AI intelligent interactive course adjusting method and system based on large language model

Abstract

The invention provides an AI intelligent interactive course adjustment method and system based on a large language model, and the method comprises the steps: obtaining original typing input rhythm data, including a keyboard event timestamp sequence, an event type and input contents, of a student in concept interpretation processes with different difficulties; pause time distribution, input speed change and error rate indexes are calculated according to the event timestamp sequence; according to the current concept understanding level of the student, dynamically adjusting the course difficulty gradient and the progress arrangement, if the understanding level of the student is high, increasing the course difficulty and the learning progress, and if the understanding level of the student is low, reducing the course difficulty and enhancing targeted practice and feedback; the typewriting input rhythm characteristics of the students are combined with other learning behavior data, a comprehensive learning effect evaluation model is constructed through multi-modal data fusion, the comprehensive learning effect evaluation model is used for evaluating mastering conditions of the students on different knowledge points, and weak links and dominant fields of the students are identified.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an AI intelligent interactive course adjustment method and system based on a large language model. Background Art

[0002] In the process of students learning concepts, by analyzing the changes in the rhythm of students when typing to explain the concepts, the depth of students' understanding of the concepts they have learned can be judged, and the course progress can be dynamically adjusted accordingly. This is considered to be a forward-looking intelligent teaching method. For example, in the game session of the AI ​​intelligent interactive course, students must complete the task by typing, and their answering method and typing rhythm can reveal their mastery of the concepts. During the game, the system will ask some questions related to the concepts learned, requiring students to enter the answers through the keyboard within a limited time. By studying the distribution characteristics of pauses when students type to answer questions, it is possible to infer the students' mastery of the concepts, thereby providing a basis for intelligently adjusting the teaching progress. However, to achieve this intelligent teaching method, several technical difficulties need to be overcome: first, how to accurately identify the distribution characteristics of students' pause time during typing, and how to eliminate the interference of factors such as typing speed and habits, and extract key feature values ​​that can truly reflect students' understanding level. This is an urgent problem that needs to be solved. Secondly, different concepts have different levels of difficulty, and the same pause time distribution may represent students' different levels of understanding of concepts of different difficulty. Therefore, how to calibrate the pause time distribution according to the difficulty of the concept itself to obtain indicators that can accurately reflect the students' level of understanding is also a technical challenge. Even if the indicators that can accurately reflect the students' level of understanding are successfully obtained, how to dynamically adjust the course progress based on these indicators is still a considerable challenge. The adjustment of the course progress needs to take into account multiple factors such as the students' level of understanding, learning interest, and cognitive load. How to find a balance between these factors and avoid excessive or insufficient adjustment of the course progress is also a technical problem that needs further research. Summary of the invention

[0003] The present invention provides an AI intelligent interactive course adjustment method based on a large language model, which mainly includes:

[0004] Obtaining students' original typing input rhythm data during the process of explaining concepts of different difficulty levels, including keyboard event timestamp sequence, event type, and input content, and calculating pause time distribution, input speed change, and error rate indicators based on the event timestamp sequence;

[0005] Preprocess the original typing input rhythm data to remove abnormal pauses caused by factors such as keyboard layout familiarity and finger flexibility differences, and construct a multi-dimensional feature vector of students' typing input rhythm based on the students' basic typing speed and rhythm stability characteristics;

[0006] The multi-dimensional feature vectors of students' typing input rhythm are clustered using a density-based clustering model. The clustering results are used to reflect the typical input patterns of students at different levels of conceptual understanding. The central feature vector of each cluster is analyzed to determine its corresponding conceptual understanding level.

[0007] When students are inputting concept explanations, their typing input rhythm features are extracted in real time, and the feature vectors are input into the clustering model for classification, so as to obtain the cluster to which the student's current typing input rhythm features belong, and then judge the student's current corresponding concept understanding level;

[0008] Dynamically adjust the course difficulty gradient and progress schedule based on students' current level of understanding of concepts. If students have a high level of understanding, increase the course difficulty and learning progress. If students have a low level of understanding, reduce the course difficulty and strengthen targeted practice and feedback.

[0009] Continuously collect typing input rhythm data during the learning process, regularly use new data to incrementally train and tune the clustering model, and dynamically update the concept understanding judgment criteria for each cluster to adapt to changes in students' typing habits and improvements in their concept understanding ability;

[0010] Combine students' typing input rhythm characteristics with other learning behavior data, and build a comprehensive learning effect evaluation model through multimodal data fusion to evaluate students' mastery of different knowledge points and identify students' weaknesses and strengths.

[0011] Based on the evaluation results of the learning effect evaluation model, the content organization, presentation method, and interaction design of the course are optimized to provide personalized learning paths and content recommendations for students with different levels of understanding, and dynamically adjust the difficulty distribution of learning resources to match the students' overall cognitive ability development level.

[0012] The present invention provides an AI intelligent interactive course adjustment system based on a large language model, which mainly includes:

[0013] Data collection and preprocessing module, used to obtain and process students' typing input rhythm data;

[0014] Feature extraction and clustering module, used to construct feature vectors and perform cluster analysis;

[0015] Real-time assessment and dynamic adjustment module to determine students’ understanding level and adjust course difficulty;

[0016] Model optimization and update module, used to continuously optimize clustering models and evaluation criteria;

[0017] Comprehensive evaluation and personalized recommendation module, used for multimodal data fusion evaluation and personalized learning recommendation.

[0018] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:

[0019] The present invention discloses an AI intelligent interactive class method based on a large language model. By acquiring the typing input rhythm data of students in the process of concept explanation, a multi-dimensional feature vector is constructed, and the feature vector is classified using a density clustering model to achieve real-time evaluation of students' concept understanding level. The system dynamically adjusts the course difficulty and progress according to the evaluation results to provide students with personalized learning paths. By continuously collecting data and updating the model, the present invention can adapt to changes in students' abilities. In addition, the present invention combines the typing input rhythm features with other learning behavior data to construct a comprehensive learning effect evaluation model, comprehensively evaluate students' knowledge mastery, and optimize the course content and presentation methods accordingly. This method can not only accurately reflect students' level of understanding, but also provide timely personalized teaching adjustments, thereby significantly improving learning effects and teaching quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The present invention is a flow chart of an AI intelligent interactive course adjustment method based on a large language model. DETAILED DESCRIPTION

[0021] The technical scheme of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] like Figure 1 In this embodiment, a method for adjusting an AI intelligent interactive course based on a large language model may specifically include:

[0023] S101. Obtain the original typing input rhythm data of students during the process of explaining concepts of different difficulty levels, including keyboard event timestamp sequence, event type, and input content, and calculate the pause time distribution, input speed change, and error rate index based on the event timestamp sequence.

[0024] The keyboard event timestamp sequence and input content of students in the process of explaining concepts are obtained. The adaptive threshold method is used to calculate the time interval between adjacent keys. If the time interval exceeds the preset multiple of the average key interval time of students, it is identified as a pause point, and the pause time distribution data is obtained. The sliding window method is used to calculate the number of character inputs in each time window to obtain the discrete speed value, and the discrete speed value is smoothed by the moving average method to generate the input speed change curve. The use of backspace and delete keys is extracted from the keyboard event sequence, and the position and number of backspace and delete operations are recorded. The error rate index is calculated in combination with the length of the input content, and the frequency of consecutive backspace or delete operations is counted to obtain the modification frequency index. Text matching is performed on the student input content according to the pre-defined difficulty level keyword list to determine the difficulty level of the current explanation concept. Based on the pause time distribution, input speed change, error rate and modification frequency characteristics, the K-means clustering algorithm is used to classify the student input behavior, and combined with the concept difficulty level, the student's understanding of concepts of different difficulty levels is evaluated.

