An interactive teaching quality continuous optimization method based on deep learning

Abstract

The application discloses a kind of based on deep learning interactive teaching quality continuous optimization method, it is related to deep learning teaching evaluation technical field, including the establishment of teaching whole cycle interaction data acquisition mechanism, capture teacher-student voice stream, screen operation trajectory, text exchange record, classroom behavior image sequence real-time interactive data, by multimodal fusion and time stamp alignment processing generation time series associated multi-dimensional teaching feature vector.Fusion deep learning evaluation model of bidirectional long short-term memory network and convolution attention module is constructed, the data deep layer time series dependence is extracted and focuses on key features, outputs classroom concentration atlas, knowledge point coverage index, interactive response heat map, and the real-time teaching quality evaluation score is obtained after normalization integration, after comparison with historical baseline, when score is low, optimization process is automatically triggered.The method realizes the complete quantitative representation of teaching interaction state and the accurate real-time evaluation, dynamic self-adaptive optimization of teaching quality.

Description

Technical Field

[0001] This invention belongs to the field of deep learning teaching evaluation technology, specifically a method for continuous optimization of interactive teaching quality based on deep learning. Background Technology

[0002] Current interactive teaching quality assessments often rely on manual classroom observation and single-data collection methods, gathering only discrete data such as classroom answers and attendance. They lack a comprehensive interactive data collection system covering the entire teaching cycle, failing to simultaneously capture four types of multi-source interactive data: teacher-student voice streams, screen operation trajectories, text communication records, and classroom behavior image sequences. Furthermore, they do not perform multimodal fusion and timestamp alignment processing on this multi-source data. Conventional teaching data processing methods can only generate independent, single-dimensional data indicators, unable to form multi-dimensional teaching feature vectors with temporal correlations, making it difficult to fully quantify the true state of classroom interaction within a given time period.

[0003] Conventional deep learning models used in teaching quality assessments often employ a single network structure, failing to integrate bidirectional long short-term memory networks with convolutional attention modules. This prevents them from simultaneously extracting deep temporal dependencies in teaching data and focusing on key features, resulting in outputs only a standardized teaching quality score. They cannot generate classroom attention maps, knowledge point coverage indices, or interactive response heatmaps. Figure 3 The existing technology cannot normalize and integrate multiple assessment results into a precise quantitative score, nor does it have a mechanism for comparison with historical teaching quality benchmarks. It also cannot automatically trigger optimization processes when teaching quality fails to meet standards. Therefore, it is necessary to realize the precise assessment application of multimodal data time-series fusion processing and composite deep learning models. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art;

[0005] To this end, this invention proposes a method for continuous optimization of interactive teaching quality based on deep learning, comprising:

[0006] Establish an interactive data collection mechanism covering the entire teaching cycle to continuously capture real-time interactive data, including teacher and student voice streams, screen operation trajectories, text communication records, and classroom behavior image sequences.

[0007] The real-time interactive data is subjected to multimodal fusion and timestamp alignment to generate a multidimensional teaching feature vector with temporal correlation. The multidimensional teaching feature vector is used to characterize the quantitative state of the classroom interaction process within a unit of time.

[0008] A deep learning evaluation model containing a bidirectional long short-term memory network and a convolutional attention module is constructed, and the multidimensional teaching feature vector is input into the deep learning evaluation model for calculation;

[0009] The deep learning evaluation model is used to extract deep temporal dependencies and focus on key features of the multidimensional teaching feature vectors, generating a classroom attention map, knowledge point coverage index and interactive response heatmap for the current teaching unit.

[0010] The classroom focus map, knowledge point coverage index, and interactive response heatmap are normalized and integrated to calculate the real-time teaching quality assessment score for the current teaching unit.

[0011] The real-time teaching quality assessment score is compared with the preset historical teaching quality score baseline to determine the quality status of the current teaching unit. When the real-time teaching quality assessment score is lower than the historical teaching quality score baseline, the teaching quality optimization process is triggered.

[0012] Furthermore, the real-time interactive data undergoes multimodal fusion and timestamp alignment processing to generate a multidimensional teaching feature vector with temporal correlation, including:

[0013] The teacher and student voice streams are processed by speech recognition and voiceprint segmentation to distinguish between teacher voice segments and student voice segments, and the voice content is converted into text records with time-series tags.

[0014] The screen operation trajectory is analyzed for coordinates, and the cursor movement speed, click frequency, and trigger time of specific teaching software interface elements are recorded.

[0015] The text communication records were segmented and analyzed for sentiment polarity to extract key question words, student answer texts and corresponding sentiment tendency tags.

[0016] The classroom behavior image sequence is processed by face detection and pose estimation to identify the students' facial orientation, head pose change frequency, and hand raising actions.

[0017] All time series labels obtained from the conversion of modal data are uniformly mapped onto the same time axis, and the aligned data is sliced ​​with a fixed time window;

[0018] Feature extraction and splicing operations are performed on the multimodal data slices within each time window to form the multidimensional teaching feature vector.

[0019] Furthermore, the construction of the deep learning evaluation model, which includes a bidirectional long short-term memory network and a convolutional attention module, includes:

[0020] A forward propagation layer and a backward propagation layer of a bidirectional long short-term memory network are established to extract forward and backward temporal context information from the multidimensional teaching feature vector;

[0021] The outputs of the forward propagation layer and the backward propagation layer of the bidirectional long short-term memory network at the same time step are merged to generate a feature representation containing a complete temporal context.

[0022] The feature representation containing the complete temporal context is input into the convolutional attention module, which is composed of a one-dimensional convolutional layer and a self-attention mechanism cascaded together.

[0023] The one-dimensional convolutional layer is used to perform local pattern scanning of features in the time dimension to capture interaction pattern features within a local time window.

[0024] The self-attention mechanism calculates the interaction pattern features within the local time window and assigns different weight coefficients to features at different times in order to focus on key time segments.

[0025] The features, after being weighted by the weighting coefficients, are aggregated to generate a high-level semantic feature vector for quality assessment.

[0026] Furthermore, the deep learning evaluation model is used to extract deep temporal dependencies and focus on key features in the multidimensional teaching feature vectors, generating a classroom attention map, knowledge point coverage index, and interaction response heatmap for the current teaching unit, including:

[0027] The multidimensional teaching feature vectors are input into the bidirectional long short-term memory network of the deep learning evaluation model in chronological order, and the bidirectional long short-term memory network extracts and outputs the hidden state features at each time step.

