching stage sequence to obtain a comparison result, and adjust the multiple of the standard deviation in the initial abnormal threshold according to the comparison result to obtain a dynamic abnormal threshold.

[0160] Exemplarily, the standard deviation of the historical group data can be compared with the standard deviation of the teaching stage sequence, that is, , wherein, represents the ratio of the standard deviation of the two, represents the standard deviation of the teaching stage sequence. If the comparison result is that the standard deviation of the current teaching stage sequence is significantly greater than the standard deviation of the historical group data, that is, <1, it means that there may be abnormal behavior or emotional fluctuations in the current teaching stage sequence, and the state of group learning may be in an unstable state, and the multiple of the standard deviation in the initial abnormal threshold can be reduced; if the comparison result is that the standard deviation of the teaching stage sequence is close to the standard deviation of the historical group data, that is, 1, the multiple of the standard deviation in the initial abnormal threshold can be maintained; if the comparison result is that the standard deviation of the teaching stage sequence is less than the standard deviation of the historical group data, that is, >1, the multiple of the standard deviation in the initial abnormal threshold can be increased. The multiple of the standard deviation in the initial abnormal threshold can be dynamically adjusted according to the comparison result, that is , wherein, is the multiple dynamically adjusted according to the comparison result, is the initial multiple (two times), , is used to control the adjustment range, and upper and lower limits (such as ) can be set to avoid excessive adjustment.

[0161] The above steps can dynamically adjust the threshold based on historical data and the standard deviation of the current stage, which can flexibly adapt to changes in group behavior, cognition and emotional characteristics, and at the same time enhance the detection ability of abnormalities.

[0162] S400, based on the teaching behavior data and biological data of the teaching stage in the abnormal detection result, a group teaching strategy is generated.

[0163] Exemplarily, based on the teaching behavior data and biological data in the teaching stage where the group anomaly exists, a machine learning model (such as a decision tree or a support vector machine (SVM)) can be used for causal analysis to determine which factors (such as teaching of a specific knowledge point, frequency of certain social interaction, etc.) have a significant impact on the group anomaly.

[0164] The goal of teaching strategy generation can be set. For example, if students show high biological response (such as increased galvanic skin response) in a certain teaching stage, the goal is to reduce student stress and improve concentration; if the interaction frequency of students is low, the goal is to improve classroom interaction and participation.

[0165] According to the analysis result and the goal, the group teaching strategy can be automatically generated through pre-set rules. For example, if the biological data (such as galvanic skin response) shows that students are too anxious, strategies such as extending the rest time, introducing meditation exercises, and reducing the difficulty of teaching can be generated; if the interaction frequency of students is low, strategies such as increasing the frequency of group discussions, setting more interactive questions, and encouraging online discussions can be generated; if the performance of students on a certain knowledge point is generally low, strategies such as providing more examples, reducing the complexity of the knowledge point, and splitting the knowledge point into smaller modules can be generated.

[0166] Optimization algorithms such as reinforcement learning or genetic algorithm can be used to dynamically generate the most suitable group teaching strategy. Reinforcement learning can simulate different teaching strategies (such as increasing interaction or slowing down teaching pace), and optimize the strategy according to feedback (such as student performance or emotional data), and finally generate an optimal teaching strategy; genetic algorithm can simulate crossover and mutation, and constantly adjust the parameters of the teaching strategy, such as teaching pace, interaction frequency, etc., until the best strategy combination is found.

[0167] After the generation of the group teaching strategy, the simulator can be used to evaluate the impact of the strategy on student behavior and biological data, and if the strategy is effective, the simulator will show that the student's mood and performance have improved. A detailed report can be generated to show the background, goals, recommended measures and expected effects of the strategy, and the group teaching strategy can be fed back to the teacher or education manager.

[0168] Through this step, specific group teaching strategies can be automatically generated, which is beneficial to the personalization and dynamic optimization of group teaching strategies, and does not rely on human intervention, helping teachers optimize teaching methods and improve students' learning experience.

[0169] In one possible implementation, please refer to Figure 3 , S400, based on the teaching behavior data and biological data of the teaching stage in the abnormal detection result, a group teaching strategy is generated, including:

[0170] S410, select the associated strategy matching the group cognitive characteristics from the strategy knowledge graph; extend the common error mode in the group cognitive characteristics to obtain the extended characteristics, and retrieve the extended strategy matching the extended characteristics from the strategy knowledge graph; integrate the associated strategy and the extended strategy to obtain the initial teaching strategy.