[0025] Specifically, according to the keyboard event timestamp sequence and input content of students in the process of explaining concepts of different difficulty levels, the adaptive threshold method is used to calculate the time interval between adjacent keys, and the interval that exceeds 1.5 times the average key interval of students is identified as a pause point to obtain the pause time distribution data. The sliding window method is used to calculate the number of characters input in each time window to obtain a series of discrete speed values, which are smoothed by the moving average method to generate an input speed change curve. The use of the backspace and delete keys is extracted from the keyboard event sequence, and the position and number of each backspace and delete operation are recorded. Combined with the length of the input content, the error rate index is calculated, and the frequency of multiple consecutive backspace or delete operations is counted as the modification frequency index. According to the pre-defined difficulty level keyword list, the text matching of the student input content is performed to determine the difficulty level of the current concept being explained. Based on the characteristics of pause time distribution, input speed change, error rate and modification frequency, the K-means clustering algorithm is used to classify the student's input behavior. Combined with the concept difficulty level as a feature, the decision tree algorithm is used to evaluate the student's understanding of concepts of different difficulty levels. The keyboard event recorder is used to obtain the student's input data in the process of explaining concepts of different difficulty levels, including timestamps, key types and input content. The pause point is calculated using the adaptive threshold method, and the interval that exceeds 1.5 times the average key interval of the students is identified as a pause. For example, if the average key interval of the students is 0.5 seconds, the interval that exceeds 0.75 seconds is marked as a pause point. A 60-second sliding window is used, which slides once every 10 seconds, and the number of characters input in the window is calculated to obtain a series of discrete speed values. These speed values ​​are smoothed using the 5-point moving average method to generate an input speed change curve. The use of the backspace and delete keys is extracted from the keyboard event sequence, and the position and number of each backspace and delete operation are recorded. If the total length of the input content is 100 characters, and it contains 10 backspace or delete operations, the error rate is 10%. The frequency of multiple consecutive backspace or delete operations, such as the number of consecutive deletions of more than 5 characters within 3 seconds, is counted as a modification frequency indicator. A difficulty level keyword list is pre-established, including three levels: elementary, intermediate, and advanced, and each level contains 20 keywords. Text matching is performed on the content entered by the students. If the advanced keyword is matched more than 5 times, the difficulty level of the concept currently being explained is determined to be advanced. Combining the characteristics of pause time distribution, input speed change, error rate and modification frequency, the K-means clustering algorithm was used to classify students' input behaviors into three categories: fluent, thinking and modification. Using these classification results and the concept difficulty level as features, the decision tree algorithm was used to evaluate students' understanding of concepts of different difficulty levels, and the understanding level was divided into four levels: excellent, good, average and needs improvement. The depth of the decision tree was set to 3, and the minimum number of leaf node samples was 10 to prevent overfitting.

[0026] S102. Preprocess the original typing input rhythm data to eliminate abnormal pauses caused by factors such as keyboard layout familiarity and finger flexibility differences, and construct a multi-dimensional feature vector of the student's typing input rhythm in combination with the student's basic typing speed and rhythm stability characteristics.

[0027] According to the timestamp sequence in the original typing input rhythm data, the sliding window method is used to count the key frequency distribution in each window to obtain the basic typing speed and rhythm stability characteristics. The average input speed and standard deviation of each key area are calculated. If the time interval exceeds the range of the standard deviation of the mean value plus or minus the preset multiple, it is marked as an outlier to identify abnormal pauses caused by the familiarity of the keyboard layout. By calculating the speed change rate during continuous input, the mean and variance of the speed change rate are obtained as the finger flexibility index. For the time interval marked as abnormal, the linear interpolation method is used to fill it to generate the cleaned typing input rhythm data. According to the cleaned typing input rhythm data, the basic typing speed, rhythm stability, keyboard layout familiarity, finger flexibility index and error correction frequency characteristics are standardized, the feature covariance matrix is ​​calculated, the eigenvalues ​​and eigenvectors are solved, and the first few principal components whose cumulative contribution rate reaches the preset threshold are selected to construct the multi-dimensional feature vector of the student's typing input rhythm.

[0028] Specifically, according to the timestamp sequence in the original typing input rhythm data, the time intervals between adjacent keys are calculated, and the key frequency distribution in each window is statistically analyzed using a sliding window method with a window size of 60 seconds and a step size of 10 seconds to obtain the basic typing speed and rhythm stability characteristics of the students. By comparing the differences in students' input speeds in different key positions, the average input speed and standard deviation of each key position are calculated, and the time intervals exceeding the range of plus or minus 2 times the standard deviation of the mean are marked as outliers to identify abnormal pauses caused by familiarity with the keyboard layout. The speed change rate during continuous input is calculated, and the mean and variance of the speed change rate are statistically used as an indicator of finger flexibility. At the same time, the error correction frequency is obtained by calculating the frequency of use of the backspace key and the delete key. The time intervals marked as abnormal are interpolated, and the outliers are filled using the linear interpolation method to generate the cleaned typing input rhythm data. Based on the cleaned typing input rhythm data, combined with the basic typing speed, rhythm stability, keyboard layout familiarity, finger flexibility index and error correction frequency, these features are standardized, the feature covariance matrix is ​​calculated, the eigenvalues ​​and eigenvectors are solved, and the first few principal components with a cumulative contribution rate of 85% are selected to construct the multi-dimensional feature vector of the student's typing input rhythm. First, the original typing input rhythm data of the students is recorded by the data acquisition device, including the timestamp and key type of each key press. For a 5-minute typing test, about 1500 timestamp data points may be collected. The time interval between adjacent keys is calculated to obtain about 1499 interval values. Using the sliding window method with a window size of 60 seconds and a step size of 10 seconds, 27 windows can be obtained in the 5-minute data. The key frequency distribution is statistically analyzed in each window, and the average key speed and standard deviation are calculated as the basic typing speed and rhythm stability features. For example, the average key speed of a window may be 60 words / minute, with a standard deviation of 5 words / minute. The keyboard is divided into three areas: the left-hand area, the right-hand area, and the middle area. The average input speed and standard deviation of each area are calculated. If the average speed of the left-hand area is 50 words / minute and the standard deviation is 8 words / minute, the intervals exceeding 66 words / minute or below 34 words / minute are marked as outliers. For the continuous input process, the speed change rate per second is calculated, such as from 55 words / minute to 60 words / minute, the change rate is 9.09%. The mean and variance of these change rates are counted as indicators of finger flexibility. At the same time, the number of times the backspace key and the delete key are used is recorded and divided by the total number of key presses to obtain the error correction frequency. For the time intervals marked as abnormal, linear interpolation is used for filling. For example, if there is an outlier between two normal values, the average of the two normal values ​​is used to replace the outlier. Finally, the cleaned data is used for feature extraction, including average typing speed, speed standard deviation, speed difference in each area, speed change rate mean and variance, error correction frequency, etc.These features are standardized, the feature covariance matrix is ​​calculated, and the eigenvalues ​​and eigenvectors are solved. Assuming that 5 principal components are obtained and their cumulative contribution rate reaches 87%, these 5 principal components are selected to construct a multi-dimensional feature vector for subsequent typing pattern analysis and evaluation.

[0029] When constructing a multidimensional feature vector of students' typing input rhythm, the students' basic typing speed is calculated as a basic indicator to measure their overall typing ability; the stability of students' typing rhythm is quantified by calculating the variance or coefficient of variation of the input interval time; the basic typing speed and rhythm stability indicators are combined with other features, including pause time distribution and input speed changes, to form a complete multidimensional feature vector.