[0028] All hidden state features in the time series are input into the convolutional attention module of the deep learning evaluation model;

[0029] In the convolutional attention module, the one-dimensional convolutional layer performs convolution operations on the hidden state feature sequence and outputs enhanced local temporal features;

[0030] The self-attention mechanism performs correlation calculations on the enhanced local temporal features to obtain an attention weight distribution that reflects the importance of features at different times.

[0031] Based on the attention weight distribution, the hidden state features are weighted and summed to obtain a context vector focused on the key segments;

[0032] The attention weight curve, knowledge point association strength curve, and interaction response strength curve that change over time are decoded from the context vector, respectively.

[0033] The attention weight curve is mapped to a classroom attention graph that reflects the change of students' overall attention over time.

[0034] The hit rate of the preset knowledge point list covered by the current teaching unit is calculated from the knowledge point association strength curve to generate the knowledge point coverage index.

[0035] By overlaying the interactive response intensity curve with spatial location information, an interactive response heatmap reflecting the distribution of interactive behavior in the virtual classroom space is generated.

[0036] Furthermore, the classroom attention map, knowledge point coverage index, and interaction response heatmap are normalized and integrated to calculate the real-time teaching quality assessment score for the current teaching unit, including:

[0037] Numerical integration is performed on the classroom attention graph to calculate the attention integral value representing the average attention level;

[0038] The knowledge point coverage index is standardized and converted into a standard coverage value between zero and one.

[0039] Peak detection and regional activity statistics are performed on the interactive response heatmap to calculate the interactive activity score;

[0040] Preset weighting coefficients are configured for the focus score, the coverage standard score, and the interaction activity score, respectively;

[0041] The weighted concentration score, the weighted coverage score, and the weighted interaction activity score are summed to obtain the real-time teaching quality assessment score.

[0042] Furthermore, the real-time teaching quality assessment score is compared with a preset historical teaching quality score baseline to determine the quality status of the current teaching unit, including:

[0043] Retrieve the history teaching quality assessment score sequence of the same teacher and the same course in the history teaching unit from the history teaching quality database;

[0044] Calculate the moving average of the history teaching quality assessment score sequence, and set the moving average as the baseline for the history teaching quality score;

[0045] Compare the real-time teaching quality assessment score with the historical teaching quality score baseline.

[0046] When the real-time teaching quality assessment score remains below the historical teaching quality score baseline for an extended period of time, the current teaching unit is determined to be in a quality state that requires optimization.

[0047] Furthermore, the process for triggering teaching quality optimization includes:

[0048] Once it is determined that the current teaching unit is in a quality state that requires optimization, the optimization strategy generation process is initiated.

[0049] The classroom attention graph identifies periods of low attention where students' attention levels are below average.

[0050] The interaction response heatmap was used to identify sparsely interacting cold spots and corresponding student groups.

[0051] The original teaching video clips corresponding to the low attention time period, the student group identifiers of the cold spot area, and the knowledge points currently being taught are associated and packaged to form a teaching problem diagnosis package;

[0052] Based on a pre-set teaching strategy knowledge base, one or more potential teaching intervention strategies are matched to the teaching problem diagnosis package. The teaching intervention strategies include at least one of adjusting the teaching pace, inserting interactive questions, and switching media presentation formats.

[0053] The matched teaching intervention strategies are then pushed to the instructor's terminal.

[0054] Furthermore, the pre-defined teaching strategy knowledge base matches one or more potential teaching intervention strategies to the teaching problem diagnostic package, including:

[0055] The teaching problem diagnostic package was analyzed to extract the core problem types, characteristics of the affected student groups, and related knowledge point attributes;

[0056] The core question type, the characteristics of the affected student group, and the related knowledge point attributes are used as joint query conditions to search the teaching strategy knowledge base.

[0057] The teaching strategy knowledge base stores historical success cases, each of which records the problem context, the teaching intervention strategy used, and the degree of quality improvement after implementation.

[0058] Retrieve a set of historical successful cases whose similarity to the joint query conditions exceeds a preset threshold;

[0059] From the set of historical success cases, the teaching intervention strategies adopted by the top-ranked cases, in descending order of quality improvement, are selected as the potential teaching intervention strategies.

[0060] Furthermore, it also includes a feedback-based model continuous optimization step:

[0061] Collect new real-time interactive data generated in subsequent teaching units after the application of the aforementioned teaching intervention strategies;

[0062] Based on the new real-time interactive data, the subsequent real-time teaching quality assessment score for the next teaching unit is calculated.

[0063] The teaching intervention strategy, the corresponding teaching problem diagnostic package, and the difference in quality assessment scores before and after application are stored as a training sample pair in the feedback database.

[0064] Sample pairs are periodically extracted from the feedback database to incrementally train the deep learning evaluation model and update the network parameters of the deep learning evaluation model.

[0065] Furthermore, sample pairs are periodically extracted from the feedback database to incrementally train the deep learning evaluation model, including:

[0066] From the feedback database, sample pairs containing complete interaction data and evaluation results before and after the application of the teaching intervention strategy are extracted;

[0067] The interaction data before the application of the teaching intervention strategy was used as the model input, and the positive change in the quality assessment score after application was used as the reward signal for reinforcement learning.

[0068] The policy gradient algorithm is used to adjust the mapping parameters from multidimensional teaching feature vectors to real-time teaching quality assessment scores in the deep learning evaluation model;

[0069] This allows the deep learning evaluation model to assign higher weights to interactive pattern features that trigger positive quality changes after incremental training, thereby optimizing its evaluation accuracy and guidance for teaching strategies.

[0070] Compared with the prior art, the beneficial effects of the present invention are:

[0071] Multimodal fusion and timestamp alignment are performed on four types of real-time interactive data: teacher-student voice streams, screen operation trajectories, text communication records, and classroom behavior image sequences. This generates multidimensional teaching feature vectors that have temporal correlations and represent the quantitative state of classroom interaction per unit time. The acquisition nodes of the four different modal data achieve precise matching in the time dimension. Fragmented information from multiple sources is integrated into a unified quantitative representation carrier. The temporal change characteristics of the classroom interaction process are fully preserved, and the information loss problem caused by single modal data is eliminated. The interactive state throughout the entire teaching cycle can be presented continuously through feature vectors.