[0171] Illustratively, the strategy knowledge graph can include knowledge points, strategies and their relationships related to teaching strategies. For example, each strategy has some basic attributes (such as applicable knowledge points, implementation methods, teaching goals, etc.), and these strategies can be associated according to different teaching situations and group characteristics.

[0172] Feature matching algorithms (such as keyword matching, semantic similarity calculation, etc.) can be used to retrieve strategies related to common error patterns and knowledge gaps in the group cognitive characteristics from the strategy knowledge graph. When matching strategies, the common error patterns in the group cognitive characteristics can be compared with the error types or learning difficulties defined in the strategy knowledge graph to match. For example, if the common error pattern of the group is formula application error, strategies containing formula application or skill practice can be extracted from the strategy graph. Select the strategy matching the group cognitive characteristics to form a preliminary associated strategy set.

[0173] The common error patterns in the group cognitive characteristics can be used to identify more knowledge points or potential learning difficulties (i.e. extended characteristics) related to the error patterns through an extension model (such as a rule-based reasoning or machine learning model). For example, if the group has a common function plotting error, the extended characteristics may involve knowledge points related to image analysis. The extended characteristics help to identify other potential problems that students may encounter in similar situations.

[0174] Based on the extended features, relevant extended strategies are retrieved from the strategy knowledge graph using a feature matching algorithm. The extended strategies can involve more extensive or complex teaching methods. For example, to help students better understand the visual association between images and functions, more practice questions can be designed to promote students' mastery of image analysis skills.

[0175] During the merging of the associated strategies and the extended strategies, a deduplication process can be performed to facilitate the inclusion of non-duplicate strategies into the initial teaching strategies. Different strategies can be assigned priorities based on their applicability, effectiveness, or cognitive difficulties of the group. Finally, an initial teaching strategy is generated, which includes teaching methods for the group's current cognitive difficulties, error patterns, and extended features.

[0176] S420, based on the group behavior characteristics, adjust the strategy form of the initial teaching strategy, and based on the group emotional characteristics, adjust the strategy attribute of the initial teaching strategy, and generate a group teaching strategy.

[0177] Illustratively, the initial teaching strategy can be adjusted in form based on the group behavior characteristics by using dynamic optimization algorithms (such as reinforcement learning) or rule-based adjustments. For example, if the group tends to have low participation, the strategy may need to increase interactivity, and more interactive sessions and collaboration tasks between students can be added; if the group behavior shows low learning progress, the pace of the strategy can be adjusted, such as by adding review sessions, extending task completion time, etc. to adapt to the learning speed of students.

[0178] The strategy attribute of the initial teaching strategy can be adjusted based on the group emotional characteristics through a rule engine or machine learning algorithm. For example, if the mean of the stress index in the group emotional characteristics is greater than the mean threshold (which can be set artificially or calculated from historical data), the teaching strategy of emotional support type, such as relaxation activities and emotional regulation exercises, can be increased; if the standard deviation of the stress index in the group emotional characteristics is greater than the standard deviation threshold (which can be set artificially or calculated from historical data), the teaching strategy may need to adjust the challenge and pace of the task to reduce anxiety and improve emotional stability, such as gradually increasing the difficulty of the task to avoid excessive challenge that may cause negative emotions.

[0179] The adjustment results of the group cognitive characteristics, group behavior characteristics and group emotional characteristics are integrated to generate the final group teaching strategy. The group teaching strategy can include teaching content and methods adjusted according to the group cognitive characteristics, teaching form and interactive mode adjusted according to the group behavior characteristics, and emotional support strategies and learning pace adjusted according to the group emotional characteristics. The final group teaching strategy output can be a specific operation guide for teachers to execute, or a strategy recommendation in the smart learning cloud platform for further teaching adjustment.

[0180] In a possible implementation, the group teaching strategy generation method based on the wisdom learning cloud platform further includes:

[0181] S401, acquire core relationship pairs and multi-modal data sources. The core relationship pairs include relationships between misunderstanding modes, knowledge points, teaching strategies, and teaching resources, and the multi-modal data sources include teaching texts and teaching videos.