[0030] According to the original typing input data, the sliding window method is used to obtain the number of characters in the time window. The window size of the sliding window is the preset time length, and the step length of the sliding window is the preset step length; the basic typing speed value is obtained by dividing the number of characters by the preset time length. The time intervals between adjacent keys in the original typing input data are calculated to obtain the input interval time series; the standard deviation of the interval time is calculated according to the input interval time series, and the coefficient of variation is obtained by dividing the standard deviation by the average interval time; the coefficient of variation is used as an indicator of rhythm stability. The input interval time series is statistically analyzed to calculate the average value of all interval times; if an interval time exceeds the preset multiple of the average value, the interval time is defined as a pause; the frequency of pauses of different lengths is counted to generate the pause time distribution characteristics. The number of times each key position is used in the original typing input data is counted, and the key position usage frequency is obtained by dividing the number of times used by the total number of key presses; the longest time period of continuous input without pause in the original typing input data is recorded as the maximum duration of continuous input. The basic typing speed value, rhythm stability index, pause time distribution characteristics, key usage frequency, and maximum duration of continuous input are numerically processed to obtain a feature vector of fixed length; the feature vector is used to characterize typing ability and typing pattern.

[0031] Specifically, based on the original data of students' typing input, the sliding window method with a window size of 60 seconds and a step length of 10 seconds was used to calculate the number of characters in each time window, and the basic typing speed was obtained by dividing it by the window length. At the same time, the speed value of each window was recorded to construct a typing speed sequence. The input interval time series was obtained by calculating the time interval between adjacent keys, and the standard deviation of the interval time was calculated, and then the coefficient of variation was obtained by dividing it by the average interval time as an indicator of rhythm stability. The input interval time series was statistically analyzed, the average value of all interval times was calculated, and the intervals exceeding 1.5 times the average value were defined as pauses. The frequency of pauses of different lengths was counted to generate the pause time distribution characteristics, and the speed difference between adjacent time windows was calculated to obtain the input speed change curve. The number of times each key position was used was counted and divided by the total number of key presses to obtain the key position usage frequency. The longest time period of continuous input without pause was recorded as the maximum duration of continuous input. The basic typing speed, rhythm stability index, pause time distribution characteristics, input speed change curve, key usage frequency and maximum duration of continuous input are digitized to form a fixed-length vector, and a complete multi-dimensional feature vector is constructed to characterize the overall typing ability and typing pattern of students. In practical applications, students' 10-minute typing input data is first collected, including the timestamp and key type of each key. A total of 55 time windows are obtained using a sliding window method with a window size of 60 seconds and a step size of 10 seconds. The number of characters is calculated for each window. If 300 characters are entered in a window, the basic typing speed of the window is 300 words per minute. The speed values ​​of all windows are recorded to form a typing speed sequence. The time intervals between adjacent keys are calculated to obtain about 3,000 interval values. The standard deviation of these interval values ​​is calculated. Assuming it is 0.2 seconds and the average interval time is 0.5 seconds, the coefficient of variation is 0.4, which is used as a rhythm stability indicator. All interval times are counted, the average value is 0.5 seconds, and intervals exceeding 0.75 seconds are defined as pauses. Count the number of pauses, such as 50 short pauses (0.75-1.5 seconds), 30 medium pauses (1.5-3 seconds), and 10 long pauses (>3 seconds), to generate pause time distribution features. Calculate the speed difference between adjacent windows, such as from 280 words / minute to 300 words / minute, the difference is 20 words / minute, and get the input speed change curve. Count the number of times each key is used, such as the letter 'e' is used 200 times, and the total number of key presses is 3000 times, then the usage frequency of 'e' is 6.67%. Record the longest period of continuous input without pause, assuming it is 30 seconds. Finally, combine these features into a fixed-length vector, such as

[0032] [300, 0.4, 50, 30, 10, 20, 6.67, 30], constitute a multi-dimensional feature vector, which comprehensively characterizes students’ typing ability and pattern.

[0033] S103. Cluster the multi-dimensional feature vectors of students' typing input rhythm using a density-based clustering model. The obtained clustering results are used to reflect the typical input patterns of students at different levels of conceptual understanding. The central feature vector of each cluster is analyzed to determine its corresponding conceptual understanding level.

[0034] The multi-dimensional feature vector of students' typing input rhythm is obtained, and the feature vector is Z-score standardized to obtain the standardized feature vector; according to the standardized feature vector, the DBSCAN algorithm is used for density clustering, and the optimal neighborhood radius and minimum sample number parameters are determined by the grid search method; if the silhouette coefficient of the density clustering is less than the preset threshold, the clustering parameters are adjusted to re-cluster until the clustering quality requirements are met; the standardized feature vector is reduced to two-dimensional space to obtain a visualized clustering result; for the visualized clustering result, the proportion of concepts of different difficulty levels in each cluster is calculated, and the difficulty level with the highest proportion is taken as the concept understanding degree corresponding to the cluster; according to the cluster center feature vector and the concept understanding degree, a mapping relationship between the feature vector and the concept understanding degree is established.

[0035] Specifically, the multi-dimensional feature vectors of students' typing rhythm were Z-score standardized, the mean and standard deviation of each feature were calculated, the original value was subtracted from the mean and then divided by the standard deviation, and each feature was converted to a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the dimensional differences between different features. The DBSCAN algorithm was used to perform density clustering on the standardized feature vectors, and the optimal neighborhood radius and minimum sample number parameters were determined by the grid search method. The neighborhood radius ranged from 0.1 to 1.0, with a step size of 0.1, and the minimum sample number ranged from 3 to 10, with a step size of 1. The clustering results and the clustering labels of each data point were obtained. The central feature vector of each cluster was calculated, and the silhouette coefficient was used to evaluate the clustering effect. The silhouette coefficient threshold was set to 0.6. If it was less than the threshold, the clustering parameters were adjusted and clustering was performed again until the clustering quality requirements were met. The t-SNE algorithm was used to reduce the high-dimensional features to 2 dimensions for visualization and present the clustering results. The feature distribution of each cluster is analyzed, the mean and standard deviation of each feature in the cluster are calculated, and the features with significant discrimination for the cluster are identified. Through expert evaluation or comprehensive judgment based on students' answering time and accuracy, pre-labeled concept difficulty samples are obtained, the proportion of concepts of different difficulty levels in each cluster is calculated, and the difficulty level with the highest proportion is used as the concept understanding degree corresponding to the cluster. The decision tree algorithm is used to construct the mapping relationship between the cluster center feature vector and the concept understanding degree. In practical applications, the multi-dimensional feature vectors of the typing input rhythm collected from 100 students are first Z-score standardized. For example, for the typing speed feature, the original mean is 50 words / minute and the standard deviation is 10 words / minute, then the original value of 80 words / minute is converted to a standardized value of 3. Similar operations are performed on all features to obtain standardized feature vectors. Next, the DBSCAN algorithm is used for density clustering, and the optimal parameters are determined by grid search, such as a neighborhood radius of 0.5 and a minimum sample number of 5, to obtain 4 clusters. The silhouette coefficient is calculated to be 0.65, which is higher than the threshold of 0.6, and the clustering result is accepted. The t-SNE algorithm is used to reduce high-dimensional features to 2 dimensions, and a scatter plot is generated to intuitively display the clustering effect. The feature distribution of each cluster is analyzed. For example, the average typing speed of cluster 1 is 70 words / minute, with a standard deviation of 5 words / minute, while the average pause frequency is 2 times per minute and the standard deviation is 0.5 times. Typing speed and pause frequency are identified as the significant features of this cluster. According to the pre-labeled concept difficulty samples, the proportion of concepts of different difficulty levels in each cluster is calculated. For example, in cluster 1, simple concepts account for 60%, medium concepts account for 30%, and complex concepts account for 10%. The corresponding simple concept understanding degree of this cluster is determined. Finally, the decision tree algorithm is used, with the cluster center feature vector as input and the concept understanding degree as output, to construct a mapping relationship, such as the feature vector [70, 2, 0.1, 0.05] (indicating typing speed, pause frequency, error rate, and modification rate) is mapped to "simple concept understanding degree", thereby realizing the prediction of the concept understanding degree of new samples.