[0072] A deep learning evaluation model integrating a bidirectional long short-term memory network and a convolutional attention module is constructed. This model simultaneously performs deep temporal dependency extraction and key feature focusing on multi-dimensional teaching feature vectors, generating a classroom attention map, a knowledge point coverage index, and an interaction response heatmap. Figure 3 The system visualizes the evaluation results, normalizes and integrates the three types of results to obtain a real-time teaching quality evaluation score, and compares it with a preset historical teaching quality score baseline. When the score is lower than the baseline, the teaching quality optimization process is directly triggered. The bidirectional long short-term memory network can fully capture the temporal correlation features of teaching interactions, and the convolutional attention module can accurately locate the core feature areas in classroom interactions. The three types of visualized results can intuitively present the multi-dimensional state of teaching interactions, and the benchmarking of evaluation scores can directly determine the teaching quality status and initiate optimization actions. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the steps of a deep learning-based method for continuous optimization of interactive teaching quality as described in this invention.

[0074] Figure 2 Flowchart for building a deep learning evaluation model;

[0075] Figure 3 A time-series graph of classroom attention levels;

[0076] Figure 4 A time-series graph of classroom attention levels;

[0077] Figure 5 Box plot of score difference for quality improvement after application of teaching intervention strategies. Detailed Implementation

[0078] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] See Figure 1 This invention provides a method for continuous optimization of interactive teaching quality based on deep learning, the specific method including:

[0080] An interactive data acquisition mechanism covering the entire teaching cycle is established to continuously capture real-time interactive data, including teacher-student voice streams, screen operation trajectories, text communication records, and classroom behavior image sequences. The captured real-time interactive data undergoes multimodal fusion and timestamp alignment processing to generate a multidimensional teaching feature vector with temporal correlation. This multidimensional teaching feature vector is used to represent the quantitative state of classroom interaction within a unit of time. A deep learning evaluation model incorporating a bidirectional long short-term memory network and a convolutional attention module is constructed, and the processed multidimensional teaching feature vector is input into this model for calculation. This deep learning evaluation model performs deep temporal dependency extraction and key feature focusing processing on the input multidimensional teaching feature vector, generating a classroom attention map, knowledge point coverage index, and interaction response heatmap for the current teaching unit. The generated classroom attention map, knowledge point coverage index, and interaction response heatmap are normalized and integrated to calculate the real-time teaching quality evaluation score for the current teaching unit. Finally, the calculated real-time teaching quality assessment score is compared with the preset historical teaching quality score baseline to determine the quality status of the current teaching unit. When the real-time teaching quality assessment score is lower than the historical teaching quality score baseline, the teaching quality optimization process is triggered.

[0081] In one embodiment of the present invention, multimodal fusion and timestamp alignment are performed on real-time interactive data to generate a multidimensional teaching feature vector with temporal correlation. Speech recognition and voiceprint segmentation are performed on the collected teacher-student voice streams to distinguish between teacher and student voice segments, and the voice content is converted into text records with time-series labels. Coordinate analysis is performed on the collected screen operation trajectories to record cursor movement speed, click frequency, and trigger times of specific teaching software interface elements. Word segmentation and sentiment polarity analysis are performed on the collected text communication records to extract key question words, student answer text, and corresponding sentiment tendency labels. Face detection and pose estimation are performed on the collected classroom behavior image sequences to identify the student group's facial orientation, head posture change frequency, and hand-raising actions. The time-series labels obtained from all modal data conversions are uniformly mapped onto the same time axis, and the aligned data is sliced ​​using fixed time windows. Feature extraction and splicing operations are performed on the multimodal data slices within each time window to form the multidimensional teaching feature vector.

[0082] In specific implementations, real-time interactive data undergoes multimodal fusion and timestamp alignment to generate multidimensional teaching feature vectors with temporal correlation. This process is performed separately for each data stream collected. Specifically, the teacher-student speech streams are processed using speech recognition and voiceprint segmentation. Speech recognition converts continuous audio signals into text, while voiceprint segmentation uses a trained voiceprint model to distinguish between teacher and student voiceprint features, thus segmenting the mixed speech stream into speech segments with clearly labeled speakers. Each identified speech segment is marked with its start and end timestamps within the entire teaching timeline, and the speech content is converted into text records with these time-series labels. In some embodiments, screen operation trajectories are analyzed using coordinates, recording the cursor's movement speed in the screen coordinate system, the click frequency on interface elements, and the moment when a specific teaching software interface element is triggered. The trigger time is recorded with millisecond precision and synchronized with the global teaching timeline.

[0083] In practical implementation, the text communication records are segmented and subjected to sentiment polarity analysis. Segmentation divides the continuous chat text into independent lexical units. The sentiment polarity analysis model calculates the sentiment polarity of each student's response text, outputting a label representing a positive, neutral, or negative sentiment tendency. Simultaneously, key question words from the teacher's questions and student responses are extracted from the text. In practical implementation, classroom behavior image sequences are processed for face detection and pose estimation. The face detection algorithm locates all student facial regions in each frame. The pose estimation model calculates facial orientation vectors based on these facial regions, estimates the yaw, pitch, and roll angles of the head in three-dimensional space, and counts the frequency of head pose changes per unit time. It also identifies hand-raising actions in the image sequence that conform to a preset pose model, and each detected event is accompanied by a timestamp.

[0084] In practical implementation, after multimodal data processing, the time-series labels obtained from all modal data are uniformly mapped onto the same time axis, with the absolute start of teaching as the zero point. In practice, the aligned multimodal data is sliced ​​using fixed-length time windows, each with a preset length, such as 5 or 10 seconds. The slicing operation ensures that each window contains time-aligned data segments from all modalities. In some embodiments, feature extraction and concatenation are performed on the multimodal data slices within each time window. Feature extraction calculates a set of statistical or semantic feature vectors for each modality. For example, features for the speech modality may include the proportion of teacher's speech duration per unit time, the proportion of student's speech duration, and speech emotion entropy; features for the screen operation modality include average cursor movement speed and click event density; features for the text modality include the distribution of sentiment values ​​and keyword frequency; and features for the behavioral image modality include average facial orientation deviation angle, head posture change variance, and hand-raising action count. The feature vectors from different modalities are concatenated along the feature dimension to form a high-dimensional feature vector, which is the multidimensional teaching feature vector for that time window. Optionally, the generation process of the multidimensional teaching feature vector can be formally represented as follows, where the speech feature vector of the t-th time window is... The screen operation feature vector is The text feature vector is The behavioral image feature vector is The resulting multidimensional teaching feature vector after concatenation Represented as:

[0085]

[0086] in: This represents the multidimensional teaching feature vector generated within the time window t. This represents the speech feature sub-vectors extracted from the teacher and student speech stream. This represents the sub-vector of operation features extracted from the screen operation trajectory. This represents a text feature subvector extracted from written communication records. Represents a sub-vector of behavioral features extracted from a sequence of classroom behavior images, with the symbol... This represents the concatenation operation of vectors along a dimensional direction. It can be understood that through the above steps, the original heterogeneous multimodal interaction data is transformed into a series of temporally continuous and feature-uniform multidimensional teaching feature vector sequences, providing structured input for subsequent deep temporal modeling.