[0182] Illustratively, the construction process of the strategy knowledge graph starts from seed knowledge generation. The seed knowledge is a core relationship pair generated by education expert annotation, template mining, or rule generation, including misunderstanding mode-knowledge point-teaching strategy-teaching resource. The core relationship pair can be converted into structured information and stored as a text, a table, or a database record. For example, the core relationship pair annotated by an expert includes: misunderstanding mode: inertia confusion -> knowledge point: Newton's first law -> teaching strategy: explain the concept of inertia -> teaching resource: video of Newton's three laws; misunderstanding mode: current direction error -> knowledge point: current direction -> teaching strategy: introduce the flow direction of current -> teaching resource: current direction exercises.

[0183] The multi-modal data sources can include teaching texts (teaching materials, lecture notes, teaching plans, etc.) for extracting language-level knowledge points, teaching videos (classroom recordings, teaching demonstrations, etc.) for extracting visual / semantic-level knowledge points, teaching AR (augmented reality teaching materials), which is temporarily used as a structured processing interface for expansion, and teaching exercises (structured exercises and answers) as misunderstanding mode-knowledge point annotation aids. The multi-modal data sources themselves exist in the wisdom learning cloud platform, and can be directly acquired when used.

[0184] S402, construct a basic node and a relationship network according to the core relationship pairs to obtain an initial graph.

[0185] Illustratively, these core relationship pairs can be extracted from a table or a database and arranged into a preliminary data structure. For the four types of objects (misunderstanding modes, knowledge points, teaching strategies, and teaching resources) in all core relationship pairs, each unique item is marked as a graph entity (such as a node ID), that is, each misunderstanding mode, knowledge point, teaching strategy, and teaching resource in each core relationship pair is regarded as a node in the initial graph.

[0186] The different nodes are connected by the core relationship between the four types of objects, and the initial graph is constructed. A graph database (such as Neo4j) can be used to store these nodes and relationships. For example, misunderstanding mode-knowledge point: misunderstanding mode and knowledge point are connected by the "cause" relationship, indicating that students may have inertia misunderstanding when learning Newton's first law; knowledge point-teaching strategy: knowledge point and teaching strategy are connected by the "need" relationship, indicating that the knowledge point needs to be explained by the strategy of explaining the concept of inertia; teaching strategy-teaching resource: teaching strategy and teaching resource are connected by the "use" relationship, indicating that the teaching strategy can be implemented by playing the video of Newton's three laws.

[0187] S403, using natural language processing technology to analyze the teaching text, extracting the knowledge points of each text in the teaching text, and using computer vision technology to analyze the teaching video, extracting the knowledge points of each video in the teaching video.

[0188] Exemplarily, candidate phrases in the teaching text can be extracted based on statistics (such as TF-IDF) or graph algorithms (such as TextRank), or context-related terms can be extracted based on deep models such as BERT; knowledge point definition class structures (such as "X is..." and "the properties of Y include...") can be identified using sentence templates; phrase nesting (such as the tangent property of a circle) can be identified using syntax analysis; extracted terms can be compared with a knowledge point library or a standard teaching knowledge system to remove duplicates or merge synonyms (such as Newton's first law = inertia law). Through these steps, the knowledge point set of each text can be obtained. For example, text ID: text_001, knowledge points include Newton's first law and inertial reference system.

[0189] The teacher's explanation voice can be converted into text using an ASR (Automatic Speech Recognition) model (such as DeepSpeech, Whisper, etc.), and subsequent processing is consistent with the extraction of knowledge points from teaching text. Frame sampling or change detection can be used to extract static key frames, locate board writing changes, PPT switching moments, and perform image text recognition (OCR) on key frames to recognize board writing text, formulas, and PPT titles. Image text and dictionary matching can be used to extract term-based phrases as candidate knowledge points. Image segmentation models (such as YOLO, LayoutLM) can be used to identify blackboard regions, diagrams, and arrow directions, and perform image classification or object detection on visual elements (such as function images, physical diagrams) to match images with a knowledge point label library (such as parabolic trajectory image → projectile motion). Text knowledge points from ASR can be combined with image knowledge points obtained from OCR / image understanding to determine whether they are the same knowledge point based on semantic similarity (such as BERT similarity) and assign a confidence score (such as voice frequency + image location + visual clue strength) to enhance knowledge point extraction accuracy. Through these steps, a set of knowledge points for each video segment can be obtained. For example, video ID: video_007, knowledge points include kinetic energy theorem, definition of work, and velocity-time image.