[0036] When analyzing the central feature vector of each cluster, analyze the numerical size of the cluster center in the dimensions of typing speed and rhythm stability, and evaluate the overall typing ability level of the student group represented by the cluster; analyze the characteristics of the cluster center in the dimensions of input pause time distribution and speed change, and combine the complexity of the corresponding concept to evaluate the activeness of thinking and the depth of thinking of the student group in the cluster when understanding the concept, so as to make an overall judgment on the degree of understanding of the concept corresponding to the cluster.

[0037] The characteristic vector of the cluster center is obtained, and the typing speed value and the rhythm stability value are extracted from the characteristic vector; the typing speed level is determined by comparing the typing speed value with the preset speed threshold; the stability level is determined by comparing the rhythm stability value with the preset stability threshold; the characteristics of the cluster center in the input pause time distribution dimension are analyzed to obtain the short pause ratio, medium pause ratio and long pause ratio; the thinking activity index is constructed by using the short pause ratio, medium pause ratio, long pause ratio and the standard deviation of speed change; the concept complexity score corresponding to the cluster is obtained from the pre-established concept complexity database; the thinking depth score is calculated by multiplying the concept complexity score and the thinking activity index; the decision tree model is constructed by using the typing speed level, stability level, thinking activity index and thinking depth score; if the maximum depth of the decision tree model is the preset value and the minimum number of samples is the preset threshold, the decision tree model is determined to be a concept understanding degree judgment model; the cluster center characteristic vector is input into the concept understanding degree judgment model to obtain the concept understanding degree category corresponding to the cluster; according to the concept understanding degree category, the overall level of the concept understanding degree of the cluster is judged.

[0038] Specifically, based on the cluster center feature vector, the values ​​of the two key dimensions of typing speed and rhythm stability are extracted, and the ability level threshold is set. The typing speed is divided into low speed less than 40 words / minute, medium speed 40-60 words / minute, and high speed greater than 60 words / minute. The rhythm stability is represented by the coefficient of variation and divided into low stability greater than 0.5, medium stability 0.3-0.5, and high stability less than 0.3. The values ​​of the cluster center in these two dimensions are compared with the threshold to obtain the overall typing ability level evaluation results of the student group represented by the cluster. The characteristics of the cluster center in the input pause time distribution dimension are analyzed, and the pause time is divided into short pauses less than 2 seconds, medium pauses 2-5 seconds, and long pauses greater than 5 seconds. The proportion of each type of pause is calculated, and the standard deviation of the speed change dimension is combined to construct a thinking activity index, and the degree of thinking activity is quantified through the set scoring rules. From the pre-established concept complexity database, the complexity score of the concept corresponding to the current cluster is obtained. The database uses expert evaluation or text analysis technology to score the complexity of the concept from 1 to 10 points. Combined with the thinking activity index, the product of the complexity score and the thinking activity index is used to calculate the thinking depth score to evaluate the thinking depth level of the clustered student group. The overall typing ability level, thinking activity and thinking depth score are combined to build a concept understanding judgment model using the decision tree algorithm. These three indicators are used as features, the concept understanding degree is used as the target variable, and the information gain is used as the splitting criterion. The maximum depth is set to 4 and the minimum number of samples is set to 10 to build a decision tree model. The cluster center features are mapped to the corresponding concept understanding degree categories to obtain the overall judgment result of the concept understanding degree corresponding to the cluster. In practical applications, the cluster center feature vector is first analyzed. For example, the typing speed of a cluster center is 55 words / minute, and the coefficient of variation is 0.35. According to the preset threshold, it is determined that the cluster represents medium-speed typing ability and medium rhythm stability. Next, we analyzed the distribution of pause time and found that short pauses less than 2 seconds accounted for 60%, medium pauses of 2-5 seconds accounted for 30%, and long pauses greater than 5 seconds accounted for 10%. The standard deviation of speed change was 5 words / minute.

[0039] =(0.6*1+0.3*0.6+0.1*0.2)*(1-5 / 60)=0.136. Query the complexity score of the current concept from the concept complexity database, assuming it is 7 points, with a full score of 10 points. Calculate the thinking depth score to be 0.54*7=3.78. Combine the three indicators of medium-speed typing ability, medium stability, thinking activity 0.54 and thinking depth 3.78, and input them into the decision tree model. The decision tree uses information gain as the splitting criterion, with a maximum depth of 4 and a minimum number of samples of 10 to construct a judgment model. It is assumed that the model first splits based on the depth of thinking, with a threshold of 3.5, then splits based on typing speed, with a threshold of 50 words / minute, and finally splits based on thinking activity, with a threshold of 0.5. According to the characteristics of the cluster, the model determines that its concept understanding level is "good."

[0040] S104. When students are inputting concept explanations, their typing input rhythm features are extracted in real time, and the feature vectors are input into a clustering model for classification to obtain the cluster to which the student's current typing input rhythm features belong, thereby determining the student's current corresponding concept understanding level.

[0041] Receive typing input data carrying a student identifier, the typing input data is collected at fixed time intervals when the student inputs a concept explanation; extract data in a specific time window from a temporary buffer according to the typing input data, the specific time window is determined by a preset feature calculation window size; use the data in the specific time window to calculate the typing speed, rhythm stability, and pause time distribution to obtain a typing input rhythm feature vector at the current moment; standardize the typing input rhythm feature vector to obtain a standardized feature vector; input the standardized feature vector into a pre-trained K nearest neighbor algorithm, and determine the cluster to which the current feature vector belongs by calculating the Euclidean distance between the feature vector and each cluster center; if the distance between the feature vector and the current cluster center exceeds a preset threshold, trigger a reclassification operation; and judge the student's current concept understanding level based on a mapping relationship between clusters and pre-established concept understanding levels.

[0042] Specifically, when students input concept explanations, a fixed time interval data collection method is used to record students' typing input data every 1 second, including key events, timestamps, and input content, and store these raw data in a temporary buffer. According to the preset feature calculation window size, such as 60 seconds, the data of the last 60 seconds is extracted from the temporary buffer to calculate features such as typing speed, rhythm stability, and pause time distribution, and construct the typing input rhythm feature vector at the current moment; when the data is less than 60 seconds, the features are calculated using the existing data and marked as incomplete samples. The constructed feature vector is standardized, and the same standardization parameters as those used in training the clustering model are used to ensure that the feature vector is consistent with the model input format. Then, the standardized feature vector is input into the pre-trained K-nearest neighbor algorithm with K=5 for real-time classification. By calculating the Euclidean distance between the feature vector and each cluster center, the cluster to which the current feature vector belongs is determined, and an update threshold is set. When the distance between the feature vector and the current cluster center exceeds the threshold, the operation of reclassification and updating the learning status is triggered. According to the mapping relationship between clustering and concept understanding level, the student's current concept understanding level is judged, and the result is updated to the student's learning status in real time. The classification results and the results of the concept understanding level judgment are stored in the time series database for subsequent learning trajectory analysis and visualization. In actual applications, the system collects students' typing input data every second, such as recording 540 keystroke events in a 60-second window. When calculating the features, the average typing speed is 54 words / minute, the rhythm stability (expressed by the coefficient of variation) is 0.25, and the pause time distribution is short pauses <1 second accounting for 70%, medium pauses 1-3 seconds accounting for 25%, and long pauses >3 seconds accounting for 5%. These features are combined into a vector [54, 0.25, 0.7, 0.25, 0.05], and then Z-score normalization is performed. It is assumed that the normalized vector [0.8, -0.5, 1.2, -0.3, -1.0] is obtained. The K-nearest neighbor algorithm (K=5) is used to classify the normalized vector and calculate the Euclidean distance to the pre-trained cluster center. Assume that the distances of the five nearest cluster centers are [0.5, 0.7, 0.8, 1.1, 1.3], and the corresponding conceptual understanding levels are [high, high, medium, medium, low]. According to the majority voting principle, the current feature vector is classified as a "high" conceptual understanding level. Set the update threshold to 0.6. Since the distance of the nearest cluster center is 0.5 < 0.6, there is no need to trigger reclassification. Store the result {"timestamp":"2023-04-0110:15:30","typ i ng_speed":54,"rhythm_stab il ity":0.25,"understanding i ng_level":"high"} in the time series database for subsequent analysis.