[0087] In one embodiment of the present invention, the constructed deep learning evaluation model includes a bidirectional long short-term memory network and a convolutional attention module. See also... Figure 2The Bidirectional Long Short-Term Memory (BSSM) network comprises forward and backward propagation layers to extract forward and backward temporal context information from the input multidimensional teaching feature vector. The outputs of the forward and backward propagation layers at the same time step are merged to generate a feature representation containing complete temporal context. This feature representation is then input into a convolutional attention module, which consists of a cascaded one-dimensional convolutional layer and a self-attention mechanism. The one-dimensional convolutional layer performs local pattern scanning of the features along the temporal dimension, capturing interaction pattern features within a local time window. The self-attention mechanism calculates the interaction pattern features within this local time window, assigning different weight coefficients to features at different times to focus on key time segments. The weighted features are then aggregated to generate a high-level semantic feature vector for quality assessment.

[0088] A multi-dimensional teaching feature vector arranged chronologically is input into the bidirectional long short-term memory network of the deep learning evaluation model. This network extracts and outputs the hidden state features at each time step. All hidden state features in the time series are then input into the model's convolutional attention module. In this module, a one-dimensional convolutional layer performs convolution operations on the hidden state feature sequence, outputting enhanced local temporal features. A self-attention mechanism calculates the correlation of these enhanced local temporal features, obtaining an attention weight distribution reflecting the importance of features at different times. Based on this attention weight distribution, the original hidden state features are weighted and summed to obtain a context vector focusing on key segments. From this context vector, the attention weight curve, knowledge point association strength curve, and interaction response strength curve, which change over time, are decoded. The attention weight curve is mapped to a classroom attention map reflecting the overall student focus over time. The hit rate of the preset knowledge point list covered by the current teaching unit is calculated from the knowledge point association strength curve, generating a knowledge point coverage index. Spatial location information is overlaid on the interaction response strength curve to generate an interaction response heatmap reflecting the spatial distribution of interactive behavior in the virtual classroom.

[0089] In the specific implementation, the constructed deep learning evaluation model includes a bidirectional long short-term memory network (LSTM) and a convolutional attention module. The LSM includes forward and backward propagation layers, used to extract forward and backward temporal context information from the input multidimensional teaching feature vector sequence. In the specific implementation, the forward propagation layer processes the multidimensional teaching feature vector sequence in chronological order, while the backward propagation layer processes it in reverse chronological order. The output vectors of the forward and backward propagation layers of the LSM at the same time step are merged. This merging operation can be vector concatenation or summation, thereby generating a feature representation containing complete temporal context. In the specific implementation, the feature representation containing complete temporal context is input into the convolutional attention module, which consists of a cascaded one-dimensional convolutional layer and a self-attention mechanism. The one-dimensional convolutional layer performs convolution operations on the feature representation sequence in the temporal dimension, with its convolution kernel sliding along the time axis to scan feature patterns within local time windows, capturing interaction pattern features within those local time windows, such as identifying brief collective silences and frequent interactive responses.

[0090] In practical implementation, the self-attention mechanism calculates the enhanced local temporal feature sequence output by the one-dimensional convolutional layer. It assigns different weight coefficients to features at different times in the sequence to focus on key time segments. In some embodiments, the self-attention mechanism obtains the attention weight distribution by calculating the correlation between the query vector, key vector, and value vector. The query vector, key vector, and value vector are all obtained from the local temporal feature sequence through linear transformation. The attention weights reflect the differences in importance of features at different time steps to the current evaluation task. In practice, the weighted features are aggregated. The aggregation operation typically involves summing the weighted value vectors of all time steps to generate a high-level semantic feature vector for quality assessment.

[0091] In practice, multidimensional teaching feature vectors arranged chronologically are input into the bidirectional long short-term memory (LSTM) network of the deep learning evaluation model. The LSM network extracts and outputs the hidden state features at each time step, and all hidden state features in the time series constitute a hidden state feature sequence. In the implementation, the hidden state feature sequence is input into the convolutional attention module of the deep learning evaluation model. In the convolutional attention module, a one-dimensional convolutional layer performs convolution operations on the hidden state feature sequence, outputting an enhanced local temporal feature sequence. A self-attention mechanism performs relevance calculations on the enhanced local temporal feature sequence. The calculation process involves generating a query matrix, a key matrix, and a value matrix, and obtaining the attention weight distribution reflecting the importance of features at different times through scaling dot product operations.

[0092] Based on the attention weight distribution, the original hidden state feature sequence is weighted and summed. The formula for the weighted summation can be expressed as a context vector. Calculation:

[0093]

[0094] in: This represents a context vector focusing on key segments. This represents the total number of time steps in the input sequence. Represents the first in the sequence Index of each time step, The self-attention mechanism is represented by the first... Attention weight coefficients assigned to the hidden state features at each time step. Representing the Each time step generates a hidden state feature vector output by a bidirectional long short-term memory network. The attention weight curve, knowledge point association strength curve, and interaction response strength curve, which change over time, are decoded from the context vector. This decoding process can be implemented using an additional fully connected layer or a specific mapping network. In a specific implementation, the attention weight curve is mapped to a classroom attention map reflecting the overall student focus over time. The mapping relationship can be linear or a non-linear calibration function. In some embodiments, the hit rate of a preset knowledge point list covered by the current teaching unit is statistically analyzed from the knowledge point association strength curve to generate a knowledge point coverage index. The preset knowledge point list is defined before the start of teaching, and the hit rate is calculated as the ratio of the number of preset knowledge points whose intensity exceeds a preset threshold on the knowledge point association strength curve to the total number of preset knowledge points in the preset knowledge point list. In a specific implementation, spatial location information is superimposed on the interaction response strength curve to generate an interaction response heatmap reflecting the spatial distribution of interactive behavior in the virtual classroom. The spatial location information comes from the coordinates or grouping information of the student's terminal in the virtual classroom. The superposition operation maps the interaction response strength value of each student at a specific time period to its corresponding spatial location. It is understandable that by capturing long-range dependencies through a bidirectional long short-term memory network and focusing on key teaching events through a convolutional attention module, the model can structurally extract three core evaluation indicators from complex multimodal time-series data: classroom attention map, knowledge point coverage index, and interaction response heatmap.