[0190] S404, fuse the text and video of the same knowledge point, and connect with the corresponding knowledge point in the initial graph to obtain the strategy knowledge graph.

[0191] It can be understood that although the core relationship pair already contains teaching resources, the role of multi-modal data sources is not to repeat the resources, but to supplement, verify, enhance, and dynamically update the relevance of teaching resources and the depth of knowledge understanding in the graph.

[0192] The teaching resources in the core relationship pair are in a structured or annotated form (such as resource ID, resource name or link), and do not contain in-depth semantic analysis of resource content. Multi-modal data sources provide content-level understanding of resource content through semantic analysis, so that the graph not only knows that resource A is used for knowledge point X, but also knows which knowledge points are specifically explained in resource A, how they are explained, and how deep the explanation is.

[0193] The core relationship pair may come from expert input, template mining or rule generation, and has lag, limitations or biases. Multi-modal data sources can be used to verify the adaptability of existing strategies or resources and knowledge points, discover missing relationships (such as knowledge point X has a video explanation, but is not associated with the strategy), and complete potential teaching misunderstandings or alternative resources.

[0194] Exemplarily, the knowledge points of the teaching text and the knowledge points of the teaching video can be respectively converted into text vectors and video vectors by BERT or word2vec embedding, and the semantic similarity between the text vectors and the video vectors can be calculated by using a method such as cosine similarity or Euclidean distance. A similarity threshold can be set, and the knowledge points of the teaching text and the knowledge points of the teaching video with a semantic similarity greater than or equal to the similarity threshold can be determined as the same knowledge points, and the text and the video of the same knowledge points can be merged into the same node to form a fusion node set. The language context (definition, example sentence) extracted from the text can be used as a text attribute, and the image frame, audio segment or image label recognized in the video can be used as a visual attribute, so that each fusion node becomes a structure with multi-modal features, which can be used for reasoning or recommendation.

[0195] The fusion node is matched with the knowledge point node in the initial graph (by the knowledge point ID or the standard name), if the knowledge point corresponding to the node already exists in the initial graph, the fusion node is used as a supplementary information node or an extended instance node, and the connection is established on the knowledge graph structure through semantic edges such as has_instance, described_by or supported_by, if the knowledge point corresponding to the node does not exist in the graph, a new knowledge point node is automatically created, and the relationship with the strategy, misunderstanding mode and the like is rebuilt. The edge between the fusion node and the knowledge point node in the initial graph can be established, and the edge type can include multi_modal_support (indicating that the knowledge point is supported by multi-modal data), text_evidence (text description), and video_evidence (video explanation). These edges can contain additional attributes such as evidence strength, timestamp, context and the like.

[0196] After the fusion node is connected to the knowledge point in the initial graph, all existing relationships of the knowledge point are automatically inherited or connected, so as to obtain a strategy knowledge graph. Each knowledge point in the strategy knowledge graph will have a multi-modal evidence node as support, and the explainability and retrievability of the teaching strategy are enhanced, and the strategy recommendation, path query and resource positioning can be performed based on the graph, and the strategy node in the graph can be reversely pointed to the fusion node supporting the strategy node, to form a causal closed loop path (such as misunderstanding mode→knowledge point→teaching strategy→video resource).

[0197] This step ensures that the strategy knowledge graph not only has structural integrity, but also integrates multi-source content understanding, so as to have stronger reasoning ability, resource scheduling ability and teaching strategy adaptation ability.

[0198] In a possible implementation, please refer to Figure 3 The group teaching strategy generation method based on the intelligent learning cloud platform further includes:

[0199] S10, obtaining feedback data.

[0200] Exemplarily, when the generated group teaching strategy is implemented, feedback data of students, such as student teaching behavior data and biological data after implementation of the group teaching strategy, can be obtained through the smart learning cloud platform.

[0201] S20, according to the feedback data, identifying a feedback error mode and calculating a feedback error strength by using a cognitive contradiction field, to obtain first data. The first data includes the feedback error mode and the feedback error strength.