[0043] S105. Dynamically adjust the course difficulty gradient and progress schedule based on students’ current level of conceptual understanding. If students have a high level of understanding, increase the course difficulty and learning progress. If students have a low level of understanding, reduce the course difficulty and strengthen targeted practice and feedback.

[0044] The real-time typing input rhythm of students is obtained, and the conceptual understanding level of students is obtained based on the input rhythm analysis; the difficulty level that matches the conceptual understanding level is selected from the pre-established course difficulty gradient library. The sliding window method is used to calculate the weighted average of students' recent conceptual understanding level; the comprehensive understanding level score of students is obtained through the weighted average. According to the comprehensive understanding level score, the course difficulty and learning progress are adjusted by a fuzzy logic controller; wherein the fuzzy logic controller includes input variables and output variables, the input variable is the comprehensive understanding level score, and the output variable is the difficulty adjustment range and the progress adjustment range. According to the adjusted course difficulty and learning progress, the exercises that meet the current difficulty level are screened from the preset exercise question bank; the exercise question bank contains exercises of multiple difficulty levels. Determine whether the student's performance in the adjusted course reaches the preset threshold; if the student's performance fails to reach the preset threshold for multiple consecutive times, manual intervention is triggered.

[0045] Specifically, according to the concept understanding level obtained by analyzing the real-time typing input rhythm of students, the appropriate difficulty level is selected from the pre-established course difficulty gradient library. The difficulty gradient library contains multiple difficulty levels, each corresponding to different concept complexity and learning objectives. The nearest neighbor algorithm is used to compare the current understanding level with the preset difficulty level standard and select the closest difficulty level. The sliding window method is used to calculate the weighted average of the student's concept understanding level of the last 10 times, and the student's comprehensive understanding level score is obtained using a linear decreasing weight. Based on the comprehensive understanding level score, the course difficulty and learning progress are adjusted using a fuzzy logic controller. The input variables, including the comprehensive understanding level score and the output variables, including the difficulty adjustment range and the progress adjustment range, are defined, and fuzzy rules are set, such as "if the score is high, the difficulty and progress are greatly increased." According to the adjusted course difficulty and learning progress, the exercises that meet the current difficulty level are screened from the exercise question bank, and an error type-question type mapping table is established to select the corresponding question type according to the student's error type; the natural language generation technology is used to generate personalized feedback text based on the error type and frequency, and these exercises and feedback content are pushed to the students. The student's performance in the adjusted course is recorded. If the performance does not improve after multiple consecutive adjustments, the manual intervention mechanism is triggered. In actual application, the system first analyzes the student's real-time typing input rhythm and obtains a conceptual understanding level of 0.75, with a full score of 1. This level is compared with the preset difficulty level standard using the nearest neighbor algorithm.

[0046] {0.2, 0.4, 0.6, 0.8, 1.0}, and select the closest difficulty level of 0.8. Then, a sliding window of 10 samples is used to calculate the concept understanding level of the last 10 times.

[0047] {0.7, 0.72, 0.68, 0.75, 0.73, 0.71, 0.74, 0.76, 0.75, 0.75} applies linear decreasing weights {1, 0.9, 0.8, ..., 0.1}, and the comprehensive understanding level score is calculated to be 0.736. This score is input into the fuzzy logic controller, which defines the fuzzy set of the input variable "understanding level" {low, medium, high} and the fuzzy set of the output variables "difficulty adjustment" and "progress adjustment" {decrease, maintain, increase}. Applying fuzzy rules such as "if the understanding level is high, the difficulty increases and the progress increases", the difficulty adjustment suggestion is obtained: increase by 10%, and the progress adjustment suggestion is: speed up by 5%. Based on these adjustments, the system selects questions of the new difficulty level from the practice question bank, and at the same time selects the corresponding "concept analysis" type questions from the error type-question type mapping table according to the student's recent error type, such as the frequency of "concept confusion" of 60%. Using natural language generation technology, feedback text is generated, such as "You need to improve your ability to distinguish between concepts X and Y. We recommend that you focus on their different application scenarios." The system continuously monitors students' performance in new courses of varying difficulty. If the level of understanding drops three times in a row, the manual intervention mechanism is triggered.

[0048] S106. Continuously collect typing input rhythm data during the students' learning process, regularly use new data to incrementally train and tune the clustering model, and dynamically update the concept understanding level judgment criteria for each cluster to adapt to changes in students' typing habits and improvements in concept understanding ability.

[0049] Acquire students' typing input rhythm data, which includes key interval time, typing speed and error rate; calculate typing feature indicators based on typing input rhythm data; execute Mini-Batch K-Means algorithm using typing feature indicators to obtain an updated clustering model; if the clustering model changes, recalculate the average concept understanding degree of each cluster; use exponential smoothing method to process historical data of typing feature indicators to obtain long-term change trend of students' typing habits; predict future typing feature changes based on long-term change trend, and use future typing feature changes to fine-tune the parameters of the clustering model.

[0050] Specifically, during the learning process, a real-time data collection mechanism is used to collect students' typing rhythm data every 1 minute, including key interval time, typing speed, error rate and other features, and store these raw data in a distributed database. The newly collected data is preprocessed and feature extracted. Two sliding windows are used: a short-term window of 1 hour and a long-term window of 24 hours. Indicators such as average typing speed, rhythm stability and error rate are calculated respectively to construct new feature vectors. The Mini-Batch K-Means algorithm is used to incrementally update the existing clustering model, input the new feature vector into the model, dynamically adjust the cluster center and boundary, and calculate the model's silhouette coefficient to evaluate the clustering quality. According to the updated clustering results, the average concept understanding of each cluster is recalculated. The weighted average method is used to comprehensively consider the average concept understanding of the cluster, the student's recent test scores and learning progress, such as the percentage of completed courses, and dynamically adjust the mapping relationship between clustering and concept understanding, and update the judgment criteria. The exponential smoothing method is used to analyze the long-term trend of students' typing habits and predict future changes in typing characteristics. Based on this, the model parameters are fine-tuned to adapt to the evolution of students' typing habits and the improvement of their conceptual understanding. In practical applications, students' typing data is collected every minute. For example, 60 keystrokes are recorded in a minute, with an average interval of 1 second, a typing speed of 60 words per minute, and an error rate of 5%. These data are stored in a distributed database in real time. At the same time, two sliding windows are maintained. The 1-hour window calculates an average typing speed of 58 words per minute, a rhythm stability (expressed by the coefficient of variation) of 0.15, and an error rate of 4.8%; the 24-hour window calculates an average typing speed of 55 words per minute, a rhythm stability of 0.18, and an error rate of 5.2%. These features form a new vector [58, 0.15, 0.048, 55, 0.18, 0.052]. The Mini-Batch K-Means algorithm is used to process 100 new samples each time and update the cluster center. Assume that after the update, 5 clusters are obtained, and the calculated silhouette coefficient is 0.68, indicating that the clustering quality is good. For a certain updated cluster, its average concept understanding is 0.75, with a full score of 1. Combined with the student's recent test score of 85 points, a full score of 100, and a learning progress of 60%, the weighted average is used [0.5, 0.3, 0.2] to obtain an adjusted concept understanding of 0.78. Finally, the exponential smoothing method is used with a smoothing coefficient of α = 0.2 to analyze the changes in the student's typing speed within 24 hours, and the typing speed in the next hour is predicted to be 62 words / minute. Based on this, the speed threshold in the model is raised from 60 words / minute to 62 words / minute.