[0095] In one embodiment of the present invention, the generated classroom focus map, knowledge point coverage index, and interactive response heatmap are normalized and integrated to calculate the real-time teaching quality assessment score for the current teaching unit. The classroom focus map is numerically integrated to calculate a focus score representing the average focus level. The knowledge point coverage index is standardized to convert it into a standard coverage value between zero and one. Peak detection and regional activity statistics are performed on the interactive response heatmap to calculate the interactive activity score. Preset weighting coefficients are assigned to the focus score, coverage standard value, and interactive activity score. The weighted focus score, weighted coverage standard value, and weighted interactive activity score are summed to obtain the real-time teaching quality assessment score.

[0096] The historical teaching quality assessment score sequence for the same teacher and the same course in a historical teaching unit is retrieved from the historical teaching quality database. A moving average of this score sequence is calculated and set as the baseline for historical teaching quality scores. The real-time teaching quality assessment score is compared with the baseline. If the real-time teaching quality assessment score remains below the baseline for an extended period beyond a preset time threshold, the current teaching unit is deemed to be in a quality state requiring optimization.

[0097] In practice, the generated classroom focus map, knowledge point coverage index, and interactive response heatmap are normalized and integrated to calculate the real-time teaching quality assessment score for the current teaching unit. The calculation process involves quantification and weighting of each indicator. Specifically, the classroom focus map is numerically integrated. Since the classroom focus map is a curve that changes over time, the numerical integration is performed by definite integral over the curve within the course's time interval. The result represents the average student focus level across the entire teaching unit, and the calculated integral value is defined as the focus score. The knowledge point coverage index is standardized. This index is a raw statistical value; standardization uses preset maximum and minimum values ​​and a linear transformation to convert the raw value into a standard coverage value between zero and one. The interactive response heatmap undergoes peak detection and regional activity statistics. Peak detection identifies local high-intensity areas in the heatmap where the response intensity exceeds a preset threshold. Regional activity statistics calculate the area percentage and average intensity of all high-intensity areas, and the interactive activity score is calculated based on the weighted sum of these area percentages and average intensities.

[0098] In practice, preset weighting coefficients are assigned to the focus score, coverage standard value, and interaction activity score, respectively. These weighting coefficients are determined based on the teaching assessment objectives, and the sum of these coefficients is one. In some embodiments, the weighted focus score, weighted coverage standard value, and weighted interaction activity score are summed to obtain the real-time teaching quality assessment score. The summation can be expressed using the following formula: Calculation:

[0099]

[0100] in: Represents real-time teaching quality assessment scores. This represents the weighting coefficients assigned to the focus score. This represents a score indicating focus level. This represents the weighting coefficients configured for the coverage standard value. Represents the standard value of coverage. This represents the weighting coefficient assigned to the interaction activity score. This represents the score for interaction and activity level.

[0101] In specific implementation, the historical teaching quality assessment score sequence generated by the same teacher and the same course in a historical teaching unit is retrieved from the historical teaching quality database. The historical teaching quality database stores the assessment results of all past teaching units, and the search criteria include teacher identifier and course identifier. In specific implementation, the moving average of the historical teaching quality assessment score sequence is calculated. The moving average can be an exponential moving average or a simple moving average, and the window size is set according to the teaching cycle. The calculated moving average is set as the historical teaching quality score baseline. In specific implementation, the real-time teaching quality assessment score is compared with the historical teaching quality score baseline. The comparison operation determines whether the real-time teaching quality assessment score is higher than, equal to, or lower than the historical teaching quality score baseline. In some embodiments, when the real-time teaching quality assessment score is continuously lower than the historical teaching quality score baseline for more than a preset time threshold, the current teaching unit is determined to be in a quality state that needs optimization. The preset time threshold can be the length of one or more teaching time windows. Continuous determination requires that the scores of multiple consecutive assessment points are lower than the baseline.

[0102] It is understandable that by normalizing and integrating the three dimensions of evaluation indicators to obtain a comprehensive score, and comparing it with a personalized historical baseline, it is possible to dynamically identify quality fluctuations relative to a teacher's usual teaching level. See Table 1, which shows example data from a comparison process, assuming a preset time threshold of three consecutive time windows.

[0103] Table 1: Example Table Comparing Real-Time Teaching Quality Assessment Scores with Benchmarks

[0104] Time window Real-time teaching quality assessment scores History teaching quality score benchmark Comparison results Window 1 85 88 Below Window 2 84 88 Below Window 3 82 87 Below Window 4 89 87 Higher than

[0105] See Figure 3 The graph of classroom attention span variation, with class time on the horizontal axis and attention score on the vertical axis, visually presents the dynamic fluctuations in student attention within a 45-minute teaching unit. The dotted line represents the average attention baseline (70.3 points), while the broken line and filled area depict the characteristics of attention changes over time: Initial stage (5-15 minutes): Attention gradually decreases from 78 points to 70 points, initially touching the average baseline, reflecting the natural decline in student attention after entering the classroom. Mid-stage fluctuation (20-35 minutes): Attention fluctuates, first rising and then falling, reaching a local peak of 85 points at 20 minutes before gradually declining, dropping to 72 points at 35 minutes, reflecting the impact of the pace of teaching content and the intensity of interaction on attention. Late recovery stage (40-45 minutes): Attention rises rapidly, reaching a peak of 92 points at 45 minutes, indicating that the reinforcement of content or interactive design in the later stages of teaching effectively activated student attention. As the core visualization output of the deep learning evaluation model, the graph provides a quantitative basis for identifying periods of low attention and triggering teaching quality optimization processes. It can be further normalized and integrated with the knowledge point coverage index and interactive response heatmap to generate real-time teaching quality evaluation scores.

[0106] In one embodiment of the present invention, when it is determined that the current teaching unit is in a quality state requiring optimization, an optimization strategy generation process is initiated. Low-attention periods where student attention is below average are identified from the classroom attention map. Sparsely interacting cold spots and corresponding student groups are identified from the interaction response heatmap. The original teaching video clips corresponding to the low-attention periods, the student group identifiers in the cold spots, and the currently taught knowledge points are associated and packaged to form a teaching problem diagnostic package.

[0107] Based on a pre-defined teaching strategy knowledge base, one or more potential teaching intervention strategies are matched to the teaching problem diagnosis package. These intervention strategies include at least one of the following: adjusting the teaching pace, inserting interactive questions, or switching media presentation formats. The matching process involves parsing the teaching problem diagnosis package to extract the core problem type, characteristics of the affected student group, and related knowledge point attributes. The core problem type, characteristics of the affected student group, and related knowledge point attributes are used as joint query conditions to search the teaching strategy knowledge base. The knowledge base stores historical successful cases, each recording the problem context, the teaching intervention strategy used, and the degree of quality improvement after implementation. A set of historical successful cases with a similarity exceeding a pre-defined threshold to the joint query conditions is retrieved. From this set of historical successful cases, the teaching intervention strategies used in the top-ranked cases, ordered from highest to lowest quality improvement, are selected as potential teaching intervention strategies. The matched teaching intervention strategies are then pushed to the instructor's terminal.