[0202] It can be understood that the cognitive contradiction field is an abstract model simulating the state of cognitive energy, which can be used to represent the difference between the student's cognitive state and the target state, and map the student's current learning state to a multi-dimensional cognitive space. If a student misunderstands a certain knowledge point, a cognitive energy trap, also known as a cognitive conflict area, will be formed at that node, indicating that there is a deviation between the student's current cognitive state and the target cognitive state.

[0203] Exemplarily, a comparison chart between the student's answer or learning path and the ideal learning path can be constructed by using the feedback data of the student, to identify the steps or sub-tasks where the deviation occurs, and to locate the cognitive obstacle point (i.e. the point where the error mode occurs) by using the path difference. For example, a student repeatedly uses an incorrect formula when solving a problem related to Newton's second law, and the student has a misunderstanding of formula misuse according to the answer process and historical records. If a student frequently confuses the difference between kinetic energy and potential energy, it can be determined that there is a concept confusion. The identified error type is the feedback error mode, which is an error performance with cognitive deviation characteristics.

[0204] After identifying the feedback error mode, the frequency of the error type appearing in the group can be calculated as the feedback error strength according to the test records of the teaching behavior data in the feedback data. If a misunderstanding repeatedly occurs in multiple tests or assignments, it indicates that the problem is deeply rooted, and the feedback error strength can be increased. Some knowledge points are at key positions in the learning path, such as the graph of a quadratic function, which is the basis for subsequent analytic geometry problems. If a misunderstanding occurs at an important node, the feedback error strength can be increased. If a certain error repeatedly occurs in the past learning cycles of the group, the error is determined to be a persistent misunderstanding, and the strength is further adjusted. It can also be analyzed whether the error can affect other knowledge points. If an error can cause students to make mistakes in multiple related knowledge points (such as confusing speed and acceleration, which will affect multiple parts of mechanics), the strength will be further increased. The above dimensions can be weighted and summarized to calculate a numerical error strength to represent the depth of the error in the student's cognitive structure.

[0205] This step forms a mapping from feedback data to a cognitive representation of errors, enabling the introduction of cognitive problems into the strategy optimization logic in a recognizable, quantifiable, and intervenable manner, which is a key cognitive basis for strategy optimization.

[0206] In S30, the multi-modal data is filtered from the strategy knowledge graph according to the first data, and the association between the feedback error mode and the multi-modal data is extracted to obtain second data. The second data includes the multi-modal data and the association between the feedback error mode and the multi-modal data.

[0207] For example, the feedback error mode in the first data is used as a key query condition, and the associated knowledge points of the feedback error mode are used to find the strategy path connected thereto in the strategy knowledge graph. Through strategy knowledge graph traversal and semantic matching, it can be identified which teaching resources are related to the error and which resources have been bound to intervene in the error type.

[0208] Among the identified resource nodes, video segments strongly related to the feedback error mode / associated knowledge points of the feedback error mode can be filtered out according to the modal type, obtained through voice text analysis (ASR) and image tagging in the video; passages semantically matched with the error mode extracted from the teaching text. Augmented reality scenes that interact with knowledge points / error modes, such as simulated kinetic energy conversion processes, can also be filtered out. Exercises related to the error can be filtered out for verification or guidance for correction. The attributes of each resource node can include modal type, knowledge point coverage, past intervention effect indicators (such as the error correction rate of the resource in the group), and semantic similarity scores (the semantic fit degree of the resource content with the feedback misunderstanding).

[0209] Based on the structure of the strategy knowledge graph, the association path between each resource and the feedback error mode can be extracted to identify whether the resource has been used to directly intervene in the error mode, indirectly support the knowledge points associated with the misunderstanding, or whether the content explicitly explains or demonstrates the knowledge misunderstanding.

[0210] Natural language processing and visual semantic embedding techniques can be used to further calculate the matching degree of the misunderstanding mode from the resource content. For example, whether the video explanation voice contains formula selection common misunderstanding, whether the text presents in the form of error examples, and whether the AR scene simulates the misunderstanding correction process through interaction. An association strength score can be assigned to each error mode-multi-modal resource relationship for subsequent ordering and strategy reasoning.

[0211] The structured modeling results of cognitive errors are semantically connected with the teaching resources in the strategy knowledge graph through the above steps, providing high-quality, multi-dimensional input support for subsequent dynamic optimization of teaching strategies.

[0212] S40, updating the strategy knowledge graph based on the feedback data, the first data, and the second data using a reinforcement learning algorithm.