[0051] S107. Combine students’ typing input rhythm characteristics with other learning behavior data, and build a comprehensive learning effect evaluation model through multimodal data fusion to evaluate students’ mastery of different knowledge points and identify students’ weak links and areas of strength.

[0052] Obtain various types of learning behavior data from the student learning system, including typing input rhythm characteristics, question completion status, online learning time and material browsing history; according to the type of learning behavior data, adopt the corresponding feature extraction method to obtain learning behavior characteristics; standardize the learning behavior characteristics, and select features with correlation coefficients greater than a preset threshold; weightedly fuse the selected features through the attention mechanism to determine the contribution of various types of data to the learning effect evaluation; use the random forest algorithm to construct a comprehensive learning effect evaluation model, the evaluation model uses the fused features as input, the student's test scores on each knowledge point as labels, and the learning effect evaluation results as output; if the feature importance is greater than the first preset threshold and the prediction accuracy is lower than the second preset threshold, it is determined to be a weak link; if the feature importance is greater than the first preset threshold and the prediction accuracy is higher than the third preset threshold, it is determined to be an advantageous area.

[0053] Specifically, various types of learning behavior data are extracted from the student learning system, including typing input rhythm features, question completion status, online learning time, data browsing records, etc. These raw data are cleaned and preprocessed, and the box plot method is used to detect outliers, and the mean filling method is used to handle missing values. For different types of data, corresponding feature extraction methods are used, such as using the sliding window method to extract statistical features for typing input rhythm data, using the accuracy rate and completion time as features for question completion, and using the autoregressive moving average model to analyze trends and periodicity for online learning time and data browsing records. The extracted features are standardized, and the Z-score standardization method is used to convert features of different dimensions to the same scale. Then, Lasso regression is used for feature selection, and features with correlation coefficients greater than 0.5 are selected. The attention mechanism is used to weightedly fuse different types of features to learn the contribution of various types of data to the evaluation of learning effects. A comprehensive learning effect evaluation model is constructed using the random forest algorithm. The fused features are used as input, and the test scores of students on each knowledge point are used as labels for training. The factors that have a greater impact on the learning effect are identified through feature importance analysis. Thresholds are set, and knowledge points with feature importance greater than 0.1 and prediction accuracy lower than 60% are judged as weak links, and knowledge points with feature importance greater than 0.1 and prediction accuracy higher than 80% are judged as advantageous areas. In practical applications, multi-dimensional data of students are first collected from the learning platform, such as an average typing speed of 60 words per minute, a correct question rate of 85%, an average daily online learning time of 3 hours, and 20 learning materials browsed per week. The box plot method is used to detect outliers, and data exceeding 1.5 times the interquartile range of the upper quartile is marked as abnormal, such as an outlier of 120 words per minute. For missing online learning time data, the average value of the student in the last 7 days is used to fill. Then, a 30-minute sliding window is used to calculate the typing rhythm features to obtain statistics such as average speed and standard deviation. The ARIMA (1,1,1) model is applied to the online learning time to predict the learning trend in the next week. After standardizing all features, 10 key features are selected through Lasso regression. The attention mechanism is used to weight these features, and the weights are obtained, such as typing speed 0.2, question accuracy 0.3, learning time 0.25, and material browsing 0.25. The weighted features are input into the random forest model, where the number of trees is 100 and the maximum depth is 10, and the mastery prediction model of each knowledge point is obtained through training. By analyzing the feature importance and prediction accuracy, it is found that the feature importance of the "data structure" knowledge point is 0.15, and the prediction accuracy is only 55%, which is determined to be a weak link; while the feature importance of the "algorithm design" knowledge point is 0.18, and the prediction accuracy is 85%, which is determined to be an advantage area.

[0054] S108. Based on the evaluation results of the learning effect evaluation model, the content organization, presentation method, and interaction design of the course are optimized in a targeted manner to provide personalized learning paths and content recommendations for students with different understanding levels, and dynamically adjust the difficulty distribution of learning resources to match the overall cognitive ability development level of students.

[0055] According to the output results obtained by the learning effect evaluation model, the course knowledge points are clustered by the hierarchical clustering algorithm, and the knowledge points with high correlation are combined to obtain knowledge units to construct a knowledge graph. The topological sorting algorithm obtains the priority of students' weak links and determines the presentation order of knowledge units based on the priority. The similarity between students is calculated by the collaborative filtering algorithm, and a personalized learning resource recommendation list is generated based on the similarity. The recommendation list contains multiple types of learning resources such as videos, texts and exercises. The deep reinforcement learning algorithm constructs an adaptive learning path generator, in which the state space includes the students' current knowledge mastery and learning progress, the action space contains learning resources of different difficulty and types, and the reward function is based on the learning effect evaluation results. The students' cognitive ability development level and learning progress are obtained, and the difficulty distribution of the entire course learning resources is globally optimized.

[0056] Specifically, according to the output results of the learning effect evaluation model, the course knowledge points are clustered and analyzed. The hierarchical clustering algorithm is used to combine the highly related knowledge points into knowledge units to construct a knowledge graph. The topological sorting algorithm is used to adjust the presentation order of the knowledge units according to the priority of the students' weak links. The user-based collaborative filtering algorithm is used to calculate the similarity between students and generate a personalized learning resource recommendation list for each student. The recommendation list contains various types of learning resources such as videos, texts, and exercises. The adaptive learning path generator is constructed through the deep reinforcement learning algorithm. The state space is defined to include the student's current knowledge mastery, learning progress, etc. The action space contains learning resources of different difficulties and types. The reward function is based on the learning effect evaluation results to train an intelligent agent that can dynamically adjust the learning path according to the student's real-time learning status. The course is divided into multiple stages, each stage corresponds to a different difficulty level. The forward dynamic programming algorithm is used to globally optimize the difficulty distribution of the learning resources of the entire course according to the students' cognitive ability development level and learning progress, ensuring that the difficulty curve can continuously challenge students without causing excessive cognitive load. Based on the students' operation behavior data, the A / B test method is used to evaluate the effects of different interaction designs, select the optimal interaction method, and optimize the user interface and learning experience. In actual applications, the system first hierarchically clusters 100 course knowledge points, sets the clustering threshold to 0.7, and obtains 15 knowledge units. Then, the topological sorting algorithm is used to sort the students' mastery of each knowledge point from 1 to 10 points, such as data structure 4 points before algorithm design 7 points. Next, the user-based collaborative filtering algorithm is used to calculate the cosine similarity between students and recommend the top 5 similarity resources that students like for the target students. For example, 4 videos, 3 articles, and 10 exercises are recommended. In the deep reinforcement learning algorithm, the state space contains 20 features, such as the mastery of each knowledge point, the learning time, etc., and the action space contains 50 learning resources of different difficulty and type. The reward function is designed based on the improvement of the learning effect, such as 1 point for a 0.1 increase in mastery. After 10,000 rounds of training, the intelligent agent can select the optimal learning path according to the student's state. The course is divided into 10 stages, with each stage ranging from 1 to 10 in difficulty. The optimal difficulty distribution is calculated using a forward dynamic programming algorithm, such as [2,3,3,4,5,6,6,7,8,9]. Finally, the two interactive designs are compared through an A / B test, with 500 students randomly selected for each design. The test period is 2 weeks, and the average learning time and knowledge point mastery are compared to select the design with better performance.