[0108] In practice, once the current teaching unit is determined to be in a quality state requiring optimization, an optimization strategy generation process is initiated. This process generates problem diagnoses and strategy matching based on specific evaluation indicator results. Specifically, the process identifies low-attention periods where student focus is below average from the classroom focus graph. This identification process compares the classroom focus graph curve with a preset focus threshold or a dynamically calculated average focus level, marking all continuous time intervals where curve values ​​are below the threshold or average level. These continuous time intervals are defined as low-attention periods. Furthermore, the process identifies sparsely interacting cold spots and corresponding student groups from the interaction response heatmap. The interaction response heatmap is spatially divided into multiple regions, and the average interaction response intensity of each region is calculated. Regions with average values ​​below a preset threshold are marked as cold spots. Based on the virtual classroom coordinate mapping, student identifiers within these cold spots are determined, and the sets corresponding to these student identifiers constitute the student groups within the cold spots.

[0109] In practice, the original teaching video clips corresponding to periods of low attention, student group identifiers in cold areas, and the currently taught knowledge points are associated and packaged to form a teaching problem diagnostic package. This associative packaging operation establishes an index relationship between timestamps, spatial coordinates, and knowledge tags at the data structure level. Based on a pre-set teaching strategy knowledge base, one or more potential teaching intervention strategies are matched to the teaching problem diagnostic package. These intervention strategies include at least one of the following: adjusting the teaching pace, inserting interactive questions, or switching media presentation formats. The matching process searches the knowledge base for similar historical contexts based on the diagnosed problem characteristics.

[0110] In practice, the matching process includes parsing the teaching problem diagnostic package, extracting the core problem type, characteristics of the affected student group, and related knowledge point attributes. The core problem type is, for example, "decreased group focus" or "insufficient local interaction." Affected student group characteristics include group size and geographical distribution. Related knowledge point attributes include the difficulty and type of the knowledge point. In practice, the core problem type, affected student group characteristics, and related knowledge point attributes are used as joint query conditions to search a teaching strategy knowledge base. This knowledge base stores historical successful cases, each recording the problem context, the teaching intervention strategy used, and the degree of quality improvement after implementation. In some embodiments, a set of historical successful cases with a similarity exceeding a preset threshold is retrieved. Similarity calculation uses methods such as vectorization and cosine similarity. The preset threshold is used to filter cases with sufficiently high relevance. In practice, from the set of historical successful cases, the teaching intervention strategies used in the top-ranked cases, ordered from highest to lowest quality improvement, are selected as potential teaching intervention strategies. In some embodiments, the quality improvement rate is defined as the relative percentage increase in teaching quality assessment scores after applying teaching intervention strategies. "Ranking high" can refer to selecting the top N cases or all cases with an improvement rate exceeding a certain level. See Table 2 for an example of a search and matching query for a teaching strategy knowledge base.

[0111] Table 2: Example Table of Retrieval and Matching of Teaching Strategy Knowledge Base

[0112] Core problem types Characteristics of the affected student group Related knowledge point attributes Matching historical teaching intervention strategies Historical quality improvement Decreased group focus All students Concept Explanation Insert interactive questions 15% Insufficient local interaction Back row students Formula Derivation Adjust the pace of the lecture and increase emphasis on blackboard writing. 8% Decreased group focus All students Case Analysis Switch media presentation format and play short videos. 12%

[0113] In practice, the selection process for teaching intervention strategies from a collection of historical successful cases can be represented by the following formula: Selection Index Determination:

[0114]

[0115] in: This represents the index of the selected case from the set of cases that meet the criteria. Representative of historical success cases The Middle Index of cases, Representing the The quality improvement values ​​recorded in each case. Indicates in set Find the degree that can improve quality Index of cases that yielded the maximum value In practice, the matched teaching intervention strategies are pushed to the instructor's terminal, and the push process is completed through the message center or notification system of the teaching software. It can be understood that by locating the temporal and spatial position of specific problems from classroom attention maps and interaction response heatmaps, and associating them with knowledge points, a precise teaching problem diagnostic package can be formed. Furthermore, based on a database of historical successful experiences, validated teaching intervention strategies can be matched to provide teachers with immediate and specific suggestions for teaching optimization.

[0116] See Figure 4 In the visualization analysis of classroom attention span changes over time, the identification of low-attention periods relies on the dynamic comparison between the real-time attention curve and the average attention benchmark. Specifically, the real-time attention curve is generated by mapping the attention weight sequence output by the deep learning evaluation model, representing the students' attention metric score (0-100) at each time step. The average attention score (67.8) is calculated as the integral mean of the attention scores across all time steps within the teaching unit, serving as the dynamic threshold benchmark. The difference between the two curves is quantified through point-by-point comparison: all continuous time intervals where the real-time attention score is lower than the average attention threshold are marked; these intervals are defined as low-attention periods (orange filled area in the figure). During parameter configuration, the minimum duration of continuous intervals is set to 2 minutes to filter out instantaneous fluctuations and ensure that the identified low-attention periods have practical significance for teaching intervention.

[0117] In one embodiment of the invention, new real-time interactive data generated in subsequent teaching units after the application of the teaching intervention strategy is collected. Based on the new real-time interactive data, the subsequent real-time teaching quality assessment score for each teaching unit is calculated. The teaching intervention strategy, the corresponding teaching problem diagnostic package, and the difference in quality assessment scores before and after application are stored as a training sample pair in the feedback database. Sample pairs are periodically extracted from the feedback database to incrementally train the deep learning assessment model and update the network parameters of the deep learning assessment model.

[0118] During incremental training, sample pairs containing complete interaction data and evaluation results before and after the application of the teaching intervention strategy are extracted from the feedback database. The interaction data before the application of the teaching intervention strategy is used as the model input, and the positive change in quality evaluation score after application is used as the reward signal for reinforcement learning. A policy gradient algorithm is employed to adjust the mapping parameters from multi-dimensional teaching feature vectors to real-time teaching quality evaluation scores in the deep learning evaluation model. This allows the deep learning evaluation model to assign higher weights to interaction pattern features that trigger positive changes in quality after incremental training.