[0213] Exemplarily, the reinforcement learning core elements include State, Action, Reward. State includes the execution state of each strategy in a specific teaching scenario, the current knowledge point, the error mode, the applied strategy and its resource path, such as on knowledge point K1, for error mode M1, using strategy S1+resource V1. Action includes the strategy operation that can be taken on the current state, such as retaining the current strategy, replacing it with other strategies, adding supplementary resources, adjusting the priority or scope of the strategy. Reward includes a quantitative indicator driven by feedback data, used to judge the effect of the strategy, such as error intensity reduction→positive reward, learning achievement improvement→positive reward, multi-modal resource intervention ineffective→negative reward, no significant improvement in learning time→weak reward or punishment.

[0214] According to the student feedback data generated after the execution of each teaching strategy, it can be analyzed whether the student has made cognitive progress under the intervention of the strategy; whether the strategy is effective according to whether the error intensity in the first data decreases; whether the resource has played the expected role in combination with the correlation of multi-modal resources in the second data.

[0215] According to the feedback effect, a reward value is assigned to each strategy path, which is not only based on the current result, but also can integrate historical performance (such as using a sliding window). The reward mechanism can include a significant decrease in error intensity, a highly effective strategy, +1 for the reward value; neutral effect, no significant change, reward value remains unchanged; misunderstanding continues or intensifies, the strategy may be ineffective or have side effects, reward value-1.

[0216] According to the cumulative reward value and the current strategy effect, the credibility, priority weight of the current strategy in the strategy knowledge graph can be improved, or the current strategy activation probability can be reduced, or the applicable scope can be reduced; other strategy paths related to the error can be introduced in the strategy knowledge graph (may be migrated from historical data, expert library or other nodes); multiple sub-strategies can be combined into a new composite strategy to deal with complex misunderstandings.

[0217] According to the output result of the reinforcement learning strategy, the structure and attributes of the strategy knowledge graph can be updated. The attributes of the strategy node can be modified, such as strategy trustworthiness, applicable misunderstanding mode range, adaptability score with resources, reinforcement weight; the edges between the strategy and the knowledge point, misunderstanding mode, multi-modal resource can be enhanced or weakened, that is, the weight of the edge is increased, indicating that the correlation is stronger, the edge with poor effect is closed / deleted, indicating that the resource is invalid, a new edge is created, indicating that a new strategy or new resource is found to be available for intervention on the error; new effective strategy paths are identified in the graph and included in the recommended strategy library for subsequent strategy recommendation or autonomous generation.

[0218] The reinforcement learning process is continuously executed in a loop, and new feedback data is generated after each round of policy execution, entering the next learning cycle, gradually realizing the personalization of policy selection, optimization and compression of policy path, accurate matching of multi-modal resources and error types, and dynamic evolution of graph structure, making it more consistent with the cognitive changes of students in real teaching scenarios.

[0219] Through the above steps, the system not only has self-learning ability, but also can continuously improve the strategy recommendation according to the real teaching results, which is beneficial to the adaptability and dynamic optimization of the teaching strategy for different student groups, and realizes a high-quality and sustainable intelligent teaching support system.

[0220] It should be understood that the size of the serial number of each step in the above embodiments does not mean the order of execution, and the execution order of each process should be determined according to its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0221] The embodiments of the present application also provide an electronic device, the electronic device of the embodiments includes at least one processor, at least one memory, and a computer program stored in the at least one memory and executable on the at least one processor, and the processor executes the computer program to enable the electronic device to implement the steps in any of the group teaching strategy generation methods based on the intelligent learning cloud platform.

[0222] By way of example, the computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present application. One or more modules / units can be a series of computer program instruction segments capable of completing a specific function, which are used to describe the execution process of the computer program in the electronic device.

[0223] The electronic device can be a computer, a notebook, a palm computer, and a cloud server, etc. The electronic device can include, but is not limited to, a processor, a memory. It can also include input / output devices, network access devices, buses, etc.

[0224] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application specific integrated circuits (ASIC), etc. The general-purpose processor can be a microprocessor, or the processor can be any conventional processor, etc.

[0225] The memory can be an internal storage unit of the electronic device in some embodiments, such as a hard disk or a memory of the electronic device. The memory can also be an external storage device of the electronic device in other embodiments, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash card, and the like. The memory is used to store an operating system, an application program, a boot loader, data, and other programs, such as program codes of a computer program.