[0057] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for adjusting AI intelligent interactive courses based on a large language model, characterized in that: The method comprises: Obtaining students' original typing input rhythm data during the process of explaining concepts of different difficulty levels, including keyboard event timestamp sequence, event type, and input content, and calculating pause time distribution, input speed change, and error rate indicators based on the event timestamp sequence; Preprocess the original typing input rhythm data to remove abnormal pauses caused by factors such as keyboard layout familiarity and finger flexibility differences, and construct a multi-dimensional feature vector of students' typing input rhythm based on the students' basic typing speed and rhythm stability characteristics; The multi-dimensional feature vectors of students' typing input rhythm are clustered using a density-based clustering model. The clustering results are used to reflect the typical input patterns of students at different levels of conceptual understanding. The central feature vector of each cluster is analyzed to determine its corresponding conceptual understanding level. When students are inputting concept explanations, their typing input rhythm features are extracted in real time, and the feature vectors are input into the clustering model for classification, so as to obtain the cluster to which the student's current typing input rhythm features belong, and then judge the student's current corresponding concept understanding level; Dynamically adjust the course difficulty gradient and progress schedule based on students' current level of understanding of concepts. If students have a high level of understanding, increase the course difficulty and learning progress. If students have a low level of understanding, reduce the course difficulty and strengthen targeted practice and feedback. Continuously collect typing input rhythm data during the learning process, regularly use new data to incrementally train and tune the clustering model, and dynamically update the concept understanding judgment criteria for each cluster to adapt to changes in students' typing habits and improvements in their concept understanding ability; Combine students' typing input rhythm characteristics with other learning behavior data, and build a comprehensive learning effect evaluation model through multimodal data fusion to evaluate students' mastery of different knowledge points and identify students' weaknesses and strengths. Based on the evaluation results of the learning effect evaluation model, the content organization, presentation method, and interaction design of the course are optimized to provide personalized learning paths and content recommendations for students with different levels of understanding, and dynamically adjust the difficulty distribution of learning resources to match the students' overall cognitive ability development level.

2. The method according to claim 1, characterized in that The method of obtaining the original typing input rhythm data of students in the process of explaining concepts of different difficulty levels, including keyboard event timestamp sequence, event type, and input content, and calculating the pause time distribution, input speed change, and error rate indicators according to the event timestamp sequence, includes: Obtain the time stamp sequence of keyboard events and input content of students during the concept explanation process; Adopting the adaptive threshold method to calculate the time interval between adjacent key presses, if the time interval exceeds the preset multiple of the average key press interval of the students, it is identified as a pause point, and the pause time distribution data is obtained; The sliding window method is used to calculate the number of characters input in each time window to obtain discrete speed values, which are then smoothed using the moving average method to generate an input speed change curve. Extract the usage of the backspace and delete keys from the keyboard event sequence, record the position and number of backspace and delete operations, calculate the error rate index based on the input content length, and count the frequency of multiple consecutive backspace or delete operations to obtain the modification frequency index; Text matching of student input is performed based on a predefined list of difficulty level keywords to determine the difficulty level of the concept currently being explained; Based on the characteristics of pause time distribution, input speed change, error rate and modification frequency, the K-means clustering algorithm is used to classify students' input behaviors, and combined with the concept difficulty level, the students' understanding of concepts of different difficulty levels is evaluated.

3. The method according to claim 1, characterized in that The original typing input rhythm data is preprocessed to eliminate abnormal pauses caused by factors such as keyboard layout familiarity and finger flexibility differences, and a multi-dimensional feature vector of the student's typing input rhythm is constructed in combination with the student's basic typing speed and rhythm stability characteristics, including: According to the timestamp sequence in the original typing input rhythm data, the key frequency distribution in each window is counted by using the sliding window method to obtain the basic typing speed and rhythm stability characteristics; The average input speed and standard deviation of each key area are calculated. If the time interval exceeds the range of the standard deviation of the average value plus or minus a preset multiple, it is marked as an outlier to identify abnormal pauses caused by familiarity with the keyboard layout. By calculating the speed change rate during continuous input, the mean and variance of the speed change rate are obtained as an index of finger flexibility; For the time intervals marked as abnormal, linear interpolation is used to fill them and generate the cleaned typing input rhythm data; Based on the cleaned typing input rhythm data, the basic typing speed, rhythm stability, keyboard layout familiarity, finger flexibility index and error correction frequency characteristics are standardized, the characteristic covariance matrix is ​​calculated, the eigenvalues ​​and eigenvectors are solved, and the first few principal components whose cumulative contribution rate reaches the preset threshold are selected to construct a multi-dimensional feature vector of the student's typing input rhythm; It also includes: when constructing a multi-dimensional feature vector of the student's typing input rhythm, calculating the student's basic typing speed as a basic indicator for measuring the student's overall typing ability; quantifying the stability of the student's typing rhythm by calculating the variance or coefficient of variation of the input interval time; combining the basic typing speed and rhythm stability indicators with other features, including pause time distribution and input speed changes, to form a complete multi-dimensional feature vector, specifically including: according to the original typing input data, using the sliding window method to obtain the number of characters in the time window, the window size of the sliding window is a preset time length, and the step length of the sliding window is a preset step length; the basic typing speed value is obtained by dividing the number of characters by the preset time length; the time interval between adjacent keys in the original typing input data is calculated to obtain the input interval time series; the standard deviation of the interval time is calculated according to the input interval time series , divide the standard deviation by the average interval time to get the coefficient of variation; use the coefficient of variation as an indicator of rhythm stability; perform statistical analysis on the input interval time series and calculate the average of all interval times; if an interval time exceeds a preset multiple of the average value, define the interval time as a pause; count the frequency of pauses of different lengths to generate pause time distribution characteristics; count the number of times each key position is used in the original typing input data, and divide the number of times used by the total number of key presses to get the key position usage frequency; record the longest time period of continuous input without pause in the original typing input data as the maximum duration of continuous input; digitize the basic typing speed value, rhythm stability indicator, pause time distribution characteristics, key position usage frequency, and maximum duration of continuous input to get a feature vector of fixed length; the feature vector is used to characterize typing ability and typing mode.