[0119] In specific implementation, the method also includes a feedback-based model continuous optimization step, which updates the deep learning evaluation model using feedback data after the application of the teaching intervention strategy. Specifically, new real-time interactive data generated in subsequent teaching units after the application of the teaching intervention strategy is collected. This new real-time interactive data uses the same acquisition mechanism and modality as the original data, covering teacher-student voice streams, screen operation trajectories, text communication records, and classroom behavior image sequences. Based on this new real-time interactive data, subsequent real-time teaching quality assessment scores for subsequent teaching units are calculated. The calculation process follows the steps described in one of the above embodiments, including processing the new interactive data, feature extraction, model evaluation, and score calculation. Finally, the teaching intervention strategy, the corresponding teaching problem diagnostic package, and the difference in quality assessment scores before and after application are stored as a training sample pair in the feedback database. The difference in quality assessment scores before and after application refers to the difference between the subsequent real-time teaching quality assessment score after the application of the teaching intervention strategy and the real-time teaching quality assessment score before application.

[0120] In practice, sample pairs are periodically extracted from the feedback database to incrementally train the deep learning evaluation model and update its network parameters. This periodicity can be triggered daily, weekly, or when the accumulated sample count reaches a threshold. Sample pairs containing complete interaction data and evaluation results before and after the application of the teaching intervention strategy are extracted from the feedback database. Complete interaction data refers to the original multimodal data required to generate multidimensional teaching feature vectors or the feature data after preliminary processing. The interaction data before the application of the teaching intervention strategy is used as the model input, and the positive change in the quality evaluation score after application is used as the reward signal for reinforcement learning. The reward signal is a scalar value, with a positive value representing quality improvement. In some embodiments, a policy gradient algorithm is used to adjust the mapping parameters from the multidimensional teaching feature vector to the real-time teaching quality evaluation score in the deep learning evaluation model. The policy gradient algorithm updates the parameters by calculating the gradient of the expected reward with respect to the model parameters.

[0121] In practice, the objective function of incremental training is to maximize the observed improvement in quality assessment scores after applying the teaching intervention strategy. The policy parameters in the policy gradient algorithm... The update can be represented as:

[0122]

[0123] in: The set of trainable parameters represents the mapping relationship between multidimensional teaching feature vectors and real-time teaching quality assessment scores in a deep learning evaluation model. This represents an assignment / update operation. The learning rate is a pre-defined positive scalar. Represents the objective function Regarding parameters gradient, objective function This refers to the expected reward, specifically the positive change in the quality assessment score. In practice, after incremental training, the deep learning assessment model assigns higher weights to the interactive pattern features that trigger positive changes in quality, thereby optimizing the model's assessment accuracy and its guidance of teaching strategies. It can be understood that by using continuously collected feedback on the effects of applied teaching intervention strategies as reinforcement learning signals to incrementally train the deep learning assessment model, the model can continuously learn which interactive feature patterns are associated with good teaching effects, thus gradually improving the model's assessment accuracy and its guiding value for subsequent teaching optimization. In some embodiments, the incremental training process uses offline learning, without interfering with the real-time operation of the online assessment system.

[0124] See Figure 5 In the feedback iteration phase of the continuous optimization method for teaching quality, the box plot of teaching quality improvement values ​​visually presents the distribution characteristics of the difference in quality assessment scores after the application of teaching intervention strategies. This box plot, with the difference in quality improvement scores on the horizontal axis, comprehensively depicts the statistical distribution of the intervention effect: the box range is approximately 5.8 to 8.4, representing the interquartile range of the core samples, reflecting the concentrated range of quality improvement in most intervention cases; the median (midpoint) is located around 7.4, indicating that more than half of the interventions achieved this level of quality improvement; the upper and lower bands extend to 4.1 and 11.7 respectively, covering the extreme value range of all samples, reflecting the overall fluctuation range of the intervention effect. From a professional perspective, this box plot can be used to quantitatively evaluate the effectiveness of continuous model optimization: Distribution stability: The compact box and stable median position indicate that the quality improvement effect of the teaching intervention strategy is repeatable and consistent, without extreme skewed distribution. Effect boundary: The band range defines the upper and lower limits of the intervention effect, providing data basis for setting the threshold of the subsequent reinforcement learning reward signal, which can be used to screen high-value training samples. Optimization Iteration Reference: This distribution can be used as a benchmark to compare the distribution changes of intervention effects after incremental training in different batches, thereby verifying the degree of optimization of the guidance of deep learning evaluation model parameter updates on teaching quality.

[0125] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for continuous optimization of interactive teaching quality based on deep learning, characterized in that, Includes the following steps: Establish an interactive data collection mechanism covering the entire teaching cycle to continuously capture real-time interactive data, including teacher and student voice streams, screen operation trajectories, text communication records, and classroom behavior image sequences. The real-time interactive data is subjected to multimodal fusion and timestamp alignment to generate a multidimensional teaching feature vector with temporal correlation. The multidimensional teaching feature vector is used to characterize the quantitative state of the classroom interaction process within a unit of time. A deep learning evaluation model containing a bidirectional long short-term memory network and a convolutional attention module is constructed, and the multidimensional teaching feature vector is input into the deep learning evaluation model for calculation; The deep learning evaluation model is used to extract deep temporal dependencies and focus on key features of the multidimensional teaching feature vectors, generating a classroom attention map, knowledge point coverage index and interactive response heatmap for the current teaching unit. The classroom focus map, knowledge point coverage index, and interactive response heatmap are normalized and integrated to calculate the real-time teaching quality assessment score for the current teaching unit. The real-time teaching quality assessment score is compared with the preset historical teaching quality score baseline to determine the quality status of the current teaching unit. When the real-time teaching quality assessment score is lower than the historical teaching quality score baseline, the teaching quality optimization process is triggered.

2. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 1, characterized in that, The real-time interactive data is subjected to multimodal fusion and timestamp alignment to generate a multidimensional teaching feature vector with temporal correlation, including: The teacher and student voice streams are processed by speech recognition and voiceprint segmentation to distinguish between teacher voice segments and student voice segments, and the voice content is converted into text records with time-series tags. The screen operation trajectory is analyzed by coordinates, and the cursor movement speed, click frequency, and trigger time of specific teaching software interface elements are recorded. The text communication records were segmented and analyzed for sentiment polarity to extract key question words, student answer texts and corresponding sentiment tendency tags. The classroom behavior image sequence is processed by face detection and pose estimation to identify the students' facial orientation, head pose change frequency, and hand raising actions. All time series labels obtained from the conversion of modal data are uniformly mapped onto the same time axis, and the aligned data is sliced ​​with a fixed time window; Feature extraction and splicing operations are performed on the multimodal data slices within each time window to form the multidimensional teaching feature vector.

3. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 2, characterized in that, The construction of the deep learning evaluation model, which includes a bidirectional long short-term memory network and a convolutional attention module, includes: A forward propagation layer and a backward propagation layer of a bidirectional long short-term memory network are established to extract forward and backward temporal context information from the multidimensional teaching feature vector; The outputs of the forward propagation layer and the backward propagation layer of the bidirectional long short-term memory network at the same time step are merged to generate a feature representation containing a complete temporal context. The feature representation containing the complete temporal context is input into the convolutional attention module, which is composed of a one-dimensional convolutional layer and a self-attention mechanism cascaded together. The one-dimensional convolutional layer is used to perform local pattern scanning of features in the time dimension to capture interaction pattern features within a local time window. The self-attention mechanism calculates the interaction pattern features within the local time window and assigns different weight coefficients to features at different times in order to focus on key time segments. The features, after being weighted by the weighting coefficients, are aggregated to generate a high-level semantic feature vector for quality assessment.

4. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 3, characterized in that, The deep learning evaluation model extracts deep temporal dependencies and focuses on key features from the multidimensional teaching feature vectors, generating a classroom attention map, knowledge point coverage index, and interaction response heatmap for the current teaching unit, including: The multidimensional teaching feature vectors are input into the bidirectional long short-term memory network of the deep learning evaluation model in chronological order, and the bidirectional long short-term memory network extracts and outputs the hidden state features at each time step. All hidden state features in the time series are input into the convolutional attention module of the deep learning evaluation model; In the convolutional attention module, the one-dimensional convolutional layer performs convolution operations on the hidden state feature sequence and outputs enhanced local temporal features; The self-attention mechanism performs correlation calculations on the enhanced local temporal features to obtain an attention weight distribution that reflects the importance of features at different times. Based on the attention weight distribution, the hidden state features are weighted and summed to obtain a context vector focused on the key segments; The attention weight curve, knowledge point association strength curve, and interaction response strength curve that change over time are decoded from the context vector, respectively. The attention weight curve is mapped to a classroom attention graph that reflects the change of students' overall attention over time. The hit rate of the preset knowledge point list covered by the current teaching unit is calculated from the knowledge point association strength curve to generate the knowledge point coverage index. By overlaying the interactive response intensity curve with spatial location information, an interactive response heatmap reflecting the distribution of interactive behavior in the virtual classroom space is generated.

5. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 4, characterized in that, The classroom attention map, knowledge point coverage index, and interaction response heatmap are normalized and integrated to calculate the real-time teaching quality assessment score for the current teaching unit, including: Numerical integration is performed on the classroom attention graph to calculate the attention integral value representing the average attention level; The knowledge point coverage index is standardized and converted into a standard coverage value between zero and one. Peak detection and regional activity statistics are performed on the interactive response heatmap to calculate the interactive activity score; Preset weighting coefficients are configured for the focus score, the coverage standard score, and the interaction activity score, respectively; The weighted concentration score, the weighted coverage score, and the weighted interaction activity score are summed to obtain the real-time teaching quality assessment score.

6. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 5, characterized in that, The real-time teaching quality assessment score is compared with a preset historical teaching quality score baseline to determine the quality status of the current teaching unit, including: Retrieve the history teaching quality assessment score sequence of the same teacher and the same course in the history teaching unit from the history teaching quality database; Calculate the moving average of the history teaching quality assessment score sequence, and set the moving average as the baseline for the history teaching quality score; Compare the real-time teaching quality assessment score with the historical teaching quality score baseline. When the real-time teaching quality assessment score remains below the historical teaching quality score baseline for an extended period of time, the current teaching unit is determined to be in a quality state that requires optimization.

7. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 6, characterized in that, The process for triggering teaching quality optimization includes: Once it is determined that the current teaching unit is in a quality state that requires optimization, the optimization strategy generation process is initiated. The classroom attention graph identifies periods of low attention where students' attention levels are below average. The interaction response heatmap was used to identify sparsely interacting cold spots and corresponding student groups. The original teaching video clips corresponding to the low attention time period, the student group identifiers of the cold spot area, and the knowledge points currently being taught are associated and packaged to form a teaching problem diagnosis package; Based on a pre-set teaching strategy knowledge base, one or more potential teaching intervention strategies are matched to the teaching problem diagnosis package. The teaching intervention strategies include at least one of adjusting the teaching pace, inserting interactive questions, and switching media presentation formats. The matched teaching intervention strategies are then pushed to the instructor's terminal.

8. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 7, characterized in that, The pre-set teaching strategy knowledge base matches one or more potential teaching intervention strategies to the teaching problem diagnostic package, including: The teaching problem diagnostic package was analyzed to extract the core problem types, characteristics of the affected student groups, and related knowledge point attributes; The core question type, the characteristics of the affected student group, and the related knowledge point attributes are used as joint query conditions to search the teaching strategy knowledge base. The teaching strategy knowledge base stores historical success cases, each of which records the problem context, the teaching intervention strategy used, and the degree of quality improvement after implementation. Retrieve a set of historical successful cases whose similarity to the joint query conditions exceeds a preset threshold; From the set of historical success cases, the teaching intervention strategies adopted by the top-ranked cases, in descending order of quality improvement, are selected as the potential teaching intervention strategies.

9. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 8, characterized in that, It also includes a feedback-based model continuous optimization step: Collect new real-time interactive data generated in subsequent teaching units after the application of the aforementioned teaching intervention strategies; Based on the new real-time interactive data, the subsequent real-time teaching quality assessment score for the next teaching unit is calculated. The teaching intervention strategy, the corresponding teaching problem diagnostic package, and the difference in quality assessment scores before and after application are stored as a training sample pair in the feedback database. Sample pairs are periodically extracted from the feedback database to incrementally train the deep learning evaluation model and update the network parameters of the deep learning evaluation model.

10. The method for continuous optimization of interactive teaching quality based on deep learning according to claim 9, characterized in that, Periodically extract sample pairs from the feedback database to incrementally train the deep learning evaluation model, including: From the feedback database, sample pairs containing complete interaction data and evaluation results before and after the application of the teaching intervention strategy are extracted; The interaction data before the application of the teaching intervention strategy was used as the model input, and the positive change in the quality assessment score after application was used as the reward signal for reinforcement learning. The policy gradient algorithm is used to adjust the mapping parameters from multidimensional teaching feature vectors to real-time teaching quality assessment scores in the deep learning evaluation model; This allows the deep learning evaluation model to assign higher weights to interactive pattern features that trigger positive quality changes after incremental training, thereby optimizing its evaluation accuracy and guidance for teaching strategies.

Images