[0226] The above embodiments are only used to illustrate the technical solutions of the present application, but not limit the same. Although the present application is described in detail with reference to the foregoing embodiments, it should be understood by those skilled in the art that the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced by equivalent replacements. The modifications or replacements do not change the essence of the corresponding technical solutions, and should be included in the protection scope of the present application.

Claims

1. A method for generating group teaching strategies based on a smart learning cloud platform, characterized in that, include: The teaching behavior data, biological data, and teaching progress data of multiple classes are obtained through the smart learning cloud platform; among them, the teaching progress data includes the knowledge points that have been taught and the corresponding teaching dates. Based on the teaching progress data of multiple classes, the teaching stages of multiple classes are spatiotemporally aligned to obtain a teaching stage sequence; wherein, the teaching stage sequence includes multiple teaching stages, and each teaching stage includes teaching behavior data and biological data of multiple classes in the same teaching stage; Based on each teaching stage of the teaching stage sequence, group anomaly detection is performed on multiple classes to obtain anomaly detection results; wherein, the anomaly detection results include teaching stages where group anomalies exist; Based on the teaching behavior data and biological data of the teaching stage in the anomaly detection results, a group teaching strategy is generated. The step of performing spatiotemporal alignment processing on the teaching stages of multiple classes based on teaching progress data to obtain a teaching stage sequence includes: The system compares the knowledge points in the teaching progress data of multiple classes with those in the standard teaching framework. If they are inconsistent, the knowledge points in the inconsistent teaching progress data are replaced with the corresponding knowledge points in the standard teaching framework according to the equivalent knowledge point table. The equivalent knowledge point table includes the equivalence relationship between the same or similar knowledge points in different textbooks and the knowledge points in the standard teaching framework. The equivalent knowledge point table is obtained through semantic similarity calculation and supplementation by a manual rule base. The standard teaching framework includes the ID of each teaching stage, the knowledge points included in each teaching stage, and the expected teaching time of each teaching stage. Calculate the distance matrix between the knowledge points in the teaching progress data of each class and the knowledge points in the standard teaching framework; based on each distance matrix, use the dynamic time warping algorithm to calculate the optimal alignment path between the teaching progress of each class and the standard teaching framework; according to the optimal alignment path between the teaching progress of each class and the standard teaching framework, map the teaching progress of each class to each teaching stage in the standard teaching framework to obtain the teaching stage sequence. For each teaching stage based on the teaching stage sequence, group anomaly detection is performed on multiple classes to obtain anomaly detection results, including: Extract teaching behavior data and biological data from multiple classes in each teaching stage sequence; The following procedures will be implemented for each teaching stage: Common error patterns are extracted from teaching behavior data of multiple classes; a cognitive network graph is constructed based on the common error patterns, and group knowledge breakpoints are identified based on the cognitive network graph; the common error patterns and the group knowledge breakpoints are integrated to obtain group cognitive characteristics; wherein, the cognitive network graph is used to characterize students' mastery of different knowledge points and the correlation between knowledge points during the learning process. Based on teaching behavior data from multiple classes, Fourier transform is used to perform frequency analysis on students' learning behaviors to obtain a behavior frequency spectrum; the resonance matching degree between the behavior frequency spectrum and the ideal learning path frequency spectrum is calculated; and non-obvious behavior patterns are identified based on the resonance matching degree to obtain group behavior characteristics. Stress index was calculated based on biological data from multiple classes to obtain group emotional characteristics; The group cognitive characteristics, the group behavioral characteristics, and the group emotional characteristics are fused to obtain fused features; The fusion feature is compared with the dynamic anomaly threshold. If the fusion feature is greater than the dynamic anomaly threshold, it is determined that there is a group anomaly in this teaching stage, and the anomaly detection result is obtained.

2. The method for generating group teaching strategies based on a smart learning cloud platform as described in claim 1, characterized in that, The method further includes: Based on text analysis technology, knowledge points are extracted from teaching content; wherein, the teaching content is pre-stored in the smart learning cloud platform; The context of the teaching content is analyzed using a relation extraction algorithm to identify the dependencies between the knowledge points and construct a knowledge point dependency graph. Based on the knowledge point dependency graph, the knowledge points are divided into various teaching stages, and each teaching stage is labeled with an ID. Based on the number of knowledge points in each teaching stage, allocate the expected teaching time for each teaching stage; The standard teaching framework is obtained by arranging each teaching stage ID, the knowledge points included in each teaching stage, and the expected teaching time of each teaching stage according to the dependencies between the knowledge points.