4. The method according to claim 1, characterized in that: The density-based clustering model is used to cluster the multi-dimensional feature vectors of students' typing input rhythm. The obtained clustering results are used to reflect the typical input patterns of students at different levels of concept understanding, and the central feature vector of each cluster is analyzed to determine its corresponding concept understanding level, including: Obtain a multi-dimensional feature vector of the student's typing input rhythm, perform Z-score normalization on the feature vector, and obtain a normalized feature vector; According to the standardized feature vector, the DBSCAN algorithm is used for density clustering, and the optimal neighborhood radius and minimum sample number parameters are determined by the grid search method; If the silhouette coefficient of density clustering is less than the preset threshold, the clustering parameters are adjusted and clustering is performed again until the clustering quality requirements are met; Reduce the dimension of the standardized feature vector to two-dimensional space to obtain a visual clustering result; For the visualized clustering results, calculate the proportion of concepts of different difficulty levels in each cluster, and take the difficulty level with the highest proportion as the concept understanding level corresponding to the cluster; According to the cluster center feature vector and the concept understanding degree, a mapping relationship between the feature vector and the concept understanding degree is established; It also includes: when analyzing the central feature vector of each cluster, analyzing the numerical size of the cluster center in the dimensions of typing speed and rhythm stability, and evaluating the overall typing ability level of the student group represented by the cluster; analyzing the characteristics of the cluster center in the dimensions of input pause time distribution and speed change, and combining the complexity of the corresponding concept, evaluating the active thinking and depth of thinking of the student group in the cluster when understanding the concept, so as to make an overall judgment on the degree of understanding of the concept corresponding to the cluster, specifically including: obtaining the cluster center feature vector, extracting the typing speed value and rhythm stability value from the feature vector; determining the typing speed level by comparing the typing speed value with the preset speed threshold; determining the stability level by comparing the rhythm stability value with the preset stability threshold; analyzing the cluster center in the dimension of input pause time distribution The features above are used to obtain the short pause ratio, medium pause ratio and long pause ratio; the short pause ratio, medium pause ratio, long pause ratio and speed change standard deviation are used to construct the thinking activity index; the concept complexity score corresponding to the cluster is obtained from the pre-established concept complexity database; the thinking depth score is calculated by multiplying the concept complexity score and the thinking activity index; the decision tree model is constructed using the typing speed level, stability level, thinking activity index and thinking depth score; if the maximum depth of the decision tree model is the preset value and the minimum number of samples is the preset threshold, the decision tree model is determined to be a concept understanding degree judgment model; the cluster center feature vector is input into the concept understanding degree judgment model to obtain the concept understanding degree category corresponding to the cluster; according to the concept understanding degree category, the overall level of the concept understanding degree of the cluster is judged.

5. The method according to claim 1, characterized in that When the student is inputting a concept explanation, the typing input rhythm feature is extracted in real time, and the feature vector is input into the clustering model for classification, so as to obtain the cluster to which the student's current typing input rhythm feature belongs, and then judge the student's current corresponding concept understanding level, including: receiving typed input data carrying a student identifier, the typed input data being collected at fixed time intervals when the student performs concept explanation input; Extracting data within a specific time window from a temporary buffer according to the typed input data, where the specific time window is determined by a preset feature calculation window size; The data in a specific time window is used to calculate the typing speed, rhythm stability and pause time distribution, and the typing input rhythm feature vector at the current moment is obtained; Performing normalization processing on the typing input rhythm feature vector to obtain a normalized feature vector; The standardized feature vector is input into the pre-trained K nearest neighbor algorithm, and the cluster to which the current feature vector belongs is determined by calculating the Euclidean distance between the feature vector and each cluster center; If the distance between the feature vector and the current cluster center exceeds the preset threshold, a reclassification operation is triggered; Based on the mapping relationship between clustering and pre-established concept understanding levels, the students’ current concept understanding level is judged.

6. The method according to claim 1, characterized in that The course difficulty gradient and progress schedule are dynamically adjusted according to the student's current level of understanding of the concepts. If the student's understanding level is high, the course difficulty and learning progress will be increased. If the student's understanding level is low, the course difficulty will be reduced, and targeted exercises and feedback will be strengthened, including: Obtain students' real-time typing input rhythm, and analyze students' concept understanding level based on input rhythm; Choose a difficulty level that matches your level of conceptual understanding from a library of pre-built course difficulty gradients; The sliding window method was used to calculate the weighted average of students’ recent conceptual understanding levels; The weighted average was used to obtain the student's comprehensive understanding level score; According to the comprehensive understanding level score, a fuzzy logic controller is used to adjust the course difficulty and learning progress; The fuzzy logic controller includes input variables and output variables. The input variable is the comprehensive understanding level score, and the output variable is the difficulty adjustment range and the progress adjustment range. Through the adjusted course difficulty and learning progress, select exercises that meet the current difficulty level from the preset exercise library; The practice question bank contains practice questions of multiple difficulty levels; Determine whether the student's performance in the adjusted course meets the preset threshold; If the student's performance fails to reach the preset threshold for multiple consecutive times, manual intervention will be triggered.

7. The method according to claim 1, characterized in that The method of continuously collecting typing input rhythm data during the student learning process, regularly using new data to incrementally train and tune the clustering model, and dynamically updating the concept understanding degree judgment standard of each cluster to adapt to the changes in students' typing habits and the improvement of their concept understanding ability, includes: Acquire students' typing input rhythm data, the typing input rhythm data including keystroke interval time, typing speed and error rate; Calculate typing feature indexes according to typing input rhythm data; The Mini-Batch K-Means algorithm is executed using the typing feature index to obtain the updated clustering model; If the clustering model changes, the average conceptual understanding of each cluster is recalculated; Exponential smoothing method is used to process the historical data of typing characteristic indicators to obtain the long-term trend of students' typing habits; Based on the long-term change trend, the future typing feature changes are predicted, and the future typing feature changes are used to fine-tune the parameters of the clustering model.

8. The method according to claim 1, characterized in that The typing input rhythm characteristics of students are combined with other learning behavior data to build a comprehensive learning effect evaluation model through multimodal data fusion, which is used to evaluate students' mastery of different knowledge points and identify students' weak links and strengths, including: Obtain various types of learning behavior data in the student learning system, including typing input rhythm characteristics, question completion status, online learning time, and material browsing records; According to the type of learning behavior data, the corresponding feature extraction method is used to obtain the learning behavior features; Standardize the learning behavior features and select the features whose correlation coefficient is greater than the preset threshold; The selected features are weighted and integrated through the attention mechanism to determine the contribution of various types of data to the evaluation of learning effects; A comprehensive learning effect evaluation model is constructed using the random forest algorithm. The evaluation model uses the fused features as input, the students' test scores on each knowledge point as labels, and the learning effect evaluation results as output. If the feature importance is greater than the first preset threshold and the prediction accuracy is lower than the second preset threshold, it is determined to be a weak link; If the feature importance is greater than the first preset threshold and the prediction accuracy is higher than the third preset threshold, it is determined to be an advantageous area.

9. The method according to claim 1, characterized in that: According to the evaluation results of the learning effect evaluation model, the content organization, presentation method, and interaction design of the course are optimized in a targeted manner, personalized learning paths and content recommendations are provided for students with different understanding levels, and the difficulty distribution of learning resources is dynamically adjusted to match the overall cognitive ability development level of students, including: According to the output results obtained from the learning effect evaluation model, a hierarchical clustering algorithm is used to perform cluster analysis on the course knowledge points, and knowledge units are obtained by combining highly relevant knowledge points to construct a knowledge graph; The topological sorting algorithm obtains the priority of students’ weak links and determines the order of knowledge unit presentation based on the priority; The similarity between students is calculated through collaborative filtering algorithm, and a personalized learning resource recommendation list is generated based on the similarity. The recommendation list includes multiple types of learning resources such as videos, texts, and exercises. The deep reinforcement learning algorithm constructs an adaptive learning path generator, where the state space includes the student's current knowledge mastery and learning progress, the action space contains learning resources of different difficulty and types, and the reward function is based on the learning effect evaluation results; Obtain students' cognitive ability development level and learning progress, and perform global optimization on the difficulty distribution of the entire course learning resources.

10. An AI intelligent interactive course adjustment system based on a large language model, characterized in that: The system comprises: Data collection and preprocessing module, used to obtain and process students' typing input rhythm data; Feature extraction and clustering module, used to construct feature vectors and perform cluster analysis; Real-time assessment and dynamic adjustment module to determine students’ understanding level and adjust course difficulty; Model optimization and update module, used to continuously optimize clustering models and evaluation criteria; Comprehensive evaluation and personalized recommendation module, used for multimodal data fusion evaluation and personalized learning recommendation.

Images