3. The method for generating group teaching strategies based on a smart learning cloud platform as described in claim 1, characterized in that, The method further includes: During the spatiotemporal alignment process, the dynamic time warping algorithm is used to detect whether there is a progress jump feature; wherein, the progress jump feature is that the knowledge point corresponding to the teaching progress of a certain class in the standard teaching framework is not covered by the teaching progress data of that class. If the progress jump feature exists, virtual knowledge points are inserted at the positions of the uncovered knowledge points in the teaching stage sequence to fill in the uncovered knowledge points; or If the progress jump feature exists, mark the uncovered knowledge points in the teaching stage sequence and indicate that the uncovered knowledge points are missing.

4. The method for generating group teaching strategies based on a smart learning cloud platform as described in claim 1, characterized in that, The method further includes: Acquire teaching behavior data of all classes in multiple historical teaching stages, and define the teaching behavior data of all classes in each historical teaching stage as a meta-task; For each meta-task, the teaching behavior data of all classes are divided into a training set and a test set; the meta-learning model is trained based on the training set, and the loss value of the meta-learning model on the training set is calculated; the model parameters of the meta-learning model are updated according to the loss value using the gradient descent algorithm; the test loss is calculated based on the test set, and the gradient of the test loss with respect to the model parameters is calculated based on the test loss. Calculate the average gradient of all meta-tasks to obtain the meta-gradient; update the model parameters of the meta-learning model according to the meta-gradient to obtain the meta-learning model after training. The teaching behavior data of multiple classes in each teaching stage are input into the trained meta-learning model so that the trained meta-learning model can extract the behavior gradient and balance coefficient of each teaching stage, and determine the group behavior characteristics of each teaching stage based on the behavior gradient and balance coefficient of each teaching stage.

5. The method for generating group teaching strategies based on a smart learning cloud platform as described in claim 1, characterized in that, The method further includes: Calculate the mean and standard deviation of the historical group data, and use the sum of twice the mean and standard deviation of the historical group data as the initial anomaly threshold; wherein, the historical group data includes the fusion features of each teaching stage in a sequence of teaching stages across multiple historical cycles; Based on the integration characteristics of all teaching stages in the teaching stage sequence, the standard deviation of the teaching stage sequence is calculated; By comparing the standard deviation of the historical group data with the standard deviation of the teaching stage sequence, a comparison result is obtained. Based on the comparison result, the multiple of the standard deviation in the initial anomaly threshold is adjusted to obtain the dynamic anomaly threshold.

6. The method for generating group teaching strategies based on a smart learning cloud platform as described in claim 1, characterized in that, The generation of group teaching strategies based on teaching behavior data and biological data from the teaching stages in the anomaly detection results includes: Select association strategies that match the group's cognitive characteristics from the strategy knowledge graph; extend the associations of common error patterns in the group's cognitive characteristics to obtain extended features, and retrieve extended strategies that match the extended features from the strategy knowledge graph; integrate the association strategies and the extended strategies to obtain the initial teaching strategy; Based on the group's behavioral characteristics, the strategy form of the initial teaching strategy is adjusted, and based on the group's emotional characteristics, the strategy attributes of the initial teaching strategy are adjusted to generate a group teaching strategy.

7. The method for generating group teaching strategies based on a smart learning cloud platform as described in claim 6, characterized in that, The method further includes: Obtain core relation pairs and multimodal data sources; wherein, the core relation pairs include relationships between misunderstanding patterns, knowledge points, teaching strategies and teaching resources, and the multimodal data sources include teaching texts and teaching videos; Based on the core relationships, a basic node and relationship network are constructed to obtain an initial graph; The teaching text is analyzed using natural language processing technology to extract the knowledge points of each text in the teaching text, and the teaching video is analyzed using computer vision technology to extract the knowledge points of each video in the teaching video; By fusing text and video with the same knowledge points and connecting them with the corresponding knowledge points in the initial graph, the strategy knowledge graph is obtained.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

Images